What Is GWAS and How Does a Genome-Wide Association Study Work?

Updated

A genome-wide association study (GWAS) tests millions of common genetic variants across many people to find statistical associations with a trait or disease. It answers the question "which positions in the genome are more common in people with this condition (or this measured trait) than in those without it?" — not "which gene causes it." GWAS is the right tool when you have a large sample, a heritable trait, and want to generate hypotheses or identify loci for follow-up; it is not designed to prove causation or to explain an individual's risk on its own.

The Core Idea

Humans differ at millions of single positions in the genome. The most commonly tested are single-nucleotide polymorphisms (SNPs) — a one-letter change at a specific position. In a GWAS you compare the frequency of each SNP between groups:

  • Case–control design: people with a disease vs. people without.
  • Quantitative trait design: a continuous measurement (height, LDL cholesterol, methylation level) across a cohort.

If a SNP is markedly more common in cases, it is associated with the trait. Association is a statistical signal, not a mechanism.

The Standard GWAS Workflow

1. Define the phenotype and design

Choose a clearly measured trait and a sample size large enough to detect realistic effect sizes. Common complex traits typically require thousands to hundreds of thousands of participants; small studies produce unstable signals.

2. Collect DNA and genotype

Genotyping arrays measure a pre-selected set of common SNPs (often 500,000–1,000,000). Whole-genome sequencing is an alternative that also captures rare variants, at higher cost and analytical complexity.

3. Quality control (QC)

Before any association test, remove or flag:

  • Samples with low call rate, extreme heterozygosity, or mismatched sex.
  • SNPs with low call rate, deviation from Hardy–Weinberg equilibrium, or very low minor allele frequency.
  • Related individuals and population outliers (see confounding below).

4. Imputation

Genotyping arrays do not measure every SNP. Imputation uses a reference panel (e.g., 1000 Genomes, Haplotype Reference Consortium) to statistically infer untyped variants, increasing the number of testable SNPs and improving comparability across studies.

5. Association testing

For each SNP, fit a statistical model:

  • Binary trait: logistic regression → odds ratio (OR) per allele.
  • Quantitative trait: linear regression → beta (effect size per allele).

Covariates usually include age, sex, and — critically — principal components of ancestry to control for population stratification.

6. Multiple-testing correction

Testing ~1 million SNPs means many will look "significant" by chance. The conventional threshold is p < 5 × 10⁻⁸ (genome-wide significance), derived from a Bonferroni-style correction for roughly one million independent tests. Results between ~1 × 10⁻⁵ and 5 × 10⁻⁸ are often reported as "suggestive" and need replication.

Reading the Output

Concept What it means Where you see it
Effect size (beta / OR) Magnitude of association per copy of the allele Association results table
p-value Probability of the signal under the null Results table, Manhattan plot
Manhattan plot Genome position (x) vs. −log₁₀(p) (y); peaks = associated loci Standard GWAS figure
QQ plot Observed vs. expected p-values; deviation at the tail = true signal, deviation across the whole line = confounding Standard GWAS figure
Linkage disequilibrium (LD) Correlated nearby SNPs; a "hit" often tags a region, not one causal variant Fine-mapping input

A clean QQ plot hugs the diagonal except at the extreme tail. A QQ plot lifted off the diagonal everywhere usually signals population stratification or another systematic bias, not biology.

Common Pitfalls

  • Population stratification: if cases and controls differ in ancestry, ancestry-linked SNPs appear associated with disease. Principal components or mixed models correct for this.
  • Linkage disequilibrium: the top SNP is usually a marker for a region; identifying the causal variant requires fine-mapping.
  • Association ≠ causation: a GWAS hit may be a bystander correlated with the true cause.
  • Winner's curse: effect sizes in the discovery sample are inflated; replication in an independent cohort is essential.
  • Small samples: underpowered studies produce non-reproducible hits.

What Happens After a GWAS

  1. Replication in an independent cohort.
  2. Fine-mapping to narrow the causal variant within an LD block.
  3. Functional annotation — eQTL analysis, chromatin data, pathway enrichment — to link the variant to a gene and mechanism.
  4. Polygenic risk scores (PRS): aggregate many small-effect variants into a single score; useful for research and, increasingly, risk stratification, though portability across ancestries is limited.
  5. Drug target discovery: GWAS loci that map to druggable genes can nominate new therapeutic targets and support repositioning.

Where Biostatistics Support Fits

GWAS generates large, messy datasets — QC, imputation, covariate modeling, multiple-testing control, and downstream fine-mapping all require careful statistical design. Consultancies such as Acclarogen provide data management and statistical analysis for medical and biological data, serving pharmaceutical companies, biotechnology companies, international consortia, and academic groups. For teams without in-house statistical capacity, that kind of support is most valuable at the design and QC stages, where mistakes are cheapest to fix and most damaging to leave in.

Practical Takeaways

  • GWAS links common variants to traits; it does not establish causation.
  • Genome-wide significance is conventionally p < 5 × 10⁻⁸; replication is mandatory.
  • Always check the QQ plot for confounding before trusting peaks.
  • A hit is a region, not a gene — follow up with fine-mapping and functional work.
  • Sample size and ancestry diversity determine what you can and cannot detect.