Running LD the Right Way: PLINK Workflow, Parameters, and LD Pruning

Need a reliable PLINK LD workflow you can trust across projects? This practical guide focuses on a production-ready LD pruning pipeline with sensible defaults, robust pairwise r2 checks, and clear sliding-window LD settings. Built for linkage disequilibrium analysis at scale, we keep theory to a minimum and zero in on execution: inputs, commands, tuning on dense data, and reviewer-ready outputs.

Quick Start Defaults for PLINK LD

When you need defensible settings fast:

• Pruning goal: r² ≈ 0.2
• Windowing: 250 kb OR 50 SNPs (whichever first), step 5–10 SNPs
• MAF for stability: start at 0.05
• Strata: compute LD within ancestry groups
• Imputation: filter on INFO/R² before LD
• Outputs: pruned marker list, small LD matrices for QA, decay plot, parameter manifest

Inputs & QC Before LD

Garbage in, garbage out. LD calculations amplify upstream errors because pairwise r2 is sensitive to missingness, batch effects, and subtle stratification. Lock down a simple QC gate before you touch LD.

Minimal, executable QC (arrays/WGS)

• Sample call rate: --mind 0.02
• Marker call rate: --geno 0.02 (tighten to 0.01 for final runs)
• MAF for LD stability: --maf 0.05
• Hardy–Weinberg by ancestry: --hwe 1e-6 midp
• Relatedness: remove up to 2nd-degree (use your preferred kinship tool prior to PLINK)
• Structure: subset or label by ancestry (PCA/labels), then run LD per stratum
• Imputed data: apply INFO/R² filters upstream; only then compute LD

Distribution of allele frequencies between data sources. (Pengelly R.J. et al. (2015) BMC Genomics)

Executable QC command

plink --bfile INPUT \

--geno 0.02 --mind 0.02 \

--maf 0.05 \

--hwe 1e-6 midp \

--make-bed --out 01_qc

Batch harmonization checklist (before merging cohorts)

• Build/coordinate alignment: confirm all batches use the same reference (GRCh37/38); liftover if needed.
• Allele consistency: recode to reference alleles; resolve strand flips (A/T, C/G sites need extra care).
• Duplicates/conflicts: remove duplicate rsIDs or colliding coordinates so each variant is unique.
• Batch labels: keep a batch field in metadata for stratified QA.
• Ancestry labels: prefer self-reported ancestry plus PCA confirmation; compute LD within strata.

Why these thresholds and checks matter is explained in the QC Gate section of Designing an LD Study.

The PLINK LD Workflow: From BED to Pruned Set

1) Standardise inputs and produce the QC'd dataset

• Harmonise alleles/IDs if merging batches.
• Split or label by ancestry to keep LD homogeneous.

plink --bfile INPUT \

--geno 0.02 --mind 0.02 \

--maf 0.05 --hwe 1e-6 midp \

--make-bed --out 01_qc

2) Prune with a sliding window

--indep-pairwise <kb> <stepSNPs> <r2> removes correlated markers as the window slides.

# Default for many GWAS QC tasks

plink --bfile 01_qc \

--indep-pairwise 250 5 0.2 \

--out 02_prune

# Keep only the independent set

plink --bfile 01_qc \

--extract 02_prune.prune.in \

--make-bed --out 03_pruned

3) Inspect LD and export matrices (targeted QA)

# Inspect local LD (pairwise r2) by distance-capped window

plink --bfile 01_qc \

--r2 gz yes-really \

--ld-window-kb 250 --ld-window 99999 --ld-window-r2 0.2 \

--out 04_ld_kb

# On dense WGS, cap comparisons by variant count

plink --bfile 01_qc \

--r2 gz yes-really \

--ld-window 50 --ld-window-r2 0.2 \

--out 04_ld_count

Choosing LD inspection modes (kb vs count)

• Use --ld-window-kb when you want to interpret LD as a function of physical distance (ideal for decay curves and reports).
• Use --ld-window (count cap) in dense WGS to limit comparisons, control runtime, and reduce file size.
• Add --r2 gz yes-really to compress outputs for easier sharing and archiving.
• Name files to reflect chromosome/stratum/window strategy (e.g., 04_ld_EUR_chr10_kb250) for traceability.

Schematic illustration for large-scale linkage disequilibrium (LD) analysis as exampled for CONVERGE cohort. (Huang X. et al. (2023) eLife)

4) Make the run reproducible

• Containerise PLINK (Docker/Singularity) and pin versions.
• Log commands, seeds, input checksums, and timestamps.
• Parallelise safely by chromosome to speed runs.
• Name files predictably (e.g., 01_qc / 02_prune / 03_pruned / 04_ld_*).

Execution Tuning: Windows, Steps, Thresholds, and Dense Data

Performance & scaling (practical guidance)

• Runtime rises with marker density and window length, and falls as step increases. Dense WGS will run slower than arrays under identical settings.
• Pilot on 1–2 chromosomes: compare step=5 vs 10 and 250 kb vs 150 kb; choose the best trade-off in retained SNPs, stability, and time.
• In very dense regions, enable count caps (e.g., --ld-window 50) and per-chromosome parallelism; record thread counts and I/O limits (shared storage can bottleneck).
• Monitor: retained SNP ratio per chromosome, peak RAM, and time per million comparisons; non-monotonic patterns usually flag missingness or batch issues.

r² thresholds (execution view)

• Start pruning at 0.2. For dense or highly structured regions, test 0.1 and 0.3 on one chromosome and compare retained SNPs and downstream model stability.
• For tagging, use ≥0.8 (fine-mapping can push ~0.9 in target regions). Keep tagging logic in your panel-design pipeline; avoid mixing with pruning defaults.

Window size and step (distance vs density)

• Default hybrid: 250 kb or 50 SNPs; step 5–10 SNPs.
• High recombination: reduce to 100–150 kb to avoid mixing distinct blocks.
• Sparse arrays: SNP-count windows stabilise comparisons when spacing is uneven.
• Dense WGS: always cap by count for LD inspection; tile by chromosome for parallelism.

Runtime levers and parallelism

• Step size is the main throttle; doubling often halves comparisons with little effect on pruning results.
• Use --threads judiciously; avoid I/O contention on shared filesystems.
• Split by chromosome; aggregate reports afterward.

Imputed and mixed data

• Filter imputed variants by INFO/R² first; then compute LD on the filtered set.
• If you must mix sources, validate LD patterns on raw versus imputed subsets during QA and only then pool.

Edge regions and practical checks

• Long-range LD areas (e.g., MHC) can dominate summaries. Exclude during genome-wide pruning and report separately with context.
• Chromosome X needs sex-aware handling; check PLINK flags for X-specific QC.

Regions with atypical LD

• Long-range LD: handle as above—separate analysis and reporting.
• Inversions/selection signals: expect extended LD; review by segment and note plausible causes (structural variation, local selection, demographic history).
• Runs of homozygosity (ROH): large ROH inflate local correlation, especially in inbred/related samples; consider removing those samples or masking ROH for global summaries.

Reporting That Reviewers Accept

Deliver an output bundle that collaborators can reuse and journals can audit.

Include

• Pruned set (.bed/.bim/.fam or variant list) with build/versions noted.
• LD decay curves per stratum (annotate MAF, r², window, and step).
• LD heatmaps for representative regions with consistent r² colour scales.
• Tag coverage tables (if tagging): r² thresholds, proxies covered, per-region summaries.
• Parameter manifest: all QC gates, windows, steps, thresholds; PLINK version; container image/tag; command history; seeds; input checksums.
• Reproducibility bundle: scripts or a Makefile/Snakemake definition.

Folder & file schema (delivery suggestion)

• 00_meta/: manifest.yaml, command logs, container/version info, input summary (samples, variants, build).
• 01_qc/: pre/post QC stats (--freq/--missing) and PCA coordinates (for ancestry verification).
• 02_prune/: *.prune.in/out, retained set bed/bim/fam.
• 03_ld/: representative LD matrices (.gz), decay raw data and plots.
• 04_reporting/: heatmaps, coverage tables (CSV/TSV), and a one-pager methods/parameters sheet.

Acceptance checks (quick sign-off)

• Retained SNP proportion aligns with expectations (too low suggests over-pruning or strict settings).
• Mean/median r² in the retained set drops versus baseline; model VIF improves; decay curves look plausible per ancestry.
• Figures label MAF, r², window, and step; colour scales remain consistent across plots.

Various linkage disequilibrium (LD) components for the 26 1KG cohorts. (Huang X. et al. (2023) eLife)

Avoid

• Over-pruning that erases structure in gene-dense areas.
• Pooled multi-ancestry LD without stratum analysis.
• LD computed on unfiltered imputed calls.

FAQs

What are pragmatic PLINK LD parameters for fast, stable pruning?

Use --indep-pairwise 250 5 0.2 as a dependable baseline. On dense WGS, trial a count-based variant (50 5 0.2) on one chromosome and compare retained markers and runtime before scaling.

How do I set sliding windows: kb or SNP count—and can I combine them?

--indep-pairwise takes kb, step (SNPs), and r². For inspection via --r2, you can cap by kb (--ld-window-kb) or by SNP count (--ld-window). In practice, use a hybrid approach: kb for biological distance, count to limit compute on dense data.

How can I limit comparisons on dense WGS and still keep signal?

Increase the step from 5 to 10, cap by SNP count in LD inspection, and parallelise by chromosome. These changes reduce compute cost with little impact on pruning quality.

How do I make runs fully reproducible across machines and analysts?

Containerise PLINK, pin versions, log commands and seeds, store input checksums, and ship a parameter manifest. Keep file naming deterministic and avoid manual edits to intermediate files.

What should I do about the MHC and other long-range LD regions?

Inspect them separately. Consider region-specific thresholds or report them as special cases so they don't dominate genome-wide summaries.

Can I mix raw genotypes and imputed variants in LD calculations?

Yes—but only after strict imputation-quality filtering. Validate LD patterns separately on raw and imputed subsets during QA, then pool if they agree.

Do I need to publish the parameter manifest?

It's recommended. The manifest records key plink ld and ld pruning pipeline parameters (MAF, r², windows, step, versions/containers, ancestry stratification, special-region handling), which helps peer review and ensures internal reruns remain consistent.

Conclusion

A disciplined ld pruning pipeline makes PLINK LD a dependable foundation for GWAS, structure analysis, and tag SNP panel design. Keep the QC gate executable and light, prune with a sliding window that respects both distance and density, and cap comparisons on dense WGS. Run a small pilot per chromosome to tune step and windows, then lock everything in a manifest so teammates and reviewers can reproduce your work.

From pipeline to downstream

• GWAS/confounding control: after linkage disequilibrium analysis, use the pruned set for PCA/GRM modelling; run association on the full set with reduced collinearity.
• PRS/panel design: for PRS or targeted panels, combine --show-tags (e.g., r²≥0.8) with ancestry-specific evaluation to improve transferability.
• Traceability: archive the pruned ID list with the manifest as the "data lineage" anchor for all downstream models.

Related reading

• Practical Guide: Designing an LD Study (MAF, r² Thresholds, Sample Size)
• LD Decay and Haplotype Blocks: Interpreting Curves for Marker Strategy
• Reporting LD Results That Reviewers Accept: Plots, Thresholds, Pitfalls

CD Genomics provides research-use services for institutions and companies. We do not offer personal or clinical testing.