Evolution at Multiple Loci

Single-locus models track one number per generation: the frequency of an allele, p. Extending to two loci requires tracking not two numbers but four haplotype frequencies — the frequencies of the four possible combinations of alleles at two biallelic loci (A/a and B/b): x₁ = freq(AB), x₂ = freq(Ab), x₃ = freq(aB), x₄ = freq(ab). If the two loci were statistically independent, each haplotype frequency would simply be the product of its constituent allele frequencies: freq(AB) would equal p_A × p_B. The extent to which reality departs from this independence is linkage disequilibrium, defined as

D = x₁x₄ − x₂x₃ = freq(AB) − p_A·p_B D is not about physical distance alone — despite the name "linkage" disequilibrium, it is fundamentally a statistical association between alleles at two loci, which can arise even between loci on different chromosomes (through selection, admixture, or drift) and can be zero even between physically adjacent loci if recombination has had enough time to erase it. "Linkage" describes physical proximity on a chromosome (measured by recombination fraction, r, ranging from 0 for completely linked loci to 0.5 for loci on different chromosomes or very far apart on the same one); "linkage disequilibrium" describes the resulting statistical non-independence of alleles. The two are related but distinct: tight physical linkage (small r) slows the decay of D once it exists, but something else — selection, drift, mutation, or admixture — must create D in the first place. A normalized version, r² (the squared correlation coefficient between the two loci' allele states), is what most genomic studies report, because D's raw magnitude depends on allele frequencies in ways that make raw D hard to compare across loci. r² ranges from 0 (no association) to 1 (perfect association, where knowing the allele at one locus tells you the allele at the other with certainty). Epistasis is the second key concept: it is any deviation from additivity in how alleles at different loci combine to determine fitness (or any other phenotype). If locus A's effect on fitness depends on which allele is present at locus B, the loci are epistatic, and selection then acts on specific combinations — creating exactly the kind of non-independence that D measures. Together, LD and epistasis explain why the two-locus (and n-locus) world can behave qualitatively differently from naive extrapolation of single-locus selection theory: alleles can be dragged to fixation or elimination not because of their own effects, but because of what they're temporarily riding alongside.

Step 1 — Recombination decays linkage disequilibrium geometrically. Each generation, a fraction r of gametes are recombinant between the two loci (r is the recombination fraction, ranging 0–0.5). Recombination takes AB and ab haplotypes and, when they pair in a heterozygote (AaBb), can produce Ab and aB gametes instead. The standard two-locus recursion shows that after one generation of random mating and recombination, disequilibrium becomes: D′ = (1 − r) D Iterating over t generations:

D_t = D_0 (1 − r)^t This is one of the cleanest quantitative results in population genetics. If loci are unlinked (r = 0.5), D decays by half every generation — four generations essentially erases any initial association. If loci are tightly linked (r = 0.001, roughly 0.1 cM apart, about 100 kb in humans), D decays by only 0.1% per generation, and an association created hundreds of generations ago can still be detectable today. This is precisely why LD mapping works in humans: a disease-causing mutation that arose on a particular ancestral chromosome remains statistically associated with nearby neutral markers for many generations, allowing geneticists to localize genes by scanning for LD between a trait and dense marker panels.

Step 2 — Selection and epistasis actively build D. Suppose fitness is epistatic: the AB haplotype has fitness 1.2, but Ab and aB each have fitness 0.9, and ab has fitness 1.0 (a case of positive/synergistic epistasis — the "good" alleles are good only together). Selection will increase the frequency of AB and ab relative to the recombinant classes Ab and aB, generating positive D even if it started at zero — because selection favors the specific combination, not the alleles independently. This is fundamentally different from single-locus selection: you cannot predict the outcome from the marginal fitness effects of A and B separately; you need the full two-locus fitness matrix (a fitness landscape in miniature). Conversely, with negative epistasis (each single mutant is more fit than the double mutant — as in cases of deleterious epistasis or "diminishing returns" cases), selection favors recombination products, because recombination lets a population escape the fitness valley between AB and ab and produce the fitter single mutants.

Step 3 — Drift and demography also generate LD, even with no epistasis at all. In a finite population, sampling is stochastic; by chance, particular allele combinations become momentarily over- or under-represented, generating LD with expected magnitude on the order of 1/(4N_e·r) for loci with recombination fraction r in a population of effective size N_e. Population admixture — mixing two previously separated populations that differ in allele frequencies at both loci — creates LD even between unlinked loci, because each source population contributes correlated bundles of ancestry-associated alleles. This admixture LD decays with each generation of interbreeding and is the basis for admixture mapping.

Step 4 — Genetic hitchhiking: selection on one locus drags along its neutral neighbors. When a new beneficial mutation arises at locus B and sweeps toward fixation, it does so on a particular chromosomal background — whatever alleles happened to be present at nearby loci on that one original chromosome. As B rises in frequency, tightly linked neutral variants rise with it purely by physical association, a process called genetic hitchhiking or a selective sweep. The closer a neutral locus sits to the selected site (smaller r), the more strongly and rapidly it hitchhikes, since recombination has less time to break the association before fixation. The result is a characteristic "valley" of reduced genetic diversity surrounding the selected site — a dip whose width is inversely related to the recombination rate and directly related to the strength and speed of selection. This diversity-reduction signature is one of the primary tools used to scan modern genomes for the footprints of past positive selection.

Step 5 — The Hill–Robertson effect: linkage interferes with selection's efficiency at multiple selected loci simultaneously. When two beneficial (or two deleterious) mutations arise on different chromosomal backgrounds in a population with limited recombination, selection cannot act on them independently — favoring one lineage while disfavoring the alternative background carrying the other beneficial mutation, effectively pitting them against each other rather than combining them. Recombination is required to bring both beneficial mutations onto the same chromosome for selection to fix them jointly and efficiently. Without recombination, this "clonal interference" or negative Hill–Robertson interference reduces the net efficiency of selection, slows adaptation, and increases the fixation load of mildly deleterious mutations (a process related to Muller's ratchet in fully asexual lineages). This effect is one of the leading hypothesized benefits of recombination and sex: it relieves interference between selected loci, letting beneficial mutations combine and deleterious ones be purged independently.

Genomic linkage disequilibrium mapping — the human HapMap/1000 Genomes demonstration that LD decays with distance as theory predicts. One of the clearest large-scale empirical confirmations of the recombination-driven decay of D comes from human population genomics. Researchers (in projects such as the International HapMap Project, 2005 onward, and subsequently the 1000 Genomes Project) genotyped millions of single-nucleotide polymorphisms (SNPs) across many individuals and measured r² between every pair of markers as a function of their physical distance. The empirical result matches theoretical prediction closely: r² is high (often > 0.8) for markers separated by a few kilobases, falls off over tens of kilobases, and approaches the background level expected from finite sample size and population history by roughly 100–500 kb in most human populations — consistent with an effective per-generation recombination fraction on the order of 10⁻⁵–10�sup⁻⁴ per kb and thousands of generations since ancestral admixture or the origin of the alleles in question. This is not merely descriptive: it directly validates the (1−r)^t decay model, since the observed decay curve's shape depends on the underlying recombination rate map (independently measured from pedigree and sperm-typing studies) exactly as the two-locus recursion predicts. This decay pattern is also precisely what makes genome-wide association studies (GWAS) work at all — a genotyped marker can "tag" an unmeasured causal variant only because LD links them over these characteristic distances, and it is why the "LD block" structure of the human genome (regions of high internal LD separated by recombination hotspots) had to be empirically mapped before GWAS panels could be designed efficiently.

Selective sweep signatures — direct field evidence for hitchhiking. A canonical case is the Drosophila melanogaster locus Adh (alcohol dehydrogenase) and the surrounding region, studied by Martin Kreitman (1983) and later population-genomic surveys, along with the classic demonstration of hitchhiking around genes under recent human selection such as LCT (lactase persistence in European and some African/Middle Eastern populations, analyzed by Bersaglieri et al. 2004 and Tishkoff et al. 2007) and G6PD and HBB (malaria resistance loci in African populations). In each case, the region immediately flanking the selected variant shows a striking reduction in nucleotide diversity and an excess of long, unusually frequent haplotypes extending unusually far from the selected site — exactly the hitchhiking signature predicted by two-locus theory, because recombination has not yet had time to break down the original association between the beneficial allele and its genetic background. The width and height of these diversity valleys have been used to estimate the strength of selection (selection coefficients on the order of s ≈ 0.01–0.1 for LCT, among the strongest documented in the human genome) purely from the population-genetic signature, an inference that only makes sense within two-locus LD theory.

Experimental evolution — clonal interference and Hill–Robertson effect in laboratory bacterial and viral populations. Richard Lenski's long-term E. coli evolution experiment (begun 1988, still running) and related microbial evolution experiments provided direct, observed evidence for the Hill–Robertson effect. In asexual bacterial populations, whole-genome and whole-population sequencing through time revealed multiple independent beneficial mutations arising on different genetic backgrounds simultaneously, competing with one another for fixation rather than combining — a directly observed instance of clonal interference. Only one lineage (or a lineage that later acquires further mutations) ultimately wins, and beneficial mutations that arose on losing backgrounds are lost along with their entire genetic background, regardless of their own individual fitness benefit. This is the empirical face of negative linkage disequilibrium between competing beneficial mutations, and it directly demonstrates why an asexual (non-recombining) population's rate of adaptation is throttled compared to a theoretical recombining population that could combine the best mutations onto a single chromosome. Complementary experiments in bacteriophage (e.g., studies of phage T7 and φX174 populations, and yeast experimental evolution by Desai, Fisher, and colleagues in the 2000s–2010s) used engineered or barcoded lineages to track competing beneficial clones directly, visually confirming that multiple adaptive lineages rise and then are outcompeted by fitter competitors before fixation — a striking, quantitative window into the very interference the Hill–Robertson theory (1966) had predicted decades earlier from population-genetic first principles alone.

Example 1 — computing D and r² from an asymmetric two-locus dataset, and predicting its decay. Suppose a population of gametes has been surveyed for two SNPs, locus 1 (alleles A/a) and locus 2 (alleles B/b), yielding these haplotype counts out of 1,000 gametes:

Allele frequencies: p_A = freq(AB) + freq(Ab) = 0.44 + 0.06 = 0.50; p_B = freq(AB) + freq(aB) = 0.44 + 0.11 = 0.55. Expected AB frequency under independence: p_A × p_B = 0.50 × 0.55 = 0.275.

D = observed freq(AB) − expected freq(AB) = 0.44 − 0.275 = 0.165. This is strongly positive — A and B are found together far more often than chance predicts. To express this on a standardized 0–1 scale, compute r²: r² = D² / [p_A(1−p_A)·p_B(1−p_B)] = (0.165)² / (0.50 × 0.50 × 0.55 × 0.45) = 0.027225 / 0.12375 ≈ 0.220 An r² of 0.22 indicates moderate-to-strong association — far from the r² ≈ 0 expected at equilibrium between unlinked, non-interacting loci, and a plausible magnitude for loci separated by a few tens of kilobases in a human-like genome. Now suppose these two loci are separated by a recombination fraction r = 0.02 (about 2 cM, roughly 2 Mb in humans) and no further selection or mutation acts on them. How many generations until D falls to 10% of its current value? D_t / D_0 = (1 − r)^t = 0.10 t = ln(0.10) / ln(1 − 0.02) = (−2.303) / (−0.0202) ≈ 114 generations Contrast this with two loci at r = 0.30 (loosely linked, far apart on the same chromosome): t = ln(0.10)/ln(0.70) ≈ 6.5 generations. The nearly 20-fold difference in decay time is the entire practical basis for using LD patterns to localize genes: an association strong enough to still be detectable after dozens of generations tells you the causal and marker loci must be very close together.

Example 2 — epistasis reverses the fate of a two-locus genotype: a Fisherian sign-epistasis scenario. Consider two loci contributing to insecticide resistance in an insect population. Locus 1: R (resistant, dominant) vs. r (susceptible). Locus 2: M (modifier, dominant) vs. m (no modifier). Fitness in the presence of insecticide is measured for all four haplotype-based genotypes (assume haploid for simplicity, as in many microbial or simplified models):

Notice the sign epistasis: R increases fitness when combined with M (1.30 vs. 0.95, a benefit of +0.35) but decreases fitness when combined with m (0.85 vs. 1.00, a cost of −0.15). Single-locus reasoning applied naively — "is R beneficial?" — has no single answer; it depends entirely on genetic background. If R arose on a genetic background lacking M, selection would initially disfavor it (0.85 < 1.00), potentially eliminating it before it ever had a chance to combine with M. Only if R and M arise together (or recombination brings them together while both are still segregating at appreciable frequency) does selection see the large fitness advantage of RM and drive the population toward that combination — building strong positive D between R and M in the process, exactly as described in Mechanism Step 2. This is why epistatic fitness interactions make multi-locus evolution history-dependent and population-structure-dependent in a way single-locus models cannot capture: whether a population ever discovers the RM combination depends on recombination rate, population size (whether R and M cosegregate long enough by drift for recombination to combine them), and the order in which mutations arise — not simply on each allele's individual fitness effect. This scenario is directly analogous to real-documented cases of multi-locus insecticide and antimalarial drug resistance evolution (e.g., multiple pfcrt and pfmdr1 mutations in Plasmodium falciparum chloroquine resistance), where resistance haplotypes show strong sign epistasis and only specific mutation combinations, not single mutations, confer clinically significant resistance.

The two-locus and multi-locus mathematical framework did not emerge from a single discovery moment the way Mendelian segregation or the double helix did; it developed gradually from the 1910s–1960s as population geneticists confronted the limits of single-locus models. J. B. S. Haldane and Ronald Fisher independently derived early two-locus recursions in the 1910s–1930s while investigating patterns of linkage in crosses. But the modern conceptual synthesis — recognizing linkage disequilibrium as a general, quantifiable population property (not just a laboratory cross phenomenon), and formally connecting it to recombination, selection, and epistasis in natural populations — is most closely associated with the mid-20th-century work of population geneticists including William G. Hill and Alan Robertson, whose 1966 paper on the effect of linkage on selection efficiency in finite populations gave the field its enduring "Hill–Robertson effect" and made the interference between selected loci a quantitative, testable prediction rather than a qualitative worry. Robertson, a quantitative and animal geneticist working largely on selection response in livestock and Drosophila at the Institute of Animal Genetics in Edinburgh, brought a strong empirical, selection-experiment sensibility to population genetics; Hill, his student and later collaborator, extended the mathematical machinery describing how finite population size and linkage jointly limit the response to selection — work with direct, continuing application in animal and plant breeding genomics today.

Genome-wide association studies (GWAS) and precision medicine depend entirely on LD: because a genotyped SNP is statistically linked to nearby untyped causal variants, GWAS panels can detect disease associations without sequencing every base — but this also means GWAS "hits" typically mark a linked region, not necessarily the causal gene, requiring fine-mapping and functional follow-up before clinical translation (e.g., in identifying causal variants for coronary artery disease or type 2 diabetes risk loci).

Antimicrobial and antimalarial resistance surveillance tracks epistatic multi-locus haplotypes directly, because resistance to drugs like chloroquine or artemisinin in Plasmodium falciparum, or multi-drug resistance in Mycobacterium tuberculosis, typically requires specific combinations of mutations rather than any single one — public health genomic surveillance now monitors haplotype (not just single-SNP) frequencies to anticipate resistance spread.

Plant and animal breeding exploits recombination and LD deliberately: marker-assisted selection and genomic selection rely on LD between markers and causal quantitative trait loci (QTLs), while breeders must also manage "linkage drag" — undesirable alleles hitchhiking along with a desired introgressed gene from a wild relative, requiring repeated backcrossing (facilitated by recombination) to break the unwanted association.

Conservation genetics uses genome-wide LD decay patterns to estimate historical effective population size and detect recent bottlenecks, since small N_e inflates drift-generated LD in a mathematically predictable way — a widely used tool for endangered species management.

Cancer genomics applies sweep and hitchhiking logic within tumors: driver mutations sweep through a cell population by clonal expansion, dragging passenger mutations to high frequency purely through physical linkage on the same genome — directly complicating the identification of true driver mutations from sequencing data, an oncological instance of the exact hitchhiking problem population geneticists first formalized in whole organisms.

Understanding the evolution of sex and recombination itself remains an active research frontier built on this framework: the Hill–Robertson effect, combined with evidence from experimental evolution (Lenski's E. coli lines, yeast and viral experimental populations), provides one of the best-supported explanations for why recombination is so widespread despite its costs — it relieves interference between selected loci that would otherwise be trapped competing on the same non-recombining genetic background, connecting directly back to core questions in Genetic Drift and Random Change and Gene Flow and Migration about how finite, structured populations actually experience selection.

Ask the tutor: How does linkage disequilibrium theory change our interpretation of a GWAS "hit" that falls in a gene desert with no obvious nearby candidate gene?