PLINK2 — GWAS and Population Genetics

Overview

PLINK2 is the high-performance successor to PLINK 1.9, designed for genome-wide association studies (GWAS) and population genetics analysis on large cohorts. It processes genotype data in PLINK binary format (.bed/.bim/.fam), VCF, and BGEN formats — performing sample and variant quality control (QC), kinship estimation, principal component analysis (PCA), and linear/logistic regression association testing. PLINK2 is 10–100× faster than PLINK 1.9 on most tasks due to multithreading and optimized I/O. Output files are compatible with downstream visualization (Manhattan/QQ plots) and meta-analysis tools.

When to Use

• Running GWAS on a case-control or quantitative trait cohort after genotyping array QC
• Performing sample QC: missingness, heterozygosity outliers, sex check, cryptic relatedness
• Computing genome-wide LD pruning for PCA or relatedness estimation
• Running PCA on genotype data to identify population stratification
• Converting between PLINK binary, VCF, and BGEN formats
• Filtering variants by MAF, HWE, missingness, or INFO score in VCF/imputed data
• Use omics-plotting SKILL to render Manhattan and QQ plots from the association results
• Use regenie or SAIGE instead for biobank-scale GWAS (>100k samples) requiring mixed model association to control for population structure
• Use VCFtools as an alternative for VCF-specific population genetics statistics

Prerequisites

• Software: PLINK2 (pre-compiled binary; no pip/conda package)
• Input: PLINK binary files (.bed/.bim/.fam) or VCF/BGEN from array genotyping or imputation

Check before installing: The tool may already be available (e.g., inside a pixi / conda env). Always run command -v plink2 first and skip the install block if it returns a path. When executing tools inside a pixi project, prefer pixi run <tool> over plain <tool>.

# Skip install if already present
if command -v plink2 >/dev/null 2>&1; then
    echo "plink2 already installed: $(plink2 --version)"
else
    # Download PLINK2 pre-compiled binary (Linux)
    wget https://s3.amazonaws.com/plink2-assets/alpha6/plink2_linux_avx2_20241112.zip
    unzip plink2_linux_avx2_20241112.zip
    chmod +x plink2
    export PATH="$PWD:$PATH"

    # macOS
    # wget https://s3.amazonaws.com/plink2-assets/alpha6/plink2_mac_20241112.zip
    # unzip plink2_mac_20241112.zip

    plink2 --version
    # PLINK v2.00a6LM
fi

# Python for downstream analysis
pip install pandas numpy matplotlib scipy

Pre-flight Interview

Settle these with the user before writing any analysis code.

decisions:
  - id: D1
    param: phenotype
    kind: required
    source: data
    ask: "Which column holds the trait being tested, and is it a measurement or a case/control status?"
    default: null

  - id: D2
    param: regressionMode
    kind: derived
    source: data
    depends_on: [D1]
    ask: "Should the association test be linear or logistic?"
    default: "linear for a quantitative trait, logistic for case/control"

  - id: D3
    param: qcThresholds
    kind: required
    source: user
    ask: "Where should variants and samples be cut for allele frequency, missingness, and Hardy-Weinberg departure?"
    default: "MAF 0.01, variant missingness 0.05, sample missingness 0.02, HWE 1e-6"

  - id: D4
    param: covariates
    kind: required
    source: data
    ask: "Which covariates belong in the model - age, sex, batch, study site?"
    default: null

  - id: D5
    param: populationStructure
    kind: required
    source: user
    ask: "How many principal components should be included to absorb ancestry differences?"
    default: "10, computed from LD-pruned variants"

  - id: D6
    param: ldPruning
    kind: optional
    source: user
    depends_on: [D5]
    ask: "Which window and correlation threshold should thin variants before computing the components?"
    default: "50 variant window, step 5, r-squared 0.2"

  - id: D7
    param: threads
    kind: never_ask
    source: data
    reason: "Affects runtime only, not the association statistics"
    default: "min(8, available_cores)"

D5 hangs on D4 because the components are themselves covariates - how many to include is a question about the covariate set, not a separate knob. Omitting them entirely is the classic way to produce a genome-wide significant result that reflects ancestry rather than the trait.

Quick Start

# Run GWAS: linear regression for quantitative trait
plink2 \
    --bfile cohort_qc \
    --pheno phenotypes.txt \
    --covar covariates.txt \
    --linear hide-covar \
    --out results/gwas_result \
    --threads 8

# View top hits
head results/gwas_result.*.glm.linear | sort -k12,12g | head -20

Workflow

Step 1: Convert VCF/Imputed Data to PLINK Binary Format

Convert input genotype data to PLINK binary format for fast processing.

# Convert VCF to PLINK binary
plink2 \
    --vcf cohort_genotyped.vcf.gz \
    --make-bed \
    --out cohort_plink \
    --threads 8

# Convert BGEN (imputed data from Michigan/TopMed imputation server)
plink2 \
    --bgen cohort_imputed.bgen ref-first \
    --sample cohort_imputed.sample \
    --make-bed \
    --out cohort_imputed_plink \
    --threads 8

echo "Files created:"
ls -lh cohort_plink.{bed,bim,fam}
echo "Samples: $(wc -l < cohort_plink.fam)"
echo "Variants: $(wc -l < cohort_plink.bim)"

Step 2: Sample QC — Missingness and Heterozygosity

Remove samples with high missingness or heterozygosity outliers.

# Compute per-sample and per-variant missingness
plink2 \
    --bfile cohort_plink \
    --missing \
    --out qc/sample_missingness \
    --threads 8

# Remove samples with > 2% missingness and variants with > 5% missing
plink2 \
    --bfile cohort_plink \
    --mind 0.02 \
    --geno 0.05 \
    --make-bed \
    --out cohort_sample_qc \
    --threads 8

echo "After sample QC:"
echo "Samples: $(wc -l < cohort_sample_qc.fam)"
echo "Variants: $(wc -l < cohort_sample_qc.bim)"

Step 3: Variant QC — MAF, HWE, and INFO Score Filtering

Filter variants by minor allele frequency, Hardy-Weinberg equilibrium, and imputation quality.

# Variant QC: MAF, HWE, missingness
plink2 \
    --bfile cohort_sample_qc \
    --maf 0.01 \
    --hwe 1e-6 \
    --geno 0.05 \
    --make-bed \
    --out cohort_variantqc \
    --threads 8

echo "After variant QC:"
echo "Variants remaining: $(wc -l < cohort_variantqc.bim)"

# For imputed data: also filter by INFO score (using VCF INFO field)
# plink2 --bfile cohort_variantqc --var-min-qual 0.8 --make-bed --out cohort_info_qc

Step 4: LD Pruning and PCA for Population Stratification

Compute principal components from LD-pruned variants.

# LD pruning: remove one variant from each pair with r² > 0.2
plink2 \
    --bfile cohort_variantqc \
    --indep-pairwise 50 5 0.2 \
    --out qc/ld_pruned \
    --threads 8

# Compute PCA on LD-pruned variants (top 20 PCs)
plink2 \
    --bfile cohort_variantqc \
    --extract qc/ld_pruned.prune.in \
    --pca 20 \
    --out pca/cohort_pca \
    --threads 8

echo "PCA files: pca/cohort_pca.eigenvec (PCs) and pca/cohort_pca.eigenval (variance)"
head pca/cohort_pca.eigenvec

Step 5: Run GWAS Association Analysis

Perform logistic regression (case-control) or linear regression (quantitative trait).

# Case-control GWAS: logistic regression with covariates
plink2 \
    --bfile cohort_variantqc \
    --pheno phenotypes.txt \
    --1 \
    --covar covariates.txt \
    --covar-variance-standardize \
    --logistic hide-covar \
    --ci 0.95 \
    --out results/gwas_cc \
    --threads 8

# Quantitative trait GWAS: linear regression
plink2 \
    --bfile cohort_variantqc \
    --pheno phenotypes.txt \
    --covar covariates.txt \
    --covar-variance-standardize \
    --linear hide-covar \
    --ci 0.95 \
    --out results/gwas_qt \
    --threads 8

echo "GWAS complete. Files: results/gwas_*.glm.*"
wc -l results/gwas_qt.*.glm.linear

Step 6: Plot Manhattan and QQ Plots

Prepare the association results as a CHR/BP/P table, then read skills/data-visualization/omics-plotting/SKILL.md and follow its "Manhattan" and "QQ plot" recipes on the exported CSV (→ figures/manhattan.png, figures/qq.png).

import pandas as pd

# Load GWAS results (PLINK2 linear output) as a CHR/BP/P table for plotting
df = pd.read_csv("results/gwas_qt.PHENO1.glm.linear", sep="\t",
                 usecols=["#CHROM", "POS", "ID", "P"])
df.columns = ["CHR", "BP", "SNP", "P"]
df = df.dropna(subset=["P"])
df["P"] = pd.to_numeric(df["P"], errors="coerce")
df = df[df["P"] > 0].copy()
df.to_csv("gwas.csv", index=False)

print(f"Variants: {len(df)}")
print(f"Genome-wide significant (p<5e-8): {(df['P'] < 5e-8).sum()}")
# Render with the omics-plotting SKILL (`skills/data-visualization/omics-plotting/SKILL.md`) "Manhattan" and "QQ plot" recipes (input columns CHR, BP, P).

Key Parameters

Parameter	Default	Range/Options	Effect
--maf	—	0–0.5	Minor allele frequency threshold; 0.01 removes singletons
--hwe	—	0–1	Hardy-Weinberg equilibrium p-value threshold; 1e-6 typical
--geno	—	0–1	Maximum variant missingness; 0.05 = 5%
--mind	—	0–1	Maximum sample missingness; 0.02 = 2%
--linear / --logistic	—	flag	Regression mode: linear for quantitative, logistic for case-control
--covar	—	file path	Covariate file (FID IID COV1 COV2...); add PCs here
--pca	—	integer	Number of PCs to compute from LD-pruned data
--indep-pairwise	—	window step r²	LD pruning: window size (variants), step, r² threshold
--threads	1	1–64	CPU threads; 8–16 typical for cluster
--ci	—	0–1	Confidence interval for effect size output (0.95 for 95% CI)
--1	off	flag	Recode phenotype: 1=control, 2=case → 0=control, 1=case (logistic)

Common Recipes

Recipe 1: Relatedness QC and Remove One from Each Related Pair

# Compute kinship coefficients (King-robust estimator)
plink2 \
    --bfile cohort_variantqc \
    --extract qc/ld_pruned.prune.in \
    --king-cutoff 0.0625 \
    --out qc/kinship \
    --threads 8

# Remove related individuals (king-cutoff auto-generates exclusion list)
plink2 \
    --bfile cohort_variantqc \
    --remove qc/kinship.king.cutoff.out.id \
    --make-bed \
    --out cohort_unrelated \
    --threads 8

echo "After relatedness QC: $(wc -l < cohort_unrelated.fam) samples"

Recipe 2: Parse GWAS Results and Report Top Hits

import pandas as pd
import numpy as np

def load_gwas_results(filepath: str, p_threshold: float = 5e-8) -> pd.DataFrame:
    """Load PLINK2 GWAS linear/logistic output and return genome-wide significant hits."""
    df = pd.read_csv(filepath, sep="\t",
                     usecols=["#CHROM", "POS", "ID", "REF", "ALT", "A1", "BETA", "SE", "P"])
    df.columns = ["CHR", "POS", "SNP", "REF", "ALT", "A1", "BETA", "SE", "P"]
    df = df.dropna(subset=["P"])
    df["P"] = pd.to_numeric(df["P"], errors="coerce")
    sig = df[df["P"] < p_threshold].sort_values("P")
    return sig

# Load results for one phenotype
hits = load_gwas_results("results/gwas_qt.PHENO1.glm.linear")
print(f"Genome-wide significant loci: {len(hits)}")
print(hits.head(10)[["CHR", "POS", "SNP", "BETA", "SE", "P"]])
hits.to_csv("gwas_top_hits.tsv", sep="\t", index=False)

Expected Outputs

Output	Format	Description
*.glm.linear	TSV	Linear GWAS results: CHR, POS, ID, BETA, SE, P per variant
*.glm.logistic.hybrid	TSV	Logistic GWAS results: OR, 95% CI, P per variant
*.eigenvec	TSV	PCA loadings: FID IID PC1 PC2 ... PC20
*.eigenval	Text	Eigenvalues for PCA variance explained
*.king.cutoff.out.id	TSV	Sample IDs to remove for relatedness QC
*.bed/.bim/.fam	Binary	PLINK binary format after QC filtering

Troubleshooting

Problem	Cause	Solution
Error: No samples left after --mind filter	Missingness threshold too strict	Relax to --mind 0.05; check input data quality
Logistic regression fails to converge	Rare variant, few cases, or collinear covariates	Use --logistic firth for Firth regression; remove correlated covariates
PCA shows clear outlier cluster	Admixed population or sample contamination	Remove outliers using PC cutoffs; check ancestry with 1000 Genomes reference panel
--hwe removes too many variants	Population structure inflating HWE test	Apply HWE filter to controls only: --hwe 1e-6 ctrl-only
BGEN import fails	Wrong ref-first/ref-last flag	Check imputation server documentation; try --bgen file.bgen ref-last
Very slow association test	Too many covariates or large file	Pre-filter to GWAS-significant window; use --threads 16
Sex mismatch warnings	X chromosome heterozygosity outside sex-specific thresholds	Run --check-sex and remove F-statistic outliers
Memory error on large dataset	RAM insufficient for 500k+ variants	Use --memory 32000 to cap RAM; split by chromosome

References

• PLINK2 documentation — official command reference and tutorials
• PLINK2 GitHub: chrchang/plink-ng — source code and releases
• Chang CC et al. (2015) "Second-generation PLINK: rising to the challenge of larger and richer datasets" — GigaScience 4:7. DOI:10.1186/s13742-015-0047-8
• GWAS best practices (Marees et al. 2018) — comprehensive QC and analysis guide