The evolution of genetic drift over 50,000 generations

Random variation in reproductive success-genetic drift-profoundly shapes genetic diversity and evolutionary trajectories. The strength of drift depends on the variance in descendant number, ${\sigma }_d^2$ , which governs key evolutionary outcomes: for instance, the establishment probability of a beneficial mutation scales inversely with ${\sigma }_d^2$ . However, whether ${\sigma }_d^2$ itself evolves over long timescales has remained unclear, because allele-frequency fluctuations depend on drift only through the effective population size, $N_e=N/{\sigma }_d^2$ , which blends census population size with descendant-number variance. Here, we disentangle these components by using model-based Bayesian inference combined with joint tracking of (i) frequency fluctuations of neutrally barcoded lineages and (ii) census population sizes across growth cycles in the E. coli Long-Term Evolution Experiment. Analyzing 33 clones spanning the ancestor through 50,000 generations in two replicate populations (Ara-2 and Ara+2), we find that the strength of genetic drift evolved markedly-and divergently-between the two replicate populations. Both census size and ${\sigma }_d^2$ changed substantially through time, with most variation in $N_e$ driven by shifts in ${\sigma }_d^2$ rather than census size. After approximately 2,000 generations, the ${\sigma }_d^2$ of the two populations diverged sharply: Ara+2 generally remained close to a bottleneck-only null expectation, whereas Ara-2 exhibited 1.5-5× stronger drift, consistent with an evolved increase in stochasticity during growth. Because establishment probability scales as $s/{\sigma }_d^2$ , a beneficial mutation of given effect is roughly twice as likely to establish in Ara+2 as in Ara-2. Our results demonstrate that the key parameter governing genetic drift can itself evolve, with direct consequences for adaptation.