py-twosamplemr

Pure-Python two-sample Mendelian randomization — a faithful port of the statistical-methods core of the R/CRAN package MendelianRandomization (Yavorska & Burgess, Int. J. Epidemiol. 2017) together with the harmonisation / diagnostics workflow of TwoSampleMR (Hemani et al., eLife 2018).

• PyPI dist name: pytwosamplemr  |  import name: pytwosamplemr
• Pure Python — numpy / scipy / pandas / matplotlib. No rpy2.
• Numerical parity with R is the design goal. Closed-form estimators reproduce MendelianRandomization 0.10.0 bit-for-bit (relative difference < 1e-5); bootstrap-based estimators agree to within Monte-Carlo error.

Installation

pip install pytwosamplemr
# or, from a checkout:
pip install -e .

Quick start

import pytwosamplemr as mr
from pytwosamplemr.datasets import ldl_chd_input

# the MendelianRandomization bundled lipid GWAS example (28 SNPs)
obj = ldl_chd_input()

mr.mr_ivw(obj)        # inverse-variance weighted
mr.mr_egger(obj)      # MR-Egger regression (+ pleiotropy intercept)
mr.mr_maxlik(obj)     # maximum likelihood
mr.mr_divw(obj)       # debiased IVW
mr.mr_conmix(obj)     # contamination mixture
mr.mr_lasso(obj)      # MR-Lasso post-selection
mr.mr_cml(obj, n=17723)   # constrained maximum likelihood

# panel + diagnostics
mr.mr_allmethods(obj, method="main")
mr.mr_heterogeneity(obj)
mr.mr_pleiotropy_test(obj)
mr.mr_singlesnp(obj)
mr.mr_leaveoneout(obj)

# plots
mr.mr_scatter(obj)    # SNP effects + fitted method lines
mr.mr_forest(obj)
mr.mr_funnel(obj)
mr.mr_loo(obj)

What is implemented

Input

Estimators (full coverage)

Diagnostics / workflow (TwoSampleMR side)

harmonise_data · mr_heterogeneity · mr_pleiotropy_test · mr_steiger / directionality_test · mr_singlesnp · mr_leaveoneout

Plots

mr_plot / mr_scatter · mr_forest · mr_funnel · mr_loo

R parity

py-twosamplemr was validated against MendelianRandomization 0.10.0 on the package's own bundled lipid GWAS example data. Closed-form estimators agree with R to machine precision:

A standout detail: mr_maxlik calls R's optim(method="Nelder-Mead"), which on this rough 29-dimensional likelihood stops at a non-global local point — so reproducing R requires a line-for-line port of R's nmmin C routine (pytwosamplemr/_nmmin.py). With that port the maximum-likelihood estimate matches R to 12 significant figures.

Bootstrap-based estimators (mr_median, mr_mbe, mr_cml with data perturbation) cannot be bit-exact across RNGs: their point estimates still match R exactly (the median / mode location does not depend on the bootstrap), and the bootstrap SE agrees to within ~10%.

The R-parity test suite (tests/test_r_parity.py) runs the R reference driver live and asserts these tolerances; it skips gracefully when R is unavailable.

Testing

python -m pytest tests/ -q

License

GPL-3, matching the upstream MendelianRandomization package (licensed GPL-2 | GPL-3).

References

• Yavorska OO, Burgess S. MendelianRandomization: an R package for performing Mendelian randomization analyses using summarized data. Int J Epidemiol. 2017;46(6):1734-1739.
• Hemani G, et al. The MR-Base platform supports systematic causal inference across the human phenome. eLife. 2018;7:e34408.