pysusie 0.1.0

Pure-Python port of the R/CRAN package susieR — Sum of Single Effects (SuSiE) regression for Bayesian variable selection and genetic fine-mapping.

py-susie

Pure-Python port of the R/CRAN package susieR — the "Sum of Single Effects" (SuSiE) regression model for Bayesian variable selection and genetic fine-mapping (Wang, Sarkar, Carbonetto & Stephens, J. R. Stat. Soc. B 2020, 82:1273-1300; Zou, Carbonetto, Wang & Stephens, PLoS Genetics 2022, 18:e1010299).

pysusie is a standalone, dependency-light implementation of susieR's complete computational core. It does not require R or rpy2 — only numpy, scipy, pandas and matplotlib.

PyPI / import name	pysusie
License	BSD 3-Clause (same as upstream susieR)
Upstream	susieR 0.14.2 (CRAN)

What is SuSiE?

SuSiE fits the sparse linear regression y = mu + Xb + e under the "susie assumption" that b is a sum of L single effects, each a vector with exactly one non-zero element. The fitting algorithm, Iterative Bayesian Stepwise Selection (IBSS), is a fast coordinate ascent over L Bayesian single-effect regressions. SuSiE summarises uncertainty with credible sets — small groups of variables, one of which is (with high probability) the true effect — which is especially useful when variables are highly correlated, as in genetic fine-mapping.

Install

pip install pysusie            # once published
# or, from a checkout:
pip install -e .

Dependencies: numpy, scipy, pandas, matplotlib.

Quick start

Individual-level data

import numpy as np
import pysusie as ps

rng = np.random.default_rng(1)
X = rng.standard_normal((400, 200))
beta = np.zeros(200); beta[[10, 50, 120]] = 1.0
y = X @ beta + rng.standard_normal(400)

fit = ps.susie(X, y, L=10)
fit.sets["cs"]                 # credible sets (0-based indices)
fit.pip                        # posterior inclusion probabilities
ps.coef_susie(fit)             # intercept + per-variable effects
ps.susie_plot(fit, "PIP", add_legend=True)

Summary-statistics fine-mapping (genetics)

# z  : p-vector of GWAS z-scores
# R  : p x p LD (correlation) matrix
# n  : sample size
fit = ps.susie_rss(z=z, R=R, n=n, L=10)

# Or from effect estimates and standard errors:
fit = ps.susie_rss(bhat=bhat, shat=shat, R=R, n=n)

Sufficient statistics

ss  = ps.compute_suff_stat(X, y)
fit = ps.susie_suff_stat(ss["XtX"], ss["Xty"], ss["yty"], ss["n"],
                         X_colmeans=ss["X_colmeans"], y_mean=ss["y_mean"])

Public API

• Fitting — susie, susie_suff_stat, susie_rss, susie_rss_lambda.
• Sufficient statistics / univariate — compute_suff_stat (compute_ss), univariate_regression, calc_z.
• Single-effect core — single_effect_regression, single_effect_regression_ss, single_effect_regression_rss.
• Inference — susie_get_cs, susie_get_pip, susie_get_objective, susie_get_posterior_mean, susie_get_posterior_sd, susie_get_lfsr, susie_get_niter, susie_get_prior_variance, susie_get_residual_variance, susie_get_posterior_samples, get_cs_correlation, n_in_CS, in_CS.
• Prediction / utilities — coef_susie, predict_susie, susie_init_coef, susie_prune_single_effects.
• Diagnostics — estimate_s_rss, kriging_rss.
• Plotting — susie_plot.

Numerical parity with susieR

pysusie is a faithful, line-by-line port of the susieR R source. On susieR's bundled N3finemapping fine-mapping benchmark the Python results are numerically identical to susieR 0.14.2:

quantity	susie	susie_rss
PIP Pearson r	1.000 (max abs diff ~1e-11)	1.000 (~1e-9)
per-effect alpha	max abs diff ~1e-11	~1e-9
ELBO	diff ~1e-12	diff ~1e-12
sigma2	diff ~1e-11	exact
credible sets	identical	identical

See tests/test_r_parity.py and examples/compare_R_vs_Python.ipynb.

Tests

python -m pytest tests/ -q

tests/test_smoke.py runs without R; tests/test_r_parity.py compares against susieR (skipped automatically when R / susieR is unavailable).

Relationship to upstream

This port mirrors susieR's R source faithfully: susie.R, susie_rss.R, susie_ss.R, single_effect_regression*.R, elbo*.R, estimate_residual_variance.R, susie_utils.R, cs/credible-set logic and susie_plots.R. The IBSS algorithm, ELBO, prior-variance estimation (EM / optim / uniroot), credible-set purity filtering and the susie_get_* family are all reproduced.