py-ldsc

py-ldsc (pyldsc) is a faithful, cleanly-rewritten modern-Python reimplementation of LDSC — LD Score Regression.

It ports the full LDSC toolkit:

• LD Score Regression — Bulik-Sullivan et al., Nature Genetics 2015, "LD Score regression distinguishes confounding from polygenicity in genome-wide association studies".
• Partitioned heritability / stratified LDSC — Finucane et al., Nature Genetics 2015.
• LDSC-SEG (cell-type / tissue-specific enrichment) — Finucane et al., Nature Genetics 2018.

The original ldsc is a Python-2-era command-line tool. pyldsc is a pure, importable, modern-Python package (numpy / scipy / pandas / bitarray — no rpy2) that matches the original numerically. The block jackknife is exact arithmetic and the weighted regressions are deterministic, so estimates agree with ldsc to floating-point tolerance.

Installation

pip install pyldsc

From source:

git clone https://github.com/omicverse/py-ldsc
cd py-ldsc
pip install -e .

Requires Python ≥ 3.9 and numpy, scipy, pandas, bitarray.

What it covers

All the LDSC machinery is reproduced: windowed bias-corrected LD Scores (--ld-wind-cm/-kb/-snps), the regression weights, the two-step estimator, chi-square filtering (--chisq-max), block-jackknife standard errors, the intercept / ratio, constrained intercepts (--intercept-h2, --intercept-gencov, --no-intercept), the --overlap-annot correction and per-category enrichment / coefficient z-scores.

Python API

import pyldsc

# 1. Estimate LD Scores from a PLINK reference panel
res = pyldsc.estimate_ldscore("reference", ld_wind_cm=1.0, out="reference")
res["ldscore"]   # DataFrame: CHR, SNP, BP, L2
res["M"], res["M_5_50"]

# 2. SNP-heritability from GWAS summary statistics
h2 = pyldsc.estimate_h2("trait.sumstats", ref_ld="baseline", w_ld="weights")
h2.tot, h2.tot_se          # heritability +/- jackknife SE
h2.intercept, h2.ratio     # LD Score regression intercept and ratio
print(h2.summary())

# 3. Partitioned / stratified heritability (--overlap-annot)
part = pyldsc.partitioned_h2("trait.sumstats", ref_ld="baseline.",
                             w_ld="weights.", frqfile="frq.")
part.overlap_results       # per-category enrichment / coefficient table

# 4. Genetic correlation between traits
rg = pyldsc.estimate_rg(["trait1.sumstats", "trait2.sumstats"],
                        ref_ld="baseline", w_ld="weights")
rg[0].rg_ratio, rg[0].rg_se, rg[0].p

# 5. LDSC-SEG cell-type / tissue enrichment
seg = pyldsc.ldsc_seg("trait.sumstats", ref_ld_cts="celltypes.ldcts",
                      ref_ld="baseline.", w_ld="weights.")
seg     # DataFrame: Name, Coefficient, Coefficient_std_error, P_value

# 6. Harmonise raw GWAS summary statistics
ss = pyldsc.munge_sumstats("raw_gwas.txt", out="trait", N=100000,
                           merge_alleles="w_hm3.snplist")

Command-line interface

A thin pyldsc CLI mirrors the original ldsc.py flags:

pyldsc l2     --bfile reference --ld-wind-cm 1 --out reference
pyldsc h2     --h2 trait.sumstats --ref-ld baseline --w-ld weights --out trait
pyldsc rg     --rg t1.sumstats,t2.sumstats --ref-ld baseline --w-ld weights
pyldsc h2-cts --h2-cts trait.sumstats --ref-ld-chr-cts celltypes.ldcts \
              --ref-ld baseline. --w-ld weights.
pyldsc munge  --sumstats raw_gwas.txt --N 100000 --out trait

Numerical parity

tests/test_parity.py drives the test data bundled with the upstream ldsc repository through pyldsc and asserts numerical agreement:

• deterministic h2 / intercept / rg / jackknife-SE values on the simulate_test data reproduce to rtol = 1e-6;
• the exact-equality identities from the original suite hold (test_twostep_h2, test_h2_M, test_rg_M, test_read_annot, PLINK .bed parsing);
• the statistical-property checks of ldsc/test/test_sumstats.py (Test_H2_Statistical) hold — averaged over simulated GWAS the estimator is unbiased: mean h2 ≈ 0.9, mean per-category h2 ≈ (0.3, 0.6), intercept ≈ 1.

pytest tests/ -q

Examples

• examples/benchmark.py — timing of the LDSC workflows on the bundled data.
• examples/compare_reference.ipynb — side-by-side comparison of pyldsc against the original LDSC reference values, with a regression-fit plot and an LDSC-SEG cell-type enrichment bar plot.

License

GPL-3.0, matching the upstream ldsc. The original LDSC was written by Brendan Bulik-Sullivan and Hilary Finucane; pyldsc is an independent modern-Python port for the omicverse project.