# Complex Traits, Heritability and GWAS

> For complex traits, population patterns partition variation and flag regions, not causes. Height, blood pressure and the risk of most common diseases vary smoothly from person to person, yet every allele behind them is inherited by Mendel's rules. This lesson shows how many small effects plus environment make a continuous trait, what heritability measures and what it does not, how twin studies and genome-wide association studies (GWAS) estimate and locate genetic effects, and why polygenic scores predict less well outside the groups they were built in. Afterwards you can compute a twin heritability estimate, explain a genome-wide threshold, and plan the path from an association to a causal gene.

Course: Genetics and Molecular Biology (Populations and complex traits). Level: core university (third year). About 24 minutes.
Lesson page (link this one when citing): https://stemhop.com/lessons/complex-traits-and-gwas/

Builds on: Probability, Crosses and the Chi-Square Test · Population Genetics: Hardy-Weinberg and Evolutionary Forces · Extending Mendel: Dominance, Multiple Alleles and Gene Interaction

## What you'll learn

- Explain how many loci of small additive effect plus environment produce a continuous trait, and partition phenotypic variance into VA, VD, VI and VE.
- Calculate broad-sense and narrow-sense heritability and Falconer's twin estimates, and state why each is a property of a population in an environment, not of a person.
- Explain how a GWAS works, why a lead SNP is usually a tag in linkage disequilibrium rather than the cause, where 5 × 10^-8 comes from, and how fine-mapping, expression data and editing move from a hit to a mechanism.
- Describe what a polygenic score is, why its accuracy falls outside the ancestry groups it was trained in, and the ethical stakes of genetic prediction, including the history of eugenics.

**Adult height in a well-fed population forms a smooth bell curve, yet each allele that affects it is passed on by Mendel’s rules. A 2022 study of up to 5.4 million people listed 12,111 independent variants associated with height.**

Traits like height are called complex or quantitative traits.

## The idea: many small Mendelian effects add up to a curve

A **locus** is a position on a chromosome, and the alternative versions at a locus are its **alleles**. A **polygenic** trait is one influenced by many loci, usually alongside the environment.

**Predict.** With six independently inherited allele copies, how many plus-allele totals are possible?

**Worked problem 1.** Suppose three unlinked loci (A, B, C) each have one allele that adds 2 cm to height and one that adds nothing, with no dominance, so effects simply add. Cross two people heterozygous at all three loci (AaBbCc x AaBbCc). Assume independent assortment and no environmental effect.

1. Each offspring receives six alleles, each a “plus” allele with probability 1/2, independently.
2. The number of plus alleles, k, is binomial with n = 6 and p = 1/2, so P(k) = C(6,k)/64. Here C(6,k) is the number of ways to choose which k of the six alleles are plus, and 64 is 26, because each particular set of six draws has chance (1/2)6 = 1/64.
3. For k = 0 to 6 the counts are 1, 6, 15, 20, 15, 6, 1 (sum 64).
4. So 20/64, about 31%, of offspring sit in the middle class (6 cm above baseline), and 1/64 at each extreme.

Three loci already give seven bell-shaped classes; with thousands of loci and environmental blur, they can merge into a curve. ([Probability and crosses](https://stemhop.com/lessons/probability-and-crosses/) has the tools.) This is the argument Ronald Fisher made formally in 1918, when he showed that many Mendelian factors of small effect predict the correlations between relatives that measurers of height had found.

Real traits are less tidy: effect sizes are unequal, and many alleles show **penetrance** below 100%, so not every carrier shows the effect ([extending Mendel](https://stemhop.com/lessons/extending-mendel/)). Autism risk, for example, includes rare variants of large effect and many common variants of small effect, as [the lesson on synapse disorders](https://stemhop.com/lessons/synapse-disorders/) describes.

> **The one idea.** For complex traits, population patterns partition variation and flag regions, not causes.
>
> *Scope:* variance components describe one population in one environment; [GWAS](https://stemhop.com/methods/gwas/) signals mark associated regions in one study population. Neither assigns one person’s cause or proves a causal variant or gene.

## Heritability: a ratio of variances in one population

Variance is the average squared distance of values from the mean, so it carries squared units (cm2 for height). Written as a sum of the first film’s four parts:

> **Equation card: Split the variation in a trait among people into genetic and environmental parts.**
>
>     VP = VG + VE, with VG = VA + VD + VI
>
> In words: Total variance is genetic variance plus environmental variance, and genetic variance is additive plus dominance plus interaction variance.
>
> - VP: phenotypic variance: the total variance of the trait in the population (cm² for height)
> - VG: genetic variance: the part that tracks differences in genotype
> - VE: environmental variance
> - VA: additive variance, from the average effects of alleles: the part that makes children resemble parents
> - VD: dominance variance, from interactions between the two alleles at a locus
> - VI: interaction (epistatic) variance, from interactions between loci
>
> Sanity check: If everyone shared one environment, VE = 0 and all the variation would be genetic. If everyone had the same genotype, VG = 0 and VP = VE. The sum assumes the parts are uncorrelated; real data can add gene-environment covariance and interaction, which it leaves out.

**Predict.** With VA = 30, VD = 0, VI = 0 and VE = 70, calculate broad- and narrow-sense [heritability](https://stemhop.com/methods/heritability-estimation/). Should they differ?

Heritability is a share of that total:

> **Equation card: What share of the variation in a trait, in one population, goes with genetic differences.**
>
>     H^2 = VG / VP and h^2 = VA / VP
>
> In words: Broad-sense heritability is genetic variance over total variance; narrow-sense heritability is additive variance over total variance.
>
> - H^2: the share of the variance, in this population, that is genetic of any kind. The ^2 is part of the name by convention; read H^2 as one symbol.
> - VG: genetic variance
> - VP: total phenotypic variance
> - h^2: the share that is additive, the part that predicts resemblance between parents and children
> - VA: additive variance
>
> Sanity check: As variance proportions in this model, both lie between 0 and 1, and h^2 cannot exceed H^2 because VA is part of VG. Empirical estimates can fall outside that range when sampling noise or model failure gives inadmissible values. The units cancel, so heritability has none. A change in VE alone changes both, with no change in DNA.
> Try one number Round values to keep the arithmetic easy (not data): VA = 30, VD = 0, VI = 0 and VE = 70, all in cm². VG = 30 + 0 + 0 = 30 and VP = 30 + 70 = 100. H^2 = 30/100 = 0.30 and h^2 = 30/100 = 0.30. With no dominance or interaction variance, the two heritabilities are equal.

**Worked problem 2.** Here is the first film’s change in reverse. In a hypothetical height population, VA = 40 cm2, VD = 10 cm2, VI = 0 and VE = 50 cm2. Then VP = 100 cm2, h2 = 40/100 = 0.40 and H2 = 50/100 = 0.50. Now improve nutrition so that VE falls to 20 cm2, with genetic variance unchanged: VP = 70 cm2 and h2 = 40/70, about 0.57. Heritability rose though no DNA changed.

*Interactive on the lesson page:* Variance partition: genes, environment and heritability

As the first film showed, an h2 of 0.60 does not mean that 60% of any individual’s height comes from genes. Two more limits:

- **High does not mean fixed.** Phenylketonuria is a single-gene disorder, yet a low-phenylalanine diet, started early and kept up, prevents most of its harm. Height is highly heritable within populations, yet average height rose across the twentieth century.
- **Within-group says nothing about between-group.** In an analogy Richard Lewontin set out in 1970, take a varied stock of seed and plant half in rich soil and half in poor. Within each tray, height differences are entirely genetic; the difference between trays is entirely environmental.

### Twins: two kinds of sibling as a natural experiment

Monozygotic (MZ) twins are genetically identical at conception; dizygotic (DZ) twins share on average half of their segregating genetic variation, like ordinary siblings. The **ACE model** splits variance into additive genetic (A), shared environment (C) and unique environment plus measurement error (E).

The twin correlations, rMZ and rDZ, measure how alike the two members of a pair are, from 0 (no more alike than strangers) to 1 (identical). Under the equal-environments, additive-effects and random-mating assumptions, the gap between them reflects half the additive variance, so doubling the gap estimates it:

> **Equation card: Estimate the shares of variance from genes, shared environment and unique environment, from two twin correlations (Falconer's estimates).**
>
>     h^2 = 2 × (rMZ − rDZ); c^2 = 2 × rDZ − rMZ; e^2 = 1 − rMZ
>
> In words: Heritability is twice the gap between the identical-twin and fraternal-twin correlations; shared environment is twice the fraternal correlation minus the identical one; unique environment is whatever the identical twins do not share.
>
> - h^2: additive genetic share of variance (A in the ACE model)
> - rMZ: the correlation between monozygotic (identical) twins for the trait
> - rDZ: the correlation between dizygotic (fraternal) twins
> - c^2: shared-environment share (C)
> - e^2: unique environment plus measurement error (E). Here e is a label, not the number e ≈ 2.718.
>
> Sanity check: If MZ and DZ twins are equally alike (rMZ = rDZ), h^2 = 0: genes add nothing to the resemblance. The three shares always add to 1. These estimates assume equal MZ and DZ environments, additive effects and random mating.
> Try one number Round values to keep the arithmetic easy (not data): rMZ = 0.60 and rDZ = 0.30. h^2 = 2 × (0.60 − 0.30) = 0.60. c^2 = 2 × 0.30 − 0.60 = 0. e^2 = 1 − 0.60 = 0.40. When DZ twins are exactly half as alike as MZ twins, the model finds no shared-environment effect. Check: 0.60 + 0 + 0.40 = 1.
> Where it comes from optional Optional: where it comes from. Under the model, MZ similarity is all of A plus C, so rMZ = h^2 + c^2. DZ pairs share half of A plus all of C, so rDZ = h^2/2 + c^2. Subtract: rMZ − rDZ = h^2/2, so h^2 = 2 × (rMZ − rDZ). Put that back into the first line: c^2 = rMZ − h^2 = 2 × rDZ − rMZ. Whatever MZ twins do not share is E: e^2 = 1 − rMZ.

**Worked problem 3.** For a trait in one population, rMZ = 0.80 and rDZ = 0.50.

1. h2 = 2(0.80 - 0.50) = 2(0.30) = 0.60.
2. c2 = 2(0.50) - 0.80 = 0.20.
3. e2 = 1 - 0.80 = 0.20.
4. Check: 0.60 + 0.20 + 0.20 = 1.00.

**Predict.** If MZ twins share more trait-relevant environment than DZ twins, which component will the model overestimate?

If MZ pairs share more trait-relevant environment than DZ pairs, h2 is overestimated. Sampling noise or a poor fit can also push estimates below 0 or above 1. Polderman and colleagues’ 2015 meta-analysis of twin studies of 17,804 traits reported a heritability of 49% across all traits: a summary of that literature under these assumptions, not a law.

A twin estimate says how much variation tracks genotype, not which DNA differences carry it; the second film looks for them.

**Film transcript: A GWAS Hit Is a Starting Point**

### Film transcript: A GWAS Hit Is a Starting Point

The first film treated heritability as a pattern of variation across people. If some variation tracks genotype, which DNA differences are involved? A genome-wide association study searches without choosing a candidate gene first. Its answer also belongs to that study population and setting.

Each person contributes a trait value and allele counts across the genome. At every marker, the study tests whether allele dosage is associated with the trait. The skyline is a computed plot of those tests, with taller points showing stronger evidence. Each point represents a statistical test, not a gene's importance. Testing many markers creates many chances for false positives. A stringent genome-wide threshold controls that burden.

Ancestry differences between cases and controls can also create false signals, so principal components help model population structure. A threshold controls false positives; it does not remove confounding. A peak rarely identifies a causal base by itself. Nearby alleles can stay correlated for generations, because recombination separates them only slowly. That correlation is linkage disequilibrium. A harmless marker may rise because it tags a causal variant in the same inherited block. The tallest marker sits beside several correlated neighbors. Must that tallest point be the causal variant?

No. Several variants correlated with the causal one can carry similar evidence. Fine-mapping uses the correlation pattern to rank variants and build a credible set, but that set still expresses uncertainty. Next ask which gene the variant affects in a relevant tissue. Colocalization tests whether trait and expression signals may share a variant. The target can be a neighboring gene, not the gene containing the marker. Editing then changes a candidate allele in the right cell type. If gene expression and a relevant cellular phenotype change, the causal explanation gains support. A cell result still does not prove an effect in a person. Association estimates can also become a polygenic score.

Each allele count is weighted by its estimated effect, then the terms are summed. The score must be tested in people who were not part of the original study. Carry those weights into a population with different linkage disequilibrium and allele frequencies. The same tag markers may follow causal variants less closely.

Prediction can weaken even when much underlying biology is shared. That matters before clinical use. Training data shape accuracy, and genetic prediction carries a history of coercion. Ask who was studied, who benefits, and who controls the data. Association narrows the search.

Functional evidence in cells strengthens the case. Establishing a cause also needs whole organisms or human genetics.

From the film

- A GWAS peak marks an associated region, not a proven cause.
- Linkage disequilibrium lets a noncausal SNP tag a nearby causal variant.
- Fine-mapping, colocalization and editing build the case; a person-level effect needs separate evidence.
- A polygenic score must be tested outside its GWAS; accuracy can drop where LD and allele frequencies differ.

## Scanning the genome: association, tags and a threshold

A **single nucleotide polymorphism (SNP)** is a base position where people carry different nucleotides. A **[genome-wide association](https://stemhop.com/methods/gwas/) study (GWAS)** genotypes hundreds of thousands to millions of SNPs in many people and regresses the trait on allele dosage (0, 1 or 2 copies) at each one: linear regression for a measurement, logistic for a disease.

The second film’s travelling block is **linkage disequilibrium (LD)**, the nonrandom association of alleles at different loci in a population, usually strongest between nearby loci. Recombination erodes it only slowly, so it can persist for many generations; mutation, drift, selection and admixture create or reshape it. It differs from linkage in a family, where you count recombinant offspring ([linkage and mapping](https://stemhop.com/lessons/linkage-and-mapping/)).

*Figure (AI-generated illustration).* Neighbouring variants travel together on a stretch of chromosome for many generations. Because of linkage disequilibrium, a SNP that does nothing can show a strong association by tagging a causal variant nearby, so a GWAS peak marks a region. Image: a chromosome stretch drawn as twelve equal tiles. A bracket over the middle tiles marks a block of neighbouring variants that travel together; a hollow circle marks a tagging SNP and a gold diamond marks the causal variant nearby.

In 2005 Klein and colleagues, with 96 cases of age-related macular degeneration and 50 controls, found a strong association with a common intronic SNP in CFH, then proposed a coding change in LD with it as a candidate cause.

**Predict.** With twice as many independent tests, should the Bonferroni threshold be larger or smaller?

**Worked problem 4.** Why 5 x 10−8? Pe’er and colleagues estimated about one million independent common-variant tests across the genome in European-ancestry HapMap data. A Bonferroni correction divides the acceptable chance of any false positive, 0.05, by the number of tests: 0.05 / 1,000,000 = 5 x 10−8. In powers of ten, 0.05 is 5 x 10−2 and a million is 106, and dividing subtracts the exponents: −2 − 6 = −8. Written out, that is 0.00000005 (see [powers](https://stemhop.com/math-toolkit/#powers)). The same study estimated about twice as many independent tests in African-ancestry HapMap data, where LD blocks are shorter. Twice the tests halves the threshold, to about 0.05 / 2,000,000 = 2.5 x 10−8. The threshold is a convention tied to those LD patterns.

A false hit can also come from **population stratification**, the second film’s ancestry imbalance: SNPs whose frequencies differ between the source populations can look associated even when they do not affect the trait. Price and colleagues showed in 2006 that principal components of genome-wide genotypes (a few summary axes of genetic variation that track ancestry) can model these differences; they are now standard covariates that reduce confounding without guaranteeing its absence.

Yang, Visscher and colleagues estimated in 2010 that common SNPs analysed jointly captured 45% of height variance in 3,925 unrelated people, far more than the significant hits alone, because most effects are too small to pass the threshold. Later whole-genome sequencing suggested that much of the remaining gap lies in rare variants that arrays tag poorly.

## From a hit to a mechanism

The second film’s three rungs, **fine-mapping**, **colocalization** and **editing**, are tidier than a real locus. Fine-mapping often uses linkage disequilibrium (LD) patterns from several ancestry groups to build a credible set likely to contain the causal variant, and colocalization often points to a gene other than the nearest.

At FTO the work relied on chromatin contacts, epigenomic maps and editing more than on formal fine-mapping. The obesity association lies in FTO introns, yet Smemo and colleagues (2014) reported that the region regulates IRX3. Claussnitzer and colleagues (2015) then studied human fat-cell precursors and linked the signal to rs1421085, a T-to-C change that disrupts a binding motif for the repressor ARID5B and is associated with doubled IRX3 and IRX5 expression early in differentiation. Correcting the allele by [CRISPR](https://stemhop.com/methods/crispr/) in cultured primary adipocytes from a risk-allele carrier restored heat production, reported as about sevenfold. That edit supports a cellular mechanism; it does not show weight change in a person. The mechanism is still debated: later work found several functional variants on the same haplotype, and mice carrying the risk base had more brown-fat heat production, not less. See [the CRISPR lesson](https://stemhop.com/lessons/crispr-and-functional-genomics/) for the tools.

## Polygenic scores, portability and the weight of history

A **polygenic score** sums weighted allele counts:

> **Equation card: Combine thousands of small GWAS effects into one predicted value for a person.**
>
>     PGS = sum over SNPs of (effect estimate × allele count)
>
> In words: The polygenic score is, for every SNP in the panel, the effect estimate times how many copies of the effect allele the person carries, all added up.
>
> - PGS: the person's score, in the trait's units (cm for height)
> - effect estimate: the GWAS estimate of how much one copy of the allele shifts the trait, on average, in the study population
> - allele count: how many copies of that allele the person carries: 0, 1 or 2
>
> Sanity check: Someone carrying no copies of any effect allele scores 0, and each extra copy adds its effect once. The effect estimates come from one study population, so the score is only as accurate as those estimates are for the person being scored.
> Try one number A made-up panel of three SNPs, to show the arithmetic: effects +0.5, +0.2 and −0.3 cm per copy; the person carries 2, 1 and 0 copies. PGS = 0.5 × 2 + 0.2 × 1 + (−0.3) × 0 = 1.0 + 0.2 + 0 = 1.2 cm. A real height panel sums thousands of such terms, each tiny.

A polygenic score must be tested outside its [GWAS](https://stemhop.com/methods/gwas/); accuracy can drop when linkage disequilibrium (LD), allele frequencies, effect sizes or environments differ. Like [heritability](https://stemhop.com/methods/heritability-estimation/), its accuracy belongs to the population and setting where it is tested. Yengo and colleagues reported in 2022 that their 12,111-SNP height panel captured about 40% of variance out of sample in European-ancestry analyses but around 10 to 20% in the other ancestry groups examined; they judged differences in LD and allele frequency likely explanations. Martin and colleagues reported in 2019 that the scores of that time were several times more accurate in European-ancestry individuals, and warned that clinical use could widen health disparities. Accuracy also varied with age, sex and socioeconomic characteristics within one ancestry group in UK Biobank, as Mostafavi and colleagues found in 2020. Allele frequencies are the subject of [population genetics](https://stemhop.com/lessons/population-genetics/).

Genetic prediction has a history. Francis Galton coined the word eugenics in 1883. In the United States, states passed laws for forced sterilization of people judged unfit, and in Buck v. Bell (1927) the Supreme Court upheld Virginia’s law. Tens of thousands of people were sterilized. That history is why the second film’s three questions come before clinical use.

### Design problem

**Predict.** Does p = 3 x 10−12 pass a threshold of 5 x 10−8? Compare the exponents first.

A genome-wide association study (GWAS) reports a lead single nucleotide polymorphism (SNP) in an intron of gene A, with p = 3 x 10−12 and an odds ratio of 1.08 per risk allele. The associated LD block spans genes A, B and C. In the relevant tissue, the risk allele associates with higher expression of C, not A. Which gene is causal, and how would you find out?

> **One way to do it.** Do not name a gene from the association. The p-value (3 x 10−12, below 5 x 10−8) says data this extreme would be very unlikely if the lead SNP had no association with the trait, not which variant or gene is responsible, because SNPs in strong LD with the lead SNP can carry a similar signal. An odds ratio of 1.08 is a small effect for any one person.
>
> 1. **Fine-map.** Use the LD structure, ideally from cohorts of more than one ancestry, to build a credible set of variants likely to contain the causal one.
> 2. **Colocalize.** Test whether the disease signal and the expression signal for C share one causal variant, rather than being two nearby signals; test A and B too.
> 3. **Edit.** In the relevant human cell type, use CRISPR to swap the credible-set variant (or delete the candidate enhancer), with matched unedited and sham-edited controls. Measure A, B and C expression and a cellular phenotype linked to the disease.
>
> The FTO region, though still debated, is the precedent. A positive edit implicates C in that cell type; linking C to the disease still needs evidence from whole organisms or human genetics.

---

From STEMhop (https://stemhop.com), free lessons by Christopher Deppmann. Lesson text © the authors. Practice questions: https://stemhop.com/lessons/complex-traits-and-gwas/
