Selection experiments and the study of phenotypic plasticity1

Authors


  • 1

    This article is the material of the US Government, and can be produced by the public at will.

 Samuel M. Scheiner, Division of Environmental Biology Rm 635, National Science Foundation, 4201 Wilson Blvd., Arlington, VA 22230, USA. Tel.: 703 292 7175; fax: 703 292 9065; e-mail: sscheine@nsf.gov

Abstract

Abstract Laboratory selection experiments are powerful tools for establishing evolutionary potentials. Such experiments provide two types of information, knowledge about genetic architecture and insight into evolutionary dynamics. They can be roughly classified into two types: (1) artificial selection in which the experimenter selects on a focal trait or trait index, and (2) quasi-natural selection in which the experimenter establishes a set of environmental conditions and then allows the population to evolve. Both approaches have been used in the study of phenotypic plasticity. Artificial selection experiments have taken various forms including: selection directly on a reaction norm, selection on a trait in multiple environments, and selection on a trait in a single environment. In the latter experiments, evolution of phenotypic plasticity is investigated as a correlated response. Quasi-natural selection experiments have examined the effects of both spatial and temporal variation. I describe how to carry out such experiments, summarize past efforts, and suggest further avenues of research.

The goals of selection experiments

Selection experiments are powerful tools for evolutionary biologists. By manipulating the process of natural selection in a controlled way we can obtain a better understanding of the evolutionary process. For example, evidence from artificial selection was central to Darwin's arguments in The Origin of Species. In this review I explore the use of selection experiments in the elucidation of phenotypic plasticity and its evolution.

Selection experiments provide two types of information. First, they are probes into the genetic architecture of organisms. Today we have a range of molecular techniques for exploring minute segments of a genome. Selection experiments, however, have the advantage of affecting large parts of a genome simultaneously. Secondly, selection experiments are mimics of natural selection. They allow us to observe evolutionary dynamics under controlled conditions. We can perform such experiments on organisms whose generation time is short enough that substantial evolution can be observed within the time span of a typical grant or dissertation.

Evolutionary models are built on a large tower of assumptions. Many of those assumptions deal with the genetic basis of traits. Is the trait controlled by one locus or many loci? Are allelic effects discrete or continuous? Are allelic effects additive, or is there dominance, or is there epistasis? Is standing genetic variation sufficient to allow a long-term evolutionary response? Is evolution constrained by indirect effects of pleiotropy and linkage? In the study of the evolution of plasticity, disagreements over genetic architecture have engendered lively, extensive, and sometimes vociferous debates (Via et al., 1995 and references therein). One method for settling these disagreements is the selection experiment. Sometimes alternative genetic architectures provide clear alternative predictions about the outcome of selection. I give details of an example below.

In this review I describe and explore the variety of goals and methods of selection experiments. First, I provide an overview of two types of selection experiment, artificial and quasi-natural. Secondly, I explore each in detail, provide examples of their use, and highlight some major findings. Finally, I look at some future directions in the use of selection experiments in this field. Selection experiments are a large topic and could be the subject of an entire book (e.g. Bell, 1997a). Thus, I do not attempt to provide an indepth review. Excellent general references on selection experiments can be found in Falconer & Mackay (1996) and Lynch & Walsh (1997). Instead I focus on their use in the study of phenotypic plasticity and the pivotal role that they have played in our current understanding.

An overview of the two types of selection experiments

Selection experiments represent a continuum. For convenience I have separated this continuum into two categories, artificial and quasi-natural. At one end of the continuum, artificial selection experiments manipulate selection in precise ways to study alternative evolutionary models. At the other end of the continuum are quasi-natural selection experiments designed to be as nearly natural as possible.

The distinction between these categories is whether the hand of the experimenter acts as the agent of selection on some particular trait or whether the experimenter sits back and allows total fitness to be the agent. Consider two scenarios for a selection experiment with Drosophila melanogaster. In both, a single mass population is established. In the first, at each generation the experimenter captures all adults and measures some key trait, say thorax length. She culls out the bottom 85% of the population, only allowing the top 15% to breed in a series of controlled matings. This is artificial selection. In the second scenario, the experimenter again measures the flies each generation. But, breeding is determined by the flies themselves through competition for food, access to mates, oviposition sites, etc. This is quasi-natural selection.

In the latter experiment, selection operates on the whole of the organism and evolution proceeds as a function of total fitness. Although we can observe evolution, we may not know what is the primary target of selection. Natural selection operates in this fashion, although we can decompose natural selection into components operating on individual traits (Lande & Arnold, 1983; Scheiner et al., 2000). In contrast, in an artificial selection experiment, ideally, fitness is the function of a single trait. Of course we know that this is rare, if ever, the case. Pleiotropy, epistasis, linkage, and so forth can create additional fitness effects through differences in longevity, fecundity, fertility, and other life history traits. These correlated effects can be minimized, however, by the experimenter. The key is that selection is brought to bear on a single trait. All other evolution is a correlated response to that primary selection.

Because of this difference in the target of selection – the whole organism vs. a single trait – the information that the two types of experiments provide differs. To overly simplify the issue, artificial selection experiments are best as probes of the genome whereas quasi-natural selection experiments are best as studies of evolutionary dynamics. In both cases, however, these experiments provide some of the best tests of the assumptions and predictions of evolutionary theory. Together they give us a powerful set of tools to understand the evolution of phenotypic plasticity. With both types of experiments we can ask, how likely is plasticity evolution? Can it occur by both direct and indirect selection? If so, how important are these different modalities? What are the effects of population and environmental structure? These questions are just beginning to be addressed by selection experiments.

This variety – architecture probes and selection mimics, artificial and quasi-natural – is not mutually exclusive. Any particular experiment may have several goals. Of course, the scientific enterprise, just like nature, consists of trade-offs. An experiment designed to best answer one question may be less good at addressing another.

Artificial selection

Experimental design

The evolution of phenotypic plasticity has been studied in three ways with artificial selection experiments. In the first, plasticity was the direct object of selection. In the second, a trait was the object of selection and plasticity evolved as a correlated response. Although this dichotomy holds for all types of artificial selection experiments, it is especially important in the study of plasticity. In addition, studies of plasticity evolution offer a third possibility, simultaneous or alternating individual selection in multiple environments.

Plasticity as a trait has special properties not shared by most other traits. The most important property is that plasticity is often not expressed by or measurable on a single individual. If development is fixed such that a single individual can express only a single trait value, then plasticity of that trait can be measured only by raising related individuals (clones, sibs, etc.) in multiple environments. Plasticity is then measured across that set of individuals. Most typically plasticity has been measured as the difference in trait values between individuals in two environments (e.g. Scheiner & Lyman, 1989), although other indices can be used (Scheiner, 1993). In this instance, individual selection cannot be applied. Rather, some form of group selection must be used. An example of this type of selection experiment is detailed below. Individual selection could be applied to plasticity if the trait were labile within an individual. Physiological traits come immediately to mind. Until now no such selection experiment has been performed.

The most typical form of group selection that is used in plasticity selection experiments is family selection (Scheinberg, 1973). This protocol involves taking siblings, typically of a single brood, and raising them in two or more environments. The mean trait value of each set of siblings in each environment is used to calculate a plasticity index for that family. Again, the most typical situation is two environments with plasticity calculated as the difference in family means. If data from more than two environments are available, more complex indices can be used such as the slope of a linear reaction norm, polynomial coefficients for nonlinear reaction norms (Gibert et al., 1998), or some measure of the variance among environments. Although all previous selection experiments have used full-sib families, other types of groups are also possible. Clonal replicates can be used with many plants and some invertebrates (e.g. Daphnia and aphids). Half-sib families could be used, although this would create a very complex experimental design.

Even a typical full-sib family selection protocol can be complicated. Complications come from two characteristics inherent to plasticity selection experiments. First, sample sizes that might be used in a single-environment experiment must be multiplied by the number of environments. Population size during selection cannot be too small, otherwise genetic variation will rapidly be lost and the response to selection will halt. But the strength of selection, and thereby the rate of response, is inversely proportional to the percentage of families selected each generation. An important consideration is the number of generations that are planned for the experiment. The greatest long-term response to selection is achieved with a selection intensity of 50% (Robertson, 1960), whereas a short-term selection experiment might use a much higher intensity. Secondly, because selection is on families, relatively complex tracking schemes must be used. Mass selection experiments (discussed later) need not keep track of individuals. Family selection experiments must do so. This tracking can be relatively easy for animals that can be raised easily in isolation or as sibships (e.g. Daphnia or Drosophila in small vials) or for plants that can be raised in individual pots. If only mass culturing is possible, then individuals must somehow be tagged. These complications may explain why direct selection on plasticity has rarely been performed.

Analysing these sorts of selection experiments are no different than those for any other type of trait. As there exists an ample literature on the statistics of selection experiments (e.g. Falconer & Mackay, 1996; Lynch & Walsh, 1997), I will not duplicate it here. Direct selection on plasticity does offer one advantage over typical mass selection experiments. Because information is kept on family, clone, or group identity, it is possible to calculate genetic variances and covariances each generation. Such information is not typically gathered in mass selection experiments.

Examples of selection experiments

Direct selection

The first type of experiment that I review is direct selection on phenotypic plasticity. All experiments of this type were performed with D. melanogaster (Waddington, 1960; Kindred, 1965; Druger, 1967; Hillesheim & Stearns, 1991; Scheiner & Lyman, 1991), with the exception of two experiments: on tobacco, Nicotiana rustica (Brumpton et al., 1977; Jinks et al., 1977) discussed in detail later, and on the butterfly Bicyclus anynana (Wijngaarden & Brakefield, 2001; Wijngaarden et al., 2002). The purpose of these experiments was to determine if plasticity is a heritable trait that could be selected upon. All used a family selection protocol. I will focus here on my own study (Scheiner & Lyman, 1991) because it is the most comprehensive in terms of treatment combinations, replicate lines, population sizes, and number of generations. The targeted trait was thorax length and the environmental variable was temperature. Over the range of temperatures used (19 and 25 °C) the reaction norm is approximately linear and negative (David et al., 1983). Flies are smaller at higher temperatures, which is typical for many insects. The experiment involved direct selection to both increase and decrease plasticity as well as equivalent family selection to increase and decrease the trait itself in each environment for a total of six treatments plus a no-selection control treatment. There were two replicate lines per treatment. For all treatments, siblings were measured in both environments even if selection occurred in only one environment. Thus, correlated responses of the trait in both environments and trait plasticity could be followed.

We found a significant response to selection on plasticity with a realized heritability of 0.088 ± 0.027 SE (Fig. 1). Thus, we established that plasticity of thorax length was a selectable trait. This conclusion was not the most important, though, as others had previously demonstrated the selectability of plasticity. More important were our conclusions about the genetic architecture of thorax length plasticity. We demonstrated that plasticity was not the result of overdominance as had been theorized for many decades (Lerner, 1954; Gillespie & Turelli, 1989). Plasticity had been posited to be a function of homozygosity, increasing heterozygosity resulting in decreasing plasticity. That model predicts that the response to selection will plateau as maximal heterozygosity or homozygosity is reached. We found no such plateau when selecting for increased plasticity. It predicts an increase in plasticity when selecting on the mean of the trait, regardless of the direction of selection, because directional selection causes fixation of loci and homozygosity. Rather plasticity evolved as a correlated response, increasing in some lines and decreasing in others. Finally we failed to find a predicted correlation between the amount of genetic variation for thorax length and the plasticity of thorax length.

Figure 1.

The response to selection for increased ( solid lines ) and decreased ( dashed lines ) phenotypic plasticity of thorax length in response to temperature in D. melanogaster . Controls shown as dotted lines . For downward selection, the response halted after approximately 11 generations at a zero slope of the reaction norm. (data from Scheiner & Lyman, 1991 ).

We demonstrated that plasticity and the response to selection was because of a complex interaction among a number of different types of loci. Various pieces of evidence led to this conclusion. One critical piece was a lower bound to selection on plasticity. During the first 11 generations we succeeded in decreasing plasticity so that flies raised at 19 °C were the same size, on average, as flies raised at 25 °C. That is, plasticity was reduced to zero. In theory, plasticity should have been reversible, larger flies at 25 °C than at 19 °C. However, during the next 11 generations we failed to push the selection lines below zero plasticity. We postulated that thorax length plasticity has two components, a response/no response component and a magnitude of response component. This, admittedly ad hoc, model explains a number of puzzling aspects of the experiment. First, estimates of the heritability of plasticity based on sib-analyses (Scheiner & Lyman, 1989) were much higher than the realized heritability. In particular, the overestimates were much greater than those for the mean of the trait. Such a consistent overestimate could result if selection operated primarily on only one of the two components of plasticity, response or magnitude, while measures of genetic variation included both components. A similar low realized heritability for plasticity of body mass in response to food level in D. melanogaster was found by Hillesheim & Stearns (1991).

Indirect selection

More commonly performed than selection directly on plasticity, is selection directly on a trait itself, with plasticity evolving as a correlated response. One of the earliest and most complete experiments of this type was carried out by Falconer (1960) looking at the effects of diet on body size in mice. The target of selection was growth rate, the increase in body weight from 3–6 weeks after birth. He selected for both increased and decreased growth rate in each of two environments, high and low food. To monitor the correlated response in the alternative environment, each generation a second litter was raised in the alternative environment. In all cases the cross-environmental correlation was <1.0 (range 0.19–0.75). Plasticity increased when selection was away from the overall mean of the population – upward at high food levels and downward at low food levels – and decreased when toward the overall mean – downward at high food levels and upward at low food levels. Falconer (1990) codified this result in a review of nine experiments. He concluded that selection towards the overall mean of the population always resulted in a lowering of plasticity relative to selection away from the overall mean. Three additional experiments were published since that review reached similar conclusions (Hillesheim & Stearns, 1991; Scheiner & Lyman, 1991; Matsumura, 1996; but see Noach et al., 1997). These results are in accord with theoretical predictions (Gavrilets & Scheiner, 1993b). Substantially more than these 12 experiments have looked for correlated responses of plasticity to direct selection (see References in Scheiner, 1993). What distinguishes these 12 experiments is the inclusion of all possible combinations of upward and downward selection in both environments. Although logistically more challenging, such experiments are necessary to test theoretical predictions. Best are those experiments which track changes in both environments each generation (e.g. Falconer, 1960; Scheiner & Lyman, 1991) because they provide the most robust estimates of realized cross-environmental correlations.

Combined experiments

The one other organism that has been subjected to selection on plasticity is tobacco, N. rustica (Brumpton et al., 1977; Jinks et al., 1977). More significantly, this is the only experiment, of which I am aware, involving simultaneous (correlational) selection on plasticity and the mean value of the trait across environments. To make the experiment even more complicated, selection was replicated for two traits, flowering time and height. The environment was sowing date. The experiment consisted of selection on all four combinations of high and low mean trait value and high and low plasticity. The realized heritability of plasticity was always lower than that of the trait mean. Mean trait value and plasticity were positively genetically correlated, and the strong response to selection on the mean overwhelmed selection on plasticity when it was opposite to the correlation. This experiment demonstrated that trait mean and trait plasticity are partially independent, providing additional evidence that plasticity is caused by a complex interaction among a number of different types of loci.

Alternating selection and disruptive selection

The group-selection protocol used to select directly on plasticity, although very useful for some questions, is limited because natural selection most often acts at the level of the individual (but see Goodnight & Stevens, 1997; Wilson, 1997). In nature, selection on plasticity will most often take the form of individual selection in multiple environments, either simultaneously in different demes or alternating among generations. Selection schemes that mimic these forms of natural selection were the basis of the quasi-natural selection experiments discussed in the next section. They also were the basis for a number of artificial selection experiments. Once again, Drosophila was the organism of choice. Simultaneous selection was used by Waddington & Robertson (1966), Scharloo et al. (1972), and Thompson & Rook (1988). Alternating selection was used by Waddington (1960). Plasticity changed in most cases, although the mode of response varied. In some studies the means responded in only one environment (Waddington, 1960, aristae morphology; Kindred, 1965), in other studies they responded in both environments but in the same direction (Waddington & Robertson, 1966; Scharloo et al., 1972), and in still others they responded in opposite directions (Waddington, 1960, eye size; Druger, 1967; Thompson & Rook, 1988).

In none of the studies were the effects of selection on plasticity contrasted with the effects of simple directional selection. An important unanswered question is which form of selection, individual or group, results in the greatest short-term and long-term change in plasticity. Even if group-level selection is relatively rare in nature, it may prove to be more important for the evolution of adaptive plasticity (Goodnight, 1985).

Integrating with other types of information

Selection experiments provide powerful evidence for issues such as trait heritability and genetic architecture. They are limited, however, in that alternative genetic models (pleiotropy vs. epistasis; Scheiner & Lyman, 1991) may predict similar outcomes, so that selection experiments are insufficient to chose between them. Other forms of genetic analysis are necessary, such as recent molecular approaches, as well as classical methods including the use of chromosome extraction in Drosophila (e.g. Schnee & Thompson, 1984; Cavicchi et al., 1989; Weber & Scheiner, 1992) and among-line crosses (e.g. Jinks & Connolly, 1973; Connolly & Jinks, 1975; Clare & Luckinbill, 1985; Bell, 1991). All of these studies confirm the conclusion of the selection experiments. Trait expression across environments is complex and results from interactions among many kinds of loci.

Quasi-natural selection

The other broad class of selection experiments is quasi-natural selection. Such experiments have a long and distinguished history, exemplified by the population cage experiments of Dobzhansky (Lewontin et al., 1981). In this section I examine three examples of quasi-natural selection experiments, one with temporal variation, one with spatial variation, and one with both effects. I highlight these experiments because they contrast selection in constant environments with selection in variable environments, thus providing the most complete tests of theoretical predictions.

Experimental design

These experiments investigate the effects of spatial or temporal variation and are similar in form to the last type of artificial selection experiment. Although there have been a number of experiments which have manipulated one or the other type of variation, only a few examined plasticity (Bell & Reboud, 1997; Reboud & Bell, 1997; Kassen & Bell, 1998; Scheiner & Yampolsky, 1998). All of the other experiments focused on the maintenance of genetic variation (e.g. Powell, 1971; McDonald & Ayala, 1974; Minawa & Birley, 1978; Powell & Wistrand, 1978; Oakeshott, 1979; Mackay, 1980; Haley & Birley, 1983; García-Dorado et al., 1991; Bell, 1997b; Goho & Bell, 2000; Fry, 2001; see review by Kassen, 2002) or the effects of environmental heterogeneity on trait evolution (e.g. Wallace, 1982; Goodnight, 1985; Ehiobu & Goddard, 1989).

The key issue in designing these types of experiments is environmental grain or predictability. The evolution of plasticity depends on two factors, the rate at which the environment changes relative to the response rate of an organism and the predictability of that change (Scheiner, 1993). If an organism can respond immediately to environmental changes (e.g. physiological responses), then plasticity is always favoured. In contrast, if the organism's phenotype becomes fixed at a single developmental point (e.g. adult size in holometabolous insects), then plasticity is favoured only if the environmental cue at the determination point accurately predicts the future environment. In addition, if the environment changes only very slowly relative to the generation time of the organism, then genetic specialization is favoured over plasticity (Orzack, 1985).

Spatial variation adds a second complication, hard and soft selection. Under soft selection, all demes contribute a fixed proportion of offspring to the next generation regardless of the environment of that deme. Relative fitness of genotypes is determined at the level of the deme. In contrast, under hard selection demes contribute to the next generation in proportion to their population sizes. Relative fitness is determined at the level of the metapopulation. Because plasticity is a trait which is expressed across environments, hard and soft selection differentially favour plasticity and genetic specialization, respectively (Van Tienderen, 1991, 1997).

Examples of selection experiments

Temporal variation

Temporal variation was explored in my experiment on the effects of temperature on the water flea, Daphnia pulex (Scheiner & Yampolsky, 1998). We compared three patterns of variation: constant temperatures, varying but predictable temperature change, and unpredictable temperature change. For the first treatment, the temperature was held at a constant 20 °C. For the second, the temperature was varied in a regular fashion consisting of 12 days at 17 °C, 6 days at 20 °C, 12 days at 23 °C, 6 days at 20 °C, then the entire cycle was repeated once more, so that the total time spent at each temperature was equal across the entire experiment. For the third treatment, the temperature was randomly switched between 17, 20, and 23 °C every 3 days with the constraint that, again, the total time spent at each temperature was equal across the entire experiment. There were two replicate populations in each treatment. Daphnia can replicate asexually, a characteristic which we exploited to create genetically identical starting populations. The six populations consisted of 34 clones, each clone with two adults and four juveniles. Selection proceeded for 72 days, or approximately for ten generations. Populations grew continuously so that the primary target of selection was the Malthusian parameter (r) and through that, life history traits of development time, offspring size, and offspring number.

We had expected the predictable treatment to select for plasticity, whereas plasticity would be selected against in the other treatments. Instead we found that plasticity increased in all treatments. Whereas populations differed in their plasticities, the differences did not follow those predicted by theory. One of the constant populations was the most plastic for seven traits, whereas one of the predictable populations was the least plastic for two, exactly opposite the predictions. Nor were the results consistent within treatments. For r, the direct object of selection, one predictable population was the most plastic, whereas the other was among the least plastic. We explained these results by observing that the heritability of trait plasticities was lower than the heritability of trait means. Because of this difference, the selection response of trait means overwhelmed the selection response of plasticity. This lower heritability of plasticity is very common (Scheiner, 1993), suggesting that our results will be typical of responses to selection in nature. These results mirror those of the correlational selection experiment described above (Brumpton et al., 1977; Jinks et al., 1977).

Spatial variation

Spatial variation was explored in an experiment by Bell & Reboud (1997). They used Chlamydomonas reinhardtii, a unicellular green alga, as the experimental organism, investigating the effects of variation in light level. This species is capable both of growing autotrophically in the light and heterotrophically in the dark. As with Daphnia, Chlamydomonas will reproduce asexually. Again, replicate lines were genetically identical at the beginning of the experiment. The experiment consisted of three treatments – constant light, constant dark, and both light and dark – with two replicate lines per treatment. The final treatment, spatial variation, was achieved by mixing light and dark grown populations after each growth cycle. Selection proceeded for 1 year, or approximately 750 generations in the light treatment and 250 generations in the dark. The allopatric lineages (light alone or dark alone) increased their adaptation in the selection environment at the expense of their ability to grow in the opposite environment. That is, growth in the light and growth in the dark were negatively genetically correlated. This negative correlation is necessary for the evolution of genetic specialization. It also sets the stage for tests of models of plasticity evolution that gauge the relative likelihood of adaptation by plasticity vs. specialization. In this experiment, selection in the spatially variable environment resulted in the evolution of genetic specialists, rather than phenotypic plasticity. This result is in agreement with theoretical models (Scheiner, 1993).

Combined experiment

In a second experiment, Reboud & Bell (1997) examined the effects of both spatial and temporal variation in a continuation of the experiment described above. An additional treatment was added, alternating cycles of light and dark (temporal variation). Each cycle was about 12 generations long. Evolution proceeded for an additional 100–200 generations. As before, spatial variation continued to select for genetic specialization. Temporal variation, however, selected for phenotypic plasticity. Lines previously specialized for growth either in the dark or the light evolved the ability to grow well under both conditions (Fig. 2). Again, these results are in accord with theoretical predictions (Gillespie, 1977; Gavrilets & Scheiner, 1993a). The plastic lines failed to reach the growth rates of the specialized lines, however, suggesting a cost or constraint on plasticity (Van Tienderen, 1991, 1997).

Figure 2.

The response to selection for growth in the light and the dark by Chlamydomonas for each of 12 lines (two each for selection in constant light or constant dark, four each for spatial or temporal variation) measured as culture turbidity after 10 days. Bars are SEs. The mean values for the spatial variation lines are misleading in that these populations consisted of a mixture of light and dark specialists (data from Reboud & Bell, 1997 ).

Implications and recommendations

Why was there a discrepancy in adaptive responses between this last experiment and the one using Daphnia? Most likely the discrepancy was because of differences in population sizes (c. 102 vs. 106) and the number of generations (10 vs. 1000) in the Daphnia and Chlamydomonas experiments, respectively. In the Daphnia experiments, the outcome was limited by initial genetic variation and the heritabilities for plasticity were low. In contrast, in the Chlamydomonas experiments new genetic variation could accumulate, thus favouring longer term evolutionary responses. The results of a follow-up experiment with Chlamydomonas (Kassen & Bell, 1998) strengthen this interpretation. This experiment contrasted constant environments (light or dark), with change every hour (five times per generation), with change once a day (every fifth generation). Plasticity did not increase in any of the treatments, although it should have in the third treatment. The third treatment was similar to the temporal variation treatment in the previous experiments. This experiment, however, only proceeded for 200 generations, perhaps not enough time for genetic variation in plasticity to accumulate.

Because selection on plasticity is based on its across-deme, global fitness, evolutionary responses will usually be slow (Scheiner, 1998). These results have implications for comparative studies of natural populations. Such studies may need to shift from closely related, local population differences to those of more distantly related populations or different species. On the other hand, a recent study of the natural evolution of phenotypic plasticity in resistance to algal toxicity in Daphnia galeata, found a significant change in the population in about 15 years (Hairston et al., 2001). This is a response to coarse-scale temporal variation as the relative abundances of palatable and toxic algal species changes over the course of a growing season.

All of these experiments share several characteristics. First, both species have a rapid generation time. This characteristic is one reason that Drosophila was the organism of choice for Dobzhansky's population cage experiments. Secondly, both species are small allowing for moderate to large population sizes. Thirdly, both species are capable of asexual reproduction. This characteristic allows the establishment of genetically identical founder populations, removing one potential source of experimental variation. Fourthly, both species are eukaryotes allowing for complex evolutionary responses. A long series of experiments, also focusing in part on plasticity, were carried out using the bacteria Escherichia coli (e.g. Bennett et al., 1992; Bennett & Lenski, 1993). They looked at adaptation to various combinations of constant and varying temperatures and generally failed to find any change in the amount of plasticity. In these experiments, however, the main form of adaptation was biochemical. There may be a limit to the extent that plasticity can evolve in such systems, in contrast to the more complex developmental responses of eukaryotes, especially multicellular organisms.

Currently few studies have directly tested theoretical predictions of plasticity evolution using quasi-natural selection experiments. I urge more such experiments, either exploiting currently well-studied systems such as Chlamydomonas, Daphnia, and Drosophila, or developing new systems. Potentially suitable systems include Arabidopsis thaliana and Caenorhabditis elegans both of which are well known genetically and developmentally.

Where do we go from here?

Selection experiments have been instrumental in our current understanding of the genetics and evolution of phenotypic plasticity. They have clearly established that phenotypic plasticity is a heritable, evolvable trait. Plasticity can be directly selected for, although in nature it most likely evolves as a correlated response to selection within single environments. Plasticity is a genetically complex trait, determined by multiple loci. Trait expression is likely determined by both environmentally responsive (plastic) loci and nonresponsive loci which interact to create complex evolutionary dynamics. Now the genetic questions are in the hands of those using molecular tools that can pinpoint these loci and explore their interactions in detail. The main role for selection experiments in coming years will be observing evolutionary dynamics and testing evolutionary theories.

Selection experiments are ideal for such theory testing. Phenotypic plasticity is an area where theory has far outstripped data. We now have nearly 20 years of theoretical advancement (Scheiner, 1993), with new papers being published constantly (e.g. Van Tienderen, 1997; Scheiner, 1998; De Jong & Gavrilets, 2000, van Dooren 2001; De Jong & Behera, 2002). Such theory development needs to be solidly anchored in data. For example, if plasticity has a cost, evolutionary outcomes can differ substantially (e.g. Van Tienderen, 1991, 1997). Currently, few studies have measured the cost of plasticity and those that have find little evidence for them (DeWitt, 1998; DeWitt et al., 1998; Scheiner & Berrigan, 1998). So, although we could continue to develop theories based on assuming costs of plasticity, such efforts might be a waste of time. On the other hand, we (Scheiner & Berrigan, 1998) showed that measuring costs is relatively easy, and many already published studies could be profitably revisited.

Selection experiments are a necessary complement to comparative studies of natural systems. For example, theory indicates that the evolution of plasticity will differ in structured, subdivided populations compared with panmictic populations (Scheiner, 1998). Finding properly structured natural systems will be difficult, however. In contrast, a quasi-natural selection experiment could test these theoretical predictions. Selection experiments on plasticity can be complex to carry out, one reason that they have been relatively rare. On the other hand, they provide a bountiful harvest of information. I urge more such studies.

Acknowledgments

I thank Kim Hughes, Steve Stearns, and an anonymous reviewer for their comments on the manuscript. The views expressed in this paper do not necessarily reflect those of the National Science Foundation or the US Government.

Ancillary