Experimental Design, Population Dynamics, and Diversity in Microbial Experimental Evolution

Cripple mutants.Cripple mutants, often lacking major regulatory or metabolic genes (49) or carrying malfunctioning essential genes (88), have been used as founders to investigate alternative evolutionary solutions of biology (Fig. 1C). For example, evolution experiments have been initiated with mutants lacking key metabolic enzymes (89–92) or global regulators (93, 94), with mutants having alterations in their central metabolism (95–97), or with mutants lacking genes that were previously shown to be targets of evolution by accumulating gain-of-function mutations, thereby opening the road to alternative paths (98, 99).

Beyond simple nutritional stress.Experimental evolution has also been carried out under more-challenging environmental conditions (Fig. 1D), for example, under extreme pH (100–103), osmotic pressure (101, 104–106), suboptimal redox status (101), oxygenation (107), extreme temperature (67, 108–110), or UV radiation (111); in the presence of antibiotics (50, 112–117), antimicrobial peptides (118–120), and alcoholic solvents (101, 121); and even under microgravity (122, 123).

Similarly to the above-mentioned abiotic stresses, biotic stress has also been imposed on microorganisms during evolution experiments. Experiments allowing predation by protists or infections with phages or, when the study focuses on predators, using new or scarce hosts (124–127) are often performed. This kind of biotic pressure seems to constrain the simultaneous adaptation to abiotic conditions (128).

Spatiotemporally changing environments.In experimental evolution, the imposed environment is not always kept constant but instead often changes over time (Fig. 1E). First, there are evolution experiments where the environment is changed progressively. By changing the environment according to evolutionary progress, researchers have tried to prevent the selection pressure from dropping over time and therefore stimulate further adaptation (129), for example, by increasing antibiotic concentrations (83, 85, 130–132) or the concentrations of other antibacterial substances such as silver nanoparticles (133) and other metals (134), the ionizing irradiation dosage (135, 136), solvent concentrations (32, 65, 137–139), hydrostatic pressure (140), or temperature (79, 141, 142) according to the emerged resistance. In a sense, these experiments were all inspired by one of the first evolution experiments ever performed: William Dallinger enabled protozoa to grow at extreme temperatures in the 19th century by gradually increasing the temperature as soon as the protozoa adapted (6).

Others used fixed rates of environmental change without evolutionary feedback, often resulting in deteriorating conditions. These experiments often aim at understanding whether species can adapt by evolutionary rescue or will go extinct in the context of global environmental changes. Generally, this is correlated with the rate of environmental change and shows that gradual changes result in greater evolutionary adaptation rather than abrupt ones (143–145). As means to study global change, several kinds of increasing stresses have been used, such as increasing antibiotic concentrations (143, 146), osmotic stress (51, 52, 147), phosphate limitation (148), or hydrostatic pressure and temperature (149–151) or, in the case of virus adaptation, changing to a novel host type (145).

Alternating between two or more contrasting environments, either in cycles (107) or randomly (100, 152) and with various frequencies (153, 154), is another category of environmental change that has been used in experimental evolution (Fig. 1E). Classical examples are shifts between carbon sources (155–157), but others have been examined as well, such as alternations between acidic and basic pH (158), different temperatures (159, 160), light and dark regimes in photosynthetic organisms (161–163), different antibiotics (130) or antibiotic treatment and recovery (26, 164), alterations between hosts and predators (126, 152, 165–167), or freeze-thaw growth cycles (168).

In some experiments, the kind and rate of environmental change are determined by the outcome of evolution itself, i.e., when two or more coevolving species interact with each other and evolve in response to each other (169, 170). The best-known examples are predator-prey systems. For example, coevolving bacteriophages and bacteria often lead to an evolutionary arms race where constant adaptation is needed not to dominate but merely to survive with respect to an ever-evolving counterpart, known as the Red Queen hypothesis (RQH) (originating from the statement that the Red Queen made to Alice in Lewis Carroll's Through the Looking-Glass [516], the sequel to Alice's Adventures in Wonderland [517], explaining why in Looking-Glass Land everyone needs to run to stay in the same place) (171–175). Also, for commensal and mutual interacting partners, reciprocal coevolution dynamics have been reported (176–179).

In a last group, the selective environment is changed not over time but rather in space, or it simultaneously contains contrasting niches. The combination of different niches, often with various grades of interconnectivity and, thus, migration, leads to a heterogeneous and/or structured environment. The use of two or more carbon sources at the same time in a well-mixed environment is an extremely simple example of such a heterogeneous environment without physical boundaries to migration (155, 156, 180–182). Experimental evolution has also been carried out under more-heterogeneous conditions and in structured environments with less mixing (43, 183–185), like biofilms (176, 186, 187), or in a patchy environment with different antibiotic concentrations (188) or differences in illumination (162).

Evolution under conditions closer to those of natural environments.In an attempt to study adaptive processes that could also take place in nature, conditions in experimental evolution have mimicked natural conditions as close as possible (Fig. 1F). While the environment is still often well defined, it combines (multiple) abiotic and/or biotic stresses, changing over time and/or in space (134, 189). Note that some of the studies cited above also combined several stresses although they did so in a generally less extreme fashion and often were not focused on the effect of complex selective forces on evolution. As an example, Pseudomonas fluorescens was evolved in a structured environment with a combination of protists, phages, and antibiotics (190), or the adaptation of Pseudomonas aeruginosa was monitored in artificial sputum medium to mimic the lungs of cystic fibrosis patients (113, 191, 192). In a less-defined setup, a Lactococcus lactis plant isolate was domesticated to a dairy niche by adaptation to milk, which resulted in properties highly similar to those of L. lactis isolates from dairy products (193). A similar domestication took place in Burkholderia cenocepacia evolved on onion extracts, which resulted in the loss of its pathogenicity to the nematode Caenorhabditis elegans (194).

Experimental microbial evolution has also been conducted in situ, for example, in eukaryotic cell lines (195, 196), and using whole-animal or plant model hosts such as mice (197–200), corn (201), Mimosa pudica (202), rabbits (203), ferrets (23), worms (204), or caterpillars (205), often with the goal of understanding pathosymbiotic adaptation or how hosts can affect the evolution of the microorganism. These kinds of experiments can serve as nice parallels to observations made based on comparing isolates from long-term infections or symbioses in real life (56, 206, 207). In an analogous way, active efforts are being made, especially by the groups of Buckling and Brockhurst, to monitor the evolution of focal species in communities and/or living in structured and complex microcosms, better resembling natural, free-living conditions. Many current experiments still use only a rather simple binary setup of two species (208–211) that often fall within the coevolving regime of predator-prey systems or beneficial interactions between two species (see also above). However, the use of more-complex communities is emerging. Here, one attempts to understand how ecological interactions can affect the evolution of the focal organisms or how evolution can affect ecosystem functioning. Local adaptation of focal species could, for example, slow down due to strong interstrain competition leading to strong population bottlenecks. Alternatively, evolution might speed up due to fast coevolution between strongly interacting species or take other directions altogether. Depending on the specific system under study and the strength and sign of the ecological interactions present in the community, both outcomes have been observed. For example, adaptation of P. fluorescens is constrained when strong competitors are present (212–214), while local adaptation of the species was shown to be potentially equally as important to community structure as the presence of the species itself (215). In contrast, evolution elicited stronger changes when 5 decomposer bacteria, all isolated from the roots of beech trees, were propagated together in a community than those elicited by evolution in monoculture. Not only did stronger metabolic interactions emerge in the form of diverged resource use and waste product cross-feeding, but communities were also more productive (216). Performing evolution experiments with communities can also lead to fairly unexpected and somewhat counterintuitive results. Interspecies gene transfer between Pseudomonas putida and P. fluorescens, for example, was recently shown to be inhibited in soil microcosms when positive selection was applied for traits encoded by the conjugative agent, a mercury resistance plasmid (217).

All these conditions together in which experimental evolution has been performed listed in this section show its power as a research tool and explain how the field exploded and diversified in many complex and specialized subdomains (Fig. 1F).

A special case: Richard Lenski's long-term evolution experiment.On 24 February 1988, Richard Lenski started culturing his famous 12 parallel Escherichia coli populations (see http://myxo.css.msu.edu/ecoli). He used a constant and simple environment. The cultures were grown in minimal medium with low levels of glucose as the sole accessible carbon source, and 1% of each population was transferred daily into new flasks with fresh medium, allowing for another cycle of overnight batch growth (with shaking at 37°C and at 120 rpm). His evolution experiment is therefore an example of a standard serial transfer setup. Being very well thought through, it contributed to current unwritten “laws” to be followed when starting evolution experiments. For example, as a common ancestor, he used two isogenic variants of an E. coli B strain that had a rare combination of sensitivity to phage T5 (confirming that it is E. coli) and resistance to T6 (most E. coli strains are sensitive) (45). Moreover, REL606 and REL607 (six populations founded by each) differed by a neutral, visual Ara marker. Therefore, not only can contamination from external sources be easily detected, but cross-contamination between the separate populations can also regularly be checked for since the handling of the cultures was always performed by alternating Ara⁺ and Ara⁻ populations. Afterwards, this same marker was also used for determining fitness in head-to-head competition experiments. Along with the additional precautions taken (keeping the transferred flasks in the fridge for one night and regularly preserving population samples by freezing), these measures would later add up to the characteristic (and, to outsiders, seemingly excessive) caution taken by anyone working in the field today.

It is not only the clever experimental design of Lenski's experiment that deserves a special mention of his work, as other ingenious evolution experiments with microbes were also carried out during that time (46, 82, 218–221). The experiment especially deserves an extended discussion because it is still running today, hence its name “long-term evolution experiment” (LTEE). This experiment led to a plethora of results ever since (more than 80 publications on the LTEE alone [see http://myxo.css.msu.edu/ecoli]). Indeed, over the past 30 years and over 68,000 generations (theoretically 6.67 a day, corresponding to more than 1,000,000 years in human terms, while our own species, Homo sapiens, is only ±7,500 generations old [222, 223]), the populations evolved and changed (224), resulting in numerous, sometimes unexpected, observations, many of which are used as examples throughout this review. For example, the cell size and growth rate increased (225, 226), while lag times became shorter (62), the cell shape changed from rods to more-spherical cells (227), mutators emerged (228), two or more genotypes coexisted for many generations (229, 230) or interfered with each other in a race to fixation (226, 231), and indirect, sometimes correlated, responses to other, naive environments occurred (232–234). Novelties evolved, such as the capacity to use citrate as a carbon source (13), and elemental stoichiometry changed such that evolved cells contained relatively more phosphorus and nitrogen than carbon, as these elements are abundant in the LTEE environment (235).

The mutations responsible for many of these changes have been identified over the years and are of all kinds, either single-nucleotide polymorphisms (SNPs), insertions-deletions (indels), or larger rearrangements (236–238), but combined, they show a strong signature of natural selection (239). Furthermore, they were often found to interact epistatically in their final, combined result on the phenotype (240–242). Sometimes, several mutations emerged seemingly simultaneously in the same background and were fixed as clades (231). As such, the LTEE combines in one experiment many of the evolutionary observations reported for all other evolution experiments together and consequently enables the testing of evolutionary theories on a longer time scale. Adaptation in the LTEE slowed down over time, but it has not stopped, and according to the power-law model, without any upper boundary that best describes its trajectory, it probably never will (45, 243).

The fitness landscape and the distribution of fitness effects.To conceptually visualize the fitness of all possible genotypes in a given environment, Sewall Wright conceived the metaphor of a fitness landscape (518). Here, a three-dimensional (3D) surface represents all possible genotypes (a simplified version of the actual multidimensional space), peaks denote fitness optima, and valleys represent genotypes with low success (Fig. 2C). Evolution can be seen as a walk on this landscape, usually toward higher fitness peaks under sufficiently strong selection. Representations of such landscapes generally contain few peaks (rugged landscape) or only one peak (smooth landscape), with many vast planes and valleys (Fig. 2C). Although the shape of the landscape points to other features as well, it proposes a plenitude of neutral and deleterious mutations in the genotype space. Microorganisms have encountered both smooth landscapes (240) and rugged ones (255, 278) in experimental evolution. Few landscapes have been reconstructed and often provide only an incomplete image, mainly because they are small or contain only combinations of mutations that were identified at the very end of an evolution experiment, thereby often ignoring genotypes with an improved fitness that did not make it as well as deleterious or neutral mutations. Still, for small genome sizes at least, genome-wide single-nucleotide fitness landscapes were established, as for poliovirus (279), and more and larger landscapes are being constructed frequently nowadays, aided by extensive sequencing and fitness determinations, for example, resulting in a large fitness landscape in yeast adapting to limiting glucose concentrations (280). MA experiments, in which all spontaneous mutations are allowed to be fixed by single-cell bottlenecks at transfer, except for lethal ones, confirm the abundance of deleterious and neutral spontaneous mutations compared to beneficial ones. Indeed, in these experiments, the fitness of most replicate lines tends to decline over time, which has been dubbed Muller's ratchet, as in these setups, the evolving species has no way to lose deleterious mutations (48, 57, 59, 60, 281). In general, the rates of spontaneous mutation are highest for neutral and then deleterious mutations and lowest for beneficial mutations (14).

Many researchers have tried to profile the distribution of fitness effects (DFE) of spontaneous mutations (58, 61, 282), a key to understanding or predicting biological adaptation (Fig. 2D). While MA experiments are generally skewed toward the abundant deleterious mutations (283, 284), marker divergence studies can be a suitable additional way to pinpoint the DFE of new beneficial mutations (Fig. 2C) (199, 245, 285). An excellent example is a recent report in yeast where the frequency of 500,000 differently DNA-barcoded sublineages of a population was monitored by deep NGS during adaptation under glucose limitation (Fig. 2E) (248). Those authors found 25,000 of them to have acquired a beneficial mutation; most of them had a small effect (1.02 to 1.05), and some carried a larger fitness benefit (peaks at 1.07 to 1.08 and 1.10 to 1.11), but none had any higher fitness. Follow-up work on isolated clones confirmed these findings in the construction of a broad landscape linking fitness to specific single mutations (280). Surely, any aspect of the fitness landscape or DFE depends on the selective conditions, the sensitivity of fitness assays, and the number of mutations examined, which might also explain the different shapes that have been reported for the DFE (286–290). Remarkably, the DFE of beneficial mutations found by deep NGS of the barcoded population does not resemble any of the previously proposed ones (248). In general, however, it seems to be true that the number of beneficial mutations drops once the effects become larger and that while beneficial mutations are rare, mutations with a strong beneficial effect are even rarer (Fig. 2D) (288).

Clonal interference.When beneficial mutations emerge in different individuals, an evolving population harbors different mutant sublineages. In the absence of recombination, only one of these mutations can ultimately sweep to fixation, while the remaining ones will be outcompeted along with the ancestor. The resulting competition among mutant clones, in addition to the competition between a clone and its ancestor, is called clonal interference (CI). It has been shown to be prevalent in evolution experiments, observed by either NGS (187, 263, 270, 293) or marker divergence studies (42, 199, 245, 248, 285) or based on phenotypes of separate clones (13, 294), and clearly influences the dynamics of evolution.

Intuitively, the mutant with the largest fitness advantage should eventually become the dominant one. However, the fate of such a genotype becomes uncertain by CI. Indeed, during the sweep to complete fixation, a more beneficial mutation can arise in a different background. In that case, the frequency of the mutation that was initially the most beneficial will increase only to a certain level in the population and then decrease again in favor of the new mutant. As such, CI promotes the fixation of genotypes with large fitness improvements, even if they are rare. Based on empirical data, beneficial mutations with small effects are indeed quickly outcompeted in favor of prior fixation of mutations with large fitness increases (231, 263, 270, 285, 294). In extreme cases, when a genotype that is almost fixed is outcompeted by a mutant that emerged in a background that was hardly detectable because of its very low frequency, this has resulted in the so-called leapfrog phenomenon, as the dominant genotype is replaced by a mutant that is more closely related to the founding ancestor than to the dominant genotype itself (79, 231, 295).

Even when no fitter mutants arise, new additional beneficial mutations will still keep arising in the competing backgrounds, thereby decreasing the benefit of the most-fit sublineage. The fixation sweep will slow down, again allowing more time for mutants with higher fitness to arise (187, 226, 248, 294). Thus, the rate of adaptation cannot be increased indefinitely by increasing the population size or the beneficial mutation rate. From a certain point on, CI will start to act as a speed limit since its strength scales well with the population size or with the beneficial mutation rate. Larger populations or higher beneficial mutation rates might increase the probability of a mutation with a large effect to occur, yet at the same time, more different mutant sublineages will be present and slow down its sweep further (243, 246, 296, 297).

Genetic hitchhiking.At times, multiple mutations also cooccur in the same background. In large populations, mutants with multiple beneficial mutations often arise and spread in the population like a cohort (on the same background) (15, 231, 270, 298). Cohorts can occur when mutations arise and start spreading in a mutant background before this background has reached fixation, so-called nested fixation (270). Sometimes, multiple mutations arise seemingly at the same time in the same background. The latter, although possible, is very unlikely and probably just a side effect of our limited ability to detect single mutants at a very low frequency. Consequently, a substantial part of the time that it takes for the frequency of single mutants to increase remains undetected, and it is during this time that additional mutations can occur (231). Indeed, in the absence of recombination, mutations in a haplotype are completely linked and therefore can use one another as a piggyback ride uphill in the fitness landscape, a process called genetic hitchhiking. The frequent occurrence of sweeping cohorts instead of single mutants also seems to indicate that the combined effect of mutations is needed for an increase in frequency, to overcome drift or CI. Cohorts arising during adaptation of Saccharomyces cerevisiae in rich glucose medium, however, were recently shown to mostly contain only a single driver mutation per cohort (299), yet the cases where multiple beneficials drive cohort expansion might have profound effects, and the number of their occurrences and their consequences potentially also depend on the specific conditions (population size, mutation rates, and selective pressure). In cases where two or more beneficial mutations that generally do not genetically interact are involved (epistasis [see below]), the term “quasihitchhiking” has been used (270), and the mutants are generally called codrivers, as they aid each other in their fixation (231).

Strictly speaking, the term genetic hitchhiking is limited to the situation where neutral or deleterious mutations are being fixed. Natural selection would normally purge single deleterious mutations from the population, while neutral mutations could be lost or are expected to be fixed only very slowly in large populations by drift (292). However, during asexual reproduction and, thus, the complete linkage of the entire genome, neutral mutations have been shown to have an increased frequency and can become fixed in the population by hitchhiking with beneficial ones (38, 175, 255, 270). Hitchhiking of deleterious mutations has been observed less frequently (148, 300). While neutral mutations are often regarded as neutral passengers (15), hitchhiking of deleterious mutations is less likely, as they probably hinder the beneficial effect of the driver. Only under specific conditions where mutations interact with each other to change their effect on fitness can deleterious mutations arise along with a beneficial one, which masks the harmful effect. Interactions of this kind are extensively discussed in the section on epistasis below.

Sexual reproduction and recombination speed up evolutionary adaptation.Sexual reproduction is common among different taxa of organisms in nature. However, the possibility of having sex comes with a substantial cost. Only recently have people started to understand the reasons why sex is so pervasive (301). Over the years, many hypotheses were formulated to explain the emergence and maintenance of sexual reproduction, such as necessity, but generally speaking, all of them can be brought down to the theoretical prediction that the absence of recombination, through both CI and genetic hitchhiking of deleterious or neutral mutations, places a (speed) limit on adaptation (295). In the absence of sex, a beneficial mutation remains linked to a specific genetic background, and hence, the overall effect of that mutation largely depends on the fitness of that background (302). Sexual reproduction could offer advantages here, and CI has therefore been proposed as one possible driver for its maintenance (303). Indeed, sex unlinks mutations from their genetic background, thereby making natural selection on beneficial mutations more efficient and releasing the speed limit on adaptation in sexual populations (the “ruby in the rubbish effect”) (304–306). This hypothesis is backed up by data that show a greater strength of selection on new mutations when recombination was present in Drosophila melanogaster populations (302). In contrast, earlier data show increased fitness in the case of sexual reproduction in an environment to which the population was well adapted but not in a new environment in which adaptation was necessary. This observation supports the hypothesis that sex aids in purging deleterious mutations rather than increasing the efficiency of selection of beneficial mutations (307). Overall, it is generally accepted that sex increases the rate of adaptation, but the underlying evolutionary dynamics remain unclear.

New technologies have recently offered the possibility of addressing the long-standing question of why sex evolves and persists despite its high cost from a genomics point of view (301). The genetics of 6 sexually evolving yeast populations were compared to those of 12 asexually evolving populations. Both groups accumulated similar proportions of synonymous, nonsynonymous, and intergenic mutations. While all types of mutations were equally likely to be fixed in asexual populations, predominantly nonsynonymous mutations were fixed in the sexual populations. Thus, sex altered the molecular signatures of adaptation, suggesting that sex and recombination indeed improve the efficiency of selection to fix beneficial mutations and purge deleterious ones (301). These results were confirmed in asexual E. coli strains that were converted to high-frequency-recombination strains by the genomic incorporation of the F-plasmid conjugation machinery and the removal of surface exclusion factors so that bidirectional DNA exchange can occur (308). Recombination within the “genderless” E. coli population accelerated adaptation in evolution experiments in comparison to a recombination-deficient wild-type strain by alleviating clonal interference and combining beneficial mutations in one background (308–311). Previously, attempts to increase the rate of adaptation in evolved strains of the LTEE by importing new genetic variation through similar rounds of bacterial sex did not result in increased adaptation (312). In contrast to the work of the group of Kao, the purpose of sex was to import en masse variation from distant E. coli strains to speed up the process of adaptation. However, as became clear recently, this can overwhelm selection and lead to the fixation of donor alleles without a positive effect on recipient fitness (313). The combination of more variation by starting from standing variation rather than clones and sexual reproduction improved selection in S. cerevisiae populations, as it resulted in sustained gradual adaptation over long periods of time by breaking down linkage disequilibrium and transitioning from a selection of genotypes to one of beneficial alleles (314). In asexual populations, in contrast, adaptation was slowed down after the first sweep to fixation that purged all diversity on which natural selection could act.

Antagonistic or diminishing-returns epistasis.The outcome of a combination of mutations can be less extreme than expected (Fig. 3E). For example, two beneficial mutations can result in a double mutant with lower-than-expected fitness, or two deleterious mutations can have less of a negative effect than calculated based on their separate effects (without resulting in a respectively lower/higher effect than any of the mutations separately). This interaction is known as antagonistic epistasis and is the most abundant form of epistasis observed so far (for specific examples, see references 266, 290, and 307–310).

Apart from specific examples, antagonistic epistasis has been found to be a general theme among beneficial mutations in experimental evolution (240, 255, 259, 317–319). Being so pervasive, it is therefore also held responsible for the generally observed declining rate of fitness evolution during adaptation in a simple environment that is held constant. Under these conditions, the gradual fitness increase is strongest in the beginning but fades over time (reaching a predicted rate limit [243, 320]), bearing a sign of diminishing returns, and therefore, antagonistic epistasis has often been dubbed diminishing-returns epistasis.

Indeed, in a recent large-scale experiment with yeast, populations founded by maladapted derivatives of a general ancestor showed the most fitness progression over time and succeeded in catching up with populations that started from more-adapted founders (259). A correlation between initial fitness and adaptivity has been seen by others as well (Fig. 3E) (317, 321–323). The different adaptation rates mainly correlated with the fitness of the founder, not its genotype, and were the result not of a differential availability of possible beneficial mutations but of a global pattern of diminishing-returns epistasis among beneficial mutations (259). As such, the effect of the beneficial mutations decreased with increasing background fitness. The same is true for the LTEE, where the effect of each of the first few mutations that were fixed in the Ara-1 population proportionally declined with the increasing number of remaining mutations in (and, thus, fitness of) the background (240), which was recently extrapolated to be true when constructed in a more diverse set of natural isolates (324) or when the LTEE was studied over a longer time scale with replay experiments (325). At roughly the same time and in a similar way as the study described above (240), the prevalence of diminishing-returns epistasis between adaptive mutations and its effect on the rate of adaptation were also reported for the experimental evolution of M. extorquens populations that grew on methanol based on a native engineered pathway (318).

Synergistic epistasis.In synergistic epistasis, mutations tend to enhance each other's effects (Fig. 3B). Thus, a mutant with two beneficial mutations will have an improved fitness that is higher than expected based on the sum of the effects of the individual mutations. Just like sign epistasis, this form of genetic interaction is not very common among fixed mutations in experimental evolution (240, 259, 318).

During adaptation to glycerol, for example, E. coli almost always acquired at least two mutations, one in the RNA polymerase and another in glpK, the gene coding for glycerol kinase (254). Their frequent cooccurrence is explained by their synergistic interaction that improves metabolic efficiency and growth on glycerol (309, 326). A similar example can be found in the adaptation of E. coli to lactate, where the effect of an 82-bp deletion in rph-pyrE was increased in a variety of genotypes that already carried mutations adaptive to lactate limitations, which also explains why it is often present in independently evolved populations (258). In the LTEE, the benefit of a mutation in pykF in the Ara-1 population increases as the background acquires more beneficial mutations and thus becomes more adapted to the LTEE environment (240). Across multiple populations of the LTEE, for multiple beneficial mutations that were fixed in pykF during the LTEE and over the course of 50,000 generations, epistasis is, however, more complex and dynamic, combining different forms of genetic interactions (327). Also, in adaptation to isobutanol, some of the adaptive mutations in the endpoints interact synergistically (137, 139). While most of the cohort mutations arising in S. cerevisiae populations growing in glucose-rich medium contained only single driving mutations, one cohort could expand due to the synergistic interaction between two mutations located in previously unidentified targets of evolution under these conditions (299). A synergistic epistatic interaction detected during experimental evolution can also have profound medical consequences. The evolution of S. aureus toward high-level resistance to colistin, a last-resort antibiotic, was determined by strong synergistic interactions among 5 contributing mutations (328).

Although all of the above-described scenarios are examples of interactions between beneficial mutations, similarly, two deleterious mutations can combine into a double mutant with an even more negative effect than expected (Fig. 3B). Out of six mutations identified to result in trimethoprim resistance, two are actually deleterious in a background already containing the four other mutations (sign epistasis). Combining them into a sextuple mutant further aggravates their separate effects and results in a mutant with lower resistance than expected because of synergistic interactions between the two mutations (278). An MA experiment in yeast suggested generalized synergistic epistasis among deleterious mutations since the rate of fitness decrease increased in time (329). Although others have made similar observations (330), the opposite observation has also been made (61, 331, 332) (an example of antagonistic epistasis), and clearly, not all combinations of random deleterious mutations show synergism, making a generally synergistic epistatic interaction among deleterious mutations questionable (333–335).

Sign epistasis.An extreme example of genetic interaction is sign epistasis (Fig. 3C). In this case, the effect of a beneficial or deleterious mutation becomes negative or positive, respectively, when arising in an alternative background. Although this kind of interaction is not frequently encountered in fixed mutations in experimental evolution (240, 259, 318), the cases that show sign epistasis have a profound impact on the dynamics and outcomes of evolution.

In the Ara-1 populations of the LTEE, a mutation in ompF was found to have a deleterious effect when introduced into the ancestral strain (238). Still, this mutation was able to be fixed in these populations, a result of competition with other mutated alleles of ompF that contended for fixation (CI). So while deleterious on its own, the ompF mutation was in fact beneficial in a more adapted background (15). Similarly, the stable coexistence of two ecotypes in the LTEE required the fixation of two mutations (in arcA and gntR) that were neutral and deleterious, respectively, when introduced into the ancestral background (see “Sustained diversity from negative frequency-dependent selection,” below). Their fixation was possible only by sign epistasis because of earlier fixed mutations (336). In yeast strains adapting to glucose-limiting conditions, reciprocal sign epistasis between two adaptive mutations transformed them in deleterious ones and made them mutually exclusive. As a result, reciprocal sign epistasis explains the separation of two distinct peaks by a valley in a rugged fitness landscape (255). During adaptation to the fungicide nystatin, reciprocal sign epistasis was more widespread and affected one-third of the pairwise combinations of first-step mutations (337). Similar reciprocal sign epistasis between fixing mutations explains the rugged fitness landscape on which E. coli evolves under glucose-limiting conditions in a chemostat (315) and when adapting to increasing ethanol concentrations (338). An extreme example of reciprocal sign epistasis can be found when two beneficial mutations combine into a double mutant that is lethal. Synthetic lethality can also occur as a specific case of synergistic epistasis between two deleterious mutations or as all-or-none epistasis (see below) between two neutral mutations. Synthetic lethality is most often reported when constructing artificial double mutants (332, 339), and it is rarely detected between mutations identified in evolving microbes in the laboratory. Most reports come from evolution experiments with viruses, where synthetic lethality in tobacco etch virus (TEV) adapting to a new host resulted in holes in the fitness landscape (340) and genetic incompatibility between mutations leading to OmpF and LamB specialists in phage lambda resulted in the first step of speciation (341). As such, synthetic lethality could more generally also underlie the mutual exclusion of beneficial mutations (338).

The possible constraints of sign epistasis on evolution are very clear in β-lactam resistance conferred by five mutations in a certain β-lactamase. Although 120 theoretical paths to reach the quintuple mutant with the highest fitness are possible, only 18 of them are adaptive (342, 343). In contrast, the fitness landscape of evolution toward trimethoprim resistance revealed more-indirect paths when also considering high-order (sign) epistasis, taking into account that the kind of interaction between a pair of mutations can depend on the presence or absence of any combination of the additional mutations. Some adaptive paths resulted in a loss of a certain mutation, which delayed the commitment to a genotypic fate (278). So far, the effects of these high-order epistatic interactions on evolutionary dynamics have been understudied, but they have been shown to be equally as important as pairwise interactions in tobacco etch virus adaptation to new hosts as well (340). On a similar note, the cost of some antibiotic resistance-conferring mutations under drug-free conditions can be completely reverted by yet another resistance-conferring mutation or can be highly dependent on the genetic background in which it appears (344–346). Even without being specifically measured, sign epistatic interactions likely also play a role in the historical contingency of adaptation of resistant P. aeruginosa mutants to new antibiotics. Indeed, some initial adaptive trajectories to resistance to the first drug limited subsequent evolution of resistance to the second, while evolution of resistance to the second antibiotic can also revert the initial resistance (28).

A more abundant form of sign epistasis potentially remains hidden in evolution experiments. Different adaptive mutations in the same gene are hardly ever fixed in the same background, very likely because the presence of a first beneficial mutation makes subsequent adaptive mutations neutral or deleterious (110, 224, 319). Indeed, except for some notable examples of intergene sign epistasis (332), sign epistasis is believed to be more prevalent between adaptive mutations in the same gene (342, 347–349).

All-or-none epistasis at the basis of innovations.In evolution experiments in simple and constant environments, the fitness trajectory usually resembles one of gradual improvement, which is in line with the stepwise process of Darwinian theory (1) and probably results from widespread diminishing-returns epistasis, as discussed above. Nevertheless, more-sudden fitness improvements also occur, and these can be explained by all-or-none epistasis (Fig. 3D). Here, a combination of different mutations is necessary for the abrupt fitness increase, while, as long as one of the mutations is missing, the others have no effect on fitness. The accumulation of multiple strictly neutral mutations in the “right combination” is possible but unlikely, and for now, the strict formulation of this kind of interaction in terms of its fitness effect has, to our knowledge, not been observed.

All-or-none epistasis still remains a valid explanation for the emergence of some innovative phenotypes during evolution experiments. In this case, initial mutations (neutral or beneficial mutations) would potentiate the emergence of a driver mutation that actually creates the new phenotype. None of the potentiating mutations show any form of this phenotype prior to actualization, but without these mutations, the driver mutation would be futile, having no or little effect (on fitness and/or the innovative phenotype) and no chance to be fixed at all. Later, the new phenotype could be further optimized in a refinement period through additional mutations.

This process of potentiation-actualization-refinement was first observed when the Ara-3 population of the LTEE evolved the capacity to grow on citrate under oxic conditions after 31,500 generations (13), caused by a rare driver amplification that made a preexisting citrate transporter oxygen responsive because of a promoter capture event (241). During the first 15,000 generations, the population accumulated mutations that created the tendency to evolve a Cit⁺ phenotype (13), while additional mutations after its actualization further refined this key innovation (241). Since the actualizing mutation and one of its refinement mutations were sufficient for growth on citrate, the qualification as all-or-none epistasis was somewhat questioned (350), although potentiating periods where necessary mutations presumably accumulated have also been reported for the fast emergence of Cit⁺ phenotypes during direct selection schemes (351). It thus seems very likely that the potentiating mutations made the very rare driver mutation (more) beneficial, and thus, at least at some point in the evolution, the Cit⁺ phenotype was dependent on its history by all-or-none epistatic interactions (13). In the usage of another carbon source by E. coli, ethylene glycol, all-or-none epistatic interactions were present between 2 mutational events. Similar to the emergence of citrate utilization, the acquirement of the actualization mutation allowing adaptation to ethylene glycol was also historically contingent on a previous mutation, selected by adaptation toward propylene glycol usage (352). In a similar way, all-or-none epistasis was responsible for the de novo evolution of stochastic switching between two different phenotypes in P. fluorescens. Here, a total of nine mutations was found, and while the last was necessary and also sufficient to cause stochastic switching, the previous eight were necessary to set the stage by improving its fitness effect (25). In Salmonella enterica, new genes were shown to evolve but only after an all-or-none epistatic type of innovation in a preexisting gene. In medium lacking histidine and tryptophan and in a mutant missing a key enzyme in the production of tryptophan, TrpF, hisA needs to acquire two mutations to take over the TrpF activity while maintaining its native function in histidine production (353). Afterwards, by the processes of amplification and divergence, new genes with segregated functions arose (a process called subfunctionalization [354]). Another innovation by all-or-none epistasis arose during the coevolution of E. coli with phage lambda, where the phage evolved the capacity to use an alternative receptor, OmpF instead of LamB, to infect E. coli. A single mutation enabled the use of OmpF, but the effect of this mutation was dependent on the presence of three others, which improved attachment to the ancestral receptor LamB, and on the coevolution dynamics of the host toward resistance to the effect of those infectivity-improving mutations (355, 356). A human pathogen, avian influenza A/H5N1 virus, gained the capability of airborne transmission in the laboratory (23) by the combination of five mutations (357), some of which were already present in natural populations (358). Finally, multicellularity, a major transition in the history of life itself, evolved in yeast by the combination of several mutations that were all necessary. Here, the low sucrose concentration in the medium used made multicellularity a beneficial trait because of the obligate extracellular digestion of sucrose before it could be used as a nutrient (359).

In contrast, other transitions from unicellular life to a multicellular state in experimental evolution might not have emerged from all-or-none epistatic interactions (129, 360, 361), for example, when E. coli starts to aggregative strongly during the evolution of a passive mutualistic interaction involving the exchange of amino acids (362), in the origin of simple multicellularity in Chlamydomonas reinhardtii to escape predation (363), or in a setup favoring faster settling (364). In S. cerevisiae genotypes that show fast sedimentation in clusters, only one frameshift mutation was sufficient to cause the multicellular trait (365). Other innovations are also less dependent on the collaboration between different mutations and do not rely on an all-or-none interaction exhibited in the above-described examples. Superior social cooperation, for example, can evolve in a population of obligate cheaters in the spore-forming bacterium Myxococcus xanthus by acquiring just a single mutation (253, 366), and in the same bacterium, de novo kin discrimination evolved gradually over time by multiple mutations or all of a sudden but probably by a single mutation (367). Likewise, flagellar locomotion could be fully resurrected in a P. fluorescens strain made immotile due to a deletion of the main regulator of flagellar synthesis, by only two mutations that each gradually increased motility (368).

Second-order selection.As stated above and because of CI, an individual that is unfit at present could outperform the currently fittest individual if it can acquire additional beneficial mutations and hence become more fit. As a result, selection in evolution experiments not only acts on direct fitness but also can contain an extra layer, i.e., second-order selection for more-evolvable individuals (Fig. 4A).

In the Ara-1 population of the LTEE, for example, second-order selection for greater evolutionary potential selected eventual winner genotypes that were, at an intermediate time point, less fit than eventual losers but prevailed nevertheless because of more-beneficial epistatic interactions with subsequent mutations (369). Similarly, in E. coli evolving in a chemostat under glucose limitation, a duplication was fixed and remained stable, not because of its own effect on fitness but probably because of its higher evolutionary potential (370).

Second-order selection could also be an explanation for the frequent emergence of hypermutators in experimental evolution and natural populations or for the existence of a stress-inducible form of mutagenesis (SIM) (38, 67, 113, 128, 371, 372). The increased mutation rate would result in a higher probability of producing a rare beneficial mutation or a combination of multiple ones, which is likely important under complex selection pressures. Subsequently, the mutator or SIM genotype could hitchhike along with the beneficial mutation. Despite the direct risk of increasing the genetic load and decreasing fitness by the higher statistical chances for the mutated offspring to accumulate deleterious mutations rather than hitting a beneficial one, such an increase in evolvability can still be indirectly beneficial to a population (197, 373). A mutator in Candida glabrata, for example, has been shown to accelerate the evolution of caspofungin resistance in a mouse infection model (374). Likewise, the emergence of mutators was key in the adaptation of E. coli toward high levels of ethanol tolerance, where they reoccurred during each adaptive step but were compensated for as soon as adaptation was reached (65). Similarly, a transient state of hypermutation was necessary to accelerate the adaptation of Ralstonia solanacearum to its new endosymbiotic role in the rhizosphere of Mimosa pudica (375). Although it is suggested that this is especially the case under stressful conditions that require quick adaptation or difficult (combinations of) mutations (65, 128, 371, 376), an increased mutation frequency has also been found in the simplest evolution experiments, like the LTEE (228). Here, mutators are likely beneficial as long as the population is not overly well adapted so that they have great potential to hitchhike to fixation on the beneficial mutations that mutators generate at an increased rate. After the Cit⁺ phenotype evolved and there was great potential for improvement, mutators emerged strikingly quickly (64, 197, 241, 377). The second-order selection rationale remains controversial, as the reasons for the existence of SIM or the emergence of mutators are far more complex, and other explanations exist as well, e.g., drift, the pleiotropic by-product of adaptation, or a shifting fitness landscape under altered environments that results in a DFE more biased toward beneficial mutations (378–380). More-in-depth overviews can be found elsewhere (380–382).

Suboptimal peaks and selection of the flattest.Evolution by natural selection does not necessarily lead to the highest peak in the fitness landscape. In rugged landscapes, for example, an adaptive walk might “miss a turn” and get stuck at a suboptimal peak surrounded by valleys of maladapted genotypes that are impossible to cross (Fig. 4B) (383, 384). So far, only very limited empirical data exist to show replicate populations trapped at such evolutionary dead ends (65, 255).

Alternatively, suboptimal peaks surrounded by less-maladapted variants can be preferred over more-optimal solutions that have highly maladapted variants in their direct neighborhood as a kind of genetic robustness and “selection of the flattest” (Fig. 4C) (385, 386). Especially when mutation rates are high in large populations, as shown in competition experiments with RNA viruses, the genotype with the highest fitness wins under normal conditions but loses when the mutation rate is artificially increased (387).

Nontransitive fitness and the Penrose staircase.Relative competitive fitness can be nontransitive, e.g., when an endpoint outcompetes its immediate predecessor but has a lower fitness than that at an earlier time point. As such, fitness landscapes might not contain an absolute peak (14, 388), and it is thus possible that the fitness landscape looks more like the illusion of the Penrose staircase, where an immediate successor is always more fit than its predecessor, just like steps in the Penrose staircase that seem to perpetually ascend in a continuous loop (Fig. 4D). Nevertheless, the ancestor can outcompete, or at least be equally as competitive as, more-distant successors, so it appears as if evolution went full circle, in analogy to climbing the Penrose stairs, where you end up at the place where you started, without getting higher.

Nontransitiveness of fitness has been observed in certain evolution experiments and indeed even caused an apparent loss of fitness compared to that at the starting point (25, 218), but it is far from a general rule. For example, it was not observed over the long time scale of the LTEE (389), and fitness is likely nontransitive only if special social interactions are present (316). For example, consider the artificial system with three E. coli strains where one produces toxin, the second is toxin resistant, and the third is neither resistant nor a producer. Since the second tops the first, the third outperforms the second, and the first outcompetes the third, competitive fitness is clearly nontransitive and could result in a rock-paper-scissors game (Fig. 4D) (390). In a similar way, loner strains of P. aeruginosa stabilize the community of cooperators and defectors in a social mechanism to take up iron by inducing rock-paper-scissors dynamics. Also, in experimental coevolution between P. fluorescens and its parasitic phage SBW25Φ6, after an initial period of arms race dynamics (ARD), evolution becomes less directional and is dominated by fluctuating selection dynamics (FSD). During the initial period of ARD, fitness increases progressively by increases in mean infectivity and resistance in phage and host, respectively, in response to each other. Under FSD, however, alternative hosts and parasites are continuously selected based not on continuously increasing fitness anymore but on differences in specific infection or resistance mechanisms (172, 391).

Cost of generalism.As foretold by the phrase “Jack-of-all-trades, master of none,” the evolution of generalism can be constrained since trade-offs can impair the extent or speed of evolution. An alternative explanation for the constrained evolution of generalists is the shorter time that each separate condition is experienced when selecting for generalists by changing environments, which should be accounted for when comparing results of evolution under changing and constant environments. The resulting lower fitness of the generalist under each separate condition than of the condition-specific specialist has been called the cost of generalism (Fig. 5A) (157).

This was clearly the case during evolution alternating between high and low oxygen concentrations (107), temperature extremes (159, 414, 415), antibiotics (130, 416), or hosts in the case of viruses (126, 165). The most extreme example is when environments fluctuate randomly between different kinds of selection pressures. The fitness of E. coli populations under these fluctuating conditions improved only marginally under the separate acid, osmotic, and oxidative stress conditions that were used (similar to experimental evolution under random temperatures chosen between two extremes [189]). Instead, their performance improved when challenged by novel, unexpected stresses (134). While extending the original evolution experiment in time and to more replicates showed improvements in the mean fitness under the separate stress conditions, it foremost resulted in a total loss of trade-off costs and decreased variance in fitness across environments, showing signs of the appearance of true generalists (417).

Still, in all these cases, the “Jack-of-all-trades, master of none” is selected over specialists, as it minimizes fitness fluctuations over time (as in the extended phrase, “Jack-of all-trades, master of none, often times better than a master of one”) by adapting not only to all the separate conditions but also to fluctuations as such. As a result, measurements made with the separate, constant components were shown to be bad predictors of the level of adaptation to the fluctuating environment itself (160).

Cost-free and superior generalism.Since side effects are not universal and the simple accumulation of mutations might be the underlying factor for trade-offs, cost-free generalists could emerge during adaptation to a fluctuating environment. Cost-free generalism would imply that fitness under each separate condition is comparable to the fitness of specialists that evolved under separate constant conditions (“Jack-of-all-trades, master of at least some”). Indeed, such cost-free generalists have often been observed (155), for example, in the evolution of the photosynthetic Chlamydomonas species in cycles between dark and light (162, 163), in some reports on viruses that alternated between different host cells (152, 167), or in the adaptation of E. coli to alternations between acid and alkaline conditions (158). Generalists are sometimes superior to the condition-specific specialists (“Jack-of-all-trades, master of all”), as observed in the evolution of vesicular stomatitis virus to fluctuating temperatures (189) or when chikungunya virus adapted to alternating hosts (166).

A recent report on evolution in the presence of fluctuating levels of glucose or lactose, however, showed that although an initial period of adaptation might be one of cost-free generalism, at longer time scales, trade-offs that constrain either the speed or the extent of evolution become inevitable (182). The authors of that report explained the somewhat counterintuitiveness of cost-free and superior generalism by the small number of generations during many evolution experiments. In contrast, pleiotropy in epistatic interactions could lift certain evolutionary constraints and open up superordinate paths that are inaccessible to adaptation in both constant environments, which would be another explanation for cost-free generalism.

Evolution of the bet-hedger, the specialist-generalist, and the importance of clonal phenotypic heterogeneity in evolution.Nature has an alternative answer to fluctuating environments besides evolving generalists or specialists, which could be regarded as genotypes that simply sense and respond, at a population-wide level, to changes in the environment. We then consider a generalist to be a genotype that can sense and respond to all condition shifts, while specialists are responsive to only one (418). Bet-hedging is a kind of phenotypic multistability in an isogenic population where the individuals present one of different phenotypes, each specialized for a certain condition but often less fit for other conditions (419). This heterogeneity by the combination of phenotypic specialists can be seen as a special form of generalism (i.e., a combination of many “Jacks-of-one-trade”). It might decrease the average fitness of the population because of the presence of suboptimal variants but instead can minimize the variance of fitness over time; i.e., it can maximize the geometric mean fitness or long-term fitness of a population in changing environments and partially avoid the cost of generalism that generalists could suffer (Fig. 5B) (420). The geometric mean is indeed a better measure for long-term success than the arithmetic mean, intuitively, as selection over many generations is a multiplicative process (384, 419), but also theoretically, since fitness over time under changing conditions is not an independent linear combination of each separate fitness but rather the result of an interacting combination. For example, when part of the population is lost under one condition, this part cannot generate offspring under the next, beneficial condition.

Bet-hedging is predicted to be advantageous over specialism or normal generalism involving sense-and-response systems under some conditions where environments change infrequently (421, 422), changes are hard to predict and to sense (422, 423), and/or changes are extreme (strong trade-offs) (423, 424). The configuration of bet-hedging (i.e., the number of different phenotypes and the rate of switching between them) needs to maximize the geometric mean fitness and therefore should be proportional to the environmental dynamics, e.g., the frequency of each condition, the rate of alternation between them, and the fitness of each variant in each of them (Fig. 5B) (421–423, 425, 426).

Although bet-hedging is widespread in nature and many observations seem to support the theoretical predictions listed above (425, 427–431), reports of experimental evolution dealing with bet-hedging behavior remain scarce. Periodic selection for competence in Bacillus subtilis maintained bet-hedging (432), while evolution in a constant environment lacking selection for sporulation led to the loss of bet-hedging behavior (433). The latter was because of mutation accumulation rather than selection, but the de novo evolution of a bet-hedging phenotype in P. fluorescens under conditions alternating between shaking and nonshaking environments provided stronger proof for the adaptive nature of bet-hedging (25). Under each of these conditions, genotypes with different colony morphotypes usually evolve, but by imposing bottlenecks and exclusion rules, i.e., transferring only a single colony of a novel morphotype as soon as it arose, to the alternative environment, a genotype that phenotypically switched between morphotypes emerged (25). These bottlenecks and exclusion rules are analogous to sudden and extreme environmental changes: the fitness of the ancestral morphotype is arbitrarily set to a value of zero, and the cost of generalism by bet-hedging is held low by limiting competition. Moreover, they were an absolute prerequisite for the evolution of bet-hedging in this case and likely operate in many natural environments (424). The configuration of bet-hedging was also shown to experimentally evolve according to the environmental dynamics. Dormancy levels in the fungus Neurospora crassa, for example, were shown to evolve according to the frequency of bad years (434). Likewise, the number of phenotypic antibiotic-tolerant cells in a population of many bacterial species can be increased to extremely high levels by frequent antibiotic treatments (26, 27, 164, 435). The adaptive process under these conditions seems to be even more specific, as the rate of switching of tolerant E. coli cells back to the normal cell type was shown to evolve according to the treatment duration (26), while the rate of switching of normal cells to antibiotic-tolerant persisters evolved according to the imposed treatment frequency (27). The evolution of antibiotic bet-hedging strategies can have profound consequences on future evolution dynamics, impacting our medical care. Indeed, these persisters have been shown to be the first stepping stone to the evolution of resistance (436), also in evolution experiments with tumor populations (437, 438). Under some conditions, it seems that the initial evolutionary response to strengthen tolerance proceeds and opens the paths to the evolution of resistance development and thereby accelerates it (439). Here, this increased evolvability likely is a combined result of a game of numbers (more survival and larger populations) and increased mutation rates, although epistatic interactions cannot be excluded. Similarly, clonal heterogeneity and stochastic expression of antifungal resistance genes increased the evolutionary potential with gradually increasing concentrations of the drug (440). This increased potential likely stems from a modulation of the fitness benefit of beneficial mutations. The single-cell lag times of switching between carbon sources in yeast have also been shown to be heterogeneous in nature and evolvable in the laboratory as a bet-hedging strategy under a frequently changing environment (441). Another recent study in yeast evolved under mixed glucose-galactose conditions resulted in the de novo emergence of clonal phenotypic heterogeneity resembling a bet-hedging strategy. Here, however, the outcome of populational diversity was likely a response to negative frequency-dependent interactions rather than a bet-hedging strategy (442).

Parallel evolution at different levels.Replica lines in experimental evolution often show similar evolutionary responses and, hence, limited interpopulational diversity. A high degree of parallelism is usually found at the level at which natural selection acts, namely, fitness in a given environment (62, 189). In addition, parallel changes are also present across lower levels of the biological organization, like phenotypic traits that might contribute to the change in fitness (66). For example, M. extorquens populations all adapted to a newly introduced pathway for growth on methanol in the same phenotypic way, by decreasing the costly overexpression of the exogenous pathway (443) and by changing their entire transcriptome in a synonymous way (97). Also, in the LTEE, gene expression changes were found to be highly similar (224). Similarly, populations evolving resistance to a certain kind of antibiotic often show similar increases in resistance, and even the changes in patterns of antibiotic susceptibility to other, not-selected-for antibiotics have been shown to be highly conserved (85, 114, 130, 131).

Indeed, traits that are only indirectly correlated with a fitness increase can also change in a parallel way between replicate populations (232). The pattern of resistance to different phages, a trait under relaxed selection, changed, for example, similarly in all populations of the LTEE throughout 45,000 generations (444). Pathways and processes underlying the fitness or phenotypic changes also show frequent parallel optimization (110). In the emergence of doxycycline or chloramphenicol resistance, transcription and translation processes are always changed along with membrane transport systems (85). Similarly, mutations or mutated genes in 640 S. cerevisiae populations adapting at a large population size showed only a low level of interpopulational parallelism. Grouping the mutated genes by biological functions, however, increased the level of parallelism between the lines and showed clear convergent evolution (259). Even down to the level of operons or genes, parallelism is frequently observed in the sense that genes are often found to be hit multiple times across replicate populations (37, 98, 259, 262, 270), and while this can be the result of mutational hot spots (273), detailed analyses of high-throughput whole-population sequencing time series showed that in the case of the LTEE, this is unlikely (260). In extreme examples, unique target genes seem to exist, which are sometimes hit repeatedly by identical mutations in parallel evolving populations (94, 175, 189, 293). Under selection for trimethoprim resistance, dihydrofolate reductase was the single target (85). Such a high level of parallelism is also known for certain other antibiotics (113, 130).

Explanations for parallelism or diversity.These forms of parallelism are usually signs of adaptive responses and have helped distinguish causal mutations from irrelevant ones. However, moving further away from the direct target of natural selection (e.g., looking at phenotypic traits that are less fitness correlated or at correlated fitness in other environments [111, 153, 445, 446]) and down to lower levels of biological organization, eventually to the DNA base pair level, parallelism decreases, and the observed diversity increases. The conceptual logic is that a similar increase in fitness can be caused by the optimization of different traits via diverse genetic networks or processes, each containing multiple possible target genes that can be changed in an adaptive way at various sites, possibly by different mutations per site (110, 447, 448). An extreme example is that genes, albeit being hit repeatedly, are not often hit twice or more by exactly the same mutations over different replicate populations (132, 237, 249), although exceptions exist (114, 175, 259, 270).

Differences between replicate lines could simply reflect differences in speeds or paths taken to climb the same peak in the fitness landscape. Alternatively, populations could indeed be climbing different peaks, leading to more-pronounced and sustained diversity over time. In both cases, diversity between replicate populations will be higher if more beneficial mutations are available (wider peaks and more paths) or if more alternative and mutually exclusive optimization solutions exist (more peaks and a rugged landscape) (74, 98, 448). Diversity should therefore increase if evolution experiments are carried out with smaller replicate populations, which would lead to a larger effect of chance and faster spreading of mutations (less CI), a better exploration of the genotype space in the fitness landscape among replicate populations, and a higher chance of ending up on different peaks (449, 450). Extreme examples are MA experiments where interpopulational diversity increases during the experiment, although natural selection is, of course, very limited, and fitness usually decreases (61, 329). However, the effect of population bottleneck sizes can be more complex, as in the adaptation of P. fluorescens toward rifampin resistance, where intense or weak bottlenecking resulted in a higher diversity of genetic resistance mechanisms than in the intermediate situation (451). Similarly, the kind of selection and its strength can influence diversity. Just as some stressful conditions might have more possible adaptive paths than others, increasing selection might decrease the number of accessible solutions and has been shown to limit diversity (132, 143, 148, 191).

Clonal interference and soft sweeps.As stated above, CI in large populations can slow down the action of natural selection (see “Clonal interference,” above). As such, hard sweeps of beneficial mutants on their way to fixation are slowed down, and simultaneously, the diversity is purged more slowly. In cases with extremely strong CI (e.g., very large populations, high mutation rates, or a large target size for adaptive evolution), selection will not be able to keep up with purging diversity, and hard sweeps will be rare. As a consequence, different adaptive mutations will arise repeatedly in different haplotypes without reaching fixation, also called soft sweeps (Fig. 6B) (452, 453).

The domination of evolutionary dynamics by a succession of soft sweeps, and thus (temporarily) preserved diversity, is frequently seen in the large populations of chemostats that do not experience population bottlenecks like the serial transfer regime does (36, 37, 42, 73, 270). Even during the adaptation of E. coli to the mouse gut, a chemostat-like in vivo system, high genetic diversity was created by many soft genetic sweeps that, on the phenotypic level, however, resulted in a hard sweep (199). Also, during serial transfer, strong CI can result in high diversity because of soft genetic sweeps (247, 248, 293), which was also observed in the LTEE (231). In extreme cases, in a fluctuating environment with strong trade-offs where conditions change too rapidly to allow for a complete takeover of the population by the current adaptive specialist, ultimate soft sweeps can occur, which act on standing genetic variation and thus cause changes only in the frequencies of specialists that are already present in the population. Standing genetic variation could initially be generated before environmental changes occur by de novo mutations. Alternatively, these mutants could already be present in the culture before the start of the evolution experiment. In both cases, this standing genetic variation has been shown to be more important in the beginning of an adaptive trajectory than de novo mutations (154, 412, 454, 455). A recent experiment on yeast adapting to antifungals showed that de novo mutations, together with standing variation, can influence the future adaptive trajectory but that standing variation clearly imposes a selective threshold for these new arising mutations (456). As such, temporally fluctuating environments have been shown to increase and/or sustain diversity (155, 457, 458), although spatial heterogeneity in the environment is still regarded as the most important factor for diversity to arise (see “Mixed environments with heterogeneity in niches,” below).

Sustained diversity from negative frequency-dependent selection.Diversity generated by strong CI, without any form of interaction, will ultimately be lost once the genotype on the highest point of the landscape sweeps to fixation. A more sustained form of diversity in a homogeneous environment arises from negative frequency-dependent selection (NFDS). Here, the fitness of a genotype is inversely correlated with its frequency in the population. Such a pattern predicts some kind of social interaction and helps to explain the high intraspecific diversity in large populations in nature, for example, in bloom-forming plankton (459).

NFDS is present in evolution experiments with communities of multiple interacting species, such as coexisting mutualistic partners or prey and predator. Here, the success of a predator or one of the mutualistic partners is high when their numbers are low compared to the numbers of prey or other partners. Combined with coevolution, in the form of ARD or FSD for the predator-prey system, this can result in a stable coexistence of the interacting partners. In fact, FSD contains another form of NFDS since predator types that target the dominant prey genotype will be selected, and this will result in a higher fitness for alternative prey genotypes that are present at a low frequency (172, 174, 176, 177, 460).

Sustained diversity because of NFDS can also evolve under well-mixed conditions limited by a single carbon source and started with only one founding genotype (73, 461). A classic example under these conditions is diversification into so-called cross-feeders, where one genotype becomes specialized in metabolizing the single carbon source and others adapt to metabolize the products generated by the first. For example, during the adaptation of E. coli to glucose limitation in chemostats, genotypes that benefit from glycerol and acetate excreted from the glucose specialist emerged, which enabled their coexistence by NFDS (46, 462, 463). Similar, strong NFDS based on cross-feeding was found in the Ara-2 population of the LTEE, with the sustained coexistence of two genotypes, L and S (for “large” and “small” based on their colony morphologies), for over 12,000 generations (229, 464). Although this situation is more complex than just simple cross-feeding on excretions and also depends on the death rates in stationary phase and cannibalism of one type on the other (465), the fitness of S or L was shown to be simply inversely proportional to its frequency in the population (466). These cross-feeders can further adapt as interacting but separate sublineages (comparable to the coevolution of two separate species) (467, 468), which causes their stable levels to fluctuate slightly over time. The fact that their community can be disassembled and reassembled because of their cooperative interactions nevertheless shows that genetic diversity is sustained by clonal reinforcement rather than by CI (469). Massive population sequencing in all the lines of the LTEE recently showed that NFDS likely underlies pervasive quasistable, long-term coexistences in many of the populations throughout 60,000 generations (260). While CI likely also explains part of the observed diversity in the LTEE, only NFDS can explain why coexisting clades emerged in which subsequent mutations were fixed.

If selection leads to the loss of costly and leaky but essential functions in part of the population, the Black Queen hypothesis (BQH) (referring to the Queen of Spades, a playing card just like the Red Queen from the RQH, which is to be avoided in the game Hearts) predicts a yet stronger form of coexistence by NFDS (470). While part of the population benefits from the loss of the Black Queen function (the beneficiaries), the remainder of the population is stuck with the function and is obliged to carry it out (the helpers), and since neither them can exclude the other (NFDS, since the function is essential), this leads to some sort of commensalism or mutualism (471) (a distinction between altruists and helpers on the one side and cheaters and beneficiaries on the other, with an overview of many possible Black Queen functions, can be found in reference 472). The BQH was initially used to explain adaptive gene loss in nature resulting in dependence on other cooccurring organisms, as for Prochlorococcus, an oxygen-producing photoautotroph in the ocean that lost many mechanisms for detoxifying molecules, leading to oxidative stress, and became dependent on other members in its community (471). Recently, however, the BQH resulted in the stable intraspecies coexistence of diversity in the laboratory. In an environment containing hydrogen peroxide, the founder E. coli population split up into two subpopulations: one genotype lost the costly detoxification process, while the other maintained the leaky Black Queen function (473). A similar adaptive loss of a Black Queen function (the assimilation of vitamin B₁₂) was also recently observed to cause a stable coexistence between a beneficiary alga that lost the capacity for vitamin B₁₂ production on the one hand and a helper bacterium and the wild-type alga that both rescued B₁₂ production on the other (474).

Mixed environments with heterogeneity in niches.While designed as a single-niche-selective environment, the long-term evolutionary process of the LTEE showed that it was in fact carried out in a multiniche environment, eventually leading to the stable coexistence of multiple specialists (241). In particular, the LTEE environment contains citrate as a potential carbon source besides glucose. The Cit⁺ phenotype that evolved is an extreme example of a niche specialist that coexisted with the Cit⁻ glucose specialist, which actually also evolved cross-feeding behavior on excreted C₄ metabolites of the Cit⁺ genotype, in further support of their coexistence (476).

Other experiments intentionally started with an environment with multiple niches to induce diversity (155, 161, 181, 477). A mixture of acetate and glucose resulted, for example, in the emergence and NFDS-based coexistence of two E. coli genotypes (180, 447), both of which were specialized either to switch to acetate after glucose depletion or to use glucose (250, 478). The latter simultaneously also excretes acetate, which is believed to have helped in the development of the acetate specialist (178, 479). Apart from different types of nutrients, diversification is also observed in the evolution of P. fluorescens simultaneously challenged by different types of predators, resulting in defense specialists (173). Diversification of a unique predator in multiple specialized predators also seems to have happened in a system that started with only one predator and one host after the host diversified under coevolving conditions (171). In the evolution of phage lambda simultaneously presented to a mixture of two hosts, both expressing different receptors for lambda infection routes, the initial generalist rapidly diversified into specialists, even to the extent that the two specialists showed the first signs of speciation (341).

Sometimes, for example, in the combination of glucose with either lactose or maltose or in complex undefined media altogether, evolution did not result in diversification and stable coexistence (155, 181, 182, 194, 477). While this could be due to the limited evolutionary time given in these experiments, another explanation is that the niches are insufficiently distinct from one another (i.e., too few trade-offs exist between specialists), and specialization to either of them does not allow for the avoidance of competition and, thus, the stable coexistence of multiple specialists (181) (see Selection for Specialists or Generalists?, above).

Spatially structured environments.In addition to heterogeneous but well-mixed environments, the presence of structure in the environment is also believed to be important for the evolution of diversity. Structured environments are nowadays frequently used in experimental evolution, and they can induce heterogeneity and thus cause potentially different niches to arise. Seminal experiments in this regard are those on the radiation of P. fluorescens. In one such experiment, 3 different morphotypes readily evolved in static broth, each adapted to a different part of the culture (the smooth morph is specialized to the bulk, the fuzzy spreader is specialized to the bottom, and the wrinkly spreader is specialized to the air-liquid interface) and stably coexisted by NFDS (98, 183). Although oxygen availability clearly is an important aspect of each niche, radiation into phenotypes can also be evoked in well-mixed environments where cultures are simultaneously challenged with predators (125), while the presence of predators in static cultures can, in fact, decrease diversity (480). Furthermore, the propensity to radiate is highly dependent on niche position and width of the ancestor. A very confined initial niche, or specialist ancestor, would allow for diversity to arise and endure, while ancestors with less-confined niches that partially overlap the potential new variants have shown less-adaptive radiation behavior (481). Similar adaptive radiation was also observed during the evolution of B. cenocepacia in biofilms, structures notorious for their medical importance and known to harbor many different niches (187, 262, 482). In P. aeruginosa populations evolving under a regime that optimizes dispersal, attachment, and biofilm formation, diversification was already apparent from the many different colony morphologies (483). One of the causes, increased diversity in the cyclic di-GMP content, enabled variants with high levels, as early colonists, to make way for increased biofilm production by the entire population. In the evolution of vesicular stomatitis virus, while the simultaneous presence of different hosts selected for generalists rather than multiple specialists, diversity in generalists within populations was also increased in the heterogeneous structured environments (484).

Even one of the simplest setups possible that still includes structure, the growth of populations on agar plates with no clear niches at the start, resulted in more-diverse populations than in the situation under mixed-liquid conditions. Another experiment using P. fluorescens showed that in simple expanding colonies, selection for increased territorial expansion already occurs. Within 10 days, de novo social behavior develops by the emergence and coexistence of a genotype that pushes the lubricant-producing ancestor forward (257). Additionally, populations on a plate can also be regarded as many small populations that evolved semi-independently in parallel, which can also lead to more diversity (43, 184). A more dynamic form of diversity in a rock-paper-scissors-like game among three E. coli strains, a toxin producer, a toxin-resistant mutant, and a mutant that was neither a producer nor resistant (see “Nontransitive fitness and the Penrose staircase,” above), could be maintained only if played in a structured environment where boundaries to diffusion and migration exist (390). Likewise, the commensal interaction between Acinetobacter excreting benzoate when growing on benzyl alcohol and Pseudomonas putida metabolizing this benzoate benefits from a structured environment in the form of a biofilm (176). Another cooperative behavior between S. enterica serovar Typhimurium and E. coli, where the former evolves increased methionine excretion to serve the latter, from which it receives carbon sources, could emerge only under a structured environment of a plate where the benefits of the cooperation are predominantly received by cooperators (485). In these structured circumstances, positive frequency-dependent selection (PFDS) can promote and sustain diversity. As such, natural isolates of M. xanthus showed increased fitness when abundant, which acts as a form of territorial behavior thwarting invasion but promoting diversity when genotypes vary patchily in structured environments (486). In the same species, structured environments were recently also shown to increase diversity between populations adapting to an environment selective for the increased expansion of swarming colonies (487). To study adaptation of large populations, microbial evolution and growth arena (MEGA) plates were designed (488). The MEGA plate incorporates, in addition to spatial structure, a sense of directionality, as selective pressure is varied not in time but in a spatial sequence (Fig. 7). When applying this method to adaptation to antibiotic pressure, resistance increased rapidly, by 5 orders of magnitude. More importantly, Kishony's group visualized and sampled the adaptive routes and their massive diversity that evolved over time and in space (488). By doing so, they also showed that the fitness of a mutant is a complex trait that cannot simply be reduced to the level of resistance but actually can also depend on the time and place of its emergence and its capacity to be mobile, etc. As a consequence, some highly resistant mutants became trapped in space behind more-sensitive lines and thereby did not contribute to the population's ultimate evolutionary path, an effect that likely occurs under many natural conditions as well (Fig. 7).

• Open in new tab
• Download powerpoint

FIG 7

Evolution in a spatially structured environment. Shown is a MEGA (microbial evolution and growth arena) plate setup as a prime example of evolution in a spatially structured environment (488). The plate consists of adjacent regions containing stepwise increased concentrations of a selective agent. The wild-type strain is inoculated at the left side and rapidly grows in the region with the lowest concentration of the selective agent. Growth stops at the intersection with the region containing a higher concentration of the selective agent. The subsequent occurrence of different adaptive mutations allows growth in this region until it reaches the next boundary. This setup enables visualization of evolution in action, demonstrates the power of mutations for adaptation under stressful conditions, and allows for the isolation and determination of individual mutants that arise. This setup also shows the impact of a spatially structured environment on the outcome of an evolution experiment. The red mutant (marked with *) becomes spatially trapped due to fast growth of the light blue and dark green mutants. (Based on data from reference 488.)

Although structure in the environment clearly often promotes diversity, exceptions exist (448). For example, for the evolution of sustained diversity by cross-feeding in E. coli, structure has been shown to have negative impacts, probably because of reduced diffusion of metabolites (489). Likewise, if a multispecies community consists of only a toxin producer and a sensitive genotype, structure promotes invasion and take-over by the toxin producer, even at very low initial frequencies (490). Clearly, the specific kinds of ecological mechanisms and interactions maintaining diversity can impact the effect that environmental structure has.