A history of phenotypic plasticity accelerates adaptation to a new environment

Authors

  • J. L. FIERST

    1. Center for Ecology and Evolutionary Biology, University of Oregon, Eugene, OR, USA, and Department of Biological Sciences, Florida State University, Tallahassee, FL, USA
    Search for more papers by this author

Janna L. Fierst, Center for Ecology and Evolutionary Biology, University of Oregon, Eugene, OR 97403-5289, USA. Tel.: +1 541 346 0519; fax: +1 541 346 2364; e-mail: jfierst@uoregon.edu

Abstract

Can a history of phenotypic plasticity increase the rate of adaptation to a new environment? Theory suggests it can be through two different mechanisms. Phenotypically plastic organisms can adapt rapidly to new environments through genetic assimilation, or the fluctuating environments that result in phenotypic plasticity can produce evolvable genetic architectures. In this article, I studied a model of a gene regulatory network that determined a phenotypic character in one population selected for phenotypic plasticity and a second population in a constant environment. A history of phenotypic plasticity increased the rate of adaptation in a new environment, but the amount of this increase was dependent on the strength of selection in the original environment. Phenotypic variance in the original environment predicted the adaptive capacity of the trait within, but not between, plastic and nonplastic populations. These results have implications for invasive species and ecological studies of rapid adaptation.

Introduction

Empirical studies suggest that invasive species tend to have a history of environmental disturbance (reviewed in Lee & Gelembiuk, 2008). Ecological disturbances constitute fluctuating selection pressures over evolutionary time, and evolutionary genetic theory predicts that patterns of fluctuating selection can cause genetic architectures to take three different paths (e.g. Meyers et al., 2005). When environmental fluctuations are rapid, fitness is maximized by genetic architectures that produce a broad, generalist phenotype or short-term phenotypic plasticity. When the environment changes less frequently, organisms that evolve rapidly are favoured and when environmental changes happen infrequently, populations maximize fitness by reliably producing a single phenotype.

Any of these scales of environmental fluctuation could facilitate adaptation to a new environment. Short-term changes may result in generalists that tolerate a range of new environments and then adapt subsequently. Phenotypic plasticity may be expressed in a new environment, with future adaptation occurring through genetic assimilation (Waddington, 1942, 1961). Intermediate fluctuations could result in evolvable genetic architectures that rapidly produce variation from new mutations and recombination (Lee & Gelembiuk, 2008). Long-term fluctuations may produce phenotypic and developmental canalization (Meyers et al., 2005), but these populations could also accumulate cryptic genetic variation (Gibson & Dworkin, 2004). The essential pieces of theory currently lacking are as follows: (i) which of these selective histories is likely to increase adaptation to a new environment and under what circumstances; (ii) what will the mechanism of adaptation be and (iii) given that organisms that originate from areas with similar disturbance regimes show differences in rates of adaptation to new environments, which traits will be associated with rapid adaptation. These predictions are necessary to understand the influence of ecological processes on evolutionary patterns, and for applied studies of invasive species (Gilchrist & Lee, 2007).

Baldwin (1896), Simpson (1953) and Waddington (1942, 1961) proposed that phenotypic plasticity may benefit populations in new environments. Baldwin (1896) outlined a scenario where plasticity allows persistence in a new environment and selection then shifts the phenotypic mean towards the environmental optimum. Waddington (1942, 1961) described a novel trait value that occurs through environmental induction. Eventually, a genetic trigger will replace the environmental one, and the phenotype will no longer be plastic. The central division between Baldwin and Waddington is that Waddington's process of genetic assimilation requires canalization of the novel trait (reviewed in Crispo, 2007). The idea that phenotypic plasticity may facilitate adaptation through these mechanisms is relatively well established in theoretical literature (Ancel, 1999, 2000; Lande, 2009). Genetic assimilation in laboratory populations is a classic example in evolutionary biology (Waddington, 1953, 1956), but there are few natural biological cases that fit the scenarios outlined by Baldwin or Waddington. The clearest example of the Baldwin effect was reported in dark-eyed juncos in California (Yeh & Price, 2004), where a small montane population persisted in a coastal environment through plasticity in both its breeding season and reproductive effort.

Phenotypic plasticity in itself is a type of fluctuating selection, as the genes that determine the plastic phenotype are subject to varying environmental cues and selection pressures. Under this mechanism, a history of phenotypic plasticity may result in an evolvable genetic architecture (through fluctuating selection, e.g. Kashtan & Alon, 2005; Meyers et al., 2005; Crombach et al., 2008; Draghi & Wagner, 2008, 2009). Populations with increased evolvability produce higher phenotypic variance in response to new mutations and recombination, separate from the genetic variation segregating in the population (Wagner & Altenberg, 1996). This is particularly relevant for invasive species, where invasion and segregating genetic variation appear to be decoupled (e.g. Lee et al., 2007; Le Roux & Rubinoff, 2009). If phenotypic plasticity results in phenotypic variances that permit persistence in novel environments and evolvable genetic architectures, invasion dynamics may be dominated by plastic species.

To address these questions, I modelled a gene regulatory network that determined a phenotypic character in a population evolving with phenotypic plasticity and a population in a constant environment. Individuals from these populations were then placed in new environments and evolved towards new phenotypic optima. Price et al. (2003) (after Fear & Price, 1998) envisioned adaptation through phenotypic plasticity occurring in two ways. Entry into a novel environment could affect both the plastic response and the adaptive landscape, or the novel environment could affect the plastic response without a change in landscape. Here, I modelled a different situation where entry into a new environment did not change the phenotypic value, but the fitness landscape changed. This research then tracked how effectively the populations were able to adapt to this new fitness landscape. I focused on adaptation of the genetic architecture separate from the adaptive value of plasticity because several authors have modelled the conditions under which segregating plasticity (Behera & Nanjundiah, 1995, 1996; Ancel, 1999, 2000; Price et al., 2003; Behera & Nanjundiah, 2004; Espinosa-Soto et al., 2011) and genetic assimilation (Masel, 2004; Lande, 2009) confers adaptive benefits. The impact of phenotypic plasticity on a genetic architecture and the potential for future adaptation had not been explored.

A history of phenotypic plasticity increased adaptation rate in a new environment, but the magnitude of this effect was determined by the strength of selection in the original environment. Individuals from phenotypically plastic populations showed a small increase in adaptation rate under strong selection. As selection weakened, phenotypically plastic individuals adapted more rapidly to new environments. Weak selection resulted in larger phenotypic variance, but the quantitative relationship between phenotypic variance and adaptive capacity differed between plastic and nonplastic populations. This model predicts that the relative invasive capacity of different traits could be assessed through phenotypic variance in the original environment.

Methods

The regulatory network follows a model introduced by Wagner (1994) and developed in subsequent papers by multiple authors (i.e. Wagner, 1996; Siegal & Bergman, 2002; Bergman & Siegal, 2003; Masel, 2004; Azevedo et al., 2006; Ciliberti et al., 2007a,b; MacCarthy & Bergman, 2007; Siegal et al., 2007; Draghi & Wagner, 2009; Palmer & Feldman, 2009; Espinosa-Soto & Wagner, 2010; Fierst, 2011). The network is a set of genes, each of which activates or represses other genes through transcriptional regulation. Individuals are haploid with M genes (i.e. G1,…,GM). For the simulations presented here, M = 10. Genes have expression states, si(t), where i indicates the gene and t is the time step in development. Gene expression state ranges from full repression (−1) to full activation (1). The set of transcriptional regulatory elements is represented as an MxM matrix, R, where rij is the effect that gene j has on gene i. When rij = rji = 0, genes i and j do not interact. The set r1j are the cis-regulatory transcriptional factors, such as promoters, that determine the transcriptional level of gene 1.

A population without phenotypic plasticity

Specifying the network

The simulated populations without phenotypic plasticity follow previous implementations of this model (Wagner, 1994, 1996; Siegal & Bergman, 2002; Bergman & Siegal, 2003; Masel, 2004; Azevedo et al., 2006; MacCarthy & Bergman, 2007; Siegal et al., 2007; Draghi & Wagner, 2009). Each population was initiated by randomly drawing the elements of R from a Gaussian distribution with N∼(0,1), and connectivity parameter c, where c is the probability that rij ≠ 0. For the simulations presented here, c = 0.75, resulting in approximately 25% of the rij set to zero. This results in a fairly dense network where the majority of genes regulate, and are regulated by, 5–9 genes. This initial R matrix was the genotype of a single founder individual, and the founder was cloned to create an initial population of N = 1000 individuals. This population had equal numbers of males and females. The weights of the nonzero connections specified in matrix R changed through mutation and recombination (details are given below) and within a few generations varied by individual. Mutations thus did not affect the topology of the network, but there was abundant individual variation in gene regulation.

Finding a stable founder

To compute an individual's developmental equilibrium, the gene expression states are calculated over developmental iterations. After a number of iterations, each individual reaches either a stable pattern of activated and repressed genes or a cyclical pattern. Developmental stability is defined as a stable equilibrium expression pattern, and all individuals must achieve developmental stability to be incorporated into the population. The expression states of the M genes are given as a vector S(t) = [s1(t),…,sM(t)]. The activation and repression states for the initial gene expression vector, S(0), were randomly assigned as either si(0) = −1 or 1, with an equal probability of choosing between the two. This initial vector was used as the founder individual's gene expression, and the expression states over developmental iterations were calculated as

image(1)

where

image(2)

When a >> 1, this function specifies a very steep sigmoid, and all genes are either fully activated (si(t) = 1) or fully repressed (si(t) = −1).

The individual achieves developmental stability when

image(3)

where

image(4)

and ε has a small value, here chosen as ε = 10−4. inline image is over the range (tτ,…,t) and τ = 10. Both ε and θ were chosen after values used in Siegal & Bergman (2002). The D value between the gene expression vectors SU and SV is

image(5)

Division by 4 corrects the D value so that it ranges from 0 to M. The equilibrium expression inline image is S(t) when ψ(S(t)) satisfies the condition in eqn (3).

If the founder attained a stable expression state, their genotype R and initial gene expression S(0) were cloned to form an initial population. If the founder did not attain a stable equilibrium, a new R and S(0) were generated.

Fitness

Following Siegal & Bergman (2002), fitness was determined by distance from the optimum expression state

image(6)

where σ was the strength of selection and the founder's equilibrium expression inline image was used as the phenotypic optimum, Sopt. I simulated populations under a range of σ values to measure the effect of selection on the evolution of the regulatory network.

Creating the population in subsequent generations

The population was replaced each generation through sexual reproduction. Females randomly selected a male mate and recombined the rows of their genotypes (i.e. r1j) with equal probability of selecting each row from each parent. Mutations changed the individual nonzero rij values with a probability inline image, or approximately one mutation, per network, per generation. Offspring that did not achieve developmental stability within 100 iterations were given fitness ′0′ and rejected from the population. The population size was constant through the simulations, and all populations evolved for 10 000 generations. Each set of parameters was replicated over 100 populations, and I averaged the results from these replicates.

Phenotypic plasticity

The model assumes that an individual's gene expression and phenotypic optima depend on the environment it develops in. There is biological support for this type of scenario, for example, developing in the presence of the predatory glass worm Chaoborus americanus causes the waterflea Daphnia pulex to develop defensive neckteeth (Krueger & Dodson, 1981). A comparative study of gene expression found that several morphogenetic factors and genes involved in the juvenile hormone and insulin pathways were differentially regulated when developing Daphnia were exposed to waterborne Chaoborus cues (Miyakawa et al., 2010). Daphnia developing with neckteeth are subject to decreased predation, but in areas without glass worms, Daphnia typically do not have neckteeth. There appear to be no energetic costs associated with developing these additional morphological structures (Tollrian, 1995; Scheiner & Berrigan, 1998). Instead, Daphnia with neckteeth are more visible to fish, and subject to increased predation in environments dominated by fish. In this model, there are similarly no direct energetic costs to phenotypic plasticity. Instead, individuals pay a fitness cost for deviating from the optimal environmental phenotype. This is comparable to Daphnia experiencing increased predation for failing to develop neckteeth in a Chaoborus dominated environment or developing neckteeth and experiencing increased predation in a fish-dominated environment. Plasticity in this model is therefore adaptive in the original environment.

Stability, fitness and reproduction

Populations with phenotypic plasticity had their regulatory networks specified in the same way as described earlier. Each simulation began with a single R matrix with randomly selected rij values. This matrix was cloned to form two founder individuals. Founder 1 is an individual developing in environment 1, and Founder 2 an individual developing in environment 2. Each of these founders had an initial gene expression vector, S(0), with randomly selected si(0) values, but S(0) was different between the two founder individuals. Throughout the evolution of the simulated population, an individual developing in environment 1 would have the S(0) of Founder 1, whereas an individual developing in environment 2 would have the S(0) of Founder 2.

Each founder was then subject to the same developmental stability criteria described earlier, and the procedure was repeated until both founders achieved stable equilibrium gene expression. Interactions in this network model were directed and signed; thus, gene regulation was determined by the initial activation or repression state si(0) and the regulatory relationship rij. This resulted in some portions of the network being similarly deployed in each founder, whereas some portions had similar topology but different regulatory relationships. An example for a single individual is shown in Fig. 1.

Figure 1.

 Each individual has an initial matrix of regulatory interactions, in which genes can activate (shown in the figure with arrows) or repress (shown in the figure with bars) other genes. In populations evolving with phenotypic plasticity, each individual (signified as a) can be put into either environment 1 or environment 2. The environment activates some genes and inactivates others, and genetically identical individuals in different environments will have slightly different regulatory networks, and different phenotypes. Interactions that differ between environment 1 and 2 are shown with a dotted line for Individual a in environment 2. Here, ‘1’ signifies an activated gene, and ‘−1’ signifies an inactivated gene.

The founder individuals had different S(0), and this resulted in each founder having a different inline image. The fitness of each individual was calculated according to eqn (6), with Sopt determined by S(0). Individuals in plastic populations therefore started off at the adaptive peak for each environment. Each generation, the entire population was replaced through sexual reproduction, as described earlier. Offspring were randomly placed in either environment 1 or environment 2 and received the S(0) for that environment. Their fitness was then calculated according to the Sopt for that environment.

Adaptation, mutational robustness and phenotypic variance

To measure adaptation, I implemented a computational protocol similar to that used in Draghi & Wagner (2009) and Fierst (2011). I randomly selected 10 individuals and cloned each to form a new population with N = 100. This new population was then evolved, with mutation, in a new environment with the original S(0) but a new environmental optima, Sopt. Populations with a history of phenotypic plasticity were evolved in the new environment with a single S(0), and there was therefore no segregating plasticity. This permitted direct comparison of the genetic architectures evolved through the population histories. To adequately test adaptation, I evolved the new small population towards every possible environmental optima. As there were 10 genes in the simulated networks, this gave a total of 1024 possibilities and 1023 new environmental optima. In Fig. 2, I plot results from simulated populations evaluating the effects of 10–1000 generations of adaptation under strong directional selection (σ = 1) in the new environment. For the other simulated populations, I evaluated the level of adaptation after 50 generations in the new environment.

Figure 2.

 Populations evolving under a history of phenotypic plasticity adapted more quickly to new environments. This increased adaptation was apparent from 0 to 1000 generations of adaptation. A history of phenotypic plasticity affected the rate, but not extent of adaptation, as populations from constant environments did reach the same level of adaptation after longer periods in the new environment. The populations were evolved for 10 000 generations with σ = 103 before being placed in the new environments.

I measured adaptation by assessing the hamming distance between the Sopt of the environment for the population of cloned individuals and the Sopt of the new environment. Hamming distance is a binary measure that equals 0 if si(t) = si(t)opt, and 1 if si(t)≠si(t)opt. I summed this across all si(t), and the hamming distance was the amount a population could adapt to the new environment. After adaptation to the new environment, I measured the hamming distance from Sopt for each individual, and the difference between initial and final hamming distance was adaptation to the new environment. I averaged adaptation across the 10 individuals selected from the simulated population. Individuals from plastic and nonplastic populations were tested under the same scenario and thus were, on average, the same distance from the new environmental optima. Under this strength of selection (σ = 1), individuals from both plastic and nonplastic populations had poor fitness in many of the new environments. This could have produced unequal results, where populations showed dramatic adaptation in some environments and no adaptation in other environments, but the mean adaptation and median adaptation were similar across the simulated populations. The adaptation measure relates the total adaptation to the possibility for adaptation, and both the plastic and nonplastic populations were therefore showing consistent levels of adaptation across environments.

I measured mutational robustness by implementing a protocol similar to Siegal & Bergman (2002) and Azevedo et al. (2006). I took each individual at the end of the simulation run, replicated it 100 times and replaced one nonzero rij with a new value. After mutation, each individual that attained developmental stability was robust. I measured the proportion of robust individuals that attained the same phenotype before and after mutation, and a different phenotype after mutation. I did this over five successive rounds of mutation to evaluate how individuals from constant environments and varying environments used the phenotypic landscape.

I measured genetic variance through the gene diversity (Graur & Li, 2000),

image(7)

where n is the number of alleles at a locus and pi is the frequency of allele i. I measured phenotypic variance in the simulated populations by measuring the hamming distance of each individual from the population optima. Because the phenotype in this model (gene activation/repression states) is discontinuous, phenotypic variance does not represent a functional mapping. Variance instead measures the dispersion of individuals across the phenotypic landscape. For individuals in populations with phenotypic plasticity, I measured the hamming distance from the optima for their environment.

Results

Populations evolving with a history of phenotypic plasticity adapted to new environments more rapidly than populations evolving in constant environments (Fig. 2). This accelerated adaptation occurred across 10–1000 generations of adaptation in the new environment. A history of phenotypic plasticity affected the rate of adaptation, but had a smaller effect on the extent of adaptation in the new environment. Populations from constant environments required longer periods of adaptation to reach similar levels, and after 1000 generations, the extent of adaptation was not equivalent between plastic and nonplastic populations.

The strength of stabilizing selection in the original environment affected the rate of adaptation to the new environment (Fig. 3). When the trait originally evolved under strong selection, there was a small increase in adaptation rate for populations with a history of phenotypic plasticity. For moderate to weak stabilizing selection in the original environment, this difference increased and populations with a history of phenotypic plasticity progressed towards the new environmental optimum at roughly 1.3–1.5 times the rate of populations from constant environments (over 50 generations of adaptation).

Figure 3.

 Populations with a history of phenotypic plasticity adapted to a new environment more quickly. This effect was small when selection was strong, and larger with moderate to weak stabilizing selection.

The difference in adaptation rate was due to evolution of the genetic architecture (Fig. 4). Populations with a history of phenotypic plasticity produced a greater proportion of robust individuals after random mutation (grey bars, Fig. 4), and a larger proportion of those individuals achieved a different phenotype after mutation when compared with individuals from populations in constant environments (black bars, Fig. 4). Under moderate to weak stabilizing selection, and with increasing numbers of mutations, this difference increased. A larger proportion of individuals from populations with a history of phenotypic plasticity were thus able to reach different phenotypic states through random mutations and recombination. Gene diversity was similar between populations but slightly lower in plastic populations (H = 0.40) than nonplastic populations (H = 0.43). This indicates that the differences in adaptation were not due to differences in segregating genetic variation.

Figure 4.

 Populations with a history of phenotypic plasticity (grey bars) showed increased robustness to mutation compared with populations from constant environments (black bars). The coloured portion of each bar represents individuals that achieved the same phenotypic state before and after mutation, and the white portion are individuals that were viable but had a different phenotype. With increased mutations, more individuals from phenotypically plastic populations were stable but achieved different phenotypes.

Under strong selection, there was very little phenotypic variance in either populations from constant environments or populations with a history of phenotypic plasticity (Fig. 5). Under weak to moderate stabilizing selection, phenotypic variance increased in all simulated populations. Phenotypic variance was higher in populations evolving in constant environments when compared with populations evolving with phenotypic plasticity.

Figure 5.

 Under strong selection, most individuals had phenotypes close to the optimum, but under moderate to weak stabilizing selection, phenotypic variance increased for individuals with a history of phenotypic plasticity and individuals from constant environments. Individuals from populations in constant environments had greater phenotypic variance.

Discussion

In this network model, populations evolving with a history of phenotypic plasticity adapted rapidly to new environments. Results from empirical studies of invasive species suggest that a history of disturbance may structure genetic architectures such that invasion and rapid adaption are possible (Lee & Gelembiuk, 2008). Although this may occur through different mechanisms, this model suggests that populations that respond to environmental change through phenotypic plasticity may have an increased ability to adapt to new environments. Phenotypic plasticity has typically been seen as an one type of environmental response, with the alternative being adaptation and genetic evolution (Pfennig et al., 2010). The results presented here demonstrate that a history of phenotypic plasticity may determine the evolution of genetic architecture and shorten the waiting time for the generation of phenotypic variance from new mutations and recombination. Rather than acting as a short-term alternative, phenotypic plasticity may facilitate future adaptation and genetic evolution.

Accelerated adaptation

In the simulated populations, a history of plasticity increased the rate of adaptation, but neither populations from constant environments nor populations with histories of phenotypic plasticity achieved complete adaptation to the environment. This is likely due, at least in part, to the details of the model. Kauffman & Smith (1986) studied a similar boolean circuit model and found that networks would adapt quickly but reach an asymptote at 60–90% of possible fitness when selected towards a new phenotype. The proportion of possible fitness the networks achieved was influenced by network size, connectivity, mutation and selection. In this model, populations from constant environments showed an average of 72.34% adaptation to new environments over 1000 generations, whereas populations with a history of phenotypic plasticity averaged 77.58% adaptation to new environments. It is likely that populations with a history of phenotypic plasticity would maintain a small difference in the extent of adaptation over even longer periods of adaptation to the new environment. Given the results presented here and those of Kauffman & Smith (1986), the difference would likely continue as only a few per cent. Consistent directional selection over 1000 generations or more is likely not a biologically realistic scenario, and thus, the shorter-term results encompassing 10–500 generations are probably more representative of biologically relevant adaptation.

The differences in rate of adaptation to the new environment occurred through evolution of the genetic architecture. Populations with a history of phenotypic plasticity had slightly lower segregating genetic variation but responded to new mutations with higher robustness for fitness and higher variance in phenotypes when compared with populations evolved in constant environments. Over a few generations of adaptation, populations with a history of phenotypic plasticity rapidly produced new phenotypes, some of which better fit the new environment. As these phenotypes conferred higher fitness, they quickly spread through the population and beneficial allelic combinations fixed over a number of generations. Previous models have suggested phenotypic plasticity may accelerate adaptation towards a target, but described mechanisms related to segregating plasticity (Behera & Nanjundiah, 1995, 1996; Price et al., 2003; Behera & Nanjundiah, 2004; Espinosa-Soto et al., 2011) or biologically unrealistic assumptions (Hinton & Nowlan, 1987). Segregating plasticity in a new environment results in a population that produces multiple phenotypes and can more rapidly search the fitness landscape and approach fitness peaks. While this can result in rapid adaptation, it is also not an unexpected result as the populations are essentially attacking a single problem with twice the efficiency. In this work, I have shown that phenotypic plasticity can influence adaptation in a new environment, separate from the effects of segregating plasticity. This did not occur through ‘learning’ or ‘sensing’ the environment, or through plasticity's role in ‘tuning’ the phenotype to match the environment. A biologically accurate description of phenotypic plasticity's role in adaptation in this work is that it increased the evolvability of the genetic architecture. The genetic architecture in these populations evolved so that new mutations and recombination allowed the population to quickly percolate through phenotypic space and reach adaptive phenotypes more readily.

The complex nature of the regulatory networks and the model means that I was not able to identify network metrics that describe mutational robustness or the potential for future adaptation. Kauffman (1969, 1974, 1993; Kauffman & Smith, 1986) first developed the type of regulatory network model I have described here and suggested that parameters like network size and connectivity determined the propensity for change. Further work has shown conclusively that evolved networks of similar size and connectivity can have different responses to mutation and selection, but the adaptive changes appear to be due to complex multi-way epistatic interactions that are not captured in current network metrics. Sevim & Rikvold (2008) analysed a similar model in a dynamical systems framework and found no conclusive descriptors separating evolved and random networks. Two studies (Munteanu et al., 2008; Draghi & Wagner, 2009) suggested that the balance of positive and negative connections evolves to determine the response to new mutations, but that metric is model-specific and difficult to interpret biologically. Network metrics were developed in mathematical graph theory and reflect the requirements of that field. We currently lack ways to measure networks that capture and describe biologically significant features like response to mutation. Networks are typically described by features of topology, but topology is less significant for studies of biological networks because the strength and nature of interactions varies widely and determines the outcome.

Implications for measuring adaptive potential in the field

The prediction from this study is that the relative adaptive potential of different species and traits could be assessed in the field through their relative phenotypic variances. The idea that phenotypic variance reflects adaptive capacity is not novel; however, the crucial point here is that it does not map linearly across populations with a history of phenotypic plasticity and those without a history of phenotypic plasticity. The adaptive capacity of the populations was qualitatively similar across strengths of selection. Strong selection in the original environment resulted in lower phenotypic variance and lower adaptive capacity, whereas weak to moderate selection resulted in higher phenotypic variance and adaptive capacity.

Adaptive potential was thus reflected in the levels of phenotypic variance, but the relationship between adaptive capacity and phenotypic variance was quantitatively different between the populations evolving in constant environments and populations evolving with phenotypic plasticity. Under strong selection, both types of populations showed very little phenotypic variance, but under moderate to weak stabilizing selection, populations evolving in constant environments had higher levels of phenotypic variance than populations evolving under phenotypic plasticity. This indicates that studying phenotypic variance and rate of adaptation across these populations may provide misleading results. Instead, the adaptive capacity of an organism may need to be studied within similar life histories and across traits within an organism.

The Baldwin effect and other models of adaptation through phenotypic plasticity

The framework of this model differs from previous work addressing phenotypic plasticity, genetic assimilation and the Baldwin effect, but it captures essential features regarding the adaptive value of phenotypic plasticity. In an attempt to explain the influence of environmentally induced characters in the absence of a Lamarckian-type process, Baldwin divided his theory into two processes (Baldwin, 1896; Simpson, 1953; Crispo, 2007). First, the trait produced under phenotypic plasticity would enable survival in a new environment through ‘organic selection’. Over time, natural selection would push the trait in the direction of the environmental optimum through a process he termed ‘orthoplasy’. Simpson (1953) returned to Baldwin's original writings after the modern synthesis and recast the theory in current genetic terms. There has been some debate over the essential aspects of Baldwin's original theory (reviewed in Crispo, 2007), and the names Simpson–Baldwin effect (Ancel, 1999) and Baldwin expediting effect (2000) have also been used to clarify the interpretation of Baldwin's original writings.

For Baldwin, the trait value could be produced under the existing range of plasticity but Waddington (1942, 1961) described a situation where environmental induction results in a completely novel phenotype. Waddington referred to his own studies in these descriptions, where he used heat shock and mutagenic agents to induce severe phenotypic abnormalities in Drosophila. He observed that over several generations of selection, the environmental trigger was no longer necessary, and the novel phenotype was constitutively expressed in a portion of the population. Waddington described this process as canalization, where developmental trajectories are separated by ridges that deepen as the organism responds to selection. Later work addressing canalization suggests that a requirement for developmental stability can produce canalization and genetic assimilation in the absence of selection (Siegal & Bergman, 2002; Masel, 2004).

Given these historical scenarios, the model presented here best fits the Baldwin effect. The one key difference is that Baldwin described a population that would not survive without the plastic trait. In these simulations, individuals are placed in a new environment and maintain some level of persistence despite the phenotypic value. Individuals are sometimes far from the adaptive peak and thus forced to adapt. Waddington's genetic assimilation rests heavily on a novel environmental induction that does not occur in these simulated populations. The phenotype that is placed in the new environment is produced in the original populations. However, these simulated populations likely experience a decrease in plasticity and some degree of canalization as they are subject to strong directional selection for a period of time. This network model has been used to study canalization by other authors (Wagner, 1996; Siegal & Bergman, 2002), and the requirement for a stable gene expression state drives evolution of the model towards a canalized architecture. Canalization is therefore an artefact of this model, and not explicitly modelled to simulate genetic assimilation.

The results presented here addressed one component of adaptation involving phenotypic plasticity, and in real populations, the situation would of course be much more complex. It is difficult to judge the accuracy of a model of this type, as the simplified regulatory network presented here maps to a single trait. Although the results presented here are robust over a large parameter space, it is unclear how these parameters may correlate to natural populations. Adding segregating plasticity in the new environment would be one way to extend this work to include greater complexity. Multiple authors have hypothesized that phenotypic plasticity and polyphenism may impact speciation, adaptive divergence and the evolution of novelty because of the one-to-many mapping for genotypes, developmental systems and phenotypes (West-Eberhard, 1989, 2003; Pfennig et al., 2010). Combining these hypotheses with the results of this study suggests that phenotypic plasticity may indeed play a large role in organismal diversity and evolution by shaping the genetic architecture, the production of phenotypic variance and the genetic and developmental mappings involved.

In this model, simulated populations were forced to react with phenotypic plasticity and could not evolve another way to deal with environmental change. Modelling the evolution of phenotypic plasticity and then measuring the evolvability of the resulting populations may also provide insight to how these results would extrapolate to greater complexity. Plasticity itself can also affect selection by altering the adaptive landscape and produce a feedback between environment and phenotype that shifts the evolutionary trajectory of a population (Price, 2006).

Using network models as a framework for testing ecological hypotheses

This modelling framework has been previously used to study the evolution of genetic architecture and levels of evolvability under fluctuating selection (Draghi & Wagner, 2009) and in populations with sexual dimorphism (Fierst, 2011). In those models, the rate of adaptation was higher than observed here under phenotypic plasticity. This is likely due to the frequency with which individual genotypes experienced fluctuating selection. Draghi & Wagner (2009) selected the entire population towards a different phenotypic optimum every 100 generations, whereas Fierst (2011) modelled a population where males and females were selected towards two separate optima. In this model, mating occurred randomly between the environments and individual genotypes experienced shifts in selection less frequently. Meyers et al. (2005) demonstrated that the evolution of genetic architecture under fluctuating selection is dependent on the period of fluctuation. The results presented here suggest that ecological scenarios may produce irregular shifts in selection, and that these irregular fluctuations may have quantitative influences on adaptive capacity. This network model, and similar modelling frameworks, has been used to study a wide variety of questions in evolutionary biology, but these papers have not addressed ecological scenarios. By building on an established model, I was able to test novel hypotheses that are relevant for both ecological and evolutionary studies and frame the results in an evolutionary context. Network thinking is common in ecological work, but typically used to study food webs, behaviour, and communities. The results presented here suggest that genetic network models can also inform ecological studies.

Acknowledgments

I would like to thank my advisors, Thomas F. Hansen and David Houle, for help in developing this research, and the Center for Ecological and Evolutionary Synthesis at the University of Oslo for hosting my research visit. The manuscript and figures benefited from comments from Jason Pienaar. This work was supported by a Norwegian Research Council, Leiv Eiriksson Mobility Grant, and a U.S. National Science Foundation Postdoctoral Fellowship in Biological Informatics.

Data deposited at Dryad: doi: 10.5061/dryad.58pq3

Ancillary