{
  "id": "genome-wide-association-study",
  "title": "Genome-Wide Association Study",
  "lang": "en",
  "source_lang": "en",
  "summary": "A genome-wide association study tests genetic variants across the genome for statistical relationships with traits, diseases, or other measured characteristics.",
  "category": "science",
  "subcategory": "biology",
  "type": "concept",
  "aliases": [
    "GWAS",
    "genome-wide association analysis",
    "whole-genome association study"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/genome-wide-association-study/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/genome-wide-association-study/",
      "json": "https://aiwiki.page/zh/genome-wide-association-study.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A genome-wide association study (GWAS) is an investigation that systematically tests genetic variants across the genome for statistical associations with a measured characteristic, or phenotype. In human research, these characteristics include diseases, physical measurements, molecular traits, and responses to environmental exposures or treatments. Most GWAS compare the frequencies or estimated effects of many variants among large numbers of individuals, rather than restricting analysis to genes selected in advance.\n\nGWAS have transformed the study of complex traits by showing that their inherited components commonly involve many genetic variants, most with individually small effects. They have identified biological pathways, connected noncoding variation to gene regulation, and supplied data for genetic prediction and causal analysis. Their central limitation is equally important: an associated variant is not necessarily a causal variant, and an associated genomic region does not automatically identify the responsible gene. Interpreting a GWAS therefore requires an understanding of population history, study design, statistical inference, and experimental biology."
    },
    {
      "heading": "Genetic foundations and the meaning of association",
      "text": "A GWAS relates differences in genotype to differences in phenotype. The most frequently analyzed markers are single-nucleotide polymorphisms (SNPs), positions at which individuals differ in a single DNA base. Studies also examine short insertions and deletions, copy-number changes, and other structural variants when suitable data are available. At a typical autosomal biallelic marker, an individual carries zero, one, or two copies of a designated allele. This allele count can be tested against a quantitative measurement, such as height, or a categorical outcome, such as disease status. The word “association” describes a statistical relationship; it does not specify the biological mechanism producing that relationship.\n\nA defining feature of GWAS is that a marker can reveal a disease-associated region without itself affecting disease. Nearby variants are often inherited together because relatively few recombination events have separated them. This nonrandom association among alleles is called linkage disequilibrium (LD). A typed SNP may therefore act as a proxy for an untyped causal variant. The strength of this tagging relationship is often described by \\(r^2\\), the squared correlation between allele counts. When \\(r^2\\) is close to one, the variants provide nearly equivalent association information in that population.\n\nLD reflects population genetics, including demographic history, migration, recombination, and genetic drift. A haplotype is a combination of alleles carried on the same chromosome segment. Different populations may share a causal variant but have different surrounding haplotypes, so the strongest associated marker can differ between studies. Association signals commonly extend across several genes or regulatory elements. Consequently, naming the nearest gene is a convenient way to label a region, not proof that the gene mediates the effect.\n\nThe distinction between association and causation also applies to the phenotype. An allele associated with a disease may influence an intermediate process, survival to recruitment, or the probability that the disease is recorded. It may tag another variant or correlate with ancestry-related exposures. GWAS results are therefore interpreted within a model of how genotype, biology, environment, and ascertainment are connected. The proportion of phenotypic variation attributable to genetic differences—heritability—is a population-specific statistical quantity, not a measure of how genetically determined an individual’s outcome is. It can change when environmental variation, age distribution, or measurement practices change."
    },
    {
      "heading": "Historical development and the expansion of scale",
      "text": "GWAS emerged from the convergence of genome mapping, inexpensive genotyping, and the recognition that common variants could tag substantial portions of human genetic variation. Earlier approaches included linkage analysis in families and candidate-gene association studies. Linkage analysis was especially effective for some rare disorders with large genetic effects, but generally localized broad regions. Candidate-gene studies tested variants chosen from biological hypotheses; many reported associations did not withstand larger replication studies. Genome-wide testing reduced dependence on prior biological guesses while making stringent control of false-positive findings essential.\n\nThe Human Genome Project provided a reference sequence, while the International HapMap Project, launched in 2002, characterized common variation and LD patterns. These resources helped researchers select sets of tag SNPs for genotyping arrays. The 1000 Genomes Project, whose initial phase began in 2008 and whose major final-phase report appeared in 2015, subsequently supplied broader reference panels for genetic variation. Together with improvements in DNA sequencing, these projects enabled studies to evaluate many more variants than were measured directly in each participant.\n\nThe designation “first GWAS” depends on how comprehensively a study surveyed the genome. In 2002, Ozaki and colleagues reported an association between variation in LTA, encoding lymphotoxin-alpha, and myocardial infarction using a large SNP-based association screen. A landmark 2005 study by Klein and colleagues connected the complement factor H gene, CFH, with age-related macular degeneration. Despite its relatively small sample, the study detected a comparatively large common-variant effect and highlighted a role for the complement system in the disease.\n\nIn 2007, the Wellcome Trust Case Control Consortium published a study of approximately 14,000 cases across seven common diseases and 3,000 shared controls. It established the practicality of a coordinated genome-wide approach and identified multiple reproducible loci. Subsequent international consortia pooled studies of diabetes, cardiovascular traits, psychiatric conditions, height, and many other phenotypes. Population resources such as UK Biobank, which recruited about 500,000 participants between 2006 and 2010, enabled analyses across thousands of measured or recorded traits. The historical progression has been from relatively small, disease-specific scans to large, reusable resources that support both discovery and systematic comparison of genetic effects across phenotypes."
    },
    {
      "heading": "Study designs, phenotypes, and statistical power",
      "text": "The two most familiar GWAS designs are case–control studies and quantitative-trait studies. A case–control study compares individuals meeting a disease definition with controls selected from an appropriate source population. A quantitative-trait study relates genotype to a measured variable, such as blood pressure, lipid concentration, or organ volume. Cohort studies can also analyze disease incidence over time using survival models. Each design has distinct assumptions: controls may later develop disease, a single laboratory measurement may poorly represent long-term exposure, and prevalent cases can overrepresent people who survive longer after diagnosis.\n\nPhenotype definition is often as consequential as genotyping quality. Diagnoses may come from specialist assessment, self-report, hospital records, or combinations of codes and prescriptions. These definitions can differ in specificity, sensitivity, age coverage, and access to care. Measurement error usually reduces power when it is unrelated to genotype, but differential error can create more complicated biases. Combining related disorders increases sample size yet may obscure subtype-specific biology. Conversely, highly precise phenotyping can improve biological interpretability even when the available sample is smaller.\n\nPower depends on sample size, allele frequency, effect magnitude, phenotype quality, and the significance threshold. Under a simple additive model, the variation in allele count is approximately \\(2p(1-p)\\), where \\(p\\) is the effect-allele frequency. Other things being equal, a rare allele provides less statistical information because fewer participants carry it. For binary traits, the balance between cases and controls also matters. A common approximation to the effective sample size is\n\n\\[N_{\\mathrm{eff}}=\\frac{4}{1/N_{\\mathrm{case}}+1/N_{\\mathrm{control}}}.\\]\n\nThis expression shows why adding controls produces diminishing returns when the number of cases remains fixed. Exact power additionally depends on the disease model and analysis method.\n\nFamily-based designs provide complementary information. Comparing transmitted and nontransmitted parental alleles, or comparing siblings within families, can reduce confounding from population structure and shared family background. Such designs often have lower power than equally large samples of unrelated individuals, but are valuable for separating direct genetic associations from parental and demographic influences. GWAS are also widely used in crops, livestock, and experimental organisms. In those settings, breeding structure, environmental blocks, and relatedness require explicit modeling. Across organisms, a larger sample does not compensate for an ill-defined phenotype or a design in which genotype and exposure are systematically entangled."
    },
    {
      "heading": "Genotyping, imputation, and quality control",
      "text": "Most large human GWAS historically measured hundreds of thousands of variants using genotyping arrays. These arrays infer alleles from hybridization signals at predefined positions in DNA. Sequencing-based studies instead read DNA fragments and infer variants from the aligned sequence data. Whole-genome sequencing can capture a broader spectrum of variation, while exome sequencing focuses mainly on protein-coding regions. Array design affects which variation is measured well: arrays optimized using one ancestry group may provide less complete coverage in another, particularly for less common variants.\n\nGenotype imputation extends the number of analyzable variants by combining observed genotypes with reference haplotypes. The data are usually phased first, estimating which alleles lie together on each chromosome copy. Imputation then estimates the probabilities of unobserved genotypes from matching reference segments. Rather than treating uncertain calls as definite, association analyses commonly use allele dosage, the expected number of effect alleles. Dosages can take any value between zero and two. Accuracy depends on allele frequency, reference-panel size and ancestry, array coverage, and local genomic complexity; rare variants are particularly difficult to impute reliably.\n\nSample-level quality control examines missing genotypes, unusual heterozygosity, duplicate samples, unexpected relatedness, contamination, and inconsistencies between recorded metadata and genetic measurements. Investigators also check whether sample processing or genotyping batches correlate with the phenotype. Related individuals can be excluded or retained using models that account for their covariance. Genetically inferred ancestry can help characterize sample structure, but it is not interchangeable with nationality, ethnicity, or self-identified race. Decisions about inclusion and grouping affect both statistical validity and the populations to which findings apply.\n\nVariant-level checks include call rate, allele frequency, imputation quality, and departures from Hardy–Weinberg equilibrium. Such departures can indicate genotyping error, but can also arise from population mixture, inbreeding, or biological selection; they require interpretation rather than automatic attribution to error. Particular attention is needed for strand alignment, reference-genome coordinates, and allele coding. For example, A/T and C/G SNPs can be ambiguous when combining datasets reported on opposite strands. A mistaken allele alignment can reverse an estimated effect or disrupt a combined analysis. Sex chromosomes, mitochondrial variants, structural variants, and regions with complex sequence similarity often require specialized calling and testing procedures beyond the standard autosomal pipeline."
    },
    {
      "heading": "Association models, significance, and replication",
      "text": "For a quantitative trait, a basic GWAS model is\n\n\\[Y_i=\\alpha+\\beta G_i+\\boldsymbol{\\gamma}^{\\mathsf T}\\mathbf{C}_i+\\varepsilon_i,\\]\n\nwhere \\(Y_i\\) is the phenotype of individual \\(i\\), \\(G_i\\) is an allele count or dosage, and \\(\\mathbf{C}_i\\) contains covariates such as age, sex, and ancestry components. The coefficient \\(\\beta\\) estimates the average change in the phenotype per additional effect allele, conditional on the model. For binary outcomes, logistic regression models the log odds of disease; exponentiating the genotype coefficient gives an odds ratio. An odds ratio is not generally equal to a risk ratio, particularly for common outcomes. More specialized methods handle relatedness, highly unbalanced case–control samples, survival outcomes, or repeated measurements.\n\nThe genome-wide search creates a multiple-testing problem. Testing one million independent null hypotheses at a nominal level of 0.05 would produce about 50,000 positive results by chance. A conventional threshold for common-variant human GWAS is \\(P<5\\times10^{-8}\\), historically motivated by approximately one million effectively independent tests. This is a convention rather than a universal constant. The appropriate threshold can differ with ancestry, variant classes, sequencing coverage, and the number of phenotypes or models examined. A small \\(P\\)-value measures incompatibility with a specified null model; it does not measure biological importance or the probability that a particular variant is causal.\n\nResults are commonly displayed in a Manhattan plot, showing genomic position against \\(-\\log_{10}P\\), and a quantile–quantile plot comparing observed and expected test statistics. Peaks in a Manhattan plot often contain many correlated markers. Selecting approximately independent lead variants reduces this redundancy, while conditional analyses test whether a region contains additional signals after accounting for an initial association. These procedures depend on an appropriate LD reference. Summary statistics usually report the variant identifier, effect and other alleles, effect estimate, standard error, allele frequency, sample size, and significance value.\n\nReplication in an independent sample helps distinguish reproducible findings from technical artifacts and chance fluctuations. Discovery effects are often overestimated because selection favors unusually large observed effects, a phenomenon known as the winner’s curse. Meta-analysis combines association estimates across studies, frequently using inverse-variance weighting. It requires consistent allele orientation, compatible phenotypes, assessment of heterogeneity, and attention to overlapping participants. A reproducible association can still reflect systematic confounding shared across studies, while a failure to replicate may reflect inadequate power, different LD patterns, or a genuinely different phenotype rather than a false original result."
    },
    {
      "heading": "Population structure, relatedness, and indirect associations",
      "text": "Population stratification occurs when allele frequencies and phenotype distributions both vary across population groups, creating associations unrelated to a direct effect of the tested variant. For example, geographic differences in ancestry can correlate with environmental exposures, recruitment centers, or socioeconomic conditions. Because many small allele-frequency differences occur across the genome, even subtle structure can matter in very large studies. Increasing sample size narrows standard errors but does not remove a systematic bias.\n\nA standard adjustment uses principal component analysis to summarize major axes of genetic variation and include them as covariates. This can reduce broad ancestry confounding but does not guarantee control of fine-scale geography or every social and demographic process. Linear mixed models and related generalized mixed models incorporate a genetic relationship matrix to account for covariance among participants. They allow many studies to retain relatives while controlling structure more effectively than ordinary regression alone. Computational implementations must also avoid absorbing too much of the local association signal into the relatedness component.\n\nTest-statistic inflation requires careful interpretation. A genome-wide excess of small \\(P\\)-values can arise from confounding, but it can also arise because thousands of variants truly influence the trait. Methods based on the relationship between association strength and LD can help distinguish these contributions under stated assumptions. Simply rescaling every test statistic to remove all inflation may erase genuine polygenic signal. Conversely, an apparently well-behaved diagnostic plot does not prove that every locus is unconfounded.\n\nGenetic associations can also include indirect pathways. Parental genotypes may influence the environment experienced by a child, producing an association between the child’s inherited alleles and an outcome partly through parental characteristics. Assortative mating can correlate trait-associated alleles across the genome and change expected relationships among genotype, family environment, and phenotype. Within-family analyses reduce several of these influences by comparing relatives who share much of their background. Selection into a study creates another difficulty: if participation depends on both a phenotype and other genetically influenced characteristics, conditioning on participation can induce associations. These mechanisms are especially important when interpreting socially patterned traits. Standard ancestry correction alone does not convert an association estimate into a purely biological or individually causal effect."
    },
    {
      "heading": "What GWAS reveal about genetic architecture",
      "text": "One of the major findings of GWAS is extensive polygenicity. Height, blood pressure, many common diseases, and numerous other traits involve contributions from large numbers of variants. For many complex traits, individual common variants explain only a small fraction of phenotypic variation, so their discovery requires large samples. A few regions can nevertheless have much larger effects, and the distribution of effect sizes varies considerably between traits. Highly polygenic architecture is not equivalent to an absence of biological organization: many small associations can converge on particular tissues, cell types, or pathways.\n\nGWAS helped clarify the problem of “missing heritability.” Early significant loci explained much less variation than family-based estimates suggested. Part of the gap reflected effects too small to cross stringent significance thresholds in the available samples. Other contributions can come from rare variants, imperfect tagging, structural variation, differences among heritability estimators, and features of family-based designs. SNP-based heritability estimates the variation captured by a specified set of variants in a particular population and model. It is not automatically equivalent to total narrow-sense heritability, and its numerical value depends on phenotype measurement and environmental context.\n\nMany variants affect more than one trait, a pattern called pleiotropy. Genetic correlation summarizes the extent to which estimated genetic influences on two traits covary. It can reveal shared architecture between disorders or between molecular and organism-level measurements. However, a genetic correlation does not show that one phenotype causes the other. It may reflect direct effects of the same variants on both traits, an intermediate biological pathway, correlated exposures, or methodological biases. Similarly, two nearby association peaks need not share the same causal variant simply because their lead SNPs are close.\n\nRare variation provides a complementary view. Because individual rare alleles are carried by few people, sequencing studies often aggregate variants within genes using burden or variance-component tests. These approaches can connect disruption of a protein-coding gene more directly with a phenotype, although variant grouping and functional annotation introduce assumptions. The major histocompatibility complex illustrates a different challenge: extensive LD, high allelic diversity, and multiple correlated biological effects complicate interpretation. Analyses of genetic architecture must therefore distinguish the number of associated markers, the number of statistically independent signals, the number of causal variants, and the number of affected genes; these are not interchangeable counts."
    },
    {
      "heading": "From associated loci to biological mechanisms",
      "text": "The first task after discovering an association is often genetic fine-mapping: identifying which variants within the associated region are most compatible with causing the signal. Statistical methods compare correlated variants and may assign posterior probabilities under explicit assumptions about effect sizes and the number of causal variants. A credible set is constructed to contain the causal variant, or variants, with a specified posterior probability under the fitted model. Its interpretation depends on accurate LD information, adequate variant coverage, and a suitable model. A credible set is not an experimentally confirmed list of causal sites.\n\nCombining populations can improve resolution because LD patterns differ across ancestries. Variants that are nearly indistinguishable in one population may be separated statistically in another. This benefit requires sufficient sample sizes and allowance for possible effect heterogeneity. Functional annotation can further prioritize variants in coding sequences, accessible chromatin, promoters, or enhancers active in relevant cell types. However, overlap with a regulatory annotation is supporting evidence, not proof. The biologically relevant tissue or developmental stage may be missing from available datasets, and regulatory effects can depend on cell state or environmental conditions.\n\nMolecular association studies provide additional links. An expression quantitative trait locus (eQTL) is a variant associated with gene expression. Related analyses examine RNA splicing, protein abundance, or DNA methylation. Colocalization methods ask whether a trait association and a molecular association are consistent with sharing a causal variant. This is stronger evidence than merely finding both signals in the same region, but shared causality at a variant does not by itself establish that the measured molecular change mediates the disease effect. A variant can influence several molecular processes independently.\n\nExperimental work tests specific mechanisms using reporter assays, cellular perturbations, animal models, and genome editing. A prominent example is the obesity-associated region within FTO. Work on particular noncoding variants implicated regulation of IRX3 and IRX5 in adipocyte biology, illustrating why the gene containing a marker need not be its only functional target. This mechanism does not imply that every association in the region acts identically across tissues. Mechanistic interpretation is strongest when statistical fine-mapping, molecular measurements, physical regulatory interactions, and targeted perturbations converge on the same variant-to-gene-to-phenotype pathway."
    },
    {
      "heading": "Genetic prediction, causal inference, and therapeutic research",
      "text": "A polygenic score combines information from many variants into a numerical predictor. In a simple form,\n\n\\[S_i=\\sum_{j=1}^{M}\\widehat{\\beta}_jG_{ij},\\]\n\nwhere \\(\\widehat{\\beta}_j\\) is an estimated effect and \\(G_{ij}\\) is the allele dosage for individual \\(i\\) at variant \\(j\\). Modern methods account for LD and shrink noisy estimates rather than simply summing every discovery coefficient. Evaluation requires data not used to estimate or select the model; hidden overlap between training and test samples can exaggerate performance. Association significance and prediction accuracy are different objectives: many variants that do not individually reach genome-wide significance can still contribute predictive information.\n\nPredictive performance is population- and setting-dependent. A score may discriminate between higher- and lower-risk groups yet be poorly calibrated for absolute risk in a new population. Age, baseline disease incidence, clinical measurements, and environmental exposures can materially affect prediction. Scores trained predominantly in one ancestry group often perform less well in others because allele frequencies, LD, effect estimates, and study environments differ. A score also reflects the phenotype and selection processes of its training data. Its value for a particular clinical use requires evaluation beyond demonstrating a statistical association with disease.\n\nGWAS summary statistics are widely used in Mendelian randomization, which uses genetic variants as instruments to investigate causal relationships between exposures and outcomes. Valid inference requires that the instruments influence the exposure, are not associated with relevant confounders, and affect the outcome only through the modeled exposure. Horizontal pleiotropy, population structure, sample overlap, and selection can violate these assumptions. Genetically proxied lifelong differences in an exposure are also not necessarily equivalent to a short-term intervention. Mendelian randomization is therefore a framework for causal inference under assumptions, not an automatic consequence of genome-wide significance or a substitute for a randomized controlled trial.\n\nTherapeutic research uses GWAS to prioritize targets and pathways. Variants affecting a gene or its regulatory elements can indicate that altering the corresponding biological process may influence disease. Genetic findings can also suggest possible adverse effects if a target influences several traits. The direction of a pharmacological intervention, however, cannot always be read directly from an associated marker: the functional allele, target gene, tissue, developmental timing, and degree of perturbation may remain uncertain. Evidence from common variants is frequently combined with rare loss-of-function variants, molecular studies, and experimental pharmacology. These applications connect genetic discovery to drug development without making every associated locus an immediately actionable target."
    },
    {
      "heading": "Diversity, data governance, and research boundaries",
      "text": "The human populations represented in GWAS have historically been unevenly distributed, with a strong concentration of participants of European ancestry. This affects both scientific discovery and the distribution of potential benefits. A variant common in one population may be rare or absent in another, and an imputation panel may represent some haplotypes poorly. Underrepresented populations can consequently have fewer well-characterized associations and less accurate prediction. Greater participation across ancestries improves the ability to distinguish shared biology from population-specific tagging patterns and can uncover variants missed by studies concentrated in a single population.\n\nGenetic ancestry is continuous, multidimensional, and shaped by migration and population history. Broad labels can be useful for particular analyses but conceal substantial variation within groups. Self-identified race and ethnicity can capture social experiences that genetic ancestry does not, including discrimination, language, and access to services. Treating these variables as interchangeable can produce misleading biological interpretations. Studies involving admixed populations require methods that account for both genome-wide ancestry and, where relevant, ancestry along individual chromosome segments. Indigenous and other communities may also have distinct governance arrangements concerning participation, data access, and collective interests.\n\nGWAS data raise substantial data privacy issues because genetic information is identifying, persistent, and informative about biological relatives. Removing names does not make genomic data intrinsically anonymous. Individual-level data are therefore commonly distributed through controlled-access systems with conditions on use, security, and redistribution. Summary statistics support reproducibility and wide scientific reuse, but some releases also require assessment of disclosure risks. Consent frameworks must address storage, secondary analyses, linkage to health records, and the possibility that future research questions differ from those anticipated at recruitment.\n\nReproducibility depends on clear phenotype definitions, transparent inclusion criteria, complete reporting of quality control, and compatible descriptions of variants and effect alleles. Public association catalogs and repositories allow findings to be compared, but a catalog entry is not independent validation of every biological interpretation attached to it. Continuing methodological challenges include reliable analysis of structural variation, repetitive genomic regions, rare alleles, sex chromosomes, interactions, and changing phenotypes over the life course. As biobanks become larger and more deeply characterized, the limiting step increasingly shifts from detecting a statistical signal to establishing what it means in a particular biological, population, and environmental context."
    }
  ],
  "links": [
    "genome",
    "phenotype",
    "genotype",
    "single-nucleotide-polymorphism",
    "allele",
    "linkage-disequilibrium",
    "population-genetics",
    "genetic-drift",
    "gene",
    "heritability",
    "dna-sequencing",
    "complement-system",
    "dna",
    "hardy-weinberg-equilibrium",
    "multiple-testing",
    "meta-analysis",
    "principal-component-analysis",
    "linear-mixed-model",
    "major-histocompatibility-complex",
    "chromatin",
    "gene-expression",
    "genome-editing",
    "randomized-controlled-trial",
    "data-privacy"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}