Explain genome-wide association studies (GWAS)

Date: 2026-07-01
Tags: #genomics #GWAS #complex-traits #statistical-genetics #SNP-analysis

Overview

Genome-Wide Association Studies (GWAS) are systematic approaches to identify genetic variants (specifically single nucleotide polymorphisms or SNPs) associated with specific traits or diseases across entire genomes by comparing allele frequencies between cases and controls.

[!intuition] Core Intuition

Think of GWAS as a genetic treasure hunt across millions of locations. Imagine you want to find out why some people are tall and others short. Instead of checking one or two genes, you check hundreds of thousands to millions of genetic positions simultaneously across many people.

The key insight: If a particular genetic variant (like having an A instead of G at position 12,345,678 on chromosome 7) appears much more frequently in tall people than short people, that variant is associated with height. It's like noticing that people wearing red shirts cluster at the tall end of a line—the red shirt (genetic variant) is corelated with tallness, even if it doesn't directly cause it.

Why GWAS matters: Most common diseases and traits (diabetes, height, schizophrenia, blood pressure) are polygenic—influenced by many genes, each with small effects. GWAS is the tool that can detect these subtle genetic contributions.

[!definition] Formal Definition

A genome-wide association study (GWAS) is an observational study of a genome-wide set of genetic variants in different individuals to determine if any variant is associated with a phenotype. GWAS typically focus on associations between single nucleotide polymorphisms (SNPs) and major traits or diseases.

Key components:

• Cases: Individuals with the disease/trait
• Controls: Individuals without the disease/trait
• SNPs: Genetic markers tested (typically 500,000 to 5+ million)
• Association test: Statistical comparison of allele frequencies

[!formula] Statistical Foundation

The2×2 Contingency Table

For each SNP, we construct a table comparing allele frequencies:

Chi-Square Test of Association

$\chi^2 = \frac{n(ad - bc)^2}{(a+b)(c+d)(a+c)(b+d)}$

where $n = a + b + c + d$ (total allele count).

Why this formula?

• The numerator $(ad - bc)^2$ measures deviation from independence. If alleles are distributed identically in cases and controls, $ad \approx bc$, making $\chi^2 \approx 0$.
• The denominator normalizes by marginal totals, accounting for overall allele frequencies and sample sizes.
• Under the null hypothesis (no association), $\chi^2$ follows a chi-square distribution with 1 degree of freedom.

P-value and Significance

$P = P(\chi^2_{(1)} \geq \chi^2_{observed})$

But we test millions of SNPs, so we need Bonferroni correction:

$\alpha_{corrected} = \frac{\alpha}{m}$

where $m$ is the number of independent tests.

Derivation of genome-wide significance threshold:

• Start with $\alpha = 0.05$ (standard significance)
• Assume ~1 million independent SNPs (accounting for linkage disequilibrium)
• $\alpha_{GW} = \frac{0.05}{1,000,000} = 5 \times 10^{-8}$

Why this matters: This is the famous genome-wide significance threshold: p < 5×10⁻⁸. Any association weaker than this is likely a false positive.

Odds Ratio

To quantify effect size, we calculate the odds ratio (OR):

$OR = \frac{a/c}{b/d} = \frac{ad}{bc}$

Why use odds ratio?

• OR = 1: No association (allele frequency identical in cases and controls)
• OR > 1: Risk allele (more common in cases)
• OR < 1: Protective allele (less common in cases)

Derivation from first principles:

• Odds of disease given allele A: $\frac{a}{c}$ (cases with A : controls with A)
• Odds of disease given allele G: $\frac{b}{d}$
• Ratio of these odds gives the odds ratio

For small effect sizes (typical in GWAS), we often report log odds ratio because it's symmetric around zero and approximately normal in distribution.

[!example] Worked Example 1: Type 2 Diabetes SNP

Setup: We genotype SNP rs7903146 in the TCF7L2 gene in 5,000 Type 2 Diabetes (T2D) cases and 5,000 controls.

Data:

• T2D cases: 3,200 have T allele, 1,800 have C allele
• Controls: 2,400 have T allele, 2,600 have C allele

Step 1: Construct contingency table

Why this step? We need to organize observed frequencies to compare against expected frequencies under independence.

Step 2: Calculate chi-square statistic

$\chi^2 = \frac{10,000(3,200 \times 2,600 - 2,400 \times 1,800)^2}{5,600 \times 4,400 \times 5,000 \times 5,000}$

$= \frac{10,000(8,320,000 - 4,320,000)^2}{616,000,000,000,000}$

$= \frac{10,000 \times 16,000,000,000}{616,000,000,000} = 259.74$

Why this step? The large $\chi^2$ value indicates a strong deviation from the null hypothesis of no association.

Step 3: Find p-value

With $\chi^2 = 259.74$ and df = 1, we get p < 10⁻⁵⁷ (far below 5×10⁻⁸).

Why this step? This extremely small p-value confirms genome-wide significance—this SNP is robustly associated with T2D.

Step 4: Calculate odds ratio

$OR = \frac{3,200\times 2,600}{2,400 \times 1,800} = \frac{8,320,000}{4,320,000} = 1.93$

Interpretation: Individuals carrying the T allele have 1.93× higher odds of developing Type 2 Diabetes compared to those with the C allele.

Why this step? While p-value tells us the association is real, OR tells us the magnitude of the effect—a 93% increase in odds is clinically meaningful.

[!example] Worked Example 2: Manhattan Plot Interpretation

Setup: You've run a GWAS for schizophrenia with 500,000 SNPs across 10,000 cases and 10,000 controls.

Results visualization: A Manhattan plot displays −log₁₀(p-value) on y-axis vs. chromosomal position on x-axis.

Key features:

1. Genome-wide significance line: Horizontal line at −log₁₀(5×10⁻⁸) = 7.3
2. Peaks above the line: Significant associations
3. Chromosome clustering: Multiple nearby SNPs showing association (linkage disequilibrium)

Example peak: Chromosome 6, position 25-35 Mb, showing 15 SNPs with p < 5×10⁻⁸.

Why this step? The clustering tells us this isn't 15 independent findings—these SNPs are in linkage disequilibrium (LD), meaning they're inherited together. This region contains one causal variant (or a few), and the others are just tagging it.

Step for fine-mapping: To identify the causal variant:

1. Calculate LD between all SNPs in this region
2. Perform conditional analysis: Test each SNP while controlling for the top SNP
3. SNPs that remain significant are likely independent signals
4. SNPs that become non-significant were just tagging the top SNP

Why this matters? This distinguishes genuine causal variants from corelated bystanders.

[!example] Worked Example 3: Power Calculation

Question: How many samples do we need to detect a SNP with OR = 1.2(small effect) at genome-wide significance (p < 5×10⁻⁸) with 80% power?

Step 1: Set parameters

• Effect size: OR = 1.2
• Risk allele frequency: f = 0.3(common variant)
• Significance level: α = 5×10⁻⁸
• Desired power: 1 - β = 0.80

Step 2: Convert OR to genetic relative risk

For a multiplicative model: $GR = \frac{OR}{1 - f + f \times OR}$

Why this formula? OR measures association in case-control studies, but the underlying genetic model uses relative risk. This converts between them.

$GRR = \frac{1.2}{1 - 0.3 + 0.3 \times 1.2} = \frac{1.2}{1.06} = 1.132$

Step 3: Apply power formula

For a chi-square test: $n = \frac{(Z_{\alpha/2} + Z\beta)^2}{2f(1-f)(GRR - 1)^2 \times K \times (1-K)}$

where K = disease prevalence.

Why this formula? Statistical power depends on:

• Signal strength: $(GRR - 1)^2$ (effect size)
• Genetic variance: $2f(1-f)$ (heterozygosity)
• Sample composition: $K(1-K)$ (case-control balance)

For schizophrenia, K ≈ 0.01 (1% prevalence), balanced design:

$n = \frac{(5.45 + 0.84)^2}{2 \times 0.3 \times 0.7 \times 0.132^2 \times 0.01 \times 0.99}$

$= \frac{39.56}{0.0000714} \approx 554,000 \text{ per group}$

Interpretation: We need ~1.1 million total samples (550K cases + 550K controls) to reliably detect a common variant with small effect (OR = 1.2) at genome-wide significance.

Why this matters: This explains why early GWAS (with 5,000-10,000 samples) missed most common variant associations. Current mega-GWAS consortia with >500,000 samples are necessary, not excessive.

[!mistake] Common Mistakes

Mistake 1: "Significant SNP = Causal Gene"

Why it feels right: The SNP shows strong association, so it must be causing the disease.

Why it's wrong: Most GWAS SNPs are not causal—they're in linkage disequilibrium with the true causal variant. The significant SNP might be in a gene desert, while the causal variant is in a regulatory region 200 kb away, affecting different gene.

The fix:

1. GWAS identifies loci (regions), not causal variants
2. Requires fine-mapping with denser genotyping/sequencing
3. Needs functional follow-up (expression QTL, CRISPR validation) to identify mechanism

Example: The FTO locus was associated with obesity, and everyone assumed FTO gene was causal. Years later, researchers discovered the causal variants regulate IRX3 and IRX5, genes 500 kb away, which control thermogenesis.

Mistake 2: "No genome-wide significant hits = No genetic basis"

Why it feels right: If GWAS finds nothing significant, the trait must not be genetic.

Why it's wrong: Four possible explanations:

1. Insufficient power: Effect sizes too small for your sample size
2. Rare variants: GWAS chips capture common variants (MAF > 5%); rare variants need sequencing
3. Structural variants: CNVs, inversions not captured by SNP arrays
4. Environmental interaction: Genetic effects only manifest in specific environments (G×E)

The fix:

• Check heritability estimates from twin/family studies (if h² > 0, there IS genetic component)
• Increase sample size (mega-consortia with 100K+ samples)
• Use sequencing-based approaches for rare variants
• Consider polygenic risk scores (PRS) that aggregate weak signals

Example: Schizophrenia has h² ≈ 80% but early GWAS found nothing. With >100,000 samples, modern GWAS identified 200+ loci, explaining ~20% of heritability. The "missing heritability" is distributed across thousands of variants with tiny effects.

Mistake 3: "P-value ranking = Effect size ranking"

Why it feels right: The SNP with lowest p-value must have the strongest biological effect.

Why it's wrong: P-value depends on both effect size AND sample size (and allele frequency). A common variant (MAF = 0.4) with OR = 1.15 can have p = 10⁻⁵⁰, while a rare variant (MAF = 0.01) with OR = 3.0 might have p = 10⁻⁶.

The fix: Always report effect size (OR or β) and confidence intervals alongside p-values. For clinical/biological interpretation, effect size matters more.

Example: In height GWAS, the top p-value SNP (p = 10⁻⁷⁰) explains 0.4 cm of height, while a rarer variant (p = 10⁻⁸) explains 2cm. The latter is more biologically interesting despite weaker p-value.

Mistake 4: "GWAS requires family studies or pedigrees"

Why it feels right: Genetics studies traditionally used families to track inheritance.

Why it's wrong: GWAS is a population-based approach using unrelated individuals. The power comes from:

• Large sample sizes (not deep pedigrees)
• Common variant discovery (not rare Mendelian mutations)
• Population allele frequency differences between cases and controls

The fix: Recognize that GWAS and linkage analysis are complementary:

• Linkage (family-based): Good for rare, high-penetrance variants (Mendelian diseases)
• GWAS (population-based): Good for common, low-penetrance variants (complex traits)

Example: Huntington's disease (caused by single CAG repeat expansion) was mapped by linkage analysis. Type 2 diabetes (polygenic, 500+ variants) was dissected by GWAS.

[!recall]- Feynman Explanation (Age 12)

Imagine you have 1,000 kids in your school, and you notice that some kids are really tall and some are short. You wonder: "Is there something in their DNA that makes them tall?"

Your DNA is like an instruction book with3 billion letters (A, T, C, G). Sometimes one kid has an "A" at position #12345 while another kid has a "G" at that same position. These tiny differences are called SNPs (pronounce it "snips").

Here's what scientists do in a GWAS:

1. They measure everyone's height
2. They check about 1 million of these SNP positions in every person
3. For each SNP, they ask: "Do more tall kids have A, while more short kids have G?"

If they find that yes, tall kids are way more likely to have A at position #12345, that SNP is associated with height. It's like finding a clue!

But here's the tricky part: Having that "A" doesn't cause you to be tall by itself. It's just a marker. Think of it like this: if all the tall kids wear red backpacks, you might notice "red backpack = tall kid" correlation. But the backpack doesn't make you tall! Similarly, the SNP might just be sitting near the actual gene that affects height.

Scientists have to do a LOT of extra work to figure out which gene is really responsible and how it works. The GWAS just gives them a map with "X marks the spot" for where to dig deeper.

Why do they need so many people? Because each SNP only makes a tiny, tiny difference—maybe 1 millimeter of height. To spot such a small effect, you need thousands of people. It's like trying to hear a whisper in a noisy room—you need many people to confirm "Yes, I heard it too!"

[!mnemonic] Memory Aids

GWAS = "Genome-Wide A-lot-of SNPs"

• Genome-wide: Scan the entire genome
• Wide: Hundreds of thousands to millions of positions
• Association: Looking for correlation, not causation
• Studies: Population-based comparison

The 5×10⁻⁸ rule: "Five in a hundred million" sounds harsh, but remember: "Million tests, micro-p" (testing millions of SNPs requires microscopic p-values).

OR interpretation:

• OR = 1: "One is none" (no effect)
• OR = 2: "Two times the trouble" (double the risk)
• OR = 0.5: "Half the hazard" (protective)

LD (Linkage Disequilibrium): "Linked DNA goes Downstream together"—nearby SNPs are inherited as blocks, so they correlate even if only one is causal.

Key Concepts

• Genome-Wide Association Study (GWAS): Systematic scan of genetic variants across entire genomes to identify associations with traits/diseases
• Single Nucleotide Polymorphism (SNP): Single-base DNA variation at a specific genomic position (e.g., A vs. G)
• Case-control design: Compare allele frequencies between individuals with trait (cases) and without (controls)
• Genome-wide significance: P-value threshold of 5×10⁻⁸ accounting for multiple testing correction
• Odds Ratio (OR): Measure of effect size; ratio of odds of disease given risk allele vs. reference allele
• Linkage Disequilibrium (LD): Non-random association of alleles at different loci due to physical proximity on chromosome
• Manhattan plot: Visualization showing −log₁₀(p-value) across all chromosomes
• Minor Allele Frequency (MAF): Frequency of the less common allele at a SNP (common variants: MAF > 5%)
• Fine-mapping: Follow-up analysis to pinpoint causal variants within associated loci
• Polygenic trait: Phenotype influenced by many genetic variants, each with small effect
• Missing heritability: Gap between heritability estimated from families vs. variance explained by discovered SNPs

Connections

• Single Nucleotide Polymorphisms (SNPs): The genetic markers interrogated by GWAS
• Linkage Disequilibrium and Haplotypes: Explains why multiple nearby SNPs show association
• Polygenic Risk Scores: Use GWAS results to predict individual disease risk
• Heritability and Variance Components: GWAS findings contribute to understanding trait heritability
• Expression QTL (eQTL) Studies: Functional follow-up to GWAS to identify mechanism
• Mendelian Randomization: Uses GWAS variants as instrumental variables for causal inference
• Population Stratification: Confounding factor requiring correction in GWAS
• Fine-mapping and Credible Sets: Statistical methods to narrow down causal variants
• Rare Variant Association Studies: Complement GWAS for low-frequency alleles
• Pharmacogenomics: GWAS applied to drug response traits

Flashcards

What is a genome-wide association study (GWAS)? :: An observational study that tests hundreds of thousands to millions of genetic variants (SNPs) across the genome to identify associations between genotype and phenotype by comparing allele frequencies between cases and controls.

Calculate the genome-wide significance threshold if testing 500,000 independent SNPs with α = 0.05 :: Using Bonferroni correction: α_corrected = 0.05 / 500,000 = 1×10⁻⁷. This is the p-value threshold below which associations are considered statistically significant after correcting for multiple testing.

What is the difference between GWAS and linkage analysis? :: GWAS uses unrelated individuals to test common variants (MAF > 5%) in population-based case-control designs. Linkage analysis uses families/pedigrees to track inheritance of rare, high-penetrance variants. GWAS is for complex polygenic traits; linkage is for Mendelian diseases.

Concept Map

Hinglish (regional understanding)