Prune variants with LD

Prune Variants With LD

LD pruning removes variants that are highly correlated with variants already kept earlier in a priority order.

magenpy pruning thresholds are based on the absolute Pearson correlation coefficient. The CLI accepts an r^2 threshold and converts it internally to sqrt(r^2).

Prune A Variant File With The CLI

Start with a PLINK-style file containing SNP IDs:

rs9605903
rs5746647
rs16980739

Run pruning:

mgp_prune_ld \
  --ld output/ld_windowed/chr_22/ \
  --variants-file variants.txt \
  --r2-threshold 0.1 \
  --output-file output/pruned_variants.tsv

The script reports how many input variants were present in the LD matrix before pruning. This is useful when the variant list comes from a different reference panel or genome build.

Use A Rank Column

When two variants are in high LD, the earlier/higher-priority variant is kept. If your file has SNP IDs and a rank column, pass both column indices:

mgp_prune_ld \
  --ld "output/ld_windowed/chr_*" \
  --variants-file ranked_variants.tsv \
  --snp-column 0 \
  --rank-column 1 \
  --r2-threshold 0.1 \
  --output-file output/pruned_ranked_variants.tsv

By default, smaller rank values are higher priority. This matches p-values. Use --rank-descending when larger values should be prioritized.

Prune Summary Statistics With The CLI

Summary statistics can be read and harmonized against LD through GWADataLoader:

mgp_prune_ld \
  --ld "output/ld_windowed/chr_*" \
  --sumstats-file gwas.tsv \
  --sumstats-format magenpy \
  --rank-column PVAL \
  --r2-threshold 0.1 \
  --output-file output/pruned_sumstats_variants.tsv

If --rank-column is omitted, the script tries common p-value column names.

Prune With Python

Load the LD matrix and map requested SNPs to LD indices:

import numpy as np
import pandas as pd
import magenpy as mgp

ld = mgp.LDMatrix.from_path("output/ld_windowed/chr_22/")

variants = pd.read_csv("variants.txt", header=None, names=["SNP"])
requested = variants["SNP"].astype(str).tolist()

ld_snps = ld.snps.astype(str).tolist()
snp_to_idx = {snp: i for i, snp in enumerate(ld_snps)}

matched = [snp for snp in requested if snp in snp_to_idx]
variant_order = np.array([snp_to_idx[snp] for snp in matched], dtype=np.int32)

print(f"{len(matched)} of {len(requested)} variants are present in LD")

Prune at r^2 <= 0.1:

corr_threshold = np.sqrt(0.1)

kept = ld.prune(
    threshold=corr_threshold,
    variant_order=variant_order,
    return_value="snps",
)

pd.DataFrame({"SNP": kept}).to_csv(
    "output/pruned_variants_from_python.tsv",
    sep="\t",
    index=False,
)

If you do not provide variant_order, LDMatrix.prune() uses the stored LD matrix order.