Demography of rare mutations (population genetics)

The classic formula for mutation frequencies in large DNA samples can now be evaluated exactly at any strength of selection. A companion technique, counting pairs of nearby mutations, reveals that gene conversion plays an important role. Includes a short film.

Two mutations at a time — a hand-drawn six-minute telling of the second paper on this page: how counting pairs of nearby mutations in 113 Gambian genomes shows that a model with gene conversion, left out of standard analyses, accounts for about 95 percent of their departure from the standard picture, and why in Zambian fruit flies the same test still fails with conversion added, while shared fly haplotypes point instead to lineages that merge many at a time.

The 1992 formula for the effect of natural selection on variant frequencies can be evaluated exactly, at sample sizes where ordinary computer arithmetic fails, and applied to the protein-coding DNA of about 730,000 people. A count of 5.7 billion pairs of nearby variants in the genomes of 113 Gambians traces most of the pairs’ departure from the textbook model of ancestry to gene conversion, a process that standard analyses leave out. Both results are the work of Matthew Schwartz and Michael Desai.

One mutation at a time

Sample $n$ chromosomes today and count the variants carried by exactly one of them, by two, and so on. The resulting histogram is the site-frequency spectrum, in use since Motoo Kimura’s work of 1969, and most inference in population genetics comes from fitting models to it. Ancestry shapes it: traced back in time, the sample descends from fewer and fewer ancestors, and in the standard description, the coalescentThe random family tree of a sample of chromosomes traced backward in time; in Kingman’s 1982 version ancestral lines merge two at a time, at a rate inversely proportional to the population size. of John Kingman (1982), ancestral lines merge two at a time, faster when the population is small. A recent expansion slows mergers near the present and inflates the rare end of the histogram; a bottleneck depletes it. Selection shapes the spectrum too. A harmful mutation is removed before it can spread, so the more harmful a class of variants, the more it concentrates at the rare end, and the distribution of fitness effects across new mutations, the DFE, can in principle be read from the shape of the spectrum.

The theory behind such readings is the Poisson random field of Sawyer and Hartl (1992). Write $S = 4N_e s$, where $s$ is the mutation’s fractional effect on fitness, negative when harmful, and $N_e$ the effective population sizeThe size of an idealized randomly mating population with the same amount of random genetic drift as the real one; the textbook long-term human value is about 15,000.. For a population of constant size and an additive mutation (one copy doing half the harm of two), the expected number of sites whose variant is carried by $i$ of the $n$ sampled chromosomes is

$$\xi_i(S) \;=\; \theta \binom{n}{i} \int_0^1 x^{\,i-1}(1-x)^{\,n-i-1}\, \frac{1-e^{-S(1-x)}}{1-e^{-S}}\, dx .$$

Here $x$ is the variant’s unobserved frequency in the whole population, integrated out, and $\theta$ sets the mutation rate. The binomial coefficient times $x^{i}(1-x)^{n-i}$ is the chance that a variant at frequency $x$ is caught exactly $i$ times, and the remaining factor is proportional to how long a variant under selection $S$ spends near frequency $x$ before it is lost or takes over. Mixing the formula over a DFE and matching an observed histogram is the core of the standard programs polyDFE, fitdadi and fastDFE, and the modern picture of the human DFE comes from such fits.

Why sixteen digits are not enough

The integral has a closed form in the confluent hypergeometric function, and neighboring entries obey a three-term recurrenceA rule that fixes each entry of a sequence from the two entries before it, so that two starting values determine the whole sequence. with both end values known exactly. On a computer the recurrence fails, because it has two independent solutions: the physical one, with entries of order one, and an unphysical companion that is astronomically larger, already at least $10^{83}$ for 100 chromosomes at $S=\pm10$. Every rounding error injects a little of the companion, which then swamps the answer. Solving for all entries at once is no better: the output passes the only check available at fixed precision, satisfying the equations to rounding error, and can still share no digits with the true values, because the companion satisfies the same equations.

Log-log plot of required interior working precision in decimal digits (vertical axis, 10 to 10,000) against sample size n in haplotypes (horizontal axis, 50 to 2000). Four blue curves with markers, for |S| = 1, 10, 100 and 1000, rise almost linearly and nearly coincide, from between about 55 and 125 digits at n = 50 to about 6,000 digits at n = 2000; a note says the curves for +S and -S coincide exactly. A dashed gray guide line labeled 3 digits per haplotype runs just above them, and an annotation marks 6332 digits at (n, S) = (2000, ±1). A solid horizontal gray line near the bottom is labeled float64 working precision (about 16 digits); every curve lies far above it at every n shown.

Decimal digits that must be carried to evaluate the whole selected spectrum by the recurrence, or by solving for all entries at once, against the number of sampled chromosomes $n$, for four selection strengths $|S|$ (curves for $+S$ and $-S$ coincide). The requirement grows by about three digits per sampled chromosome (dashed guide) and reaches 6,332 digits at $n=2000$. Ordinary double precision supplies about sixteen (solid horizontal line), too few already at $n=50$, so fixed-precision arithmetic cannot evaluate the spectrum at modern sample sizes. (Figure 1 of the first paper.)

The spectrum can instead be evaluated without any rounding. For rational $S$, every entry $M_i$ of the hypergeometric vector behind the spectrum has the form

$$M_i \;=\; \alpha_i + \beta_i\, e^{-S}, \qquad \alpha_i,\ \beta_i \in \mathbb{Q},$$

a fraction plus a fraction times one exponential ($\mathbb{Q}$ denotes the rational numbers). With the two end values known, the linear system splits into two systems over the rationals, which are solved exactly. The companion reappears as the size of the fractions’ numerators and denominators, where it does no harm; computed along with the solution, that size fixes in advance how many digits to carry when the two parts are combined. This route is practical to about 2,000 chromosomes. Out to $n=10^5$ the recursion is run instead in ball arithmetic, in which every number carries a rigorous error radius, and above a million chromosomes each needed entry is summed as an arbitrary-precision series to 30 significant digits with its own error bound. The construction extends, with controlled error, to arbitrary dominance, the share of a mutation’s harm already felt in one copy.

Broken genes in 730,000 exomes

A harmful variant at strength $S$ circulates at frequencies near $1/|S|$, so at classical sample sizes everything from $|S|=10^2$ to $10^5$ falls in the first cell or two of the histogram. The v4.1 exomeThe protein-coding part of the genome, about one percent of its length; exome sequencing reads only those regions. release of the public gnomAD database pools 730,947 people, up to 1,461,894 sampled chromosomes per position, and at that depth the characteristic frequency for every strength in that range lies inside the observable spectrum. The deleterious DFE is fitted to the rare end for missense variants, which swap one amino acid, and for high-confidence loss-of-functionA variant predicted to destroy the gene’s product outright, for example a premature stop codon or a broken splice site. variants, with population history absorbed by calibrating on synonymous variants, which change the DNA without changing the protein. The strong tail is modeled as eight free bins of mass between $|S|=10^2$ and $10^5$; simulated data showed that single bins are not identifiable, so only totals over three wide windows of $|S|$ are reported.

Grouped bar chart. Horizontal axis: scaled strength of selection against the mutation, |S| = 4 N_e |s|, in three windows labeled 100 to 1,000, 1,000 to 10,000 and 10,000 to 100,000. Vertical axis: share of the class's harmful mutations, 0 to 100 percent. Two bars per window, slate for missense (one amino acid changed) and orange for loss of function (gene disabled), each with a black whisker for its 95 percent interval. First window: missense 32.0 percent with a whisker from 0 to 86 percent, loss of function 15.0 percent with a whisker from 0 to 40 percent. Middle window: both bars have zero height, labeled 0 at best fit, with whiskers reaching 57 percent (missense) and 33 percent (loss of function). Last window: missense 28.5 percent with a whisker from about 8 to 43 percent, loss of function 73.6 percent with a whisker from about 57 to 84 percent.

Where the harmful mutations of two classes fall on the scale of selection strength, from the fit to the gnomAD v4.1 exomes. Bars give the best-fit share of each class’s deleterious DFE in three windows of scaled strength $|S|$; whiskers are 95% intervals with the model’s measured misfit built in. About three quarters of the loss-of-function mass (orange) and somewhat more than a quarter of the missense mass (slate) sit in the strongest window, $|S|$ between $10^4$ and $10^5$. The middle window is empty only at the best fit, since its intervals reach 57% and 33%, and the weakest window is the one most entangled with population history. (Redrawn from Table 1 and Figure 2b of the first paper.)

For loss-of-function variants, 73.6% of the fitted deleterious DFE lies at $|S|$ between $10^4$ and $10^5$ (95% interval 57.4% to 83.6%); for missense the same window holds 28.5% (7.7% to 42.7%). Converting the window to a fitness cost requires a choice of $N_e$, and that choice is left open. At the textbook long-term human value, $1.5\times10^4$, most gene-disabling mutations would be close to lethal in one copy; at the larger effective sizes appropriate to recent growth, the edge of the same window corresponds to a cost near 2% in one copy.

Dominance is harder to extract, because at such low frequencies essentially nobody carries two copies: ordinary chromosomes measure only the product of dominance and strength, and the X chromosome, where males carry one copy, separates the two only partly. With the ratio of X to autosome effective size fixed at its neutral reference value of 0.75, the data give only a one-sided lower bound on the baseline dominance coefficient $h_0$, and a coarse scan brackets where that bound falls. For loss-of-function variants it lies between 0.5 and 0.934, so a gene-disabling mutation appears to do more than half its harm in the first copy; for missense, between 0.15 and 0.5. Neither bracket is a confidence interval, and both move with the assumed ratio.

When population history mimics selection

Expansions inflate rare variants just as selection against harmful mutations (purifying selection) does, and Myers, Fefferman and Patterson showed in 2008 that different histories can give identical neutral spectra. Take sample size $n=20$ and ask whether any size history of a well-mixed population with Kingman ancestry reproduces the expected spectrum of a selected site. It can be proved that the neutral spectrum of every such history is a weighted average of points on one explicit polynomial curve. Whether some history mimics a selected spectrum therefore comes down to whether that spectrum lies inside the curve’s convex hullThe region made up of all weighted averages of a set of points, with nonnegative weights that sum to one., and exact arithmetic decides this unless the margin is too small to resolve. At every weak strength sampled, from $S=-10$ to $+1.125$, there is an eighteen-epoch neutral history whose spectrum matches the selected one to within about $10^{-11}$: history genuinely mimics weak selection. For stronger selection the opposite is proved, at every sampled beneficial strength from $S=1.375$ to $100$ in the additive case and at twenty-five combinations of strength and dominance in all. Each proof exhibits a weighted sum of the spectrum’s entries plus a constant that is nonnegative along the whole curve and negative at the selected spectrum, so no history in the class reproduces it. These statements hold for well-mixed populations with Kingman ancestry at $n=20$ and say nothing about structured populations or multiple-merger genealogies.

Exact values also make it possible to measure the accuracy of polyDFE, fitdadi and fastDFE, by refitting identical data with each program’s numerics and with the exact likelihood. The three agree with the exact fit to within a small fraction of a confidence interval at the sample sizes for which they were developed. They lose accuracy in three settings: polyDFE with beneficial mass above $S\approx38$ at more than about 90 chromosomes, fitdadi under its 2017 rule for DFE mass beyond its selection grid, and all three at samples of order $10^5$ chromosomes, which none of them handles in the versions tested. The authors see no reason to revisit published deleterious-DFE estimates made with current defaults.

Two mutations at a time

The single-site histogram has a blind spot: different ancestries can produce the same spectrum. Several recent studies argue that multiple mergers, genealogies in which one ancestor occasionally leaves many descendants at once, are widespread and can mislead Kingman-based inference. In 2025 Fenton, Rice, Novembre and Desai proposed a statistic that can detect them. For pairs of nearby variable sites, record how many sampled chromosomes carry the mutated, or derived, allele at each site, $i$ and $j$; the table of pair fractions over $(i,j)$ is the two-site frequency spectrum. Two close sites share a genealogical tree, and how often their mutations fall in the same individuals depends on its branching structure, which multiple mergers alter. Applied to the fruit fly Drosophila melanogaster, the test rejected the Kingman coalescent genome-wide.

The same test can be applied to people: 113 unrelated Gambians from the high-coverage 1000 Genomes panel. Pairs of variable sites on the non-sex chromosomes, the autosomes, up to 100,000 base pairs apart are counted, 5.7 billion pairs in all, and binned by genetic-map distance, which measures how often the reshuffling of chromosomes between generations separates two points. The primary analysis takes the 590 million most closely linked pairs, whose map distance is indistinguishable from zero. Each pair’s counts among the 226 chromosomes are averaged exactly over every subsample of four people, $n=8$, which leaves a table of 28 cells. The expected spectrum of fully linked sites is computed exactly for any piecewise-constant history of up to 24 epochs and for any mixture of such histories across genomic regions. Because this class of histories includes every mixture of its members, the search for the member closest to the data has no local optima, and the smallest misfit it finds is the least that any Kingman history in the class can achieve.

Two 7-by-7 heat maps side by side sharing one color bar. In each, the horizontal axis is copies of the mutation at the second site, j, from 1 to 7 of 8, and the vertical axis is copies at the first site, i, from 1 to 7 of 8. The color bar runs from dark slate (one quarter of the prediction) through white (equal to the prediction) to dark orange (four times the prediction) and is labeled observed pairs divided by predicted pairs. Left map, titled data vs. best Kingman history: cells at low counts in the lower left are white to pale slate; the color deepens to orange toward the upper right, and cells pairing two high counts are dark orange, darkest at (6,7) and (7,6), while the diagonal cells (3,3), (4,4), (5,5), (6,6) are markedly paler than their off-diagonal neighbors. Right map, titled data vs. Kingman plus gene conversion: every cell is near white, with only a faint orange tint in the top row and right column.

The two-site spectrum of the 590 million most closely linked pairs in 113 Gambian genomes, drawn as observed pairs divided by predicted pairs in each cell. Row and column give how many of the eight sampled chromosomes carry the derived allele at each of the two sites; orange cells hold more pairs than predicted, slate cells fewer, white cells match. Left: against the best fit from the entire class of Kingman histories. The excess grows toward pairs of common alleles, passes four times the prediction in the darkest cell, $(6,7)$, and is largely absent on the diagonal, where the two sites have equal counts. Right: against the fitted Kingman model with gene conversion and ancestral mislabeling, on the same color scale; little structure remains. (Redrawn from the data of Figure 2 of the second paper.)

No Kingman history fits the Gambian pair spectrum: the best member of the whole class misses the data by a chi-squaredSum over cells of the squared gap between data and prediction in units of the measured noise; noise alone averages one per degree of freedom. misfit of $\chi^2 = 7{,}697$ on 27 degrees of freedom, more than two hundred times the rejection threshold of 29.2 set by the measured sampling noise. The margin survives every analysis choice that was varied, and the misfit points the same way on each of the nineteen autosomes used.

Seven candidate explanations

Seven candidate processes were then simulated: gene conversion, crossover-map error, Beta-coalescentA one-parameter family of family trees in which a single ancestor can leave many sampled descendants at once; the parameter sets how often that happens. multiple mergers, sequencing error, population structureDeparture from random mating because a population is divided into partly isolated groups that exchange few migrants., linked selectionDistortion of variation at a site caused by natural selection acting on nearby sites that are inherited along with it., and mislabeling of the ancestral allele. From each simulated spectrum its own best Kingman refit is subtracted, which removes whatever a change of history could absorb, and the strength of each process is capped at a direct measurement where one exists and otherwise at a deliberately strong scenario.

Horizontal bar chart of alignment with the observed deviation on a scale from 0 to 1, where 1 means exactly the same direction, with seven bars. Gene conversion is a long orange bar at 0.913. The others are slate: crossover-map error at its best-scoring rate 0.63, crossover-map error at plausible rates 0.14, multiple mergers (Beta coalescent) 0.10, sequencing error 0.33, population structure 0.35, ancestral allele mislabeled 0.14. A vertical dotted line at 0.19 is labeled spread of a random direction (one s.d. = 0.19).

How closely the change each candidate process makes to the two-site spectrum points along the observed deviation, scored from 0 (unrelated) to 1 (the same direction), with each process’s own best Kingman refit removed and every cell weighted by its sampling noise. The dotted line marks the typical score of a random direction, 0.19. Gene conversion (orange) reaches 0.913. An error in the crossover map reaches 0.63 only at its best-scoring rate and 0.14 at physically plausible rates; the multiple-merger family scores 0.10, no better than chance; every other process scores 0.35 or less. (Redrawn from Figure 4 of the second paper.)

Only gene conversion moves the spectrum in the direction of the observed deviation. The Beta-coalescent family scores no better than a random direction, even though the test is well powered against it, so these human data, unlike the fly data, give no reason to invoke multiple mergers. The six other processes together, each within its bound, remove 74.0% of the deviation, against 95.2% for the gene-conversion model.

Crossover. When sperm and eggs are made, the two copies of a chromosome can break at the same point and swap everything to one side of it, so positions far apart often end up with different ancestries while close neighbors travel together.

Gene conversion. A short patch, tens to a few hundred DNA letters long, is copied from one chromosome onto its partner with no swap in return. When the patch covers one of two neighboring variants but not the other, the neighbors end up with different ancestries, just as they would after a crossover between them. (Both from the film above.)

Gene conversion comes from the same machinery that makes crossovers when sperm and eggs form. Instead of swapping long segments between the two copies of a chromosome, it copies a short patch, tens to hundreds of base pairs, from one copy onto the other. Analyses of linked variation usually omit it. The fitted model keeps Kingman ancestry under a freely fitted history $\mathbf{w}$ and adds conversion in two tract-length classes plus a mislabeling rate:

$$\phi^{\rm model} \;=\; F_{\varepsilon}\!\left[\, \phi^{\rm K}(\mathbf{w}) \;+\; \lambda_{\rm short}\, D_{L_{\rm short}} \;+\; \lambda_{500}\, D_{500} \,\right].$$

Here $\phi^{\rm K}(\mathbf{w})$ is the exact Kingman two-site spectrum under history $\mathbf{w}$, and $D_L$ is the simulated shift in the 28 cells from conversion tracts of mean length $L$: 110 base pairs for the short class, the value measured in sequenced families (pedigrees), and 500 for the long class. Each rate $\lambda$ is the per-base-pair, per-generation probability that a site lies inside a transferred tract, and $F_\varepsilon$ reverses the ancestral label at a fraction $\varepsilon$ of sites, sending count $i$ to $8-i$. Fitted jointly with the history, so that nothing demography could explain is credited to conversion, the model removes 94.2% to 96.2% of the deviation across analysis choices; in the reference analysis $\chi^2$ falls from 7,697 to 370.2, with a mislabeling rate of $\varepsilon = 1.67\%$.

The fitted total conversion rate, $8.33\times10^{-6}$ per base pair per generation, is of the same order as pedigree and sperm-typingMeasuring recombination directly by reading the DNA of many single sperm cells from one donor. measurements. The total is known in absolute terms only to roughly a factor of two, and every rate scales with the assumed mutation rate. With its parameters fixed on the most closely linked pairs, the model predicts the deviation at other separations: in a separate pair set binned by physical distance, the measured deviation matches the prediction within 3% at 200 and 1,500 base pairs and falls 9% short at 650. At a few thousand base pairs, where crossovers dominate, the model misses by 10 to 30%, and no choice of crossover rate removes the discrepancy. The 1000 Genomes Yoruba cohort and an independently processed callset give the same deviation and parameters within about a standard deviation; since both samples are West African and share most variant sites, this is more a consistency check than an independent confirmation.

The papers

Supplementary material

The importable popcorn library implements the first paper’s exact computation; the package and its guide live in the BootLoops repository at tools/popcorn.

The gnomAD, 1000 Genomes, HGDP and deCODE data are public.

References

M. Kimura, The number of heterozygous nucleotide sites maintained in a finite population due to steady flux of mutations, Genetics 61 (1969) 893the expected site-frequency spectrum under a steady flux of new mutations
J. F. C. Kingman, The coalescent, Stochastic Process. Appl. 13 (1982) 235the standard model of ancestry, with lines merging two at a time
S. A. Sawyer and D. L. Hartl, Population genetics of polymorphism and divergence, Genetics 132 (1992) 1161the Poisson random field: the formula for the selected spectrum that every fit evaluates
W. Gautschi, Computational aspects of three-term recurrence relations, SIAM Rev. 9 (1967) 24why a recurrence run at fixed precision loses its small solution to the large one
P. Tataru, M. Mollion, S. Glémin and T. Bataillon, Inference of distribution of fitness effects and proportion of adaptive substitutions from polymorphism data, Genetics 207 (2017) 1103polyDFE, the first of the three standard programs compared with the exact values
B. Y. Kim, C. D. Huber and K. E. Lohmueller, Inference of the distribution of selection coefficients for new nonsynonymous mutations using large samples, Genetics 206 (2017) 345fitdadi, with the 2017 rule for DFE mass beyond its selection grid
J. Sendrowski and T. Bataillon, fastDFE: fast and flexible inference of the distribution of fitness effects, Mol. Biol. Evol. 41 (2024) msae070fastDFE, the third program compared with the exact values
S. Chen et al. (Genome Aggregation Database Consortium), A genomic mutational constraint map using variation in 76,156 human genomes, Nature 625 (2024) 92the gnomAD database, whose v4.1 exome release the first paper fits
S. Myers, C. Fefferman and N. Patterson, Can one learn history from the allelic spectrum?, Theor. Popul. Biol. 73 (2008) 342proof that different population histories can give identical neutral spectra
E. F. Fenton, D. P. Rice, J. Novembre and M. M. Desai, Detecting deviations from Kingman coalescence using 2-site frequency spectra, Genetics 229 (2025) iyaf023the two-site test for multiple mergers, and its rejection of Kingman ancestry in fruit flies
M. Byrska-Bishop et al., High-coverage whole-genome sequencing of the expanded 1000 Genomes Project cohort, Cell 185 (2022) 3426the sequence data from which the 113 Gambian and the Yoruba genomes are drawn
A. Bergström et al., Insights into human genetic variation and population history from 929 diverse genomes, Science 367 (2020) eaay5012the HGDP callset, the independently processed callset of the replication check
G. Palsson et al., Complete human recombination maps, Nature 639 (2025) 700the deCODE genetic map used to bin the pairs, and the pedigree tract length of 110 base pairs

← back to the web summaries