Rapid evolution: from genes to communities, and back again?


Correspondence author. E-mail: spe2@cornell.edu


  1. Ecologists now recognize that ecologically important traits sometimes undergo large heritable changes on the same time-scale as ecological change (‘rapid’ or ‘contemporary’ evolution). We know much less about the meanings of ‘sometimes’ (how often?) and ‘important’ (how important?). I review some of what I and my collaborators have learned about rapid evolution and suggest some of its implications.
  2. Rapid evolution is common: when conditions change, traits evolve. But the rate of change is hard to predict, even across a set of published laboratory predator–prey experiments, and very similar trait changes can result from very different molecular mechanisms, even within one species.
  3. Evolutionary rescue (adaptive response to a detrimental environmental change) is common, but not always effective. The full range is observed, from little to almost complete compensation for environmental change.
  4. Cases of limited compensation show that strong selection is necessary, but not sufficient for rapid adaptation. The response of the copepod Onychodiaptomus sanguineus to fluctuating predation shows that the time-scale of environmental variation is important. The response of Daphnia galeata to eutrophication shows that life-history trade-offs and constraints can prevent evolution from buffering the entire life history against environmental changes.
  5. Eco-evolutionary dynamics are harder to predict when traits affecting multiple interactions evolve. Theory predicted, and experiments confirmed, a much larger range of possible evolutionary outcomes when a prey species evolves defences against two predators rather than one. These and other recent findings suggest that the devil is in the details: many subtle factors, such as how trade-offs depend on resource availability, can have large effects on eco-evolutionary dynamics.
  6. Ecologists are increasingly engaged in predicting the ecological effects of human impacts such as species introductions and climate change. Forecasting evolutionary responses is part of that challenge. One implication of the ‘newest synthesis’ is that rapid adaptation to changing conditions is nothing new, so the past and present can tell us a lot about the future. If prediction is our goal, then developing predictive theory for eco-evolutionary dynamics, and testing it in tractable model systems, should be a priority.

A large part of the future work of plant ecology will, I think, be concerned with life form and its relation on the one hand to environment and on the other to heredity

A.G. Tansley, 1939 President's Address, British Ecological Society Annual Meeting


I had not intended to study rapid evolution. I wanted to study how evolutionary optimization of life histories affected population dynamics. Lamont Cole taught us that life-history parameters such as age at maturity and clutch size are evolved traits just like coat colour or leg length. I felt that the same must be true of population dynamics parameters such as stability versus cycles versus chaos. At the time (the late 1990s), analyses of long-term population data seemed to say that cycles were common, but chaos was very rare (Ellner & Turchin 1995; Kendall, Prendergast & Bjornstad 1998). If chaos were completely absent, we could ask why it might be impossible. But it isn't impossible, in nature or in the laboratory (e.g. Turchin & Ellner 2000; Dennis et al. 2001). And one recipe for chaos is to put an oscillating system into a time-varying environment, such as any habitat with strong seasonality or large effects of multiannual oscillations such as ENSO or NAO. So if population cycles are common, chaos should be common. Why isn't it?

Clever theoreticians had already shown that life-history evolution in theoretical organisms can lead to a wide range of population dynamics, depending on your assumptions (e.g. Doebeli 1993; Doebeli & Koella 1995, 1996; Ferriere & Fox 1995). But real organisms, I thought, would have life-history trade-offs causing them to gradually evolve away from chaos. Henry Wilbur and I were discussing this over lunch at a conference, imagining fields of artificial ponds with experimental amphibian communities. From the other end of our picnic table, Nelson Hairston, Jr., said ‘You should do that with rotifers’.

That seemed a lot more practical. Building on classic experiments of Halbach (1970), Nelson and I created chemostat microcosms with a rotifer predator, Brachionus calyciflorus, and algal prey, Chlorella vulgaris. To test our understanding of the system and figure out how to push it into chaos, we and our postdoc Gregor Fussmann modelled the system and tested whether our model could accurately predict the transitions between stable equilibrium and population cycles (Fussmann et al. 2000). The model worked well, making the correct prediction for 16 out of 18 experiments at different combinations of nutrient supply and washout rates.

But there were problems. The models predicted cycles with 7- to 10- day periods; the actual periods were 20–60 days (Fig. 1a). And our model, like all standard predator–prey models, predicted that prey peaks would lag those of the predator by about 25% of the cycle period. We actually observed a lag of half the cycle period, so that predator and prey were exactly antiphase. This can't happen in any conventional predator–prey model, except as the result of long maturation delays (de Roos et al. 1990; de Roos & Persson 2003) that were absent in our study species (our rotifers matured within about 1 day; Gregor Fussmann, unpublished data). Antiphase cycles are impossible because they imply that predator population growth rate is positive at some level of prey abundance and then negative at the very same level of prey abundance, at some later time (e.g. days 20, 30, 40 and 55 in Fig. 1a). If predator birth and death rates are a function of prey and predator abundance, the antiphase cycles in Fig. 1a canWt happen.

We eventually hypothesized that we were watching an effect of rapid prey evolution (Shertzer et al. 2002). During one population cycle (a few dozen algal generations), the algae gain and then lose a heritable trait providing defence against rotifer predation. High predation selects for defence, which leads to predators becoming rare. Selection then favours prey genotypes with poor defence, but higher competitive ability, allowing predators to increase again. The response to selection is slow enough to lengthen the cycles, but fast enough to change the ecological dynamics as they happen. Our postdoc Takehito Yoshida verified this hypothesis experimentally (Fig. 1b) by showing that shorter, normal predator–prey cycles return when evolution is prevented by eliminating genetic variability in the algae (Yoshida et al. 2003).

It's hard to keep evolution from happening: when conditions change, traits often change

Planning to study slow evolution, we wound up studying rapid evolution instead. And we were not alone. Our postdoc Teppo Hiltunen asked: ‘if rapid evolution is so important and general, why didn't many other people see it before in their predator-prey experiments?’ It turns out that it had been seen, but not recognized. Scouring the literature for predator–prey experiments or other long-term data, he found many studies in which populations either evolved towards long antiphase cycles or had such cycles from the outset (e.g. Fig. 1e,f; T. Hiltunen, N. G. Hairston, Jr., G. Hooker, L. E. Jones & S.P. Ellner, unpublished).

We were not alone in another, more important sense: ecologists had discovered, and were continuing to discover, rapid evolution in natural populations of a wide variety of taxa: single- and multicellular, plant and animals, vertebrates and invertebrates, aquatic and terrestrial (see, e.g., Thompson 1998; Hendry & Kinnison 1999; Reznick & Ghalambor 2001; Pelletier, Garant & Hendry 2009). It is now documented beyond doubt that rapid evolution is common in natural populations, so we need to take the next step of understanding the general importance and consequences of rapid evolution (Schoener 2011). Is each instance of selection and response unique, or are there some general principles?

My goal in this paper is to suggest a few principles, based mainly on projects I have been involved with and following the actual lecture fairly closely. The study of rapid evolution is young enough that generalization is risky. But given the potential importance of evolutionary responses to changes in climate and land use, and the arguably more pressing goal of understanding the natural world while it still exists, it's a risk worth taking [and you can also get several second opinions, e.g., Hendry et al. (2011); Morris (2011); Hanski (2012); Reznick (2013)]. My first general principle is the title of this section. The second is as follows:

But it's hard to predict how fast traits will change

Substantial evolution takes at least one generation. But selective mortality can shorten the generation time, so rapid evolution can occur in less than one average generation time. For example, a sea fan coral's susceptibility and response to infection by a fungal pathogen changed markedly within a decade, following an epizootic in which local populations experienced up to 90% mortality (Bruno et al. 2011). So rapid evolution is potentially possible in many species, and of course, very slow evolution is also possible. Predicting when each of these will happen would be an important first step in understanding the ecological importance of rapid evolution.

But this appears to be difficult, even in the simple setting of laboratory populations. For their compilation of ≈20 laboratory predator–prey experiments, Hiltunen et al. (unpublished manuscript) quantified the rate of prey evolution based on the change from classical to antiphase predator–prey cycles. The method is illustrated in Fig. 1c,d. Smoothed population data are plotted in the prey–predator phase plane, and the ‘evolutionary dynamics index’ (EDI) is defined as the ratio of the second to first principal components of the data. Antiphase cycles approximate a line in the phase plane (EDI≈0, Fig. 1c), while classical predator–prey cycles give a limit cycle with larger EDI (Fig. 1d). In about half of the data sets, the decrease in EDI between the start and end of the experiment was statistically significant. If the change is due to prey evolution, we should see more change in experiments that ran for more prey generations and more change with larger prey population size (so there would be more opportunities for defended mutants to arise). Both of those were true and statistically significant.

Figure 1.

Classical and evolutionary predator–prey cycles. (a) Evolutionary (antiphase) cycles in rotifer–algal chemostats (Yoshida et al. 2003). Solid red circles are the rotifer predator, and open green circles are the algal prey. Solid curves are spline smooths of the data. (b) Classical cycles with quarter-period lag, in rotifer–algal chemostat with genetically uniform prey (Yoshida et al. 2003). (c) Illustration of the evolutionary dynamics index (EDI) for the data in panel (a). The smoothed data, omitting initial transients prior to cycles, are plotted in the prey–predator phase plane. EDI is the ratio of the second to the first principal components of the plotted data. (d) As in panel (c), for the data in panel (b). (e) Population cycles of E.coli (open green circles) and Bdellovibrio bacteriovorus (solid red circles), from Dulos & Marchand (1984). (f) Antiphase population cycles of Alcaligenes faecalis (bacterial prey, open green circles) and Colpidium campylum (ciliate predator, solid red circles), from Sudo, Kobayashi & Aiba (1975).

However, regression on prey generations and population size only explained 39% of the variance among data sets in the amount of change in EDI. An math formula of 0·39 sounds pretty good, but a 39% reduction in variance is only a 22% reduction in the standard deviation. To experience this first-hand, extend both arms straight out at your sides. The fingertip-to-fingertip distance represents the interspecific variation in the total change in EDI. The unexplained variation then goes from one wrist to the other, or an inch or two in from each wrist if your fingers are short. That's what math formula really means, and that's why expressing it in terms of variance (squared error) reduction is so popular. The analysis by Hiltunen et al. has limitations (e.g. the predictors are estimated rather than known exactly). But it's hard to imagine that we could do better with natural populations than with laboratory experiments where confounding factors were held to a minimum.

And it's not just selection to escape predation for which the amount of evolutionary response is so variable. ‘Despite long-term selection with a relatively large population’, Low-Decarie et al. (2013) ‘failed to detect any form of evolutionary response’ in seven species of phytoplankton from three major groups after selection for over 700 generations under increased CO2 (1000 ppm); Collins & Bell (2004) found essentially the same in a green alga. In contrast, in another phytoplankton species, the coccolithophore Emiliania huxleyi, 500 generations of exposure to increased CO2 resulted in higher growth rates and improved calcification at increased CO2 (Lohbeck, Riebesell & Reusch 2012).

Prediction of evolutionary responses is further complicated by the fact that…

Selection acts on phenotypes, not genes

We began to investigate the molecular basis for the repeated gain and loss of algal defence traits in our chemostats, using microarrays to track changes in gene expression during the eco-evolutionary cycles (Becks et al. 2012). This was possible because we had switched to a different alga, Chlamydomonas reinhardtii, whose genome had been sequenced. We expected to see predictable, repeatable gene expression changes paralleling the predictable, repeatable oscillations of the evolving defence trait [formation of clumps too large for the predator to eat, in the case of Chlamydomonas (Becks et al. 2010)]. Our postdoc Lutz Becks used a custom microarray, covering all annotated genes on 87% of the known transcriptome, to assay gene expression at 8 time points during two complete cycles. As we expected, within each of the population cycles, genes that were upregulated as the fraction of unclumped cells increased, were then downregulated as the fraction of clumped cells increased and vice versa. But comparing between cycles, the changes in gene expression during two successive periods when the fraction of unclumped cells increased were not positively correlated (in fact there was a very small negative correlation). The same was true for the two successive periods when cell clumping increased. So when the same selection pressure was repeated, the algal population evolved the same trait for a second time, but it did that in a completely different way. The first round of selection may have changed the gene pool, even though the trait returned to its original state, or perhaps by chance different genotype(s) with the selected phenotype happened to increase during the second cycle. The phenotypic response to selection was predictable, but the genetic response was not, because selection is blind to how the phenotype is produced.

However, we already know that unpredictability at the genetic level is not a general principle – predictability can also be high. Meyer et al. (2012) exposed 96 replicate phage populations to the same selection pressure: an E. coli strain lacking the membrane protein that the phage uses to attach themselves. In 24 of the lines, the phage evolved the ability to use a different membrane protein. All 24 lines used the same new protein, and the four mutations that were found to be necessary and sufficient for using the new protein were strikingly similar: two were identical in all 24 lines and the other two had slight variations. Similarly, Herron & Doebeli (2013) found many parallel (and a few unique) changes in three independent populations of E. coli diversifying into two ecotypes when competing for two different carbon sources. In both these cases, the phenotypic and genetic changes were very closely aligned, exactly the opposite of our chemostat experiments. Can predictability at the genetic level (or its absence) be predicted? For example, might it depend on trait complexity (predator defence by hook or by crook versus the structure of a protein for binding to E. coli?) Experiments by Toprak et al. (2012) suggest that this will be challenging. Their experiments involved exposing replicate lines of E. coli to different antibiotics and following the evolution of antibiotic resistance. For some antibiotics, the course of evolution after exposure to the antibiotic was highly repeatable across 5 replicate lines, but for others, the evolutionary path was variable. We can hypothesize that this difference relates to the fitness landscape for resistance to the different antibiotics. If many different single-gene mutations confer some degree of resistance, then several different mutational sequences might lead to high resistance, while others turn out to be dead ends even if the first few steps are helpful.

For ecologists, these findings suggest that predicting long-term evolutionary responses is complicated by the contingency of evolution. If each step in adapting to new conditions can occur in several different ways at the molecular level, then even with full knowledge of the fitness landscape and the species’ evolutionary potential, the long-term path of evolution may be unpredictable.

The evolution of antibiotic resistance is also an important example of my next generalization:

Evolutionary tracking and rescue are rampant, but they aren't always effective

Evolutionary rescue refers to situations where a population decreases towards extinction because of a harmful environmental change (less rain, an emerging pathogen, etc.), but it evolves an adaptation to the new challenge and population size rebounds (Gomulkiewicz & Holt 1995). Evolutionary tracking is when adaptation is so fast that the population size never declines by much.

Table 1 lists populations for which the relative contributions of evolutionary versus environmental change have been estimated. The method, which is originally due to Monica Geber (Hairston et al. 2005), is based on writing

display math(eqn 1)

where X is some ecological property of interest, k is the environmental variable whose change is affecting X and z is the trait that evolves in response. The ‘ECO’ and ‘EVO’ terms in Eqn. (eqn 1) measure the contribution of the environmental change and the evolutionary response to the ecological property. The key idea is that the importance of evolution is measured with an ecological ‘yardstick’. In equation (eqn 1), there is one environment variable and one evolving trait with 100% heritability, but more general versions allow for imperfect heritability (e.g. phenotypic plasticity), population structure, multivariate k and z, and indirect effects mediated by interactions with other species or environmental variables (Ellner, Geber & Hairston 2011). X can be any ecological property (butterfly abundance, denitrification rate, species richness, etc.),and the importance of an evolving trait will depend on which property is considered. Strictly speaking, for evolutionary rescue, X should be population fitness (per capita rate of increase), but population size was not always measured. The table therefore also includes major fitness components, such as juvenile survival, or a trait, such as body size, that correlates strongly with fitness, when these were the measured property.

Table 1. Estimates of the importance of evolutionary response, relative to an environmental or ecological change, for population growth rate or a component of growth rate, using the methods of Hairston et al. (2005) and Ellner, Geber & Hairston (2011). ECO:EVO is the ratio between the average absolute values of the corresponding terms in the appropriate version of equation (eqn 1). The value for Soay sheep was derived from annual average values reported in Ozgul et al. (2009): EVO = 153g (viability selection) +32g (fecundity selection), ECO + EVO  = −81g (net change). Other values are from analyses described in Hairston et al. (2005) or Ellner, Geber & Hairston (2011)
SpeciesResponse XEnvironmental factor kTrait zEVO:ECO
Medium ground finch Geospiza fortis, GalapagosPopulation growth rateRainfall, seed sizeBeak size2·2
Rotifer Brachionus calyciflorus, chemostatsPopulation growth rateAlgal densityAlgal defence2·0
Alga Chlamydomonas reinhardtii, chemostatsPredation riskRotifer densityAlgal defence1·4
Great tits Parus major, OxfordSurvival to adulthoodCompetition, temperatureJuvenile body mass1·6
Soay sheep, HirtaBody massCompetition, climateGrowth rate0·7
Copepod Onychodiaptomus sanguineus, Bullhead Pond RI, USAAnnual egg productionTiming and intensity of fish predationTiming of seasonal diapause0·27
Daphnia galeata, Lake ConstanceAdult body sizeCyanobacteria blooms due to P enrichmentJuvenile growth rate0·25

The results in Table 1 say that it's not just microbes who can evolve fast enough to adapt their way out of trouble. And even the Soays, which are famously not evolving fast enough, would be shrinking more than three times as fast without the evolutionary response to their increasingly inhospitable island. But at the bottom of Table 1 are examples where EVO:ECO is small. Those tell us that…

Selection has to be strong, but that's not enough: 1

Selection and response are easy to understand. But when strong selection is observed in a population, but no response occurs, that is a puzzle calling out for explanation (Merilä, Sheldon & Kruuk 2001). So the cases where EVO:ECO is small should be especially informative for understanding when and to what extent rapid evolution can buffer species against environmental change (e.g. Chevin, Lande & Mace 2010; Gonzalez et al. 2013).

The bottom of the heap in Table 1 is Onychodiaptomus sanguineus in Bullhead Pond RI, USA (Hairston & Walton 1986; Hairston 1988). The copepods are active in fall and winter, producing eggs that hatch quickly to start a new generation. In spring, females switch to making diapausing eggs that remain in the pond sediments and hatch (unless they get too deeply buried) in the fall. Diapause provides escape from fish predation, which quickly intensifies when fish become active in the spring. The switch is predictive, cued by photoperiod and temperature rather than by the presence of fish, because diapausing eggs become advantageous before fish activity starts. If the offspring from an immediately hatching egg encounters strong fish predation before it can mature and breed (which takes several weeks), a diapausing egg would have yielded more grandchildren.

The threshold for switching is heritable (N.G. Hairston & Dillon 1990), and a natural experiment showed that selection on switch date is very strong (Hairston & Walton 1986). Bullhead is a permanent pond, but neighbouring Little Bullhead nearly dried out and then froze completely (January 1981), killing all the fish. In 1982, it refilled and copepods re-established by hatching from the sediments. Prior to this (1979), the distributions of switch dates in Bullhead and Little Bullhead were nearly identical. In 1983, after two years without fish predation, the mean switch date in Little Bullhead was later by 26 days, roughly twice the standard deviation of switch date prior to selection. Because selection acted in one generation per year, the average rate of evolution in the generations when selection occurred was roughly 1 haldane, one of the fastest rates ever observed in nature (Hendry & Kinnison 1999). Additional evidence for strong selection is that in Bullhead Pond, the year-to-year changes in mean switch date can be predicted from the intensity and timing of fish predation in the prior year (Ellner et al. 1999).

Nonetheless, the year-to-year evolution of switch date z in Bullhead failed to buffer female fitness X against the fluctuations in fish predation k (Fig. 2). In fact, roughly half the time evolution made things worse for the next generation rather than better, because predation risk changed so rapidly (Hairston 1988) that the direction of selection often flipped between one year and the next. In contrast, at the top of the heap in Table 1, Geospiza fortis experienced two distinct periods of environmental change, followed by periods of evolutionary ‘catch-up’.

Natural populations are affected by environmental variation, with the time-scale of variation ranging from seconds (sun flecks) to years (ENSO) to decades (PDO, climate warming). The Bullhead case study shows that a mismatch between the environmental and evolutionary time-scales can limit the ecological impact of rapid evolution, even when traits are evolving rapidly in response to changing environmental conditions.

The ecological significance of a rapidly evolving trait might also depend on its context within the community: which species in the food web are impacted by the environmental variation (producers or consumers?), which species is evolving and which species are affected by the trait.

Figure 2c shows a theoretical example of these predictions, for a nutrient–prey–predator chemostat model with prey defence (EVO, z) and prey abundance (ECO, k) affecting predator population growth rate X. With top-down environmental forcing (variation in predator mortality rate), EVO:ECO is highest when the environmental change is slowest. With bottom-up forcing (variation in nutrient supply rate), EVO:ECO peaks at an intermediate time-scale of environmental variation, and the failure of adaptive response to high-frequency oscillations is much stronger. Postdoc Masato Yamamichi and I are working on mathematical theory to understand and generalize this computational experiment and on chemostat experiments to test the predictions. The two different relationships between the forcing period (the mean time between peaks in nutrient supply or predator mortality) and the magnitude of response in Fig. 2c are typical of the modes of oscillation associated with real versus complex eigenvalues in a stable linear system. This suggests that environmental forcing of different species in a food web can lead to different modes of oscillation being dominant, with resulting differences in the eco-evolutionary dynamics. Note that in this example, the prey are evolving, but I chose to use the predators as the ecological ‘yardstick’; the same kind of analysis could also be done to compare how prey defence evolution and predator abundance affect prey population growth.

Figure 2.

Annual estimates of the ECO (green squares) and EVO (blue circles) terms for (a) Onychodiaptomus sanguineus in Bullhead Pond and (b) the medium ground finch Geospiza fortis. The methods used to calculate the estimates are described in Hairston et al. (2005). (c) Ratio of EVO:ECO effects as a function of the characteristic time-scale of environmental variation in a nutrient–prey–predator model with prey defence evolution, for bottom-up and top-down environmental forcing.

The general importance of fluctuating selection is a long-standing controversy (as Bell (2010) notes, Fisher and Wright argued about it). ‘It is clear to even the most casual observer that the environment is in a constant state of flux. It must be the case that fitness differences between genotypes are also in a constant state of flux’ (Gillespie 1991, p. 142). G. fortis and O.sanguineus are good examples: fluctuating selection was driven by climatic variation, specifically rainfall, which affected the seeds available to G. fortis (Grant & Grant 2002), and drove the fluctuations of copepod predation risk in Bullhead Pond (Hairston 1988). But proving the obvious has not been easy, because fluctuating selection and response are hard to distinguish from random drift and sampling error. A recent meta-analysis by Morrissey & Hadfield (2012) suggested that most observed changes in the sign of estimated directional selection coefficients (Siepielski, DiBattista & Carlson 2009) were probably artefacts of sampling error.

Selection has to be strong, but that is not enough: 2

Rapid evolution is more evident when a sustained change in conditions leads to a sustained directional trait change, such as evolutionary responses to establishment of an invasive species (Carroll, Klassen & Dingle 1998; Strauss, Lau & Carroll 2006). But even then strong selection might not be enough, as illustrated by our second failure of adaptive response to changing conditions, Daphnia galeata, in Lake Constance.

The environmental change was a rise in cyanobacteria, which are low-quality food for Daphnia, in the 1970s caused by an increased phosphorus loading (Hairston et al. 1999). After the pollution was abated (starting in 1980), the cyanobacterial bloom ended gradually between 1990 and 2000. Because Daphnia produce diapausing eggs each year, their evolution could be tracked by collecting eggs from dated sediment cores, hatching them in the laboratory and monitoring individuals’ performance on ‘poor’ and ‘good’ food sources that mimic the conditions in the lake before and during the cyanobacterial bloom.

Adult body size was chosen as the response variable X because adults’ size determines their trophic links in the lake food web: their rate of grazing on phytoplankton and their vulnerability to predation by fish. Food quality affects both juvenile growth rate and adult body size. There was rapid evolution of the norm of reaction for juvenile growth rate as a function of food quality (Fig. 3a), with almost perfect evolutionary rescue (the dashed horizontal line shows that the post-selection norm of reaction produced a growth rate on bloom period ‘poor’ food nearly equal to the pre-selection growth rate on ‘good’ food). The norm of reaction for adult size also evolved (Fig. 3b), but not by enough to ‘rescue’ adult body size from the decline in food quality, resulting in the low EVO:ECO value in Table 1.

Why did one trait return to its initial value, but not the other? Apparently this would have been impossible. Adult body size is predicted well by food quality and juvenile growth rate (Hairston et al. 2001). Moreover, the Daphnia genotypes present before, during and after the bloom all appear to follow the same relationship (Fig. 3c); interactions between time period and other variables were all non-significant in a variety of alternative models (P > 0·33 in all cases). Trait values are apparently constrained to be near the linear regression plane in Fig. 3c: they can evolve on it, but not off it by much.

Figure 3.

Reaction norms of (a) Daphnia galeata juvenile growth rate, and (b) Daphnia galeata adult body size, as a function of food quality for genotypes from before (black circles) and after (red triangles) the cyanobacterial bloom period in Lake Constance. Data from Hairston et al. (2001). The blue line in panel (a) shows how the reaction norm evolved so that juvenile growth rate was “rescued” from poor food quality. (c) Reaction norm of Daphnia adult body size as a function of juvenile growth rate and food quality, for genotypes before (black circles), during (blue diamonds) and after (red triangles) the cyanobacterial bloom. The grey plane is the linear regression model that describes all three sets of genotypes.

In theoretical terms (Gomulkiewicz & Houle 2009, p. E218), the ‘nearly null space’ of a population's additive genetic variance–covariance matrix G defines ‘the multivariate directions effectively inaccessible to it via adaptive evolution’. To rescue both juvenile growth rate and adult body size, Daphnia would have needed to evolve in an inaccessible direction. It's important to note that this is a property of the G matrix as a whole. Looked at individually, each trait would be found to have the heritable variation required for an evolutionary response that compensated for the decline in food quality. Looked at together, it's either one or the other, but not both.


Continuing the theme that interactions complicate the picture, eco-evolutionary dynamics become much harder to predict when trait evolution affects more interactions in the food web. Our theory and experiments centre on a three-species food web, in which a predator–prey food chain is augmented by an intermediate predator (Fig. 4a). The additional predator means that the prey face a three-way evolutionary trade-off between resource capture ability and defence against two different predators.

In the absence of evolution, there is a simple and very general theoretical prediction about the purely ecological cycles that can occur in this food web (Ellner & Becks 2011): peaks in the intermediate predator lag prey peaks by less than one-quarter cycle period, while peaks in the top predator lag prey peaks by more than one-quarter cycle period (Fig. 4b). Experimental results for a rotifer–flagellate–algae chemostat system (Hiltunen et al. 2013) were in very close agreement with the theoretical prediction; Fig. 4c shows one of the three replicates, all of which matched the predicted pattern.

But when prey defence traits can evolve, the options multiply. A prey genotype can be defended or undefended against each of the two predators, creating four qualitative categories of prey. There are then 11 different possible sets of two or more prey genotypes, even if only one genotype per category can be maintained, allowing many different eco-evolutionary dynamics (Ellner & Becks 2011). Figure 5 shows some of the possibilities when defence against both predators at once is either impossible or too costly. In (a), the two predators ‘take turns’ at being exactly out of sync with the prey. This happens because the prey population evolves to have two dominant genotypes, each with effective defence against one of the predators. Each prey outbreak is dominated by one prey genotype, so it gets eaten down by only one of the predators. But heavy grazing by the one predator favours the genotype it can't eat, so the other prey genotype dominates the next prey outbreak. In the same situation with different parameter values [panel (b)], one prey genotype dominates two successive prey peaks before giving way to the other. In (c), one of the two dominant prey genotypes is vulnerable to both predators. The two genotypes are antiphase with each other, and each predator is in sync with one of the prey genotypes. In (d), all three prey genotypes persist (the totally undefended and the two types with defence) and produce chaotic dynamics. These examples all assume clonal selection in an asexual prey species; with sexual prey and defence modelled as a quantitative trait, other kinds of dynamics become possible, such as panel (e). Here, the predators again ‘take turns’ as in panel (a), but top predator peaks are no longer synchronized with prey troughs.

Figure 4.

Three-species food web with intraguild predation. (a) Diagram of the experimental food web. (b) Theoretical predictions for the phase lags between prey (open green circles), intermediate predator (solid purple triangles) and top predator (solid red circles) cycles (Ellner & Becks 2011). (c) Experimental results from Hiltunen et al. (2013) with top predator Brachionus plicatilis (rotifer), intermediate predator Oxyrrhis marina (flagellate) and prey Chlorella autotrophica (alga). Population series were scaled relative to their maximum value.

Figure 5.

Examples of theoretically possible evo-evolutionary dynamics in the three-species food web with intraguild predation (symbols as in Fig. 4).

These theoretical predictions create a conundrum: what does experimental replication mean if the prediction is that many different outcomes are possible? And even during one experiment, the dynamics might change as new prey genotypes arise by random mutation (e.g. defence at lower cost). But to the extent that it's possible, experimental results (T. Hiltunen, S. P. Ellner, G. Hooker, L. E. Jones, N. G. Hairston, Jr, unpublished) do support the prediction that adding one more species greatly expands the range of possible eco-evolutionary dynamics. Our first successful run (Fig. 6a) was a close match to Fig. 5a. We were very happy and set out to replicate it, which we did – sometimes. Other times, the results were different (Fig. 6b). Because these experiments will be the focus of another paper, I show here only these two examples. But we also observed other outcomes, including cases where the system settled into long-term dominance by one of the two predators. Those might be replicates that followed, by chance, different evolutionary pathways. Or, they might be the result of evolutionary change producing a system with bistable dynamics, whereas the system without prey evolution invariably produced the pattern in Fig. 4.

Figure 6.

Eco-evolutionary dynamics in the experimental three-species food web with intraguild predation (from T. Hiltunen, S. P. Ellner, G. Hooker, L. E. Jones, N. G. Hairston, Jr, unpublished). Symbols as in Fig. 4).

The devil is in the details, but which details?

In the three-species food web, the crucial detail is which combinations of defence traits are feasible and not so costly that lower predation risk isn't worth the cost. This is a statement about the trade-off between defence and other fitness components. An important but relatively neglected aspect of such trade-offs is how they depend on the states of the organism and its environment. One observed pattern is that defence cost may become unmeasurably small when resources are abundant, and large when resources are scarce. Kraaijeveld & Godfray (1997) observed this for the trade-off between parasitoid defence and larval competitive ability in Drosophila, and Yoshida et al. (2004) observed this for the trade-off between predator defence and resource uptake rate in Chlorella vulgaris. But the opposite pattern has also been observed, even without changing species: in different genotypes of Chlorella vulgaris, Meyer et al. (2006) found that the cost of defence was a decrease in the maximum population growth rate when resources are abundant. When resources are scarce, the cost disappears.

By coincidence, Yoh Iwasa and I simultaneously had PhD students studying how these two different forms of resource-dependent trade-offs would affect rapid predator–prey co-evolution. Tien & Ellner (2012) showed that if defence cost is high when resources are scarce, rapid prey evolution (relative to the predator) is destabilizing (i.e. it promotes cycles rather than coexistence at a stable equilibrium), while rapid predator evolution (relative to the prey) is stabilizing. In a very similar model, but with high defence cost when resources are abundant, Mougi & Iwasa (2010, 2011) found exactly the opposite effect of predator and prey evolutionary rates when predator handling time is low. And remarkably, that's exactly what happened experimentally. With prey evolving much faster than predators, Yoshida's Chlorella genotypes exhibited the predator–prey cycles shown above, while Meyer's converged to a steady state (Meyer et al. 2006).

Other important details are interactions between plastic and heritable trait change (Chevin, Lande & Mace 2010; Cortez 2011; Yamamichi, Yoshida & Sasaki 2011; Kovach-Orr & Fussmann 2013), population size and genetic variance (Martin et al. 2013; Chevin, Lande & Mace 2010), spatial heterogeneity and gene flow (Schiffers et al. 2013), and recent exposure to the same selective force (Bruno et al. 2011; Gonzalez & Bell 2013). On top of these details about the trait are specifics of the ecological context and properties of interest. Bassar et al. (2010) evaluated how multiple ecosystem properties were affected by the evolutionary and ecological (density) differences between populations of Trinidad guppies from streams with and without predators. The EVO:ECO ratio (computed from the treatment means in their Table S6) varied from ≈0·2 for NO3 and Total N flux to 2 or larger for community respiration, decomposition rate, PO4 flux, chirnonomid abundance and large benthic organic matter particles (>250μ m). Similarly, Turley et al. (2013) found that Rumex acetosa growth rate decreased 30% during 26 years without herbivory, but there was no change in competitive ability. Given that so many details have been found to matter (at least in theory) so quickly, we can worry that many more are yet to come or hope that the list is nearly complete. And we can hope that the list of truly important details will begin to shrink, as more experimental evidence comes in.


Prediction is an increasingly important goal for ecologists. ‘Predicting the impacts of climate change on species is one of the biggest challenges that ecologists face’ (Gilman et al. 2010); ‘…ways to foresee critical transitions ranging from epileptic seizures to the collapse of fish stocks or tipping elements of the Earth climate system, rank high in their importance to humanity’ (Biggs, Carpenter & Brock 2009).

Sometimes forecasting is easy. Fig. 7 shows the forecast based on recent trends for one (real) population. It appears to be a recent invasive or maybe a species that likes a warming climate. Using a standard count-based population viability analysis, bootstrapped projections give us 90% confidence that the population will continue to grow, exceeding 100 within the next 30 years.

But even when forecasting is easy, the prediction can be wrong. The population in Fig. 7 is the percentage of papers concerned with climate change in the basic science journals of the Ecological Society of America (Ecology, Ecological Monographs), which can never exceed 100. The UK trend (Journal of Ecology, Journal of Animal Ecology, Functional Ecology) is similar, but it takes a decade or two longer until climate change is projected to crowd out every other topic.

Figure 7.

Percentage of papers in Ecology and Ecological Monographs concerned with global warming. The filled circles and solid line are the data. The open circles are a projection using the linear regression of log-transformed data against time. Dashed lines without symbols are the 10th and 90th percentiles of 25 000 replicate projections obtained by sampling (with replacement) 5-year finite rates of ’population growth’ from the set of observed values. Papers were identified by searching the Thomson Reuters Web of Science data base (accessed 12 October 2012) for papers in those journals with Topic = ’climate change’ OR ’global warming’ OR ’climate warming’ in 5-year periods (1986–1990, 1991–1995, ⋯, 2006–2010), and searching for all papers in those journals during the same periods. Only a very small fraction of the identified papers dealt with historical (pre-industrial) climate change, and no attempt was made to remove those papers from the total.

What these data really show is something we already all know: an increasing fraction of ecological research, experimental and theoretical, is motivated or justified with reference to climate change. The study of rapid evolution is no exception:

  1. ‘… little is known about the adaptive capacity of species to respond to an acidified ocean, and, as a result, predictions regarding future ecosystem responses remain incomplete’ (Pespeni et al., 2013, p. 6937);
  2. ‘… we need operational predictors of when adaptation will promote [species] persistence or compromise it’ (Gonzalez et al. 2013, p. 6).
  3. ‘… What determines the maximum rate of environmental change that populations can cope with? Understanding this will inform both models and management plans’ (Chevin, Lande & Mace 2010, p. 1).

To what extent ecological discoveries will actually inform management plans and climate policy remains to be seen.1 But relative to other aspects of ecological response to climate change, rapid evolution has the advantage that it is happening now and has been happening for a long time. We don't need to imagine a future in which species must quickly adapt to new conditions. We can study how it happens now (as many are doing) and how it happened in the past [e.g. Hairston et al. (1999); Lavergne et al. (2013); Orsini et al. (2013); Quintero & Wiens (2013)]. Another advantage is that rapid evolution would be an important ecological phenomenon even without a climate meltdown, if we avert it.

The abundance of eco–evo sessions at recent meetings suggest that many ecologists see opportunities for useful research on rapid evolution. I agree, because the principles I've proposed in this paper are mostly a litany of contingency, complexity and ignorance. Hanski (2012) concludes that ‘Eco-evolutionary dynamics may facilitate the persistence of species in changing environments, but typically the evolutionary response only partially compensates for the negative ecological consequences of adverse environmental changes’. I think it's actually too soon to conclude anything. Adaptive evolution will certainly be part of the response to climate change, but its impact spans the range from failure to complete compensation (Table 1). And the sample size is so small that doubling it could overturn any generalizations we make now.

To close this gap in our knowledge, Schoener (2011) recommends an extensive experimental research effort, especially long-term field experiments, in a variety of communities. Up to now, this inductive approach is the primary one that has been used. It is essential and it must continue. But I don't think it's enough. What it gives us is, in effect, a histogram of effect sizes for rapid evolution. With enough case studies, we can eventually estimate the mean and variance. If a wide enough variety of communities and species is studied, a meta-analysis might tell us which species, in which communities or other circumstances, are most likely to exhibit large effects. But this will only be successful if the important covariates get measured, even though we don't yet know what they are (effective population size? life history? habitat fragmentation?). And it would still be a statistical ‘fishing expedition’ with the attendant risks and difficulties of interpretation that result from testing many possible correlations (Ioannidis et al. 2001).

A more rapidly productive approach, I believe, would begin with theoretical work leading to robust a priori predictions of circumstances where rapid evolution would and would not have important effects, both in general and specifically in response to climate change. Experiments can then be targeted at testing specific predictions, rather than scattered across biomes and taxa (as would be optimal for a meta-analysis approach). Laboratory systems with short-lived organisms are probably the best place to start, as in the recent experimental verification of the theoretical prediction that eco-evolutionary dynamics can favour cooperating genotypes at the edge of an expanding population (Datta et al. 2013).

One important outcome of this research programme would be finding out whether genuinely predictive theory is actually possible. If we fail in the laboratory, success in the field is unlikely. Success in the laboratory would motivate and inform the more difficult, but more meaningful tests in the field. Like Benton et al. (2007), I believe that microcosms can play an important role in addressing global-scale environmental issues. Using microcosms to test theory has the tremendous advantage that there is no room for excuses. If your data don't match your theory, it's not because of species that you didn't monitor or unexpected exogenous factors impinging on your experiment. It's because your theory is wrong, and you need to fix it. So when answers are needed fast, an experiment on single-celled green trees might be the best option. Michael Turelli once said, Before you ask the world to believe your theory, make sure that your computer does. And then, before you do a big expensive field study, make sure that some microbes or insects believe you. The results may be as surprising to you as ours were to us, and they might change your world view as much as our experiments changed ours, in ways that you never expected.


Many people contributed to the research presented here, but none of them should be held accountable for my interpretations and opinions. All of the research has been done in collaboration with Nelson G. Hairston Jr., and with the outstanding postdocs (Gregor Fussmann, Takehito Yoshida, Laura Jones, Lutz Becks, Teppo Hiltunen, Brooks Miner, Masato Yamamichi), students (Kyle Shertzer, Paul Schliekelman, Justin Meyer, Rebecca Tien, Michael Cortez) and colleagues (including Monica Geber and Giles Hooker) who we have been fortunate to work with on rapid evolution, and many undergraduates who assisted with the experiments. Colleen Kearns ran the first two chemostats to prove that we (i.e. she) could actually do it. Funding was provided by the Andrew W. Mellon Foundation, the James S. McDonnell Foundation and the US National Science Foundation. I thank Nelson G. Hairston Jr., Andrew Beckerman and an anonymous referee for very helpful comments on the paper. Finally, I thank the BES and the organizers of the 2012 Annual Meeting for the opportunity to give the Tansley Lecture. The author has no competing interests to declare.

Data Accessibility

Data used in the paper and R scripts to generate the figures are archived on figshare. http://dx.doi.org/10.6084/m9.figshare.744841(Ellner 2013)


  1. 1

    Given this uncertainty, evolutionary theory suggests that ecologists should hedge our bets by also allocating some time and effort to other ways of affecting climate policy.