How deterministic trajectories and their fluctuations underlie allele-frequency statistics in the strong-selection regime

Abstract

In many biological contexts selection is strong relative to genetic drift. Working under the diffusion approximation, we define $R = 2N_{e}|s|$, where $N_{e}$ is the effective population size and $s$ is the selection coefficient associated with a focal allele. Strong selection corresponds to $R \gg1$ and can occur for relatively modest parameter values. For example, $N_e = 10^3$ and $s = 10^{-2}$ yield $R = 20$. The focus of this work is on statistics of the allele frequency distribution in the strong selection regime. In this regime, standard approximations often break down. For instance, under strong positive selection ($R\gg1$ and $s>0$), trajectories that proceed to fixation seem to dominate allele frequency statistics. However, omission of trajectories that ultimately achieve loss can lead to large errors. Over timescales where mutation can be neglected, all allele frequency trajectories fall into one of two classes: those that eventually achieve fixation and those that eventually achieve loss. We ensure that both types of trajectory contribute to time-dependent statistics, by separately conditioning on the eventual fixation and eventual loss of the focal allele. For large $R$, we determine approximations for allele-frequency statistics under a small noise approximation that derives contributions from two deterministic trajectories, one achieving fixation the other loss, along with fluctuations around these two trajectories. Such an approach yields the correct long time values of the mean allele frequency and its variance. Numerical comparisons with the Wright-Fisher model indicate that the contribution of the deterministic trajectories alone, to a statistic, may not always be sufficient for good accuracy, but with the inclusion of fluctuations, reasonable accuracy is obtained.