Advancedadvanced biology
Quantitative Genetics: Heritability, Selection Response, and Polygenic Traits
Learning objectives
- Partition phenotypic variance into genetic and environmental components.
- Distinguish broad-sense from narrow-sense heritability and explain why only the latter predicts selection response.
- Apply the breeder's equation R = h^2 S to predict evolutionary change.
- Explain how twin and adoption studies estimate heritability and their limitations.
- Evaluate GWAS, polygenic scores, and the missing heritability problem.
Partitioning variance: the quantitative framework
Most traits that matter -- height, yield, disease risk -- vary continuously and are influenced by many genes plus environment. R. A. Fisher's 1918 infinitesimal model showed that Mendelian inheritance of many small-effect loci produces continuous variation, reconciling Mendel with Darwin. The framework begins by partitioning phenotypic variance: V_P = V_G + V_E (+ V_GxE), where genetic variance further splits into additive (V_A), dominance (V_D), and epistatic (V_I) components. Only additive variance -- the average effects of alleles -- is reliably transmitted from parent to offspring, because meiosis shuffles allele combinations each generation, breaking up dominance and epistatic configurations. This is why resemblance between relatives follows predictable patterns: parent-offspring regression slopes estimate half the narrow-sense heritability, and the correlation structure among twins, siblings, and cousins each constrains the components differently. The variance-partitioning mindset is quantitative genetics' core move: it replaces arguments about nature versus nurture with measurable components.
Heritability: what it is and is not
Heritability is among the most misunderstood quantities in biology. Broad-sense heritability H^2 = V_G/V_P measures the total genetic contribution to variation; narrow-sense heritability h^2 = V_A/V_P measures only the additive portion -- the part that responds to selection. Both are population statistics, not individual properties: a heritability of 0.8 for height means 80% of height variation in that population at that time is genetic, not that any individual's height is 80% genetic. Heritability depends on environment: in uniform environments, genetic differences explain more variation (higher h^2); in variable environments, less. It says nothing about immutability -- high heritability traits can change dramatically with environmental shifts, as the 20th-century height increase shows. Twin studies estimate heritability by comparing monozygotic and dizygotic correlations, but the equal-environments assumption is imperfect, and gene-environment correlation inflates estimates. Modern SNP-based heritability from unrelated individuals avoids some biases but captures only tagged common variants.
The breeder's equation and selection response
The breeder's equation, R = h^2 S, is quantitative genetics' most elegant prediction: the response to selection (R, the change in mean) equals narrow-sense heritability times the selection differential (S, the difference between selected parents and the population mean). Artificial selection provides spectacular confirmations -- decades of selection on oil content in maize, or the famous Russian farm-fox experiment selecting for tameness, which inadvertently reshaped morphology and reproduction through correlated responses. In nature, the equation predicts evolutionary trajectories: Peter and Rosemary Grant's work on Darwin's finches showed beak dimensions evolving measurably within generations as droughts imposed selection, with responses matching h^2 S predictions. Complications arise: selection depletes additive variance (the Bulmer effect), genetic correlations between traits cause indirect responses (selecting larger size may reduce fertility), and genotype-by-environment interaction means h^2 measured in one environment may not transfer. The multivariate breeder's equation, delta-z = G beta, generalizes to correlated traits using the genetic variance-covariance matrix G.
From QTLs to GWAS: finding the loci
The infinitesimal model assumes many loci but does not identify them. QTL mapping in experimental crosses localized trait-affecting regions to chromosomal intervals, often broad and containing hundreds of genes. Genome-wide association studies (GWAS), enabled by HapMap linkage-disequilibrium maps and cheap genotyping, tested millions of SNPs in thousands of individuals, finding that most complex traits are influenced by thousands of variants of tiny effect -- height involves on the order of 10,000 loci. This confirmed Fisher's infinitesimal vision quantitatively. GWAS hits cluster in regulatory DNA, echoing evo-devo's lesson that regulation drives phenotypic change, and effect sizes are small because large-effect variants are removed by selection or kept rare. Fine-mapping the causal variant within an LD block remains difficult, and most GWAS loci lack functional validation -- the field's persistent gap between statistical association and biological mechanism.
Polygenic scores and missing heritability
Summing GWAS effect sizes across the genome yields polygenic scores -- predictors of individual trait values and disease risk. For height, scores now explain a substantial fraction of variance; for diseases like coronary artery disease or schizophrenia, they stratify risk usefully at population level. But the missing heritability problem looms: GWAS-significant SNPs typically explain far less variance than twin studies suggest. Proposed resolutions include many undetected small-effect variants (now being found as sample sizes reach millions), rare variants missed by genotyping arrays, inflated twin estimates, and non-additive or interaction effects. Polygenic scores also transfer poorly across ancestries because LD patterns and allele frequencies differ -- a serious equity problem as clinical use grows. The frontier combines whole-genome sequencing, diverse cohorts, and functional genomics to close the gap. Quantitative genetics thus ends where it began: Fisher's mathematics works beautifully, but connecting its variance components to actual DNA remains the grand challenge.
High heritability does not mean unchangeable. It describes variation in one population and environment. Change the environment -- nutrition, education, medicine -- and both the trait mean and its heritability can shift dramatically.
Key points
- V_P = V_G + V_E; only additive variance V_A is reliably transmitted and responds to selection.
- Narrow-sense heritability h^2 = V_A/V_P is a population statistic, not an individual property and not a measure of immutability.
- The breeder's equation R = h^2 S predicts selection response; the multivariate form handles correlated traits.
- Twin studies estimate heritability but face equal-environment and gene-environment correlation biases.
- GWAS shows complex traits involve thousands of small-effect, mostly regulatory variants.
- Polygenic scores predict traits but suffer missing heritability and poor cross-ancestry transfer.
Common misconceptions
✓ Truth: Heritability partitions variation across a population, not causation within an individual. It cannot be divided per person, and it changes with the population and environment studied.
✓ Truth: Heritability says nothing about malleability. Height is highly heritable yet average height rose dramatically with nutrition; heritability describes current variation, not future potential.
✓ Truth: GWAS finds associated regions, usually in regulatory DNA with tiny effects, and most loci lack identified causal variants or mechanisms. Finding statistical associations is the beginning of biology, not the end.
Frequently asked questions
Why does only additive variance respond to selection?
Because meiosis breaks up dominance and epistatic combinations each generation -- offspring inherit alleles, not genotypes. Only the average, additive effects of alleles are transmitted predictably, so only V_A contributes to the parent-offspring resemblance that selection exploits.
What causes missing heritability?
Partly many tiny-effect variants below detection thresholds, partly rare variants missed by arrays, partly inflated twin-study estimates, and partly non-additive effects. Million-person GWAS are steadily recovering the missing portion, suggesting the infinitesimal model was essentially right.
Why do polygenic scores work poorly across ancestries?
Scores are built from European-ancestry GWAS, but linkage-disequilibrium patterns, allele frequencies, and even causal variants differ between populations. A score tuned to European LD tags causal variants poorly elsewhere -- a major equity concern for clinical use.
Did the Grants really watch evolution in real time?
Yes. On Daphne Major, drought-induced selection on finch beaks produced measurable evolutionary responses within a generation, and the responses matched breeder's-equation predictions using field-estimated heritabilities -- among the cleanest demonstrations of natural selection ever documented.