Polygenic adaptation and GWAS: quantitative trait loci and the infinitesimal model

Intuition Beginner

Why is height so variable? No single gene determines whether you are 160 cm or 185 cm tall. Instead, hundreds or even thousands of genes contribute, each nudging the trait by a millimetre or less. This is the infinitesimal model: most biological traits are influenced by a very large number of genetic variants, each with a tiny effect, plus the environment.

Genome-wide association studies (GWAS) are the tool for finding those variants. A GWAS scans millions of positions in the DNA of thousands (or hundreds of thousands) of people, testing each position for a statistical association with a trait — height, disease risk, blood pressure, or any measurable characteristic. The result is a list of genetic variants that are more common in taller people, or in people with a disease, than in those without.

Each variant typically explains a minuscule fraction of the total variation. The largest known single variant for human height accounts for less than half a percent of the variation. But collectively, the variants identified by GWAS can explain a substantial fraction. A polygenic risk score (PRS) adds up the effects of all discovered variants into a single number that predicts an individual's predisposition to a trait or disease.

The key insight is that the genetic architecture of most traits is polygenic: many loci, small effects, cumulative impact. This is why GWAS needs enormous sample sizes to detect individual variants — the signal from each one is faint — and why polygenic scores that aggregate thousands of variants can be informative even when no single variant is decisive.

Visual Beginner

A GWAS produces a Manhattan plot: each dot represents one genetic variant, positioned along the chromosomes (x-axis) with height proportional to the strength of its association with the trait (y-axis, $-{\log }_{10}p$-value). Variants that reach genome-wide significance rise above the threshold line like skyscrapers above a cityscape.

The threshold line at $p=5\times 10^{-8}$ corrects for the fact that roughly one million independent statistical tests are performed. Most variants fall well below the line (no association), but a few tower above it. The pattern of many sub-threshold signals and few supra-threshold peaks is the hallmark of a polygenic trait.

Worked example Beginner

A GWAS of human height identifies three SNPs with the following effect sizes:

SNP	Effect allele frequency	Effect on height (per allele, cm)
rs1	0.40	+0.30
rs2	0.60	+0.20
rs3	0.30	-0.15

An individual carries the following genotypes (number of effect alleles):

• rs1: 1 copy
• rs2: 2 copies
• rs3: 0 copies

The polygenic score is:

$$\text{PRS}=(1\times 0.30)+(2\times 0.20)+(0\times (-0.15))=0.30+0.40+0=0.70\text{ cm}.$$

This individual's PRS predicts them to be 0.70 cm taller than the population mean, based on these three variants alone. In a real GWAS, the score would sum over thousands of variants. The prediction is imperfect — environment and undiscovered genetic variants also contribute — but the PRS captures part of the genetic predisposition.

Check your understanding Beginner

Formal definition Intermediate+

Fisher's infinitesimal model

Fisher's 1918 paper ^{[Fisher 1918]} resolved the conflict between Mendelians (who believed in discrete, particulate inheritance) and biometricians (who observed continuous trait distributions) by showing that continuous variation follows naturally from Mendelian inheritance when many loci contribute small effects. The infinitesimal model assumes an infinite number of loci, each with infinitesimal effect, such that the sampling of alleles from parent to offspring produces a normal (Gaussian) distribution of breeding values.

Under the infinitesimal model:

1. The breeding value of an offspring is the midparent breeding value plus a Mendelian sampling deviation: $A_o=(A_f+A_m)/2+\varphi$, where $\varphi \sim N(0,V_A/2)$.
2. The additive genetic variance $V_A$ remains constant across generations (because allele frequencies change negligibly when each locus has infinitesimal effect).
3. The distribution of breeding values is always normal.

The infinitesimal model is an idealisation. Real traits have a finite number of loci with non-infinitesimal effects, so $V_A$ does change across generations as allele frequencies shift. But for traits with thousands of contributing loci — as GWAS reveals for human height, BMI, and many diseases — the approximation is excellent over short evolutionary timescales.

Quantitative trait loci (QTL) mapping

A quantitative trait locus (QTL) is a genomic region that contributes to variation in a quantitative trait. QTL mapping identifies these regions through two principal approaches:

Linkage mapping uses controlled crosses (e.g., F2 intercross between inbred lines) to associate genetic markers with trait variation within families. The LOD (logarithm of the odds) score at each position tests for a QTL:

$$\text{LOD}(d)={\log }_{10}\frac{L(\text{QTL at position }d)}{L(\text{no QTL})}.$$

A LOD threshold of 3.0 (odds of 1000:1) is the conventional significance cutoff. Interval mapping (Lander and Botstein 1989) evaluates LOD scores at positions between markers using conditional genotype probabilities given flanking markers. The resolution of linkage mapping is limited by the number of recombination events in the cross, typically producing confidence intervals of 10--20 cM.

Association mapping (the basis of GWAS) uses natural populations with historical recombination to achieve higher resolution. Because recombination has occurred over many generations, the genomic segments shared among individuals are short, localising associations to small regions (often tens of kilobases). The trade-off is that population structure (systematic ancestry differences) can create spurious associations.

GWAS design and methodology

A GWAS tests each SNP $j$ for association with the trait using a regression model:

$$z_i=\mu +{\beta }_j{x}_{ij}+{\mathbf{c}}_i^T\mathbf{\gamma }+{ϵ}_i,$$

where $z_i$ is the phenotype, $x_{ij}\in \{0,1,2\}$ is the allele count, ${\mathbf{c}}_i$ are covariates (principal components of genetic ancestry, age, sex), and ${ϵ}_i$ is the residual. The effect-size estimate ${\hat{\beta }}_j$ measures the per-allele change in the trait.

For case-control studies (binary traits like disease status), logistic regression replaces linear regression:

$$\log \frac{P(\text{case})}{P(\text{control})}=\mu +{\beta }_j{x}_{ij}+{\mathbf{c}}_i^T\mathbf{\gamma }.$$

The effect size is reported as an odds ratio $\text{OR}=e^{{\hat{\beta }}_j}$. An OR of 1.10 means each copy of the effect allele increases disease odds by 10%.

Genome-wide significance threshold

The conventional genome-wide significance threshold is $p<5\times 10^{-8}$, a Bonferroni correction for approximately one million independent tests in the human genome. Given the effective number of independent SNPs $m_{\text{eff}}\approx 10^6$ (accounting for linkage disequilibrium), the Bonferroni threshold is:

$$p_{\text{GW}}=\frac{0.05}{m_{\text{eff}}}=\frac{0.05}{10^6}=5\times 10^{-8}.$$

This stringent threshold controls the family-wise error rate but also limits power: SNPs with small effect sizes require very large sample sizes to reach significance.

Linkage disequilibrium and tagging SNPs

Linkage disequilibrium (LD) is the non-random association of alleles at different loci. Two SNPs in high LD ($r^2$ close to 1) are highly correlated: genotyping one effectively tags the other. GWAS exploits this property — the associated SNP may not be causal, but it tags a haplotype block containing the causal variant.

The decay of LD with physical distance varies across the genome and across populations. African populations have shorter LD blocks (more generations of recombination), providing finer mapping resolution but requiring denser SNP arrays. European and East Asian populations have longer LD blocks (bottleneck effects), providing coarser resolution but greater tagging efficiency.

Effect sizes and the missing heritability problem

The effect-size distribution of GWAS hits is heavily skewed toward small effects. For human height, the median effect size of genome-wide-significant SNPs is approximately 0.03 standard deviations per allele. The cumulative variance explained by all genome-wide-significant SNPs often falls well below the heritability estimated from twin and family studies — the missing heritability gap ^{[Manolio et al. 2009]}.

Sources of missing heritability include:

1. Rare variants (minor allele frequency < 1%) with potentially larger effects, poorly tagged by standard SNP arrays.
2. Structural variants (copy-number variations, insertions, deletions) incompletely captured by SNP genotyping.
3. Many common variants below genome-wide significance that collectively explain substantial variance but are individually undetectable at moderate sample sizes.
4. Gene-gene and gene-environment interactions not captured by the additive GWAS model.
5. Potential inflation of twin-study estimates by shared environmental effects.

The gap narrows as sample sizes grow. The Yengo et al. (2022) height meta-analysis of 5.4 million individuals identified ~12,000 independent SNPs jointly explaining ~40% of height variance, compared to the twin-study estimate of $h^2\approx 0.80$.

Polygenic risk scores

A polygenic risk score (PRS) aggregates GWAS effect sizes into a single predictor:

$${\text{PRS}}_i=\sum _{j\in S}{\hat{\beta }}_j{x}_{ij},$$

where $S$ is a set of SNPs (often filtered by a $p$-value threshold or weighted by a Bayesian prior) and $x_{ij}$ is the allele count for individual $i$ at SNP $j$.

The simplest PRS uses all SNPs below a $p$-value threshold ($p_T$) from the discovery GWAS (the $p_T+\text{clumping}$ method). More sophisticated methods include:

• LDpred (Vilhjalmsson et al. 2015): infers posterior mean effect sizes using a Bayesian prior that accounts for LD structure and the fraction of causal SNPs.
• PRS-CS (Ge et al. 2019): uses a continuous shrinkage prior on effect sizes, automatically adapting to the unknown sparsity of the genetic architecture.

The predictive accuracy of a PRS is measured by the incremental $R^2$ (variance explained) or the AUC (area under the ROC curve) for binary traits, evaluated in an independent hold-out sample. PRS accuracy depends on the discovery sample size, the heritability of the trait, the LD structure, and the ancestry match between discovery and target populations.

Counterexamples to common slips

• A GWAS hit is not necessarily the causal variant. It is a tag in LD with the causal variant. Fine mapping and functional validation are needed to identify the true causal locus.

• Missing heritability does not mean the heritability estimates are wrong. The gap between GWAS-explained variance and family-study heritability is closing as sample sizes increase and methods improve; it reflects incomplete discovery, not nonexistent genetic effects.

• A PRS is not a deterministic prediction. It captures part of the genetic predisposition but cannot account for environment, rare variants, or gene-gene interactions. A high PRS for a disease elevates risk but does not guarantee the disease will develop.

Key theorem with proof Intermediate+

Theorem (Expected cumulative variance explained by GWAS SNPs). Under a polygenic architecture with $M$ causal variants of equal effect size $\beta$ drawn from a normal distribution, the expected cumulative variance explained by the $k$ genome-wide-significant SNPs is approximately $k\cdot 2p_j{q}_j{\beta }^2$, where $p_j,q_j$ are the allele frequencies at locus $j$. As sample size $N$ increases, $k$ grows and the cumulative variance explained approaches the total additive variance $V_A={\sum }_{j}2p_j{q}_j{\beta }_j^2$.

Proof sketch. For a single SNP $j$ with allele frequencies $p_j,q_j$ and per-allele effect ${\beta }_j$, the additive variance contributed is $V_{A,j}=2p_j{q}_j{\beta }_j^2$. The GWAS test statistic for SNP $j$ in a sample of size $N$ has non-centrality parameter ${\lambda }_j=N\cdot 2p_j{q}_j{\beta }_j^2/{\sigma }^2$, where ${\sigma }^2$ is the residual variance. Genome-wide significance is achieved when the $p$-value falls below $5\times 10^{-8}$, which requires ${\lambda }_j$ to exceed a threshold proportional to $\log (m_{\text{eff}})\approx \log (10^6)\approx 13.8$.

The power to detect SNP $j$ is:

$${\text{Power}}_j=P\left({\chi }_1^2({\lambda }_j)>{\chi }_{1,1-5\times 10^{-8}}^2\right),$$

where ${\chi }_{1,1-\alpha }^2$ is the $1-\alpha$ quantile of the null ${\chi }_1^2$ distribution. For a given effect size, power increases with $N$. As $N\to \infty$, all causal SNPs eventually reach significance, and the cumulative variance explained converges to $V_A$. The rate of convergence depends on the effect-size distribution: SNPs with larger effects are detected first, and progressively smaller effects are captured as sample sizes grow.

For the simplest case of $M$ causal variants with equal effect $\beta$ at equal frequency $p=q=0.5$, the number of genome-wide-significant SNPs at sample size $N$ is approximately:

$$k(N)\approx M\cdot \Phi \left(\sqrt{\frac{N{\beta }^2}{2{\sigma }^2}}-z_{\text{thresh}}\right),$$

where $\Phi$ is the standard normal CDF and $z_{\text{thresh}}\approx 5.45$ corresponds to $p=5\times 10^{-8}$. As $N$ increases, the argument of $\Phi$ grows and $k(N)\to M$. $\square$

This result explains the empirical observation that GWAS variance explained grows roughly linearly with $\log (N)$ for polygenic traits: each doubling of sample size detects a roughly constant number of additional SNPs, each explaining a progressively smaller fraction of variance.

Bridge. The theorem connects the statistical power of GWAS to the underlying additive variance $V_A$ introduced in 19.05.01. As sample sizes grow, GWAS-explained variance converges to $h_{SNP}^2$ (the SNP-based heritability), which itself is bounded above by the narrow-sense heritability $h^2=V_A/V_P$. The gap between $h_{SNP}^2$ and $h^2$ reflects variants not captured by the SNP array — rare variants, structural variants, and imperfect tagging through LD. This bridge connects the variance-component framework to the molecular reality revealed by GWAS.

Exercises Intermediate+

Lean formalization Intermediate+

Mathlib has no formalisation of GWAS statistics, QTL mapping, polygenic risk scores, or the infinitesimal model. The specific gaps include: the regression framework for single-SNP association testing; Bonferroni correction and multiple-testing adjustment; LOD scores for QTL interval mapping; the additive variance decomposition $V_A={\sum }_{j}2p_j{q}_j{\alpha }_j^2$ across multiple loci; polygenic risk score computation as a weighted sum; and LD score regression. The foundational prerequisites are the variance-component algebra from 19.05.01 and the heritability framework from 19.05.02 pending; once those exist in Mathlib, the multi-locus extension and GWAS-specific constructions follow. This unit ships without a Lean module and is reviewer-attested, following the prose-first contract for biology units.

Polygenic adaptation, biobank-scale genomics, and the limits of GWAS Master

Polygenic adaptation signals

Selection on polygenic traits shifts allele frequencies at many loci simultaneously, each by a small amount. Detecting such polygenic adaptation requires methods that aggregate frequency changes across GWAS-identified loci rather than scanning for individual sweeps.

Polygenic height selection in Europeans. Turchin et al. (2012) and later Berg et al. (2019) found that height-increasing alleles are systematically higher in frequency in northern Europeans than in southern Europeans, consistent with polygenic selection on height over the past ~4,000 years. The signal is detected by computing the mean frequency of height-increasing alleles across populations and testing whether the north-south gradient exceeds the null expectation under genetic drift alone. However, the interpretation remains contested: population stratification within the GWAS discovery sample can produce artefactual polygenic selection signals, and controlling for structure attenuates the evidence in some analyses.

Skin pigmentation. Multiple independent signals of selection on skin pigmentation genes (SLC24A5, SLC45A2, KITLG, OCA2/HERC2) have been identified across human populations, driven by the need to balance UV-induced vitamin D synthesis against UV-induced folate degradation. These are unusually strong signals for polygenic traits because individual pigmentation variants have relatively large effects (odds ratios of 2--10 for skin colour categories), making them detectable as individual loci as well as in aggregate.

GWAS limitations

Population stratification. Even after controlling for principal components, fine-scale population structure can produce residual confounding. The LD score regression intercept provides a diagnostic: values significantly above 1.0 indicate residual stratification or cryptic relatedness. Genomic control (dividing test statistics by the genomic inflation factor ${\lambda }_{GC}$) was an early correction but is now considered inadequate because it does not distinguish between true polygenicity and confounding.

Winner's curse. Effect sizes of GWAS-significant SNPs are systematically overestimated because only SNPs whose estimated effects exceed the significance threshold are reported. The true effect is smaller than the estimated effect by an amount that depends on the power of the study: at lower power (smaller effects, smaller samples), the bias is larger. Replication in independent samples provides unbiased effect-size estimates.

Transferability across populations. PRS accuracy drops substantially when applied to populations of different ancestry than the discovery sample, as demonstrated in Exercise 7. The causes are differences in LD structure, allele frequencies, causal variants, and gene-environment interactions. Multi-ancestry GWAS and transfer-learning methods (e.g., PRS-CSx) partially address this but require large non-European samples that are currently underrepresented in GWAS databases.

Biobank-scale GWAS

The UK Biobank (500,000 participants with genotype and extensive phenotype data) transformed GWAS by enabling association studies of unprecedented scale. Key findings from biobank-scale GWAS include:

1. Trait breadth: GWAS are no longer limited to a few diseases. The UK Biobank enables GWAS for thousands of phenotypes — biomarkers, imaging traits, dietary preferences, and self-reported conditions — revealing the pervasive polygenicity of nearly all human traits.

2. Pleiotropy: Many SNPs are associated with multiple traits. The FTO locus is associated with BMI, type 2 diabetes, and educational attainment; the APOE locus is associated with Alzheimer's disease, lipid levels, and longevity. Genetic correlation between traits, estimated by LD score regression or genomic SEM, quantifies the shared genetic architecture.

3. Phenome-wide association studies (PheWAS): Rather than testing one trait against all SNPs, PheWAS tests one SNP (or a PRS) against all available traits, revealing unexpected pleiotropic effects.

4. Within-family GWAS: By comparing siblings who differ at specific SNPs, within-family GWAS control for population structure and shared-family environment. These studies typically yield smaller effect sizes than population-based GWAS, suggesting that a fraction of the population-based estimate is confounded by structure or indirect genetic effects (parents' genotypes influencing offspring through the environment).

Pleiotropy and genetic correlation

Pleiotropy — one variant affecting multiple traits — is pervasive in GWAS results. Two forms are distinguished:

• Biological pleiotropy: the variant directly affects two traits through the same molecular mechanism (e.g., a variant in the insulin-signalling pathway affecting both glucose levels and BMI).
• Mediated pleiotropy: the variant affects trait A, which in turn influences trait B (e.g., a variant increases BMI, and higher BMI increases cardiovascular disease risk).

LD score regression (Bulik-Sullivan et al. 2015) estimates the genetic correlation $r_g$ between two traits from GWAS summary statistics without requiring individual-level data. The cross-trait LD score regression slope estimates the genetic covariance, which divided by the square root of the product of the two SNP heritabilities gives $r_g$. A genetic correlation of $r_g=0.70$ between type 2 diabetes and BMI, for example, indicates that 70% of the genetic effects on one trait are shared with the other.

Mendelian randomization

Mendelian randomization (MR) uses genetic variants as instrumental variables for causal inference. If a genetic variant $G$ is associated with an exposure $X$ (e.g., LDL cholesterol) and affects the outcome $Y$ (e.g., coronary heart disease) only through $X$, then the ratio ${\hat{\beta }}_{G\to Y}/{\hat{\beta }}_{G\to X}$ estimates the causal effect of $X$ on $Y$.

The three core assumptions of MR are:

1. Relevance: $G$ is associated with $X$ (testable from the GWAS of $X$).
2. Independence: $G$ is unconfounded with respect to the $X$--$Y$ relationship (guaranteed by random Mendelian assortment of alleles at conception, modulo population structure).
3. Exclusion restriction: $G$ affects $Y$ only through $X$ (i.e., no horizontal pleiotropy).

Violation of the exclusion restriction — horizontal pleiotropy, where $G$ affects $Y$ through a pathway other than $X$ — is the central threat to MR validity. Methods to detect and correct for pleiotropy include MR-Egger regression (which estimates and adjusts for directional pleiotropy), weighted median MR (robust to invalid instruments if fewer than 50% are pleiotropic), and MR-PRESSO (which detects and removes outlier instruments).

MR has provided evidence for causal relationships that are difficult to establish observationally: the causal effect of LDL cholesterol on coronary heart disease (confirmed by MR and later by clinical trials of statins and PCSK9 inhibitors), the lack of a causal effect of HDL cholesterol on coronary heart disease (despite observational correlation, refuted by MR and clinical trials), and the causal effect of BMI on osteoarthritis risk.

Polygenic selection in humans: recent examples

Beyond height and skin pigmentation, polygenic selection signals have been reported for:

• Educational attainment: A polygenic selection signal was reported by Beauchamp (2016) and Kong et al. (2017), who found that the polygenic score for educational attainment has been decreasing across recent birth cohorts in Iceland. However, this signal likely reflects indirect genetic effects (parents with higher polygenic scores delaying reproduction) rather than direct selection on educational attainment itself.

• BMI: Selection on BMI-related alleles has been investigated in multiple populations, with evidence for both positive selection on fat-storage alleles in historically food-scarce environments and recent changes in the direction of selection as obesogenic environments spread.

• Immune function: The strongest and most replicated signals of recent positive selection in the human genome are at immune-related loci (HLA region, TLR genes, interferon pathways), driven by pathogen-mediated selection across millennia.

GWAS for behavioral traits

GWAS of behavioral traits — educational attainment, cognitive ability, risk tolerance, subjective well-being — are scientifically powerful but ethically fraught. The GWAS of educational attainment by Lee et al. (2018) ($N\approx 1.1$ million) identified 1,271 genome-wide-significant SNPs explaining ~10% of variance. The effect sizes are tiny (the largest explains 0.02% of variance), and the biological pathways are largely unknown.

Key ethical concerns include:

1. Misuse for genetic determinism: The small effect sizes and large environmental contributions are often ignored in popular interpretations that treat PRS as destiny.
2. Eugenics history: The study of genetic contributions to intelligence and behavior has a dark history of misuse for racist and eugenic policies. Modern researchers emphasise that PRS for behavioral traits explain very little variance and are not clinically actionable.
3. Embryo selection: Polygenic scores for behavioral traits could, in principle, be used for embryo selection in IVF settings. The predictive accuracy is currently too low for meaningful selection, but the ethical framework for future use is undeveloped.
4. Between-group inference: PRS for behavioral traits differ across ancestry groups, but these differences do not imply genetic causes of between-group phenotypic differences. The transferability problems documented for physical traits are even more severe for behavioral traits, where gene-environment correlation and interaction are substantial.

PRS in clinical practice

Despite current limitations, PRS are entering clinical practice for specific applications:

• Coronary artery disease: The PRS for CAD identifies individuals at 3-fold increased risk, comparable to monogenic risk factors like familial hypercholesterolemia but affecting a much larger fraction of the population. Clinical guidelines are beginning to incorporate PRS for risk stratification.
• Breast cancer: PRS combined with family history and mammographic density improves risk prediction and could guide screening frequency.
• Atrial fibrillation: PRS can identify individuals who would benefit from earlier cardiac monitoring.

The clinical deployment of PRS faces challenges: ancestry-specific calibration is essential (a PRS validated in Europeans may misclassify risk in Africans or East Asians), the clinical utility threshold (does the PRS change clinical management?) varies by disease, and integration with existing risk factors (family history, biomarkers, imaging) requires prospective validation.

Connections Master

• Quantitative genetics and the breeder's equation 19.05.01. GWAS and the infinitesimal model are the molecular realisation of the variance-component framework. The additive variance $V_A$ that fuels selection response in $R=h^{2}S$ is the same quantity that GWAS decomposes into locus-specific contributions $V_{A,j}=2p_j{q}_j{\alpha }_j^2$. The GREML estimate of $h_{SNP}^2$ from 19.05.01 connects directly to the GWAS machinery in this unit.

• Twin studies and heritability estimation 19.05.02 pending. The heritability estimates from twin studies provide the upper bound that GWAS variance explained approaches as sample sizes grow. The missing heritability gap — first documented by comparing GWAS results to twin-study estimates — is the empirical link between the two units.

• Hardy-Weinberg equilibrium 19.02.01. GWAS association tests assume Hardy-Weinberg genotype frequencies within ancestry groups. Departures from HWE at a SNP may indicate genotyping error, natural selection at that locus, or population stratification — and SNPs with extreme HWE deviations are typically filtered out of GWAS.

• Natural selection 19.03.01. Polygenic adaptation is the quantitative-genetics analogue of the directional selection covered in 19.03.01: selection on a polygenic trait shifts allele frequencies at many loci simultaneously, each by a small amount, rather than sweeping a single beneficial mutation to fixation.

• Genetic drift 19.04.01. The null hypothesis for polygenic adaptation tests is that allele frequency shifts across loci are consistent with drift alone. Distinguishing selection from drift at the polygenic level requires aggregating signals across many loci, just as single-locus selection tests distinguish selection from drift at individual sites.

• Mendelian genetics 19.01.01. GWAS effect sizes reflect the average effect of allele substitution, a concept rooted in Mendelian inheritance. The polygenic architecture that GWAS reveals — many loci, small effects — is the quantitative-genetics consequence of Mendelian segregation across thousands of independent loci.

Historical & philosophical context Master

Ronald Fisher's 1918 paper, "The correlation between relatives on the supposition of Mendelian inheritance" ^{[Fisher 1918]}, is the foundational document of both the infinitesimal model and quantitative genetics. Fisher showed that the continuous variation observed by biometricians (Galton, Pearson, Weldon) was fully compatible with Mendelian inheritance when many loci of small effect contributed to a trait. The paper decomposed phenotypic variance into additive, dominance, and environmental components and derived the expected correlations between relatives — the framework that underpins every GWAS and heritability estimate today.

The first QTL mapping studies appeared in the late 1980s, enabled by the development of molecular markers (RFLPs, microsatellites). Lander and Botstein's 1989 paper on interval mapping transformed the field by providing a rigorous statistical framework for scanning the genome for QTL in experimental crosses. These early studies were limited to model organisms and agricultural species where controlled crosses were feasible.

The GWAS era began with the completion of the Human Genome Project (2003) and the International HapMap Project (2005), which catalogued common genetic variation and LD structure across human populations. The first successful GWAS — Klein et al. (2005) on age-related macular degeneration, identifying the CFH locus — demonstrated that genome-wide scanning could discover novel disease genes. The Wellcome Trust Case Control Consortium (2007) confirmed the approach across seven common diseases, identifying dozens of novel loci.

The discovery that most GWAS hits have tiny effect sizes was initially disappointing and led to the missing heritability debate (Manolio et al. 2009) ^{[Manolio et al. 2009]}. Where was the genetic variation that twin studies estimated? The answer emerged gradually: most heritability is distributed across thousands of variants with effects too small to detect individually at moderate sample sizes. Yang et al. (2010) demonstrated with GCTA that common SNPs collectively explain a large fraction of heritability, and the subsequent growth of GWAS sample sizes — from tens of thousands to hundreds of thousands to millions — has progressively captured more of the missing variance.

The UK Biobank (recruitment 2006--2010, genotype data released 2015) and similar biobanks (FinnGen, All of Us, Biobank Japan, China Kadoorie) transformed GWAS by providing massive samples with linked health records. The biobank era revealed the pervasiveness of pleiotropy, the transferability problem across ancestries, and the power of within-family designs to control for confounding.

Polygenic risk scores moved from research tools toward clinical applications in the late 2010s, with the first commercial PRS tests appearing around 2019. The American College of Cardiology and the American Heart Association mentioned PRS in their 2018 cholesterol guidelines, and the NHS began piloting PRS for breast cancer screening in 2024. The ethical debate — particularly around PRS for behavioral traits and the risk of exacerbating health disparities through ancestry-dependent accuracy — remains active and unresolved.

The philosophical dimension of GWAS concerns the relationship between statistical association and biological mechanism. GWAS identifies correlations between genotypes and phenotypes; it does not reveal the causal pathway from variant to trait. The majority of GWAS hits lie in non-coding regions, suggesting regulatory effects, but connecting a non-coding variant to the gene it regulates, the tissue in which it acts, and the developmental stage at which it matters requires extensive functional follow-up. The GWAS-to-function pipeline — eQTL mapping (GTEx), chromatin accessibility (ATAC-seq), CRISPR perturbation screens, and single-cell genomics — is the next frontier, moving from statistical association to mechanistic understanding.

Bibliography Master

1. Fisher, R. A., "The correlation between relatives on the supposition of Mendelian inheritance", Trans. R. Soc. Edinburgh 52 (1918), 399--433.

2. Lander, E. S. & Botstein, D., "Mapping Mendelian factors underlying quantitative traits using RFLP linkage maps", Genetics 121 (1989), 185--199.

3. Klein, R. J., Zeiss, C., Chew, E. Y. et al., "Complement factor H polymorphism in age-related macular degeneration", Science 308 (2005), 385--389.

4. Wellcome Trust Case Control Consortium, "Genome-wide association study of 14,000 cases of seven common diseases and 3,000 shared controls", Nature 447 (2007), 661--678.

5. Manolio, T. A., Collins, F. S., Cox, N. J. et al., "Finding the missing heritability of complex diseases", Nature 461 (2009), 747--753.

6. Yang, J., Benyamin, B., McEvoy, B. P. et al., "Common SNPs explain a large proportion of the heritability for human height", Nat. Genet. 42 (2010), 565--569.

7. Visscher, P. M., Goddard, M. E., Derks, E. M. & Wray, N. R., "Evidence-based psychiatric genetics, AKA the false dichotomy between common and rare variant hypotheses", Mol. Psychiatry 17 (2012), 474--485.

8. Futuyma, D. J. & Kirkpatrick, M., Evolution, 4th ed. (Sinauer, 2017).

9. Hartl, D. L. & Clark, A. G., Principles of Population Genetics, 4th ed. (Sinauer, 2007).

10. Lee, J. J., Wedow, R., Okbay, A. et al., "Gene discovery and polygenic prediction from a genome-wide association study of educational attainment in 1.1 million individuals", Nat. Genet. 50 (2018), 1112--1121.

11. Yengo, L., Vedantam, S., Marouli, E. et al., "A saturated map of common genetic variants associated with human height", Nature 610 (2022), 704--712.

12. Bulik-Sullivan, B. K., Loh, P.-R., Finucane, H. K. et al., "LD Score regression distinguishes confounding from polygenicity in GWAS", Nat. Genet. 47 (2015), 291--295.

13. Berg, J. J., Harpak, A., Sinnott-Armstrong, N. et al., "Reduced signal for polygenic adaptation of height in UK Biobank", eLife 8 (2019), e39725.

14. Vilhjalmsson, B. J., Yang, J., Finucane, H. K. et al., "Modeling linkage disequilibrium increases accuracy of polygenic risk scores", Am. J. Hum. Genet. 97 (2015), 576--592.