{
  "id": "linkage-disequilibrium",
  "title": "Linkage Disequilibrium",
  "lang": "en",
  "source_lang": "en",
  "summary": "Linkage disequilibrium is the nonrandom association of alleles at different genetic loci, shaped by recombination, evolutionary history, and population structure.",
  "category": "science",
  "subcategory": "biology",
  "type": "concept",
  "aliases": [
    "Gametic disequilibrium",
    "Allelic association"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/linkage-disequilibrium/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/linkage-disequilibrium/",
      "json": "https://aiwiki.page/zh/linkage-disequilibrium.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Linkage disequilibrium (LD) is the nonrandom association of alleles at different positions, or loci, in a population’s genome. It occurs when particular combinations of alleles are more or less frequent than expected from their individual frequencies. These combinations are usually described as haplotypes: sets of alleles carried together on a single chromosome or inherited genomic segment. LD is a property of a specified population and set of variants, not an intrinsic property of a DNA sequence.\n\nDespite its name, linkage disequilibrium does not require physical linkage. Nearby loci on the same chromosome commonly exhibit LD because recombination separates them infrequently, but associations can also occur between distant loci or different chromosomes. Conversely, physically linked loci can be in linkage equilibrium. The distinction between physical linkage and statistical association is fundamental to interpreting genetic data.\n\nLD connects population genetics with the practical analysis of inheritance, evolutionary history, and complex traits. Its patterns reflect recombination, mutation, genetic drift, migration, selection, and mating structure. They also make it possible to detect disease-associated regions without observing every causal variant, infer unmeasured genotypes, and reconstruct aspects of population history. The same correlations that enable these applications can obscure which variants are biologically responsible for an observed association."
    },
    {
      "heading": "Definition and mathematical description",
      "text": "Consider two loci, each with two alleles: \\(A\\) and \\(a\\) at the first locus, and \\(B\\) and \\(b\\) at the second. Let their allele frequencies be \\(p_A\\), \\(p_a=1-p_A\\), \\(p_B\\), and \\(p_b=1-p_B\\). There are four possible haplotypes, with frequencies \\(x_{AB}\\), \\(x_{Ab}\\), \\(x_{aB}\\), and \\(x_{ab}\\). Under linkage equilibrium, the probability of a haplotype is the product of its component allele frequencies. Thus, \\(x_{AB}=p_Ap_B\\), with analogous expressions for the other three haplotypes. This is statistical independence between alleles at the two loci.\n\nThe standard two-locus disequilibrium coefficient is\n\n\\[D=x_{AB}-p_Ap_B.\\]\n\nThe four haplotype frequencies can therefore be written as\n\n\\[\\begin{aligned} x_{AB}&=p_Ap_B+D,\\\\ x_{Ab}&=p_Ap_b-D,\\\\ x_{aB}&=p_ap_B-D,\\\\ x_{ab}&=p_ap_b+D. \\end{aligned}\\]\n\nEquivalently, \\(D=x_{AB}x_{ab}-x_{Ab}x_{aB}\\). Positive \\(D\\) means that \\(AB\\) and \\(ab\\) are overrepresented relative to independence; negative \\(D\\) means that \\(Ab\\) and \\(aB\\) are overrepresented. The sign depends on allele labels: exchanging \\(A\\) with \\(a\\), for example, reverses it. The existence and magnitude of association do not depend on which allele is designated the reference.\n\nAs a numerical example, suppose both \\(A\\) and \\(B\\) have frequency 0.5, while the haplotype frequencies are \\(0.4, 0.1, 0.1,\\) and \\(0.4\\) for \\(AB, Ab, aB,\\) and \\(ab\\), respectively. Independence would predict a frequency of 0.25 for every haplotype, whereas the observed value gives \\(D=0.15\\). The population contains an excess of the two matching combinations. This association says nothing by itself about whether the alleles influence the same biological process.\n\nLD must also be distinguished from Hardy–Weinberg equilibrium. Hardy–Weinberg proportions concern the pairing of alleles within diploid individuals at a locus; LD concerns associations across loci within inherited chromosomes or gametes. A population can satisfy Hardy–Weinberg proportions at each locus while retaining substantial LD between loci. Random mating can restore single-locus Hardy–Weinberg proportions in one generation under standard assumptions, whereas multilocus associations disappear at rates constrained by recombination.\n\nThe definition extends beyond two biallelic loci. Multiallelic loci require a set of association coefficients, and multilocus haplotypes can contain interactions not captured by pairwise statistics. Pairwise independence among all locus pairs does not necessarily imply complete independence among three or more loci."
    },
    {
      "heading": "Measures of strength and statistical interpretation",
      "text": "The raw coefficient \\(D\\) is useful for describing inheritance dynamics, but its possible range depends on allele frequencies. If one allele is rare, even its strongest possible association with another allele produces a small absolute value of \\(D\\). Comparisons across locus pairs therefore commonly use standardized measures, especially \\(D'\\) and \\(r^2\\). These quantify different aspects of association and should not be treated as interchangeable.\n\nFor a signed version of Lewontin’s \\(D'\\),\n\n\\[D'= \\begin{cases} \\dfrac{D}{\\min(p_Ap_b,p_ap_B)}, & D>0,\\\\[8pt] \\dfrac{D}{\\min(p_Ap_B,p_ap_b)}, & D<0,\\\\[8pt] 0, & D=0. \\end{cases}\\]\n\nFor polymorphic loci, this lies between \\(-1\\) and \\(1\\). Many applications report only \\(|D'|\\). A value of \\(|D'|=1\\) means that \\(D\\) has reached the largest magnitude allowed by the observed allele frequencies; for two biallelic loci, at least one of the four haplotypes is absent. This does not establish that recombination has never occurred. Historical recombinants may have been lost, never sampled, or produced at frequencies too low to detect.\n\nThe squared allelic correlation is\n\n\\[r^2=\\frac{D^2}{p_Ap_ap_Bp_b}.\\]\n\nIt ranges from 0 to 1 and is the squared correlation between indicator variables recording the alleles carried at the two loci on a randomly sampled haplotype. It has a direct predictive interpretation: a high value means that knowledge of one allele gives substantial information about the other. Under random union of haplotypes, the correlation between diploid allele dosages has the same underlying relationship. Neither \\(r^2\\) nor \\(D'\\) is defined in the usual way when a required locus is monomorphic.\n\nFrequency differences explain why the measures can disagree. Suppose \\(p_A=0.1\\), \\(p_B=0.5\\), and every copy of \\(A\\) occurs with \\(B\\). Then \\(x_{AB}=0.1\\), \\(D=0.05\\), and \\(D'=1\\), but\n\n\\[r^2=\\frac{0.05^2}{0.1\\times0.9\\times0.5\\times0.5} \\approx0.111.\\]\n\nAlthough \\(A\\) perfectly predicts the presence of \\(B\\), most copies of \\(B\\) do not carry \\(A\\). Prediction is therefore asymmetric, and the overall correlation is modest.\n\nStatistical significance is separate from effect size. With sufficiently large samples, weak LD can be highly significant; with small samples, strong-looking LD can be uncertain. For \\(n\\) independently sampled, fully observed haplotypes, the Pearson statistic for a two-by-two table is \\(nr^2\\), with an approximate chi-squared null distribution under suitable cell counts. Diploid samples with unknown haplotype phase require different treatment. Rare alleles and missing haplotype classes particularly affect \\(D'\\), while finite samples can produce positive estimated \\(r^2\\) even when population LD is zero."
    },
    {
      "heading": "Recombination and the decay of association",
      "text": "The principal mechanism breaking down LD between loci on the same chromosome is genetic recombination during meiosis. In many eukaryotes, chromosomal crossing over exchanges segments between homologous chromosomes. A chromosome carrying \\(AB\\) can consequently contribute an \\(Ab\\) or \\(aB\\) haplotype in later generations, depending on the chromosome paired with it and where exchange occurs. Recombination changes associations among alleles without necessarily changing the allele frequencies themselves.\n\nIn an idealized, infinitely large, randomly mating population without selection, mutation, or migration, the two-locus recursion is\n\n\\[D_{t+1}=(1-c)D_t,\\]\n\nwhere \\(c\\) is the recombination fraction between the loci. Iterating gives\n\n\\[D_t=(1-c)^tD_0.\\]\n\nFor small \\(c\\), this is approximately \\(D_0e^{-ct}\\). The approximate half-life is \\(\\ln 2/c\\) generations. At a recombination fraction of 0.01, LD generated by a single historical event has a half-life of about 69 generations under these assumptions. For loci assorting independently, \\(c=0.5\\), so the association halves each generation rather than disappearing immediately.\n\nRecombination fraction is not identical to physical distance. A genetic map measures recombination, conventionally in centimorgans, whereas physical maps measure base pairs. For short intervals, 1 centimorgan corresponds approximately to a 1% recombination fraction. Over larger intervals, multiple crossovers complicate that correspondence because some restore the original arrangement of the marker alleles. Recombination fraction cannot exceed 0.5, even when loci lie very far apart.\n\nRates also vary considerably along chromosomes, between sexes, among populations, and among species. Recombination hotspots concentrate crossover events into short intervals, while regions near some centromeres or within chromosomal inversions may exhibit reduced effective recombination. In humans and several other mammals, the DNA-binding protein PRDM9 helps specify many hotspot locations. Non-crossover gene conversion can disrupt associations over short tracts and contributes to the relationship between LD and distance.\n\nThese processes produce the familiar average decline of \\(r^2\\) with increasing separation, but not a smooth decline for every pair of variants. Nearby pairs can have little LD, while more distant pairs remain strongly associated. Their histories may differ in allele age, recombination exposure, selection, and sampling. Consequently, an LD-decay curve describes an aggregate pattern, not a universal law for individual loci.\n\nLD-based recombination maps use historical associations to infer where recombination has occurred frequently. Such maps integrate events over ancestral generations and depend on demographic and evolutionary models. Pedigree maps instead observe transmissions within families. Differences between the two can reflect changing hotspot activity, population history, or uncertainty in either method, rather than a simple measurement error."
    },
    {
      "heading": "Mutation, drift, population structure, and selection",
      "text": "A new mutation initially appears on a particular chromosome and therefore on a particular local haplotype. At its origin, the mutant allele is associated with the surrounding variants carried by that chromosome. As descendants inherit and recombine the region, those associations change. Young mutations often retain a relatively restricted set of backgrounds, but mutation age alone does not determine LD: allele frequency, reproductive history, and local recombination also matter.\n\nGenetic drift continually generates random associations in finite populations. Even if haplotypes were independent before reproduction, the chromosomes contributing to the next generation constitute a finite sample. Their combinations need not preserve independence. Recombination counteracts these stochastic departures, producing a balance that depends on effective population size and recombination rate. A frequently used rough relationship is\n\n\\[E[r^2]\\approx\\frac{1}{1+4N_ec},\\]\n\nwhere \\(N_e\\) is the effective size of a diploid population. This expression is not a universal estimator: mutation, allele-frequency conditioning, finite-sample bias, demographic change, and the particular definition of the expectation affect the result. Its main value is to show why smaller populations and lower recombination tend to produce stronger LD.\n\nPopulation subdivision can generate association without any interaction between the loci. Suppose two source populations are each in linkage equilibrium, but have allele frequencies \\(p_1,q_1\\) and \\(p_2,q_2\\). Pooling them in proportions \\(m\\) and \\(1-m\\) produces\n\n\\[D=m(1-m)(p_1-p_2)(q_1-q_2).\\]\n\nIf both alleles are more frequent in the same source population, the pooled association is positive. More generally, the mixture also retains a weighted contribution from LD already present in the source populations. Gene flow and recent admixture therefore create LD over long genomic distances, including between different chromosomes. In an interbreeding admixed population, recombination progressively erodes these associations.\n\nNatural selection can alter LD in several ways. A beneficial allele rising rapidly in frequency can carry neighboring variants with it, a process called genetic hitchhiking. A recent strong sweep may therefore leave extended haplotypes and reduced nearby variation. Selection against deleterious alleles also modifies the histories of linked neutral variants. Interactions among fitness effects can generate associations between selected loci, while recombination changes which combinations selection encounters.\n\nEpistasis, or dependence of an allele’s effect on genetic background, is conceptually distinct from LD. Epistasis describes a relationship among biological or statistical effects; LD describes the frequencies of allele combinations. Either can exist without the other. Assortative mating can also generate associations among trait-increasing alleles, even at unlinked loci. Because several mechanisms can produce similar patterns, high LD alone is insufficient evidence for selection, recent admixture, or functional interaction."
    },
    {
      "heading": "Haplotype structure and variation among genomes",
      "text": "LD is often displayed as a triangular heat map in which each cell represents a pairwise statistic for two variants. Such plots can reveal clusters of correlated variants, abrupt transitions, and long-range associations. However, a heat map depends on which variants were measured, their frequency distribution, the statistic chosen, and the population sampled. A region that appears strongly correlated using common variants can contain numerous rare variants with different histories.\n\nThe term haplotype block describes a genomic segment within which a limited number of common haplotypes account for much of the variation and historical recombination appears relatively restricted. Blocks became prominent in human genetics around the development of large-scale SNP surveys. Their boundaries are not unique physical landmarks: algorithms based on confidence intervals for \\(D'\\), observed haplotype diversity, or recombination evidence can assign different blocks to the same data. Recombination hotspots may coincide with boundaries, but block structure also reflects drift, mutation, and ascertainment.\n\nThe four-gamete test illustrates both the usefulness and limitations of haplotype patterns. If all four haplotypes \\(AB, Ab, aB,\\) and \\(ab\\) occur, at least one recombination event is required under an infinite-sites model in which each site mutates only once. Without that assumption, recurrent or reverse mutation can also generate all four combinations. Conversely, observing only three combinations does not prove that recombination is absent; a fourth may be rare, lost, or missing from the sample.\n\nHuman populations differ in LD because their histories differ. Many African populations show shorter average LD for common variants than populations whose histories include the out-of-Africa bottleneck, consistent with greater long-term effective population sizes and deeper genealogical diversity. This is a broad comparison rather than a fixed continental rule. Africa contains extensive demographic diversity, and local founder events, recent admixture, allele frequencies, and sampling choices can substantially alter comparisons. Population-specific estimates are therefore more informative than a single label attached to a broad ancestry category.\n\nGenome structure and reproductive biology create further variation. Chromosomal inversions can maintain long haplotypes when crossing over in heterozygotes fails to yield viable or balanced recombinant products. Self-fertilization reduces effective recombination because homologous chromosomes are more often genetically similar. Predominantly clonal lineages can retain associations across much of the genome.\n\nIn bacteria, homologous exchange and other forms of horizontal gene transfer generate patterns unlike those expected from regular sexual reproduction. Imported DNA tracts may break local associations, while clonal expansion preserves broad genomic correlations. Consequently, measures developed for randomly mating diploid populations require careful reinterpretation in microbial datasets, polyploids, and genomes containing extensive structural variation."
    },
    {
      "heading": "Estimation from genetic data",
      "text": "LD is most directly estimated from observed haplotypes, but many datasets provide diploid genotypes without identifying which alleles occupy the same chromosome. An individual heterozygous at both loci, \\(AaBb\\), can carry either \\(AB/ab\\) or \\(Ab/aB\\). This ambiguity is called phase uncertainty. Counts of unphased genotypes therefore do not always translate directly into haplotype counts, especially when double heterozygotes are common.\n\nStatistical phasing estimates haplotypes using population patterns, often supplemented by family relationships or reference panels. An expectation–maximization algorithm can estimate two-locus haplotype frequencies by alternating between possible phase assignments and frequency updates under a specified sampling model. Modern genome-wide methods use hidden Markov models and related approximations to represent each chromosome as a mosaic of previously observed haplotypes. Their accuracy depends on sample size, relatedness, variant frequency, reference diversity, and genomic complexity.\n\nExperimental data can resolve phase more directly. Parent–offspring transmissions identify many parental haplotypes, while long reads and linked molecular observations can place distant variants on the same DNA molecule. DNA sequencing nevertheless introduces its own uncertainties: low read depth, mapping errors, duplicated regions, and variant-calling mistakes can produce false combinations. A single erroneous observation can create an apparently missing fourth haplotype and substantially change \\(D'\\).\n\nSome analyses estimate LD from correlations between allele dosages without explicitly reconstructing haplotypes. Under random pairing of haplotypes and appropriate sampling assumptions, dosage correlations estimate the corresponding allelic correlation. Departures from these conditions—including relatedness, population mixture, or nonrandom mating—can change the interpretation. Dosages from imputation are uncertain expectations rather than directly observed allele counts, so their correlations need not behave exactly like those of accurately measured genotypes.\n\nSampling design is consequently central. Close relatives can make recently shared chromosome segments disproportionately influential. Combining differentiated populations can create long-range LD not characteristic of any source population. Selecting samples because they have a particular disease or extreme phenotype can also alter associations among loci. Variant filters, missing-data rules, and allele-frequency thresholds change which pairs enter an analysis and can shift apparent LD decay.\n\nReliable reporting therefore distinguishes the target population, sample composition, statistic, phasing method, and variant-selection criteria. Large studies often use LD pruning or clumping to reduce redundancy, but these procedures answer different questions: pruning selects a relatively weakly correlated subset, whereas clumping commonly groups association signals around a lead variant. Neither procedure establishes that the retained variants are independent causal influences."
    },
    {
      "heading": "Association studies, imputation, and causal interpretation",
      "text": "The central role of LD in a genome-wide association study is to allow measured variants to act as proxies for unmeasured causal variants. If a marker allele is correlated with an allele that influences a trait, the marker can show a statistical association even when it has no biological effect itself. This principle made large-scale association mapping feasible before complete sequencing of every participant was practical.\n\nFor a simple additive model with one causal variant, and under comparable sampling and analysis conditions, the association information available at a proxy marker is approximately reduced by its \\(r^2\\) with the causal variant. A marker with \\(r^2=0.5\\) may therefore require roughly twice the sample size to provide similar detection power. This is an approximation, not a general rule for all traits: allele frequencies, case–control sampling, imputation uncertainty, multiple causal variants, and model specification can change the relationship.\n\nGenotype imputation exploits the same correlations at a larger scale. Observed markers identify likely reference haplotypes, which are then used to infer alleles at unobserved sites. Accuracy is usually highest when reference panels contain closely matching haplotypes and sufficient examples of the relevant variants. Rare alleles are harder to impute because they may occur on few reference chromosomes, while structural variants can have relationships with nearby markers that are incompletely represented by standard SNP panels.\n\nLD also creates the central ambiguity of an association locus: many correlated variants may display similar evidence even when only one, or a few, influence the trait. Genetic fine-mapping uses association statistics, an LD matrix, and assumptions about causal architecture to assign probabilities to candidate variants. A credible set is a model-based collection intended to contain causal variation with a stated probability. Its interpretation depends on the assumed number of causal variants, statistical calibration, variant coverage, and the suitability of the LD reference.\n\nConditional analysis tests whether an association remains after accounting for one or more variants. Closely correlated variants can be difficult to separate, and an imperfectly matched reference LD matrix can generate misleading secondary signals. Studies across populations can improve resolution because the same causal allele may occur alongside different correlated markers. This advantage requires adequate representation of each population and does not imply that effect sizes or environmental conditions are identical.\n\nBiological annotation and experiments provide evidence distinct from statistical correlation. A variant affecting gene expression, protein sequence, or another molecular process may help explain a signal, but functional activity alone does not establish that it causes the studied trait. Likewise, overlap between two association signals does not prove a shared causal variant. Colocalization analyses explicitly assess that possibility while accounting, to varying degrees, for local LD and multiple signals."
    },
    {
      "heading": "Evolutionary inference, breeding, and polygenic analysis",
      "text": "Because recombination progressively rearranges chromosomes, LD carries information about time. Recent shared ancestry tends to preserve longer associated segments, while older ancestry has experienced more opportunities for recombination. Population-genetic methods use these relationships to investigate population-size change, bottlenecks, admixture, and selection. Such inferences depend on a model connecting observed correlations to ancestral reproduction rather than on a direct reading of elapsed time from physical distance.\n\nAdmixture dating provides a clear example. A pulse of mixing between differentiated populations generates associations among ancestry-informative alleles. Under a simplified model, the admixture-related covariance declines approximately as \\(e^{-td}\\), where \\(t\\) is the number of generations since admixture and \\(d\\) is genetic distance in morgans. Fitting this decay can estimate a historical time scale. Continuous migration, multiple pulses, source-population LD, and inaccurate recombination maps can produce more complex curves, so a fitted date may represent an effective time rather than one discrete historical event.\n\nSelection scans also use haplotype persistence. A beneficial allele that rises rapidly may reach high frequency before recombination has extensively diversified its surrounding haplotype. Statistics based on extended haplotype homozygosity compare the lengths or frequencies of such backgrounds. The inference is strongest when evaluated against demographic alternatives, local recombination patterns, and genome-wide empirical distributions. An unusually long haplotype is evidence requiring explanation, not a unique signature of adaptation.\n\nIn animal and plant breeding, marker–trait LD allows genome-wide markers to predict breeding values without identifying every causal locus. Prediction depends partly on how well markers track causal alleles in the breeding population. Small effective population sizes and recent shared ancestry can preserve substantial LD, but prediction across breeds or distantly related populations may deteriorate as marker–causal relationships change. Recombination across generations also means that a marker’s usefulness can decline unless prediction models are updated or causal variants are directly represented.\n\nHuman polygenic scores similarly combine information across many variants. LD-aware methods account for correlations so that a single underlying signal is not repeatedly counted as independent evidence. Differences in LD, allele frequency, study representation, and other aspects of genetic architecture contribute to differences in score performance across populations. A score trained in one population does not acquire equal predictive validity elsewhere merely because the same markers can be measured.\n\nLD also underlies methods that estimate genome-wide trait architecture from association summary statistics. LD score regression, for example, relates a variant’s association statistic to its LD score, broadly the sum of its squared correlations with surrounding variants. Under its model assumptions, polygenic signal increases with the amount of correlated variation tagged by a marker. Extensions estimate genetic covariance between traits. Population structure, reference-panel mismatch, selection-related architecture, and violations of modeling assumptions can affect the results."
    },
    {
      "heading": "Historical development and unresolved analytical problems",
      "text": "The mathematical study of nonrandom allele combinations developed within early population genetics as researchers extended single-locus inheritance models to linked loci. A major milestone was Richard Lewontin and Ken-ichi Kojima’s 1960 paper, “The Evolutionary Dynamics of Complex Polymorphisms,” which analyzed multilocus selection and helped establish the modern use of linkage disequilibrium. Lewontin introduced the standardized \\(D'\\) measure in 1964. Subsequent work examined the stochastic balance between drift and recombination and the consequences of selection at multiple loci.\n\nIn the 1980s, restriction fragment length polymorphisms and other DNA markers expanded opportunities to study associations in human populations. In 1989, the identification of the cystic fibrosis gene illustrated the power of combining family linkage evidence, population association, and physical mapping. These approaches were complementary: family linkage followed cosegregation within pedigrees, while LD exploited associations accumulated through population history. Neither should be confused with simply finding two variants close together on a chromosome.\n\nThe International HapMap Project, whose first major map was published in 2005, systematically characterized common human variation and LD in selected population samples. Its data supported the use of tag SNPs to represent correlated sets of variants and helped establish the infrastructure for genome-wide association studies. The 1000 Genomes Project later broadened the catalog through sequencing; its 2015 principal reference analysis described 2,504 individuals from 26 populations. These resources made both the utility and the population dependence of LD increasingly explicit.\n\nA persistent problem is that pairwise summaries discard information. Two genomic regions can have similar average \\(r^2\\) yet very different haplotype genealogies, rare-variant distributions, or higher-order associations. Conversely, a small set of common haplotypes can make a region appear simple while concealing many low-frequency combinations. Methods based on ancestral genealogies seek to represent this history more directly, but recombination means that different positions have different genealogical trees.\n\nReference matching remains another major analytical issue. Fine-mapping and summary-statistic methods often require correlations estimated outside the original study. Differences in ancestry composition, recent relatedness, sample size, variant filters, or genome assembly can make the supplied matrix inconsistent with the association data. Rare variants are particularly sensitive because their estimated correlations may depend on only a few carriers.\n\nIncreasingly complete sequencing also challenges SNP-centered descriptions of LD. Insertions, deletions, repeat expansions, copy-number changes, and complex structural rearrangements can be poorly captured by nearby common markers. Pangenome references and long-read sequencing improve access to this variation, but representing multiallelic structures, estimating phase, and defining comparable association statistics remain active areas of research. The underlying question is unchanged—how observed allele combinations depart from independence—but the genomic objects and historical models used to answer it are becoming more comprehensive."
    }
  ],
  "links": [
    "allele",
    "genome",
    "chromosome",
    "population-genetics",
    "hardy-weinberg-equilibrium",
    "meiosis",
    "chromosomal-crossover",
    "genetic-map",
    "gene-conversion",
    "mutation",
    "genetic-drift",
    "gene-flow",
    "natural-selection",
    "epistasis",
    "bacteria",
    "horizontal-gene-transfer",
    "genotype",
    "algorithm",
    "dna-sequencing",
    "phenotype",
    "genome-wide-association-study",
    "gene-expression",
    "pangenome"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}