pyglmGamPoi

Pure Python/C++ reimplementation of sctransform::fit_glmGamPoi_offset for trackcell SCTransform v2, without an R runtime.

Architecture

pyglmGamPoi/offset.py          Python API (AnnData-friendly DataFrame output)
        │
pyglmGamPoi/_core.so           pybind11 thin binding (numpy ↔ std::vector)
        │
src/glm_gp_offset.cpp          glm_gp loop per gene (no Rcpp::List)
src/offset_fit.cpp             Newton-Raphson + Brent fallback
src/overdispersion.cpp         Cox-Reid overdispersion MLE (intercept-only)

The C++ core is rewritten from the algorithms in const-ae/glmGamPoi; R headers and Rtatami are not used.

Install (editable)

git clone https://github.com/seqyuan/pyglmGamPoi.git
cd pyglmGamPoi
pip install -e ".[test]"

Requires a C++17 compiler and CMake.

Usage

import numpy as np
from pyglmGamPoi import fit_glmGamPoi_offset

umi = np.random.poisson(3, size=(100, 50)).astype(float)
log10_umi = np.log10(umi.sum(axis=0) + 1e-9)
model_pars = fit_glmGamPoi_offset(umi, log10_umi)

trackcell integration

Replace trackcell.tl._r_sctransform.fit_glmGamPoi_offset_r with:

from pyglmGamPoi import fit_offset_model

See docs.

Status

• Per-gene offset NB fit (design ~ 1 + offset)
• Overdispersion MLE (Cox-Reid; local search + Newton)
• allow_inf_theta / Poisson-boundary theta = Inf
• Cross-gene overdispersion_shrinkage (R glm_gp default)
• R parity on GSE288946 SCT step1 (3 samples, ~2000×2000 each)

Latest release: v0.2.1 — see changelog.

See evaluation docs for full benchmark report.

Docs & benchmark

python scripts/run_all_benchmarks.py    # real + synthetic → docs/_generated/
pytest tests/test_theta_od_parity.py -q

License

MIT (this package). Algorithm reference: glmGamPoi (MIT in recent releases).