Quick answer: A genome-wide association study (GWAS) asks a simple question across an enormous scale: at which genetic positions do people carrying one version of the variant differ — on average — from carriers of the other version, for a trait you care about? The output is a giant table of variants, effect estimates, and p-values, most usefully summarized by the Manhattan plot. A GWAS result is always a population-level statistical association, not a deterministic genetic verdict, and the classic beginner mistakes are mistaking a strong-looking p-value for a large effect or forgetting that population structure can generate entire forests of false signals.
Natural companion to the biostatistics series already on this blog: GWAS is hypothesis testing applied thousands of times at once, with everything that implies about multiple testing and confounding.
Table of Contents
The GWAS question, concretely
Take a trait — type 2 diabetes risk, height, milk yield in cattle. At a single variant, say rs123456, with alleles A and G, GWAS asks: among many people, does genotype G statistically associate with a different phenotype level than genotype A? The estimation is a regression: phenotype as outcome, genotype coded 0/1/2 alleles as the predictor, plus covariates. Repeat for millions of variants and you get a GWAS.
The core mechanics:
- Input data: hundreds to hundreds of thousands of genomes with SNPs genotyped, plus phenotype and covariates (age, sex, batch, ancestry components).
- Single-variant tests: for each marker — usually a linear or logistic regression, or a linear-mixed-model surrogate when relatedness matters — an effect size beta and a p-value.
- Local significance: the multiple-testing burden across millions of tests produces the “genome-wide significance” threshold p < 5 × 10⁻⁸: the value at which you would expect roughly one false positive per scan.
- Locus, not mutation: an associated SNP tags a region — the causal variant may be a nearby variant correlated with it in the population (linkage disequilibrium). A lead SNP’s position is a starting point for fine-mapping, not an answer.
This pattern tells you the three attributes a GWAS locus has: statistical signal, population-level effect estimate, and a probability that which variant is causal remains genuinely uncertain.
Reading a Manhattan plot without folklore
The plot’s x-axis is the genome marching along chromosomes; the y-axis is −log₁₀(p) for each variant:
- The floor is variants with weak evidence, forming a broad cloud whose heights track sample noise and test statistics.
- The signal is dots poking above the genome-wide line, ideally clustering locally in a small number of “knees” rather than sprinkled uniformly across the genome.
- A uniform skyline — dots distributed evenly across every chromosome — is a tell-tale of inflation, not biology: hidden population structure, batch effects, or relatedness create correlations that apply genome-wide rather than at a few loci.
- The gradient below the line is not a weakness: a real GWAS of a polygenic trait shows signals spread across many loci, many below the significance threshold but consistent in direction and frequency.
The right reflex when seeing one: check the genomic inflation factor (λ), usually near 1.00–1.05 for a properly controlled GWAS; then look at whether strong signals cluster into a few loci rather than sprinkling evenly. Both tell you more than counting dots.
Why the “population average” caveat matters so deeply
A GWAS effect is the average phenotype shift between allele carriers, estimated across entire populations. It does not tell you the probability that any one individual with genotype G will develop diabetes. Common misunderstandings with real consequences:
- “The variant explains the trait” — the reported effect sizes are small. For common complex-trait variants, per-allele effects that each explain a small fraction of trait variance are typical: real signals, useful at population scale, but far from sufficient for individual prediction alone.
- Polygenic scores aggregate thousands of tiny GWAS effects into one per-person number, which is what makes risk prediction possible — but the score’s value is a probability distribution, not destiny, and its accuracy decays across ancestries not represented in the training GWAS.
- Ancestry and stratification: differences in allele frequencies between populations mix with trait differences on average (environment, social factors, genotype batching), generating associations with no biological localization at all. This is the classic confounding issue applied at genome scale, and the standard response is mixed-model regression or structure covariates.
Why most “hits” are indirect
Because alleles travel in haplotype blocks across generations, the variant you measure correlates with other nearby variants. Practical implication: causal localization requires a fine-mapping step, and the confidence interval of a causal variant is typically a set — a 95% credible set of, say, five variants — not one. The standard outputs that turn hits into hypotheses:
- Conditional analysis asks whether a locus still has signal after accounting for the best variant in that region; colonized regions with two independent signals are common.
- Credible sets / posterior probabilities from Bayesian fine-mapping rank variants by how likely each is causal.
- Functional annotation overlay asks which credible variants sit in promoters, enhancers, eQTLs, or coding regions — usually via the same annotation resources used in enrichment analysis, such as GTEx or ENCODE — bridging the signal to a gene.
What makes a GWAS credible: the design checklist
A first GWAS lives or dies on issues that happen before any analysis:
- Sample size is not optional for modest effects. Common variants with small effects need tens of thousands of samples for a locus to detect at genome-wide significance; the power analysis intuition is the same as any other study, scaled up.
- Phenotype quality beats phenotype tally. A clean, well-defined phenotype measured the same way for everyone is better than a noisy or inconsistently recorded one across a bigger sample.
- Relatedness and cryptic duplication, if uncorrected, produce spurious loci and overconfident standard errors; a kinship matrix or a linear-mixed-model framework is expected in serious analyses.
- Batch to be modeled or removed. Genotyping arrays, sequencing centers, and sample plates each encode processing noise; a GWAS that never models batch inherits a systematic bias no downstream test can remove.
From GWAS table to biology
A credible hit list becomes a paper by assembling evidence, never by a single distance:
- Colocalization: does the same locus’s GWAS signal and an eQTL for a nearby gene share their causal variant, or are they two distinct signals? Statistical colocalization tests (coloc) answer this, and they are often the step that turns a list entry into a named gene.
- Pathway/annotation enrichment: does the set of associated loci enrich for biological functions you would expect? Same caveats as any gene-list enrichment: background matters.
- Replication in an independent cohort is the non-negotiable step that separates a discoverable, robust signal from a one-sample accident.
Key takeaways
- GWAS is millions of parallel association tests, and its outputs are estimates with population-level meaning: effects are averages, not individual predictions.
- The Manhattan plot’s two honest skeletons are a cluster of loci and inflation metrics (λ); a uniform skyline is a warning sign, not a discovery.
- The 5 × 10⁻⁸ threshold exists because millions of tests were run; anything softer is suggestive, not significant.
- Lead SNPs tag regions, not mechanisms: fine-mapping, conditional analysis, and colocalization with functional data are what name genes.
- Sample size, ancestry-aware modeling, and phenotype quality decide more about GWAS quality than which regression software you run.
Challenge
Choose a published GWAS summary-statistics table from the GWAS Catalog for a trait you care about. Find the lead variant, check what its effect size (beta) actually says about phenotype difference per allele, then look up how many variants passed genome-wide significance versus how many were tested. Converting one locus from “dot on a plot” to “variant set with an effect direction and a plausible mechanism” is the working skill behind every GWAS-based paper.
Want guidance turning summary statistics into a defensible write-up? Ampersand Academy offers one-to-one mentoring in bioinformatics and biostatistics — part of Mahadhi, a development and digital marketing company.
FAQ
What exactly does the 5 × 10⁻⁸ genome-wide significance threshold mean?
With roughly one million effectively independent common variants tested per scan, p 5 × 10⁻⁸ limits the expected number of false positives to about one per GWAS. It is a multiple-testing correction applied genome-wide, not a declaration that a locus is biologically causal.
How large does a GWAS sample need to be?
Large enough for the smallest effect you care to detect at genome-wide significance. For common complex-trait variants with small effect sizes, tens of thousands — often hundreds of thousands — of samples are typical; a power calculation that accounts for allele frequency and expected effect size is the honest starting point rather than a rule of thumb.
What is a Manhattan plot showing, really?
A per-variant p-value picture of the whole genome: genomic position on the x-axis and −log₁₀(p) on the y-axis. Dots above the genome-wide line are candidate loci; the pattern of where they cluster, and the λ inflation statistic, tell you more than the number of dots.
Does a GWAS hit mean someone with the allele will get the trait?
No. A hit is a population-level average difference in the trait between genotype groups. For individual predictions, GWAS effects must be combined into polygenic scores evaluated with explicit prediction metrics — and even then the score is a risk estimate, not a certainty.
Why can the lead SNP in a locus not be the causal variant?
Variants are correlated with nearby variants (linkage disequilibrium), so the measured SNP typically tags a haplotype containing the causal variant, which may lie many kilobases away. Fine-mapping and colocalization are what narrow the credible set.
What is population stratification and how is it handled?
Allele frequencies at many variants differ between ancestries; if ancestry is also correlated with the trait (environment, social or clinical differences), those variants show up as genome-wide false associations. Standard practice is principal components as covariates, or linear-mixed models that account for relatedness; ignoring it produces exactly the skyline Manhattan plot that signals inflation.
What data do I need to start a GWAS?
A genotype matrix for thousands of individuals with consistent QC, a phenotype measured identically across samples, and a pre-registered analysis plan for covariates, ancestry handling, and batch modeling. Software such as PLINK 2, GCTA, or REGENIE handles the testing; the quality of inputs and design decides the credibility of outputs.