Stabilizing and Disruptive Selection

Consider a continuous, quantitative trait — birth weight, beak depth, flower tube length — whose distribution in a population can be described by a mean and a variance. Selection acts on this distribution through a fitness function, w(z), that maps each phenotypic value z onto the expected survival or reproductive success of individuals with that phenotype. Natural Selection and the Concept of Fitness established that fitness differences among phenotypes drive changes in trait means. What determines whether the population's variance grows, shrinks, or stays the same is the curvature of w(z) around the current mean. Three canonical shapes exist. Directional selection has a fitness function that rises (or falls) monotonically across the observed range of z; the mean shifts, and if the fitness function is convex in the right way the variance can shift too, but the defining feature is a shift in mean. Stabilizing selection has a fitness function that peaks at some intermediate value and declines on both sides — individuals near the mean have the highest fitness, and both tails are selected against. This is sometimes called purifying or centripetal selection because it centers the distribution and shrinks its spread. Disruptive selection (also called diversifying selection) has a fitness function that dips in the middle and rises toward both tails — the two extremes each do better than the middle. This is centrifugal: it pushes mass away from the mean and can produce a bimodal distribution. Formally, if fitness is approximated by a quadratic function of phenotype, w(z) ≈ 1 + αz + βz²/2, the linear term α gives the directional selection differential (Lande's β for the mean), while the quadratic term β gives the selection differential on variance. When β is negative, fitness curves downward away from an optimum — stabilizing selection. When β is positive, fitness curves upward away from the mean — disruptive selection. Statisticians estimating these coefficients from field data (following Lande & Arnold 1983) fit exactly this quadratic regression of relative fitness on phenotype, extracting both the linear (directional) and quadratic (stabilizing/disruptive) gradients simultaneously, because real fitness surfaces often contain both components at once.

Step 1 — the fitness surface acts on standing phenotypic variance. Any sexually reproducing population has phenotypic variance (V_P) built from additive genetic variance (V_A), dominance and epistatic variance, and environmental variance (V_E). Selection each generation samples individuals according to w(z) before they reproduce; the shape of w(z) determines which slice of the distribution contributes disproportionately to the next generation's gene pool.

Step 2 — stabilizing selection culls both tails. If w(z) peaks at the population mean z̄ and falls off symmetrically, individuals more than roughly one or two standard deviations from z̄ in either direction are removed at a higher rate before reproduction. Because both the upper and lower tails lose representation, the offspring generation is drawn from parents clustered nearer z̄. Under the infinitesimal (polygenic) model, this reduces the realized heritable variance passed on, so V_A (and hence V_P) declines generation over generation — though mutation, recombination, and migration continually resupply variance, so populations under long-term stabilizing selection typically reach a mutation–selection balance rather than variance = 0.

Step 3 — disruptive selection favors both tails at the expense of the center. Here w(z) is lowest at z̄ and highest at the phenotypic extremes. Intermediate individuals are selected against — often because they fall between two resources, habitats, or mating niches and are poorly suited to either — while individuals near either extreme thrive by specializing. If the trait has an additive genetic basis and mating is random with respect to phenotype, disruptive selection alone tends to increase phenotypic variance and can produce a bimodal offspring distribution, but it does not by itself, in a randomly mating population, split the population into two reproductively isolated genetic clusters — because intermediate genotypes keep being regenerated by matings between the two extreme phenotypes. Full bimodal genetic divergence generally requires an additional ingredient: assortative mating, habitat/temporal separation of the extremes, or strong linkage between the trait loci and mating-preference loci, so that the two phenotypic modes breed largely within themselves. This is precisely the link forward to models of speciation with gene flow.

Step 4 — the fate of variance is quantifiable. Under simple models, the change in additive genetic variance per generation under quadratic selection is proportional to −β·V_A² (stabilizing, β<0, variance shrinks) or +β·V_A² (disruptive, β>0, variance grows), moderated each generation by the input of new variance from mutation (on the order of 10⁻³ V_E per generation for many polygenic traits) and by recombination reshuffling genetic variance that was temporarily reduced by selection (the Bulmer effect). This is why stabilizing selection does not drive genetic variance to zero even over very long timescales — a dynamic equilibrium is reached between selection removing variance and mutation/recombination replenishing it.

Step 5 — heritability determines the response. As in directional selection, the actual generational change in the mean or variance depends not just on the strength of selection (the selection differential, S) but on the trait's heritability (h²), via the breeder's equation R = h²S for the mean, and an analogous but more complex relationship for variance that also depends on higher genetic moments (kurtosis of the breeding-value distribution). Traits with low heritability respond more sluggishly to either selection mode because environmental variance masks the genetic variance that selection can act on.

Bumpus's house sparrows (1898) — the founding demonstration. After a severe winter storm in Providence, Rhode Island, ornithologist Hermon Bumpus collected 136 house sparrows (Passer domesticus) that had been stunned or killed by the storm and brought to his laboratory. Of these, 72 recovered and 64 died. Bumpus had measured multiple morphological traits — total length, wingspan, weight, and skeletal dimensions — on all birds before their fates were known. When he compared survivors to non-survivors, he found that survivors clustered nearer the population mean for several traits, while birds with extreme values (unusually long or unusually short bones, for instance) were overrepresented among the dead. This was among the first quantitative demonstrations that selection could act against both phenotypic tails simultaneously rather than uniformly favoring one direction — the empirical fingerprint of stabilizing selection. Later statistical reanalyses (notably by Bumpus's data being revisited across the 20th century, including work by O'Donald and by Johnston, Niles & Rohwer in the 1970s–80s) confirmed a stabilizing pattern for several traits, although some traits in Bumpus's dataset actually showed directional components, a reminder that real fitness surfaces are often mixtures rather than pure textbook shapes.

Human birth weight (Karn & Penrose, 1951) — the classic quantitative case. Mary Karn and Lionel Penrose analyzed birth weight and early mortality in over 13,000 human births in London. They plotted neonatal mortality as a function of birth weight and found a strongly U-shaped (or J-shaped combined) mortality curve: infants born much lighter than about 3.3 kg suffered sharply elevated mortality from complications of prematurity and immaturity, while infants born much heavier also suffered elevated mortality, historically often linked to complications of delivery. The minimum mortality — hence the fitness optimum — sat close to the population's actual mean birth weight, slightly above it in their data. This is the textbook illustration of stabilizing selection in humans: a continuously distributed trait, a symmetric or near-symmetric fitness cost for departing from an intermediate value in either direction, and an observed optimum close to the population mean. Modern obstetric care has changed the height and shape of this curve (heavy babies now survive far more often via cesarean delivery), which itself illustrates how the ecological or technological context — not the trait itself — sets the location and steepness of a stabilizing fitness function.

African seedcracker finches (Pyrenestes ostrinus) — disruptive selection on a discrete resource axis. Thomas B. Smith's field studies in Cameroon (starting in the late 1980s and continuing through the 1990s) documented a bimodal distribution of lower mandible (bill) width in this seed-eating finch, with almost no birds at the intermediate width. The two morphs specialize on different sedge seeds of differing hardness: small-billed birds crack soft sedge seeds efficiently, large-billed birds crack hard seeds, and intermediate-billed birds are relatively inefficient at both. Smith measured survival directly and found that intermediate-billed juveniles had markedly lower survival than birds of either extreme morph during periods of seed scarcity, when birds were forced to rely more heavily on one seed type or the other. This is a rare case where disruptive selection was caught in the act, quantified with actual mortality data linking bill morphology to survival, rather than inferred solely from a static bimodal distribution (which could in principle arise from other causes, such as mixing of migrant populations).

Lande & Arnold's (1983) statistical framework — how ecologists actually measure these forces in the wild. Russell Lande and Stevan Arnold formalized the estimation of selection acting on quantitative traits by regressing relative fitness (individual fitness divided by mean population fitness) on phenotype, fitting both linear terms (directional selection gradients, β) and quadratic terms (γ, the stabilizing/disruptive selection gradients) simultaneously, along with correlational terms for multiple traits. A significantly negative quadratic gradient γ indicates stabilizing selection on that trait; a significantly positive γ indicates disruptive selection. This method, now standard in field quantitative genetics, has been applied to hundreds of wild populations — Peter and Rosemary Grant's long-term studies of Darwin's finch beak morphology on Daphne Major being one of the most famous applications, where selection gradients have been shown to fluctuate between directional and stabilizing regimes across drought and non-drought years.

Example 1 — quantifying stabilizing selection on human birth weight. Suppose in a maternity cohort of 10,000 births, the birth weight distribution is approximately normal with mean 3.3 kg and standard deviation 0.5 kg. Suppose neonatal survival probability follows a Gaussian fitness function peaking at exactly the population mean: w(z) = exp[ −(z − 3.3)² / (2 × 0.7²) ] so that survival is 100% at z = 3.3 kg, but falls to about 68% at z = 3.9 kg or z = 2.7 kg (0.6 kg from the optimum), and to about 21% at z = 4.7 kg or z = 1.9 kg (1.4 kg away). To find the mean fitness and the effective selection pressure, we combine the Gaussian phenotype distribution (variance V_P = 0.25 kg²) with the Gaussian fitness function (variance ω² = 0.49 kg²). A standard result for Gaussian selection on a Gaussian phenotype distribution gives the post-selection variance: V_P' = (V_P × ω²) / (V_P + ω²) = (0.25 × 0.49) / (0.25 + 0.49) = 0.1225 / 0.74 ≈ 0.1655 kg² So the phenotypic variance among survivors shrinks from 0.25 kg² to about 0.166 kg² in a single generation of selection — a roughly 34% reduction in variance among the selected group. If birth weight has heritability h² ≈ 0.3 (a commonly cited estimate for human birth weight, reflecting substantial maternal-environmental influence), only 30% of this narrowing translates into reduced additive genetic variance transmitted to offspring; the rest reflects environmental (largely maternal) variance that is not heritable. Even so, the calculation shows quantitatively why birth weight distributions do not drift toward ever-larger or ever-smaller values: each generation, selection actively compresses the distribution toward 3.3 kg, and this compression is continuously counteracted by new environmental and mutational variance so that the population-level variance observed each generation (~0.25 kg²) stays roughly stable rather than shrinking to zero.

Example 2 — a disruptive-selection cross with asymmetric genetics. Imagine a bill-size locus in a finch-like bird with two alleles, L (small bill) and H (large bill), showing simple additive action: genotype LL gives bill depth 8 mm, genotype LH gives 11 mm, and genotype HH gives 14 mm (each H allele adds 3 mm). Two ecologically available seed types require bill depths near 8 mm (soft seeds) or near 14 mm (hard seeds); the 11 mm heterozygote cracks neither efficiently and survives at only 40% the rate of either homozygote (relative fitnesses w_LL = 1.0, w_LH = 0.4, w_HH = 1.0 — classic underdominance/disruptive selection at a single locus). Starting from an allele frequency of p(L) = 0.7, q(H) = 0.3 (asymmetric, reflecting a population that recently colonized a habitat richer in soft seeds), Hardy–Weinberg gives genotype frequencies before selection: LL = p² = 0.49, LH = 2pq = 0.42, HH = q² = 0.09 After selection, relative genotype contributions are: LL: 0.49 × 1.0 = 0.49 LH: 0.42 × 0.4 = 0.168 HH: 0.09 × 1.0 = 0.09 Mean fitness w̄ = 0.49 + 0.168 + 0.09 = 0.748 New allele frequency of L after selection (each LL contributes two L alleles, each LH contributes one L and one H): p' = [2(0.49) + 0.168] / (2 × 0.748) = [0.98 + 0.168] / 1.496 = 1.148 / 1.496 ≈ 0.767 So L rises slightly (from 0.70 to 0.767) simply because it started more common and the heterozygote is disfavored regardless of which allele is rarer — this is the hallmark of underdominance: selection tends to amplify whichever allele already has the numerical edge, driving the population toward fixation of one allele rather than maintaining both. If a second, geographically separated subpopulation started instead with H more common, that population would drift toward fixing H — precisely the divergence-promoting property of disruptive selection at a single locus, and a miniature illustration of how disruptive selection, combined with limited gene flow between patches specializing on different resources, can produce divergent local fixation rather than an indefinitely maintained bimodal polymorphism within one freely mating deme.

Hermon Carey Bumpus (1862–1943) was an American zoologist working at Brown University when a catastrophic winter storm struck Providence, Rhode Island in February 1898. Dozens of house sparrows were found stunned and near death on the ground; Bumpus, who had already been measuring sparrow morphology for other reasons, seized the opportunity presented by this natural experiment. He carefully measured nine morphological traits on the collected birds — total length, wingspread, weight, length of beak and head, humerus length, femur length, tibiotarsus length, sternum length — and recorded which birds survived after being brought indoors to recover and which died. His 1899 publication analyzing these data (drawing on statistical methods then being pioneered in the biometrical school associated with Karl Pearson) became one of the earliest attempts to quantify natural selection acting on a wild population using real survival data rather than theoretical argument alone. Bumpus himself interpreted his results cautiously, noting that selection appeared to favor birds of intermediate size on some measurements — an interpretation that later 20th-century biometricians would formalize as one of the first empirical portraits of stabilizing selection, even though the term itself was not coined until decades later (largely through the work of Ivan Schmalhausen and George Gaylord Simpson in the mid-20th century, and given rigorous quantitative-genetic footing by Russell Lande in the 1970s–80s).

In medicine and public health, the stabilizing-selection logic behind birth weight remains directly relevant to obstetric risk assessment: clinicians still track deviations from optimal birth weight as a predictor of neonatal complications, even though modern interventions (cesarean delivery, neonatal intensive care) have flattened the historical mortality curve and shifted the effective optimum. Understanding that fitness costs are curved, not linear, also underlies genetic counseling for polygenic disease-risk traits such as blood pressure and blood glucose, where extreme values in either direction carry elevated risk. In agriculture and animal breeding, breeders exploit the logic of disruptive selection deliberately: selecting simultaneously for both a high-yield line and a distinct trait combination (e.g., splitting a herd into a fast-growth line and a lean-meat line) can be modeled as artificially imposed disruptive selection on a formerly unimodal population, accelerating divergence between breeding lines far faster than natural disruptive selection typically would. Conversely, breeders selecting for uniform, predictable product traits (egg size, kernel dimensions) impose deliberate stabilizing selection, culling both undersized and oversized individuals from breeding stock. In conservation biology, recognizing whether a threatened population is under stabilizing or disruptive selection changes management strategy. A population under strong stabilizing selection around a narrow optimum (for instance, precise timing of migration or breeding tied to a narrow seasonal resource window) may have little standing variance to respond if climate change shifts that optimum — a direct link to the limits of adaptation. A population showing signs of incipient disruptive selection on a resource-use trait (as documented in some cichlid fish radiations in African rift lakes, and in some stickleback populations exploiting benthic versus limnetic prey) may represent an early stage of ecological speciation that conservation planning should treat as two evolutionarily significant units rather than one homogeneous population. In evolutionary and developmental biology research, quantitative estimates of selection gradients (γ, from Lande–Arnold regressions) are now routinely computed for wild populations — Darwin's finches on Daphne Major, Soay sheep on Hirta, Trinidadian guppies, threespine sticklebacks — to test whether observed morphological stasis over long timescales reflects ongoing stabilizing selection (an unmoving optimum) versus simple lack of genetic variance, a distinction central to interpreting the fossil record's frequent pattern of long-term morphological stability punctuated by rapid shifts.

Ask the tutor: If a trait shows a bimodal distribution in a wild population today, what specific field data would you need to collect to distinguish ongoing disruptive selection from a historical artifact of population mixing?