Experimental Evolution as a High-Throughput Screen for Genetic Adaptations

Experimental evolution is a method in which populations of organisms, often microbes, are founded by one or more ancestors of known genotype and then propagated under controlled conditions to study the evolutionary process. These evolving populations are influenced by all population genetic forces, including selection, mutation, drift, and recombination, and the relative contributions of these forces may be seen as mysterious.

methods now pales in comparison to today's genome-enabled screens that harness simple population genetic processes to precisely identify genetic adaptations.
Evolution experiments are influenced by all population genetic forces: selection, genetic drift, mutation, and recombination. In conducting a genetic screen, we can usually ignore the process of recombination, given that we are mostly interested in "first-step" adaptive mutants arising in different clones. Given that mutations occur inevitably, the relative balance of the remaining forces is governed by the effective population size, N e , and the strength of selection, s, acting on these mutations (Table 1). N e is a property that is typically much less than the census population size (N) at any given time and is biased toward bottlenecks that occur during transfers. In practice, microbial populations are transferred at sizes ranging from a single cell to~10 8 individuals or more. Meanwhile, values for s that have been measured during experimental evolution range from~0.01 (e.g., a minor improvement in resource uptake [6]) to~4 (9) (involving the absolute gain in ability to colonize a new niche), with smaller values more common. Mutants with s Ͻ 0.01 are likely much more abundant but remain largely undetected because they are lost by drift or outcompeted by more beneficial mutants in large experimental populations, a process known as clonal interference (Table 1) (2,10). However, technologies that rarify cells and limit clonal interference could detect even smaller fitness effects (13,42).
When evolving populations are maintained under conditions in which the product of these two properties N e and s is clearly greater than 1 (N e s Ͼ Ͼ 1), then selection becomes the dominant force in the population (2,21,43). To understand these properties more practically, consider a typical serial dilution evolution experiment in which bacterial populations are grown from a single clone in medium that supports a population of 10 8 cells/ml (Fig. 1). Let us also assume that the ancestor is not preadapted to this environment so there is opportunity for improvement. Population growth on the first day involves~10 8 cell divisions and generates~10 5 mutations given the approximate per-genome, per-generation mutation rate of 10 Ϫ3 (11,12). Selection could act upon any of these~10 5 mutations, and the probability that any one of them would reach a detectable frequency (say, 0.01) depends on the mutant's frequency and its selective value. So, the earlier a mutation occurs on this first day of population expansion, the more likely it will reach high frequency. Yet most mutations arise in the final division of population growth, i.e., when the population increases from 5 ϫ 10 7 to 10 8 cells/ml, involving 5 ϫ 10 7 cell divisions. The key point is that after this first growth cycle, there are already many mutations present in the population, but nearly all are very rare, having arisen in the last generation (16).
In a side note, in considering these estimates, one may wonder if some sites or genome regions may be more mutation-prone than others, thus increasing the number of certain mutations. While mutation rates do vary among sites, such rate variation typically involves fold changes (2ϫ to 5ϫ) and rarely more than an order of magnitude (10ϫ) (11,12). These processes can be influential over hundreds of generations but This is the size of an ideal population, in which all individuals reproduce equally and experience no fluctuation in size, which experiences the same genetic drift as the actual population. It is typically calculated as the harmonic mean of the population sizes at the time of transfer and the fraction that is transferred, which is strongly biased toward the bottleneck size. Selective coefficient (s) The fitness difference between a given genotype and typically, a wild-type genotype, in units of time Ϫ1 . Commonly in experimental evolution, s is defined as the difference in Malthusian parameters between the mutant and wild type as follows: ln (N m 1/N m 0) Ϫ ln (N WT 1/N WT 0), where N is cell number and m is mutant and WT is wild type at time 0 or 1 (e.g., N m 1 is the number of mutant cells at time 1). Establishment The process whereby a mutation rises to a high enough frequency to escape loss by drift, i.e., greater than experimental population bottlenecks and greater than 1/s.

Clonal interference
Competition between beneficial mutants in an asexual population resulting in the loss of less beneficial mutants and the delayed rate of fixation of the most beneficial mutant (10). Also known as the Hill-Robertson effect in sexual populations, where the process may be overcome by recombination that assembles beneficial mutations within the same genome (41).
typically contribute little over short time scales because they would add only a handful of new mutations to a sample of~10 5 . Nonetheless, if these mutation hot spots encode a trait under strong selection, such as phase variation of membranes affecting antibiotic resistance (14,17,18), they may greatly accelerate the evolutionary rate.
On the next day of the experiment, we transfer some fraction of the first population to fresh medium, say 1/100, and growth resumes (Fig. 1). For example, 10 6 cells/ml harboring~10 3 mutations (simple dilution) now regrow to reach 10 8 cells/ml, which involves another 10 8 cell divisions (10 6 divisions to reach 2 ϫ 10 6 cells/ml, 2 ϫ 10 6 divisions to reach 4 ϫ 10 6 cells/ml, and so on) and produces another 10 5 mutations. Despite the many individual cell divisions, it is important to realize that the number of generations the entire population undergoes during the next cycle is log 2 (10 8 /10 6 ) ϭ 6.67 generations. This is the number we commonly reference in describing the duration of an evolution experiment over any interval. It is also noteworthy that this demographic measure of generation number does not depend on the population size but rather on the dilution itself. Under these conditions, the N e of a population cycling between 10 6 and 10 8 cells/ml each day is~2 ϫ 10 6 (Table 1). Now, revisiting the condition in which the force of selection dominates, N e s Ͼ Ͼ 1, any mutation with a selective coefficient s much greater than 1/N e is likely to rise in frequency by selection provided it becomes established in the population. By establishment (Table 1), we mean the mutant must also be lucky enough to reach a frequency where it escapes genetic drift and then is guided deterministically by selection (Fig. 1). This frequency is typically 1/s individuals (2,43). Thus, a fairly typical big-benefit mutation that increases fitness by 10% (s~0.1) (19,20) exceeds the general N e s threshold by 4 orders of magnitude. If its population size reaches a mere 10 individuals, it is nearly guaranteed to be governed by selection and is likely to completely take over (fix) the population in the absence of competition. However, should this big-benefit mutation cooccur in a population with an even more beneficial mutation (s ϭ 0.2 or more), it may lose out to this superior competitor by the process of clonal interference (Table 1 and Fig. 2). In fact, even though beneficial mutations may amount to only 1% of all mutations (or less), large populations guarantee that many beneficial mutations will arise and compete under these conditions, with only the best of them ever reaching a detectable frequency of 0.01 or more ( Fig. 2) (2,19,21).
As the experiment progresses, suppose mutants with some potentially adaptive phenotype (a change in colony morphology indicative of altered secretions, growth parameter, metabolic capacity, host colonization, etc.) are observed in a population after seven transfers or~46 generations. You pick these mutants to subculture them, and their phenotypes appear heritable. What is the probability that these mutants involve adaptations driven by selection and did not become frequent by the random forces of mutation and drift? Virtually certain. Here is why: any given mutation with no effect on fitness (and likely no phenotype) will be governed by genetic drift and will reach a potentially detected frequency of 0.1 in the following number of generations, where g is the number of generations, p is the mutant frequency, and is the mutation rate (22) This would equal 3,322 generations, or 499 days in this experimental design. This is the estimated time required for any mutation to reach a frequency of 10% by genetic drift alone. For a mutant to become detected 100 times faster than this in your evolution experiment (and this is not uncommon), the mutant must have been subject to strong positive selection because it is adaptive. Plot of log mutation frequency over time, which demonstrates that adaptive lineages must rise in frequency by several orders of magnitude before they may be easily detected. All lineages may be influenced by transfer bottlenecks that randomly increase or decrease their frequency (see erratic changes in the lineages that went extinct). Dashed lines denote mutations that did not reach a conventionally detected frequency and were driven extinct by competition with more-fit lineages. Dynamics are based on simulations generated by FFPopSim (40) with asexual lineages and periodic 1:100 population bottlenecks.

Opinion/Hypothesis
As evolution experiments become commonplace for many microbial species in a range of environments, reports of parallel mutations-even identical nucleotide changes in the same gene-are becoming more frequent (15,23). This observation is often met with the question-is this particular site hypermutable? We need to consider three more questions. First, can a high mutation rate alone drive a mutation to high frequency? Second, what role might selection have played? Third, under what conditions would parallelism occur?
Let us suppose that this position is hypermutable, not the usual 1 ϫ 10 Ϫ10 mutations/bp/generation but 100-fold more mutable at 1 ϫ 10 Ϫ8 mutations/bp/ generation. Would a high mutation rate alone cause it to rise to high frequency in multiple populations?
Using the same equation above, a specific mutation would reach a frequency of 10% by drift alone by Ϫlog 2 0.1 10 8 ϭ 33,219,289 generations, which is clearly implausible, so mutation pressure alone cannot lead to parallelism. However, let us imagine this mutant is strongly selected with s ϭ 0.2, which is much greater than the mutation rate.
It will rise to 10% frequency in Ϫlog 2 0.1 0.2 ϭ 16.6 generations, or less than 3 transfer days, which is fast! Selection can cause a highly beneficial mutation to sweep very rapidly, in the typical time required for an efficient laboratory genetic screen. However, the probability that the same mutation becomes rapidly detected in another population depends on whether this same mutation actually occurs in time. The evolution experiment we have outlined here generates~10 5 new mutations per day, with only 10 3 on average surviving each bottleneck. The probability of this same mutation even recurring is the approximate per-site mutation rate of 2 ϫ 10 Ϫ10 , which will occur on average in 10 8 cell divisions/day Ϸ 39 days or 263 generations. Sitespecific parallelism is therefore still highly improbable. However, strongly selected mutations are much more likely to survive the bottleneck and may also become enriched both by demography (becoming slightly more common in the transfer by chance). Estimating the exact probability of parallelism is not straightforward because it depends on these chance effects as well as on the distribution of effects of other selected mutations within the population, but in essence, if the same mutation does indeed occur by chance in another population and escapes drift and outcompetes other mutants, it will also rapidly reach this frequency of 0.1.
To sum up, any single mutation rising from a frequency of 10 Ϫ7 to 10 Ϫ1 in an evolution experiment lasting weeks or months (100 to 1,000 generations) must be under positive selection, and equally important, the contribution of other population genetic processes is negligible. Parallelism at the level of nucleotides, amino acids, or even genes in replicate populations adds overwhelming evidence of strong selection on that target.
Why are these inferences useful? Given that the force of selection dominates evolution experiments conducted in large populations, mutants rising to high frequency must have acquired adaptive traits in the selective environment. Furthermore, the large population sizes typically used in these experiments often generate multiple beneficial mutants that compete with one another (Fig. 2), meaning that those that ultimately become detectable (e.g., as colonies on a plate) are among the most fit available. Genome sequencing of these mutants can therefore be used as a forward genetic screen for traits that enhance fitness in any environment, including new host organisms (9,15,(24)(25)(26)(27). Sequencing multiple mutants from independent populations or diverse samples of the evolved populations can expand the sample of adaptive mutations. Any parallelism at the level of gene or pathway provides powerful inference that alterations in this pathway provide an ideal adaptive solution. As one example from our screen of adaptations to biofilm growth, we observed four independent D652N mutations in the response regulator protein WspE in different populations of Burkholderia cenocepacia selected for only 32 to 64 generations (15). These mutants produced a characteristic wrinkly colony morphology and increased biofilm. No specific features of the nucleotide sequence indicate that this site is hypermutable, but this remarkable parallelism is consistent with the fitness advantage of s Ϸ 1.0 that rapidly displaces the ancestor (15). These results also demonstrate that this aspartate residue (and not the adjacent aspartate at position 653) that is outside the primary receiver domain is functionally important for signaling that the cell is in contact with the surface (28,44). Another example of an evolution experiment serving as a genetic screen involved E. coli selected for growth in test tubes in minimal medium containing glucose (19). Some of the most fit mutants acquired partial loss-of-function mutations in yfgA and prc, genes that encode a cytoskeletal protein and a protease involved in peptidoglycan synthesis. We reasoned that both mutations could result in the synthesis of a more limited peptidoglycan layer to facilitate transport but also alter cell shape, which we indeed observed as the production of more spherical cells (19). Put simply, this evolve-and-resequence approach is likely to find the best mutations (or mutation combinations) to respond to any selective challenge and can inform how the affected proteins function and enable the mutant to outcompete its ancestor.
Researchers with diverse interests have recognized this opportunity, leading to an explosion of the kinds of questions addressed by experimental evolution and two successful American Society for Microbiology (ASM) meetings focused on the subject. For example, experimental evolution has been used to discover: (i) how bacteria evolve resistance to antibiotics (29,30) or environmental pollutants (31), (ii) how microbes evolve mutualistic or antagonistic social interactions with other microbes (32)(33)(34)(35), (iii) how bacteriophage undergo host range expansion (36,37), (iv) how microbes adapt to current hosts (38) or novel hosts (9), and (vi) how pathogens evolve during infections (39). These imaginative studies have identified new pathways that reveal how these microbes adapt to new conditions and some recurring themes, such as the selection for mutations in regulators that produce multiple adaptive traits at once. With the wealth of knowledge and experience from almost 150 years of evolution experiments and an influx of creative young investigators, the future of using evolution experiments to reveal unknown mechanisms of adaptation could not be brighter.

ACKNOWLEDGMENTS
I thank Caroline Turner and Chris Marshall for helpful discussions, two anonymous reviewers for useful recommendations, and Richard Neher for assistance with simulations.
V.S.C. is supported by NASA NAI CAN-7 NNA15BB04A and by NIH R01GM110444 and U01AI124302-01.