Genetics
Genome-Wide Association Studies (GWAS)
Scanning millions of SNPs in cases vs controls — Manhattan plots, odds ratios, polygenic risk
A genome-wide association study (GWAS) is an unbiased scan of common genetic variation across the whole genome, testing hundreds of thousands to millions of single-nucleotide polymorphisms (SNPs) one at a time to find which variants differ in frequency between people who have a trait and people who don't. Each SNP gets a per-allele odds ratio and a p-value; results are drawn as a Manhattan plot, and a variant must clear the genome-wide significance threshold of 5×10⁻⁸ — a Bonferroni correction for roughly a million independent tests — to count. Associations flag a tag SNP in linkage disequilibrium with the real causal variant nearby, not the culprit itself. The first clear success, on age-related macular degeneration, was published by Klein and colleagues in 2005; the Wellcome Trust Case Control Consortium's seven-disease study of 14,000 cases followed in Nature in 2007. Summing effects across the genome yields polygenic risk scores, yet common variants still leave much of a trait's heritability unexplained — the missing heritability problem.
- Markers tested~10⁵–10⁷ SNPs per scan
- Significancep < 5×10⁻⁸ (−log10 > 7.3)
- Signature figureManhattan plot
- First successAMD / CFH, Klein et al. 2005
- LandmarkWTCCC 7 diseases, 2007
- Catalog>500,000 associations (GWAS Catalog)
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
Why GWAS matters
- It cracked open complex disease genetics. Before GWAS, common diseases like type 2 diabetes, coronary artery disease, and schizophrenia had almost no reproducible genetic hits. By the mid-2020s the NHGRI-EBI GWAS Catalog logged well over half a million trait-associated variants across thousands of traits, converting a decade of frustrated candidate-gene work into a reproducible map of the genome's contribution to health.
- It is hypothesis-free. A candidate-gene study can only find what you already suspect. GWAS tests every common variant agnostically, so it repeatedly implicates genes and pathways nobody would have nominated — famously pointing at autophagy and the innate-immune sensor NOD2 in Crohn's disease, and at the complement cascade in age-related macular degeneration.
- It nominates drug targets. Drugs whose target has human genetic support from GWAS or related genetics are roughly twice as likely to survive clinical development. PCSK9 inhibitors for cholesterol grew directly out of human genetic variation linking the gene to LDL, and GWAS hits now routinely seed target-validation pipelines.
- It powers polygenic risk scores. Because most complex traits are polygenic, summing thousands of small-effect alleles into a polygenic risk score can identify people at the tails of genetic liability — for coronary artery disease, the top few percent of a PRS carry risk on par with a monogenic familial hypercholesterolemia mutation, a stratification impossible from any single variant.
- It quantifies heritability from unrelated people. Methods like GCTA/GREML estimate how much trait variance is tagged by all SNPs jointly, letting researchers measure heritability without twins or pedigrees and pin down how much of the classic family-study heritability common variants can, in principle, explain.
- It exposes the biology of shared risk. Cross-trait analyses and LD score regression reveal genetic correlations — schizophrenia and bipolar disorder share substantial common-variant architecture, and autoimmune diseases cluster around the MHC — reframing disease categories around underlying genetics rather than symptoms alone.
- It scaled with biobanks. Resources like the UK Biobank (~500,000 participants with genotypes and deep phenotypes) turned GWAS into an industrial pipeline, enabling well-powered scans of quantitative traits from height to blood-cell counts and phenome-wide analyses across thousands of conditions at once.
Common misconceptions
- A GWAS hit is the causal gene. Almost never. The associated marker is a tag SNP in linkage disequilibrium with the real functional variant, which is frequently a non-coding regulatory change tens of kilobases away — and the nearest gene is often not the effector gene. Assigning a hit to biology requires fine-mapping, eQTL/colocalization analysis, and functional experiments, not just reading off the closest annotation.
- The odds ratio tells you your personal risk. A per-allele odds ratio of 1.1–1.3 is typical for common-variant hits and describes a population-average shift in odds, not a deterministic outcome. Individual risk depends on the whole polygenic background plus environment; a single genome-wide significant allele rarely changes an individual's absolute risk meaningfully.
- More SNPs on the chip means more discovery. Discovery is driven by sample size and effect size, not marker density beyond the point where LD is well tagged. Doubling markers on a chip does little; doubling cases does a lot. This is why height GWAS kept finding new loci as cohorts grew into the millions.
- Failing to hit 5×10⁻⁸ means there's no effect. The threshold controls false positives across a million tests, so many true small-effect variants sit just under it. That is precisely why SNP-heritability methods recover far more heritability than the significant hits alone, and why the missing heritability was largely hiding, not absent.
- GWAS results transfer across ancestries. Linkage disequilibrium patterns, allele frequencies, and effect estimates differ by population. A tag SNP validated in Europeans may not tag the same causal variant in African-ancestry genomes, whose LD blocks are shorter, and polygenic risk scores trained in one ancestry lose accuracy in others — a serious equity problem given the historical European skew of GWAS cohorts.
- Association implies the DNA causes the trait, not confounding. Population stratification — systematic ancestry differences between cases and controls — can create spurious associations. This is controlled with principal-component covariates, mixed models, and genomic-control checks; an inflated genomic inflation factor lambda on the Q-Q plot is the red flag that stratification, not biology, is driving signal.
How a GWAS works, step by step
A GWAS starts by defining a phenotype and assembling two groups: cases (people with the disease or an extreme of a quantitative trait) and controls (people without it), ideally matched for ancestry. DNA from every participant is genotyped on a SNP array that assays a chosen panel of a few hundred thousand to a few million tag SNPs — markers picked because, through linkage disequilibrium, each represents a whole haplotype block of correlated variants. Raw genotypes then pass strict quality control: SNPs with low call rates, out-of-Hardy-Weinberg-equilibrium frequencies in controls, or very low minor-allele frequency are dropped, as are samples with excess missingness, sex mismatches, or cryptic relatedness.
To move from the assayed markers to the tens of millions of common variants in the genome, genotypes are imputed against a reference panel such as 1000 Genomes or the Haplotype Reference Consortium: shared haplotypes let the software infer alleles at positions the chip never measured, filling in the gaps between tag SNPs. Ancestry is then estimated by principal-component analysis, and the top components are carried forward as covariates to guard against population stratification — the confounding that arises when cases and controls differ subtly in ancestry.
Now comes the association test itself, performed independently at every SNP. For a binary disease, logistic regression models case/control status against the count of the effect allele (0, 1, or 2 copies under an additive model), adjusting for principal components, age, sex, and batch. The coefficient exponentiates to a per-allele odds ratio with a confidence interval, and a Wald or score test gives a p-value. For a quantitative trait like height or LDL cholesterol, linear regression yields a per-allele effect size (beta) instead. Modern analyses use linear or logistic mixed models (BOLT-LMM, SAIGE, REGENIE) to absorb relatedness and stratification while preserving power.
Because roughly a million effectively independent tests are being run, the multiple-testing burden is enormous, so a SNP must reach the genome-wide significance threshold of 5×10⁻⁸ — 0.05 divided by one million — equivalent to −log10 p greater than 7.3. Results are visualized as a Manhattan plot: chromosomal position on the x-axis, −log10 p on the y-axis, with real associations rising as towers of LD-correlated SNPs above the significance line. A Q-Q plot confirms the null distribution is well-behaved, and the genomic inflation factor lambda flags residual confounding.
A single-study hit is provisional. Robust GWAS demand independent replication in a separate cohort, and most large studies fold discovery and replication together by meta-analysis — combining per-SNP effect estimates across many cohorts with inverse-variance weighting to reach the sample sizes needed for small effects. Finally, significant loci are fine-mapped to narrow the credible set of causal variants, integrated with expression QTL and chromatin data via colocalization, and aggregated into polygenic risk scores that weight each variant genome-wide by its estimated effect and sum across an individual's genotype.
GWAS vs candidate-gene study
| Feature | Genome-wide association study (GWAS) | Candidate-gene study |
|---|---|---|
| Hypothesis | Hypothesis-free — scans the whole genome | Hypothesis-driven — tests pre-chosen genes |
| Markers tested | ~10⁵–10⁷ SNPs genome-wide | A handful of variants in 1–few genes |
| Significance threshold | 5×10⁻⁸ (Bonferroni for ~10⁶ tests) | Often nominal p < 0.05 |
| Sample size needed | Thousands to hundreds of thousands | Hundreds to a few thousand |
| Can find novel loci? | Yes — implicates unsuspected genes | No — limited to the prior guess |
| Replication record | Strong when threshold + replication enforced | Notoriously poor (many false positives) |
| Output | Odds ratio / beta + Manhattan plot per SNP | Effect estimate for chosen variants |
| Main pitfall | Tag SNP ≠ causal variant; stratification | Missing the real gene entirely |
Linkage analysis vs association (GWAS)
| Property | Linkage analysis (family-based) | Association / GWAS (population-based) |
|---|---|---|
| Study units | Pedigrees / affected families | Unrelated cases and controls |
| Signal source | Co-segregation of markers with disease in families | Population-level allele-frequency differences |
| Resolution | Coarse — megabase-scale regions | Fine — locus to single-variant with fine-mapping |
| Best for | Rare, high-penetrance Mendelian variants | Common, small-effect complex-trait variants |
| Recombination used | Meiotic recombination within families | Ancestral recombination captured by LD blocks |
| Classic success | Cystic fibrosis (CFTR), Huntington (HTT) | AMD/CFH, T2D/TCF7L2, height polygenicity |
Famous studies and history
- Klein et al. (2005) — the first GWAS. Studying just 96 age-related macular degeneration cases and 50 controls across ~116,000 SNPs, Robert Klein and colleagues found an overwhelming association at complement factor H (CFH), a common Y402H variant conferring several-fold increased risk. It implicated the complement cascade in AMD and proved the whole-genome scan could work. Published in Science 308: 385–389.
- Wellcome Trust Case Control Consortium (2007). The field-defining study: ~14,000 cases across seven common diseases (bipolar disorder, coronary artery disease, Crohn's disease, hypertension, rheumatoid arthritis, type 1 and type 2 diabetes) versus 3,000 shared controls, ~500,000 SNPs. Published in Nature 447: 661–678, it established the GWAS design, the 5×10⁻⁸ threshold in practice, and rigorous quality control, and delivered dozens of robust loci including NOD2/autophagy in Crohn's and TCF7L2 in type 2 diabetes.
- TCF7L2 and type 2 diabetes. Variants in TCF7L2 emerged as the strongest common-variant risk factor for type 2 diabetes (per-allele odds ratio ~1.4), a locus no candidate-gene program had flagged, illustrating GWAS's power to surprise.
- Height as the polygenic archetype. Height GWAS grew from thousands of participants finding a handful of loci to a 2022 GIANT-consortium meta-analysis of ~5.4 million people identifying over 12,000 near-independent signals that together capture most of the common-variant heritability — the definitive demonstration that a highly heritable trait is spread across thousands of small-effect variants.
- The missing heritability debate (2008–2010). A 2008 Nature commentary crystallized the puzzle: genome-wide significant SNPs explained only a sliver of family-based heritability. Yang, Visscher and colleagues' 2010 GCTA analysis showed that common SNPs considered jointly tagged around 45% of height's heritability — the variants were real but individually below threshold, reframing the problem as heritability hiding below the significance line.
- Schizophrenia at scale. The Psychiatric Genomics Consortium's schizophrenia GWAS grew to over 100 significant loci by 2014 and hundreds more since, including a landmark signal at the MHC region traced to structural variation in complement component 4 (C4) driving synaptic pruning — a striking case of GWAS leading to concrete neurobiology.
Frequently asked questions
What is a genome-wide association study (GWAS)?
A genome-wide association study is an unbiased, hypothesis-free scan of common genetic variation across the entire genome to find which positions differ, statistically, between people who have a trait and people who do not. A genotyping array measures hundreds of thousands to a few million single-nucleotide polymorphisms (SNPs); imputation against a reference panel like 1000 Genomes or the Haplotype Reference Consortium fills the gaps to tens of millions of markers. Each SNP is then tested one at a time — typically by logistic regression for a disease or linear regression for a quantitative trait — asking whether carrying one allele shifts the odds of being a case. The output for a disease is a per-allele odds ratio with a confidence interval and a p-value. Unlike a candidate-gene study, a GWAS makes no prior assumption about which genes matter; it lets the whole genome nominate suspects. The approach was enabled by the HapMap project (2005) and cheap SNP chips, and the first clear success was the 2005 age-related macular degeneration study identifying complement factor H.
Why is genome-wide significance set at 5×10⁻⁸?
Because a GWAS runs about a million effectively independent statistical tests at once, the conventional p < 0.05 threshold would produce roughly 50,000 false positives per study. The 5×10⁻⁸ threshold is a Bonferroni correction: 0.05 divided by one million independent tests. The figure of one million comes from linkage disequilibrium — although chips genotype several million SNPs, blocks of correlated markers behave as single tests, and the number of independent common-variant tests in European-ancestry genomes is estimated at roughly one million. A SNP must therefore reach p < 5×10⁻⁸, equivalently −log10 p > 7.3, to be called a genome-wide significant hit. On a Manhattan plot this line is drawn horizontally and only peaks that punch above it are reported. A suggestive threshold of 1×10⁻⁵ is often shown as a lower line for hypothesis generation, and whole-genome-sequencing studies that assay rarer variants use an even stricter cutoff near 5×10⁻⁹.
What is a Manhattan plot?
A Manhattan plot is the signature figure of a GWAS. The x-axis lays out the genome from chromosome 1 to 22 (then X), position by position; the y-axis is the −log10 of each SNP's association p-value, so smaller p-values rise higher. Each dot is one SNP, and colours alternate by chromosome to make the skyline readable — the resemblance to the Manhattan skyline is where the name comes from. Because a real association drags along dozens of neighbouring SNPs in linkage disequilibrium, true signals appear as towers of stacked points rather than lone spikes, which helps distinguish signal from noise. A horizontal line at −log10(5×10⁻⁸) = 7.3 marks genome-wide significance; towers above it are the loci reported. A companion quantile-quantile (Q-Q) plot checks that the bulk of p-values follow the null expectation, with the genomic inflation factor lambda flagging confounding from population stratification when it drifts above 1.
What is a tag SNP and how does linkage disequilibrium affect GWAS?
Linkage disequilibrium (LD) is the non-random co-inheritance of nearby alleles: because recombination is rare over short distances, variants sitting close together on a chromosome tend to travel as a haplotype block. A tag SNP is a single marker chosen to represent a whole block — genotyping it tells you, with high probability, the alleles of every correlated variant around it. This is what makes GWAS affordable: a chip of a few hundred thousand well-chosen tag SNPs captures most common variation without sequencing every base. The flip side is that a significant association almost never points at the causal variant itself. The tag SNP is merely in LD with the true functional change, which may be a regulatory variant tens of kilobases away or in a different gene. Turning an associated locus into a mechanism requires fine-mapping, expression quantitative trait locus (eQTL) analysis, and functional follow-up. LD also differs by ancestry, which is why a tag SNP validated in Europeans may fail to capture the same signal in African-ancestry genomes with shorter LD blocks.
What is a polygenic risk score?
A polygenic risk score (PRS), or polygenic score, collapses the effects of thousands to millions of variants into a single number predicting an individual's genetic liability for a trait. It is computed by taking each person's genotype at every relevant SNP, weighting it by the effect size (the log odds ratio) estimated from a large GWAS, and summing across the genome. Most heritable traits are highly polygenic, so no single variant predicts much, but the aggregate can be informative: for coronary artery disease, individuals in the top few percent of a PRS carry a threefold-or-higher risk, comparable to a monogenic familial hypercholesterolemia mutation. PRS are already used to stratify screening and are being trialed clinically. Two major caveats limit them: prediction accuracy still falls short of total heritability, and because most GWAS were done in European-ancestry cohorts, scores transfer poorly to other populations, raising real equity concerns if deployed without diverse training data.
What is the missing heritability problem?
Twin and family studies estimate that human height is roughly 80% heritable, yet the genome-wide significant SNPs from early GWAS explained only about 5% of the variance — the gap between family-based heritability and the variance captured by discovered variants is the missing heritability problem, named in a 2008 Nature commentary. Several factors fill the gap. Many causal variants have effects too small to cross 5×10⁻⁸ in a modest sample, so ever-larger cohorts keep uncovering them: height GWAS now number millions of participants and capture most of the common-variant heritability. Rare variants with larger effects are missed by common-SNP chips and need sequencing. Gene-by-gene and gene-by-environment interactions, structural variants, and inflated twin-study estimates all contribute. Methods such as GCTA that estimate heritability tagged by all SNPs together (SNP heritability) closed much of the gap, showing the variants were there but individually undetectable — a phenomenon nicknamed heritability that was hiding, not truly missing.
How is a GWAS different from a candidate-gene study?
A candidate-gene study picks a handful of genes the investigators already suspect — based on biology, a pathway, or a prior finding — and tests variants only in those genes. It is cheap and focused but hostage to the guess: if the real culprit lies in an unsuspected gene, the study cannot find it, and the small-sample candidate-gene literature of the 1990s and 2000s is now notorious for false positives that failed to replicate. A GWAS is the opposite: it tests the whole genome agnostically, so it can implicate genes nobody would have nominated, and its punishing 5×10⁻⁸ threshold plus mandatory independent replication make surviving hits far more robust. The cost is statistical power — testing a million SNPs demands tens of thousands of samples — and interpretive work, since an unbiased hit lands on an anonymous tag SNP that must be fine-mapped back to a mechanism. In practice GWAS has largely replaced candidate-gene designs for common complex traits, while candidate approaches remain useful for targeted follow-up of a known locus.