The Selection Coefficient

Relative fitness, w, rescales the absolute reproductive output (survival × fecundity) of each genotype so that the best genotype has w = 1. The selection coefficient, s, is simply the complement:

s = 1 − w If a genotype has relative fitness w = 0.9, its selection coefficient is s = 0.1 — it produces, on average, 10% fewer surviving offspring than the fittest genotype. If w = 0 (the genotype dies before reproducing, or is completely sterile), then s = 1: total selection against that genotype. If a genotype is more fit than the current reference genotype, s is written as negative (s < 0), meaning that genotype is favored and will increase in frequency. Because diploid organisms carry two alleles per locus, s alone is not enough to describe a whole genetic system — you also need dominance, denoted h, which specifies how the fitness deficit is expressed in the heterozygote relative to the homozygote. For a locus with alleles A (favored) and a (disfavored), a standard fitness scheme is:

Here h = 0 means a is fully recessive (heterozygotes are as fit as AA); h = 1 means a is fully dominant (heterozygotes suffer the full cost); h = 0.5 means additive/codominant effects (heterozygote fitness deficit is exactly half of the homozygote's). This single table — s and h — is the entire quantitative machinery of one-locus directional selection, and it plugs directly into the discrete-generation recursion for allele frequency covered in Allele Frequency Dynamics.

Step 1: Genotype fitnesses generate differential survival/reproduction. Within one generation, individuals of genotype AA, Aa, and aa are born at frequencies given by Hardy–Weinberg proportions (p², 2pq, q², where p and q are allele frequencies of A and a). Selection acts by differentially removing or reducing the reproduction of certain genotypes before the next generation's gene pool is counted. The three fitness values (1, 1 − hs, 1 − s) determine what fraction of each genotype's zygotes make it to the reproductive gene pool.

Step 2: Mean population fitness is the weighted average. The population's mean fitness after selection is: w̄ = p²(1) + 2pq(1 − hs) + q²(1 − s) This number tracks how much the whole population's productivity has been dragged down by the deleterious genotypes present. It's central to the recursion because each genotype's relative contribution to the next generation is its own fitness divided by w̄.

Step 3: New allele frequency after one generation of selection. The standard recursion (derived by allele counting after weighting by fitness) for a recessive deleterious allele (h = 0) is: q′ = q(1 − sq) / (1 − sq²) For a fully general case, the change in allele frequency per generation, Δq, is approximately: Δq ≈ −spq[q + h(p − q)] / w̄

Step 4: Iterate. Because q itself has changed, you recompute w̄, apply the recursion again, and repeat — generation by generation, this is exactly how allele frequencies drift toward fixation or elimination under constant selection pressure. The speed of this process is entirely governed by the magnitude of s (and, for diploids, by h).

Step 5: The dominance trap for recessive alleles. When a is recessive (h = 0) and rare, most copies of a sit inside heterozygotes Aa, which have fitness 1 — indistinguishable from AA. Selection can only "see" and remove a when it is homozygous, an increasingly rare event as q shrinks (probability q² of homozygosity). This produces the famous asymptotic slowdown: strong selection against a recessive lethal (s = 1) still takes hundreds of generations to reduce a rare allele from q = 0.01 to q = 0.001, because at low frequency the allele is almost entirely shielded by heterozygosity.

Notice the shape of the s = 1 curve in the second figure: rapid loss of the allele while it's common (many aa homozygotes exposed to selection), followed by a long, slow tail — this is the recessive-allele "hiding in heterozygotes" effect described above, and it is one of the most important quantitative facts in population genetics: selection against rare recessive deleterious alleles is extremely inefficient, which is why such alleles (including many human genetic disease alleles) persist in populations for very long times even under strong selection.

Kettlewell's peppered moth experiments (1950s). H. B. D. Kettlewell tested whether the rise of the melanic (dark) form of Biston betularia in industrial England was driven by differential bird predation against the pale form on soot-darkened tree bark. He performed mark–release–recapture experiments in polluted Birmingham woodland and in unpolluted Dorset woodland, releasing known numbers of marked pale and melanic moths and recapturing survivors. In polluted woodland, melanic moths survived and were recaptured at substantially higher rates than pale moths; in unpolluted woodland the reverse held. Direct observation (and later film) documented birds selectively picking pale moths off blackened trunks. From recapture ratios, biologists could back-calculate relative fitness and hence s for the pale genotype in polluted habitat — estimates from these and later studies (e.g., Bishop 1972, Cook et al.) suggested selection coefficients on the order of s ≈ 0.2–0.5 against the disadvantaged morph per generation, strong enough to explain the observed multi-decade frequency shift from records of moth-trap surveys across England. Later criticisms (moth resting-site behavior, some staged photographs) refined methodology but did not overturn the core finding — subsequent large-scale field studies (Cook, Grant, Saccheri et al. 2012 in Biology Letters) using more naturalistic release methods reproduced the same qualitative and quantitative result, with recorded gene-frequency declines matching predictions from selection-recursion models using estimated s.

Endler's guppy predation experiments (1980s). John Endler transplanted Poecilia reticulata (guppies) between pools with differing predation intensity in Trinidad streams and also ran replicated artificial-pond experiments with and without predators. He tracked the frequency of conspicuous color-pattern spots over generations. Populations exposed to the predatory fish Crenicichla evolved toward drab coloration within roughly 15–20 generations (a few years), while populations moved to predator-free habitats evolved brighter coloration on a similar timescale. Because Endler could measure survival directly by recapturing marked fish, he estimated selection coefficients on spot number and size — differential survival differences of predator-exposed versus predator-free guppies implied s values in the range of roughly 0.1–0.3 per generation for conspicuous coloration under high-predation regimes. This was one of the first field demonstrations that natural selection's strength could be measured quantitatively and used to predict the rate of evolutionary change, which Endler then confirmed by comparing predicted versus observed changes in color-pattern frequency.

Bacterial chemostat and fluctuation-test measurements. In modern microbiology, s is measured with even more precision using competition experiments: a resistant and a susceptible bacterial strain are mixed at a known starting ratio and grown together (with or without antibiotic), and the ratio is tracked by plating or fluorescent labeling over generations. Because bacteria divide asexually and generation number is known exactly, fitting the exponential-growth-ratio data to w = 1 − s gives extremely precise selection-coefficient estimates — routinely used to quantify the fitness cost of antibiotic-resistance mutations (often s ≈ 0.01–0.20 cost in drug-free media, flipping to strongly negative, i.e. advantageous, s in the presence of drug). This experimental logic traces back to the design principles of the Luria and Delbrück fluctuation test tradition of quantitative bacterial population genetics, extended into precise fitness assays. Together, these three lines of evidence — field mark-recapture (moths), field/experimental-pond manipulation (guppies), and laboratory competition assays (bacteria) — establish the same general point from different angles: s is not a theoretical abstraction but a directly measurable quantity, and populations change allele frequency at rates that match the values of s estimated from independent fitness/survival data.

Example 1 — a genetics cross: recessive lethal allele in a laboratory population. Suppose a recessive allele a causing embryonic lethality when homozygous (aa fitness = 0, so s = 1, h = 0) starts at frequency q₀ = 0.20 in a large randomly mating population (no new mutation, no drift). Using the recursion: q′ = q(1 − sq) / (1 − sq²), with s = 1 q′ = q(1 − q) / (1 − q²) = q(1 − q) / [(1 − q)(1 + q)] = q / (1 + q) Iterate:

Notice the pattern: for a fully recessive lethal, the exact solution is qₜ = q₀ / (1 + t·q₀). Plugging t = 100, q₀ = 0.2: q₁₀₀ = 0.2 / (1 + 20) = 0.0095, matching the table. The allele is never fully eliminated in finite time under this deterministic model — it just gets rarer and rarer, hiding increasingly in heterozygous carriers who are fully fit. This is precisely the mechanistic point from the section above: complete selection (s = 1) against a recessive genotype produces slow, asymptotic decline, not rapid extinction — directly relevant to why recessive Mendelian disease alleles (e.g., many autosomal recessive disorders) persist at low but nonzero frequency in human populations despite being severely deleterious in homozygotes.

Example 2 — a clinical/epidemiological scenario: a drug-resistance mutation in a pathogen. An antiretroviral drug is introduced to treat a chronic viral infection. Before treatment, a resistance-conferring point mutation (call it allele R) exists at very low frequency (p₀ = 0.001) among the viral quasispecies within a patient, because in the drug-free environment R carries a small replication cost: its relative fitness without drug is w = 0.95, i.e. s = 0.05 disadvantage (asexual/haploid population, so no dominance term needed — every copy is "exposed"). Once the drug is started, the fitness relationship flips: wild-type virus is now strongly suppressed (relative fitness w_wildtype ≈ 0.30 under drug pressure) while resistant virus keeps replicating close to its pre-drug rate (w_R ≈ 0.95). Rescaling so the fittest type (R) has w = 1, the wild type now has s ≈ 1 − (0.30/0.95) ≈ 0.68 — strong selection against wild type, equivalently strong selection for R. For a haploid asexual population, allele frequency change per generation follows the simpler recursion: p′ = p·w_R / [p·w_R + (1 − p)·w_wildtype] With p₀ = 0.001, w_R = 0.95, w_wildtype = 0.30: p′ = (0.001)(0.95) / [(0.001)(0.95) + (0.999)(0.30)] = 0.00095 / 0.30065 ≈ 0.00316 Already the resistant frequency has tripled in one viral generation. Because HIV replicates roughly once per day (with generation-scale turnover on the order of 1–2 days) this kind of ratio compounding produces resistant-strain dominance of the within-host population within a few weeks — which is exactly why clinical guidelines require multi-drug combination therapy: a single-drug regimen imposes such a large s against the wild type that a pre-existing low-frequency resistance mutation sweeps to near-fixation before the drug can clear the infection.

The mathematical formalization of s within a full population-genetic framework is usually credited jointly to the architects of the Modern Synthesis's mathematical wing: J. B. S. Haldane, Ronald Fisher, and Sewall Wright, who independently derived the allele-frequency recursions under selection in the 1920s–30s. Haldane in particular is famous for explicitly calculating selection coefficients from field and historical data — his 1924 paper A Mathematical Theory of Natural Selection derived the very q′ = q(1 − sq)/(1 − sq²)-type recursions used above, and his later analysis of the peppered moth's rate of spread (estimating that s on the order of 0.2–0.5 was required to explain the observed 50-year shift from pale to melanic moths in industrial England) was one of the first attempts to reverse-engineer a selection coefficient from real ecological/historical data, decades before Kettlewell's direct field experiments confirmed selection was actually operating at that intensity.

Antibiotic and antiviral resistance management. Measuring s for resistance mutations, both with and without drug present, lets clinicians and public-health agencies predict how fast resistant strains will sweep under a given treatment regimen and design combination therapies (multi-drug HIV cocktails, tuberculosis regimens) that keep the effective selection coefficient for any single resistance mutation low enough that resistance cannot fix before the pathogen is cleared.

Agricultural pest and herbicide resistance. Selection coefficients estimated from field trials of insecticide- or herbicide-resistant pest populations guide "resistance management" strategies — refuge crops, rotation of chemical classes — explicitly designed to keep s for resistance alleles close to neutral in refuge populations, slowing the spread predicted by the same recursion equations used above.

Conservation genetics. Estimating s for deleterious recessive alleles in small, inbred populations (e.g., endangered species with reduced heterozygosity) informs genetic-rescue and management decisions, since the asymptotic-decline mathematics shows why purging deleterious recessives from a small population is far slower than naive intuition suggests.

Genomics and detecting selection in humans. Modern statistical methods (site-frequency spectra, extended haplotype homozygosity, ancient-DNA time-series analysis) estimate s for alleles like lactase persistence or high-altitude adaptation variants, quantifying historical selection strength directly from genomic data rather than direct fitness measurement — a computational descendant of the same logic Kettlewell used in the field.

Cancer evolution. Tumor clones with driver mutations can be modeled with selection coefficients analogous to those in population genetics, where s quantifies the proliferative advantage of a mutant clone within the tissue — used to predict how fast resistant subclones will dominate a tumor under targeted therapy, directly paralleling the HIV drug-resistance example above.

Ask the tutor: How does the selection coefficient change when we move from constant directional selection to frequency-dependent or heterozygote-advantage scenarios?