Genome-Wide Association Study (GWAS)

GWAS is fundamentally agnostic to prior biological knowledge. Unlike candidate gene studies that test specific genes based on existing hypotheses, GWAS scans millions of genetic variants across the entire genome simultaneously. This unbiased approach enables the discovery of novel loci in unexpected genomic regions, including non-coding regulatory elements and intergenic spaces, that would never have been prioritized a priori.

Statistical power in GWAS depends on sample size, often requiring tens or hundreds of thousands of individuals. This is because:

• Individual variant effect sizes are typically very small (odds ratios of 1.05–1.20)
• A stringent multiple testing burden requires p-values below 5 × 10⁻⁸ for genome-wide significance
• Large biobanks like UK Biobank (n=500,000) and FinnGen have become essential infrastructure

The core statistical engine of GWAS tests each single nucleotide polymorphism (SNP) independently for association with the phenotype. For quantitative traits, linear regression is used; for binary disease outcomes, logistic regression. Each test yields:

• Effect size (beta/OR): Magnitude and direction of the variant's influence
• Standard error: Precision of the estimate
• P-value: Statistical significance of the association This produces the summary statistics that feed downstream PRS models.

Systematic ancestry differences between cases and controls can produce spurious associations that are genetic in origin but unrelated to the trait. GWAS addresses this through:

• Principal Component Analysis (PCA) on genotype data to model continuous ancestry axes
• Inclusion of principal components as covariates in regression models
• Genomic control using the lambda inflation factor (λ) to quantify and correct residual stratification
• Mixed linear models that account for cryptic relatedness via a genetic relationship matrix

GWAS results are shaped by the non-random correlation structure between nearby variants. A significant SNP at a locus is often not the causal variant but a tag SNP in linkage disequilibrium (LD) with the true functional variant. This necessitates:

• Fine-mapping to statistically prioritize causal variants within associated loci
• LD score regression to distinguish polygenic signal from confounding
• Careful interpretation that an associated locus, not a single SNP, is the unit of discovery

A single GWAS is rarely definitive. The gold standard requires:

• Independent replication in a separate cohort with the same phenotype definition
• Meta-analysis combining summary statistics across multiple studies using fixed or random effects models to increase power
• Cross-ancestry validation to ensure associations generalize beyond the discovery population This multi-stage framework protects against winner's curse and false positives inherent in high-dimensional screening.

Systematic differences in allele frequencies between subpopulations due to ancestry, not disease status. This confounding can generate spurious associations in GWAS.

• Arises from divergent ancestral mating patterns
• Corrected using Principal Component Analysis (PCA) or mixed linear models
• Uncorrected stratification inflates the Genomic Inflation Factor (λ)

The non-random association of alleles at different loci within a population. LD is the physical basis for mapping disease genes without directly genotyping the causal variant.

• Measured by r² and D' metrics
• LD Score Regression leverages this correlation structure to estimate heritability
• Fine-mapping disentangles LD to prioritize causal variants

Aggregated GWAS results for each variant including the effect allele, beta coefficient, standard error, and p-value. These serve as the base data for polygenic risk score construction without requiring access to individual-level genotypes.

• Enable privacy-preserving meta-analysis
• Used directly by Bayesian PRS methods like PRS-CS and LDpred2
• Require Winner's Curse correction to adjust for effect size overestimation

The proportion of phenotypic variance attributable to the additive effects of all measured single nucleotide polymorphisms. Estimated via methods like LD Score Regression or GREML.

• Sets the theoretical upper bound for polygenic risk score prediction accuracy
• Typically lower than twin-study heritability due to imperfect tagging of rare variants
• Essential for understanding the genetic architecture of complex traits

A statistical technique combining summary statistics from multiple independent GWAS cohorts to increase statistical power for variant discovery. Fixed-effects and random-effects models are used depending on heterogeneity.

• Enables sample sizes exceeding millions of individuals
• Improves generalizability of derived polygenic scores
• Requires careful harmonization of phenotype definitions and genotyping platforms

A statistical process that prioritizes the most likely causal variants within a GWAS-associated locus by modeling LD patterns. Distinguishes the true functional variant from those merely correlated with it.

• Uses Bayesian methods like SuSiE or FINEMAP
• Outputs credible sets containing the causal variant with specified probability
• Critical for translating GWAS hits into drug targets and functional experiments