Fixation probability and time, and why a better mutant often loses
Sandbox Physics ·
A mutant that grows five per cent faster sounds like a sure winner. In a population of fifty it is lost about nine times out of ten. Early on it exists as one or two copies, and a single unlucky generation removes it before its advantage can count. This chapter computes that probability three ways, exactly and approximately, and checks them against simulations.
Give every A individual a relative fitness 1+s, where s is the selection coefficient. In the Wright–Fisher draw, a child now picks an A parent with probability p′ instead of p. A positive s favours the mutant and a negative one harms it.
The mean change per generation is about sp(1−p), a push towards fixation. The random part is the same as before, with variance p(1−p)/N. Selection wins when its push over one generation is large compared with the random step, which happens when ∣Ns∣ is large.
p′=1+spp(1+s),E[Δp]≈sp(1−p),Var[Δp]≈Np(1−p)
Mean and variance of the change in one generation, to first order in the selection coefficient.
The Moran model changes one individual at a time. In each step one individual is chosen to reproduce, with probability proportional to fitness, and one is chosen uniformly to die. The number of mutants moves up by one, down by one, or stays put, so the process is a birth-death chain.
Let ϕi be the probability that i mutants eventually take over. Conditioning on the first step that changes i gives a recursion with ϕ0=0 and ϕN=1. Its solution depends only on the ratio of the down and up probabilities, which equals 1/r for constant fitness r. For neutral mutants it reduces to ϕi=i/N, as in the previous chapter.
One Moran generation is N steps. Traced backwards, its lineages merge twice as fast per generation as Wright–Fisher lineages, so its drift is twice as strong. The same s then gives about half the fixation probability.
ϕi=1−r−N1−r−i,r=1+s
Probability that i mutants of relative fitness r take over a Moran population of N.
The Wright–Fisher chain has no closed form with selection, but it can be solved numerically, and for large N it has a good approximation. Let u(p) be the fixation probability from frequency p. After one generation the frequency has moved by a small random step, and u must equal its own average over that step. Expanding to second order in the step gives the equation below.
Divide by p(1−p)/(2N) to get u′′=−2Nsu′, so u′ decays as e−2Nsp. Integrating once and imposing u(0)=0 and u(1)=1 gives the result on the right. A single mutant in a large population fixes with probability close to 2s when Ns is large, and close to 1/N when Ns is small.
Worked example: a five per cent advantage among fifty
Problem. One mutant with s=0.05 appears in a population of N=50. How likely is it to fix in the Wright–Fisher model, and how long does a successful mutant take?
Diffusion. With 2Ns=5 and p=0.02, u=(1−e−0.1)/(1−e−5)=0.0958.
Exact chain. Writing the 51×51 transition matrix and solving the linear equations for absorption gives 0.0941. The diffusion formula is 1.8 per cent high. The Moran formula with the same N and r=1.05 gives 0.0522, about half, as expected from its stronger drift.
Time. The exact chain gives a mean of 75.8 generations for mutants that fix, against 97.0 for a neutral mutant in the same population. Successful favoured mutants rise faster.
Check. In 12000 seeded replicate populations of the Lab preset Better, yet usually lost, the mutant fixed in a fraction 0.0948±0.0027 and took 77.6±1.1 generations. The fraction is within a quarter of a standard error of the exact value; the time is 1.6 standard errors high, which is within ordinary scatter.
Answer. More than nine favoured mutants in ten are lost. The 2s rule of thumb, 0.1, is already close.
The product Ns decides. In a population of 20 a harmful mutant with s=−0.05 starting at p=0.3 still fixes with exact probability 0.1255, far above zero; the Lab measures 0.1304±0.0031. In a population of 100 a mutant with s=0.05 starting at p=0.1 fixes with probability 0.625, six times its starting frequency.
Across the five Lab presets, the simulated fixation fractions and times scatter around the exact values as the standard errors predict. That check only tests the simulation against the exact chain of the same model. Whether real populations behave like Wright–Fisher sampling is a separate question.
Choose Better, yet usually lost. Predict the fraction of populations in which the mutant fixes before reading the exact value.
Set the selection coefficient to zero and keep everything else. Predict how much the fraction drops, then check it.
Choose Selection wins in a larger group, then lower the population size towards ten. Explain with the product of population size and selection coefficient why the advantage stops mattering.
Exercises with explained answers
In a Moran population of N=10, one mutant has relative fitness r=1.1. What is its fixation probability, and how does it compare with a neutral mutant?
Show the answer
ϕ1=(1−1/1.1)/(1−1.1−10)=0.0909/0.6145=0.148. A neutral mutant fixes with probability 0.1, so the advantage raises the chance by about half.
Use the diffusion formula for one mutant with s=0.01 in a Wright–Fisher population of N=1000.
Show the answer
Here p=0.001 and 2Ns=20. u=(1−e−0.02)/(1−e−20)=0.0198, which is almost exactly 2s: the mutant is lost about 98 times in 100.
The Lab estimates a fixation probability near 0.094 from 12000 populations. What is the binomial standard error, and how many populations would bring it down to 0.001?
Show the answer
0.094×0.906/12000=0.0027. The error falls as one over the square root of the number of runs, so 0.001 needs 0.094×0.906/0.0012≈85000 populations.
Version 2, full text. Section II, Population genetics in well-mixed populations: genetic drift as random sampling of a finite population, the heterozygosity as the chance that two sampled individuals differ, and its loss with time.
Version 1, full text. Section 1.3.1: the Moran process as a birth-death chain with transition probabilities. Section 1.3.2, Eqs. (1.7)-(1.20): fixation probabilities from a recursion, with the closed form of Eq. (1.20), which reduces to the starting fraction for neutral selection. Eqs. (1.26)-(1.35): mean and conditional fixation times.
Version 1, full text. Section 2.1.1: the Kingman coalescent, in which each pair of lineages merges at rate one. Theorem 2.3: the Moran genealogy becomes this coalescent after rescaling time. Theorem 2.4: the Wright-Fisher genealogy does too when time is measured in units of N generations. Theorem 2.8, Eqs. (36)-(38): a neutral allele fixes with probability equal to its frequency, with the diffusion formula for the mean absorption time; the formula after it gives the time conditioned on fixation.
Version 2, full text. Eq. (41): a selective advantage gives sector boundaries a mean drift on top of their diffusion. Eq. (42): the drift is tan(Φ/2) for opening angle Φ. Eq. (46): cos(Φ/2) is the ratio of wild-type to mutant front speed. Eq. (49): survival probability of a beneficial sector at a linear front; its footnote 4 notes the formal similarity to the Kimura fixation probability in a well-mixed population.
Keep reading
What this page links to directly
You are hereFixation probability and time, and why a better mutant often loses