Chapter 3. Coevolution of polygenic traits Biological Motivation In the previous chapter we considered the process of coevolution for those cases where a single gene of major effect mediated the outcome of interactions. In many other well-studied cases, however, interactions are more likely mediated by quantitative traits. Take, for example, the interaction between the seed boring weevil, Curculio camelliae, and its host plant the Japanese camellia, Camellia japonica. Female weevils attempt to deposit their eggs into the fruit of the Japanese camellia by first boring through the protective pericarp of the fruit with their rostrum. If the female weevil is successful, she lays an egg (EGGS?) into the fruit which then provides food for the developing larvae (REFS). Thus, this interaction has the potential to exert strong reciprocal selection on weevil and plant. Studies exploring this interaction over its native range have identified substantial variation in rostrum length and pericarp thickness among populations (REFS), with some populations showing extremely exaggerated trait values (REFS). Another interesting observation made by these studies is that weevil rostrum length and camellia pericarp thickness are well-matched in only a subset of populations (REFS). Because similar patterns have been demonstrated within a wide range of systems (REFS), this chapter will focus on evaluating the extent to which simple mathematical models of coevolution based on quantitative traits can explain these commonly observed empirical patterns. Key Questions: When should coevolution cause arms races and escalation? What factors determine which species wins? Should we expect coevolution to cause traits of interacting species to match? Building a Model of Weevil-Camellia Coevolution Although the genetic basis of rostrum length in the weevil and pericarp thickness in its host plant is unknown, the continuous nature of phenotypic variation in these traits is incompatible with a single gene of major effect. Instead, the continuous nature of phenotypic variation suggests these may be quantitative traits, controlled by the action of many genes of small effect. Although we could, in principle, begin to develop a coevolutionary model which attempted to track the frequencies of all possible genotypes, such a model would be ridiculously complicated. For instance, even if as few as five diploid loci determine rostrum length and pericarp thickness, we would need 35 = 243 equations for each species even if only two alleles segregated at each locus! Obviously we will need a simpler approach if we are to gain any insight into the process of coevolution for such genetically complex traits. An obvious avenue to explore in such a case is the theoretical framework provided by quantitative genetics. The key assumption made by the quantitative genetic approach is that phenotypes are determined by the action of very many genes, each of which has only a small effect on the phenotype {Lynch, 1998 #1188;Turelli, 1994 #255}. Under such conditions, it is often reasonable to assume that the distribution of phenotypes is Gaussian and can thus be described by only its mean and variance {Lande, 1976 #500;Lande, 1979 #499;Turelli, 1994 #255}. Equally important, as long as selection is sufficiently weak for mutation to replace the genetic variation it erodes, the variance may remain approximately constant, at least over relatively short evolutionary timescales. Remarkably, when these conditions are met, we can predict evolutionary change in the mean phenotype of a population, 𝑧̅, using the classic equation: ̅ 1 𝜕𝑊 𝜕𝑧̅ ∆𝑧̅ = 𝐺𝑧 𝑊 ̅ (1) ̅ is the mean fitness of the population, and 𝑧̅ is the where 𝐺𝑧 is the additive genetic variance for trait z, 𝑊 average value of the trait, z, within the population {Lande, 1976 #500}. Because the quantitative genetic framework assumes that additive genetic variance 𝐺𝑧 is a fixed parameter, using equation (1) to predict ̅ , be calculated. Thus, the key to evolutionary change requires only that the population mean fitness, 𝑊 making coevolutionary predictions using quantitative genetics is to develop expressions for population mean fitness in each of the interacting species. Although myriad factors undoubtedly determine the fitness of weevils and camellias in the real world, our goal here is not to be comprehensive, but rather to be insightful. With this goal in mind, we can focus on writing down the simplest possible expressions for fitness in weevil and Camellia that capture the essence of their interaction. The first step in writing down these fitness expressions is to predict the outcome of an encounter between a weevil with rostrum length x and a camellia with pericarp thickness y. Fortunately, elegant experimental studies by Toju and Sota {, 2006 #1457} have quantified how weevil and camellia traits interact to determine the outcome of the interaction. Specifically, this work shows that the probability a weevil succeeds in boring through the protective pericarp of the plant fruit and depositing an egg depends on the thickness of the pericarp relative to the length of the weevil’s rostrum {Toju, 2006 #1457}. If the weevil’s rostrum is too short to penetrate the plant’s pericarp, oviposition fails; if the weevil’s rostrum is long enough to penetrate the plant’s pericarp, oviposition succeeds. Thus, the probability that a weevil with rostrum length x succeeds in an encounter with a camellia with pericarp thickness y could be reasonably modeled using the following function: 1 𝑃(𝑥, 𝑦) = 1+𝐸𝑥𝑝[𝛼(𝑥−𝑦)] (2) where the parameter α determines how sensitive weevil oviposition success is to differences in pericarp and rostrum thickness (Figure 1). Many other interactions, such as those between toxic newts and garter snakes {Brodie, 1999 #104} or between parsnips and parsnip webworms {Berenbaum, 1998 #73} also appear to be well described by a phenotype differences function like (2). Although equation (2) captures an important component of fitness — the probability that a weevil succeeds or fails to deposit eggs in the camellia —the impact weevil success (or failure) has on the fitness of weevil and camellia is also critical. This additional component of fitness can be easily modeled by assuming that successful weevil oviposition increases weevil fitness by some amount sX but reduces camellia fitness by some amount sY. Putting these two fitness components together yields expressions for the fitness consequences of an encounter between an individual weevil with rostrum length x and an individual camellia with pericarp thickness y: 𝑊𝑋 = 1 + 𝑠𝑋 𝑃(𝑥, 𝑦) (3a) 𝑊𝑌 = 1 − 𝑠𝑌 𝑃(𝑥, 𝑦) (3b) We now have simple expressions for the fitness outcome of encounters between individual weevils and camellias at our disposal. What we really need in order to capitalize on the classical framework of quantitative genetics, however, is expressions for the population mean fitness of both weevil and camellia. Population mean fitness is nothing more than the average or expected value of fitness within a given population and can thus be calculated by integrating over all possible values of fitness multiplied by the frequency with which they occur (i.e., a weighted average). As long as we are willing to assume weevils and camellias encounter one another at random, calculating population mean fitness is particularly straightforward and can be accomplished in two discrete steps. First, calculate the expected fitness of particular phenotypes within the focal species by integrating over the phenotype distribution of the interacting species: 𝐸[𝑊𝑋 (𝑥)] = ∫ 𝑊𝑋 𝜑𝑦 𝑑𝑦 = 1 + 𝑠𝑋 ∫ 𝑃(𝑥, 𝑦)𝜑(𝑦)𝑑𝑦 (4a) 𝐸[𝑊𝑌 (𝑦)] = ∫ 𝑊𝑌 𝜑𝑥 𝑑𝑥 = 1 − 𝑠𝑌 ∫ 𝑃(𝑥, 𝑦)𝜑(𝑥)𝑑𝑥 (4b) where 𝜑(𝑥) is the frequency of weevils with phenotype x and 𝜑(𝑦) is the frequency of camellias with phenotype y. Next, calculate the expected value of (4) over the phenotype distribution of the focal species: ̅𝑋 = ∫ 𝐸[𝑊𝑋 (𝑥)]𝜑𝑥 𝑑𝑥 = 1 + 𝑠𝑋 ∫ ∫ 𝑃(𝑥, 𝑦)𝜑(𝑥)𝜑(𝑦)𝑑𝑥 𝑑𝑦 𝑊 (5a) ̅𝑌 = ∫ 𝐸[𝑊𝑌 (𝑦)]𝜑𝑦 𝑑𝑦 = 1 − 𝑠𝑌 ∫ ∫ 𝑃(𝑥, 𝑦)𝜑(𝑥)𝜑(𝑦)𝑑𝑥 𝑑𝑦 𝑊 (5b) Although equations (5) look a bit frightening, remember that the double integral in each is nothing more than two weighted averages: one taken over the phenotype distribution of the interacting species and the other over the phenotype distribution of the focal species. Our challenge now, is to evaluate the integrals in (5). Remarkably, this integration is not as challenging as you might think, and does not even require reaching for a calculus text. As long as we are willing to make a simplifying assumption, the integrals boil down to nothing more than an application of basic statistical rules for taking expectations of random variables. One way to make progress, and the one I personally favor, is to assume the probability of a successful interaction, 𝑃(𝑥, 𝑦), does not depend too strongly on the phenotypes of the interacting individuals. This, of course, begs the obvious question: what do you mean by “too strongly”? Although this question can be answered in several ways, the essence of all answers is that it must be possible to approximate the logistic function (2) with the linear function: 1 𝛼 𝑃(𝑥, 𝑦) ≈ 2 − 4 (𝑥 − 𝑦) (6) while maintaining a sufficient level of accuracy. Formally, the linear approximation (6) can be derived by taking a first order Taylor Expansion of 𝑃(𝑥, 𝑦) around the point α = 0. This approximation ignores all terms of order α2 and higher which will be negligible for small values of α (Figure 2; change function). Figure 2 makes clear that this approximation is quite accurate as long as α is small and the phenotypes of the two interacting individuals are not too far apart. Remembering the latter is very important when analyzing coevolutionary models which may generate sustained evolutionary change and drive the traits of the interacting species outside the range over which the approximation is valid! Our task now is to use the linear approximation (6) to facilitate evaluation of the integrals in (5). Substituting (6) into (5) shows that the integral can now be written as: ̅𝑋 ≈ 1 + 𝑠𝑋 ∫ ∫ (1 − 𝛼 (𝑥 − 𝑦)) 𝜑(𝑥)𝜑(𝑦)𝑑𝑥 𝑑𝑦 𝑊 2 4 (7a) ̅𝑌 ≈ 1 − 𝑠𝑌 ∫ ∫ (1 − 𝛼 (𝑥 − 𝑦)) 𝜑(𝑥)𝜑(𝑦)𝑑𝑥 𝑑𝑦 𝑊 2 4 (7b) where each integral is simply taking the expectation of the approximate fitness function over one of the random variables (either x or y). Taking expectations with respect to the non-focal species yields: ̅𝑋 ≈ 1 + 𝑠𝑋 ∫ (1 − 𝛼 (𝑥 − 𝑦̅)) 𝜑(𝑥)𝑑𝑥 𝑊 2 4 (8a) ̅𝑌 ≈ 1 − 𝑠𝑌 ∫ (1 − 𝛼 (𝑥̅ − 𝑦)) 𝜑(𝑦)𝑑𝑦 𝑊 (8b) 2 4 where overbars indicate expectations. Next, taking expectations over the focal species yields: ̅𝑋 ≈ 1 + 𝑠𝑋 (1 − 𝛼 (𝑥̅ − 𝑦̅)) 𝑊 2 4 (9a) ̅𝑌 ≈ 1 − 𝑠𝑌 (1 − 𝛼 (𝑥̅ − 𝑦̅)) 𝑊 2 4 (9b) with overbars again indicating expectations. Voila! Doing nothing more than taking statistical expectations has allowed us to derive approximate expressions for mean fitness in the interacting species. Before proceeding to use mean fitness to predict coevolutionary change, it is worth taking a moment to think a bit about what these simple expressions tell us about the way in which weevil and camellia influence each other’s population mean fitness. Although crude, studying population mean fitness provides a tie to population growth and insight into how the phenotype distributions of the interacting species might impact their ecology. Two realizations are key: First, the population mean fitness of weevil and camellia depends on the extent to which their population mean egg colorations match. The closer the average egg coloration of the two species, the lower the fitness of the warbler and the greater the fitness of the cuckoo. This result is quite intuitive and suggests that close phenotype matching may have negative impacts on warbler population size but positive impacts on cuckoo population size. With expressions for population mean fitness of weevil and camellia in hand, we can now make the final step toward predicting the dynamics of coevolution between these species. Specifically, we need only substitute (10) into the classical quantitative genetics expression (1), evaluate the partial derivative, and ignore all higher order terms in α, which are negligible under our assumption that α is small. Making these steps yields the following expressions for coevolutionary change in population mean values of weevil rostrum length, x, and camellia pericarp thickness, y: ∆𝑥̅ ≈ 𝑆𝑋 𝐺𝑋 (10a) ∆𝑦̅ ≈ 𝑆𝑌 𝐺𝑌 (10b) 𝑠 𝑋 where 𝑆𝑋 = 𝛼 2(2+𝑠 𝑋) 𝑠 𝑌 and 𝑆𝑌 = 𝛼 2(2−𝑠 . ) 𝑌 Analyzing the Model We are now in a position to study the process of coevolution between weevil rostrum length and camellia pericarp thickness. Because it is the difference between traits which matters for the outcome of interactions, we define the new variable 𝛿 = 𝑦̅ − 𝑥̅ as we did in the previous section, yielding a very simple expression for the change in δ which occurs over a single generation: ∆𝛿 = ∆𝑦̅ − ∆𝑥̅ = 𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 = 𝐾 (11) Because this recursion equation is so very simple, we can derive a complete solution for the entire timecourse of coevolution by simply solving it directly: 𝛿1 = 𝛿0 + 𝐾 𝛿2 = 𝛿1 + 𝐾 = 𝛿0 + 2𝐾 𝛿3 = 𝛿2 + 𝐾 = 𝛿0 + 3𝐾 ⋮ 𝛿𝑡 = 𝛿0 + 𝐾𝑡 = 𝛿0 + (𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 )𝑡 (12) Inspection of (12) shows that coevolutionary dynamics are even simpler than they were for the case of the cuckoo and warbler we studied previously, although they continue to depend solely on the quantity, 𝐾 = 𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 . If this quantity is positive, the mean pericarp thickness of the camellia population is increasing more rapidly than is the mean rostrum length of the weevil population and the weevil is losing the coevolutionary race (Figure 3a). In contrast, if this quantity is negative, the mean rostrum length of the weevil population is increasing more rapidly than is the mean pericarp thickness of the camellia population and the camellia is losing the evolutionary race (Figure 3b). Specifically, equation (12) shows that no matter what the value of K, trait values increase in perpetuity for both camellia and weevil; there is no possibility for decreases in trait values, nor is there any possibility of a coevolutionary equilibrium. Instead, our model predicts that camellia and weevil will become inextricably linked in a perpetual coevolutionary arms-race with the winner being the species which can increase its trait values most rapidly. Answers to Key Questions: When should coevolution cause arms races and escalation? Our analysis of coevolution inspired by the interaction between the seed boring weevil C. camelliae and its host plant C. japonica suggests that coevolution should invariably lead to a coevolutionary arms race and perpetual increases in rostrum length and pericarp thickness (Figure X). The rate at which this arms race proceeds and traits become exaggerated depends on the strength of reciprocal selection and the amount of additive genetic variance for the traits. What factors determine which species wins? The results of our simple analysis suggest that the species which wins the coevolutionary arms race is the one with the largest response to reciprocal selection, measured by the quantity, 𝑆𝐺. Thus, there are two factors that need to be considered before predicting which species will win: additive genetic variance and the Should we expect coevolution to cause traits of interacting species to match? Although there is some intuitive reason to expect traits of interacting species will tend to evolve well-matched phenotypes, our results do not support this idea for interactions like those between C. camelliae and C. japonica. Instead, our results suggest that the traits of the interacting species will be ever increasing, with a general trend toward being more dissimilar over time. New Questions Arising: Our simple model of weevil camellia coevolution provides us with valuable insights into the dynamics and outcome of coevolution mediated by quantitative traits. We now have some idea which biological conditions favor perpetual coevolutionary arms-races and determine which species is likely to ultimately win. We have also learned that there may be little reason to expect interactions such as those between weevils and camellias to result in closely matched traits. At the same time, however, our model analysis raises several important new questions: What keeps traits from zooming off to infinity? Do our results hold for antagonistic interactions with a different mechanistic basis? Do our results extend to other types of ecological interactions? In the next three sections, we will develop generalizations of our simple model which allow us to answer these questions and gain further insight into the process of coevolution mediated by quantitative traits. Generalizations Generalization 1: What keeps traits from zooming off to infinity? As you read through the analysis of the previous sections and studied Figure 3, you likely noticed that trait values were unbounded and could increase or decrease in perpetuity. Almost surely, this prediction struck you as unrealistic and made you suspect something important was missing from our model. One of the most obvious omissions is, of course, selective constraint. For instance, producing a very thick protective pericarp or a freakishly long rostrum might very well be energetically expensive and result in reduced fitness. Under such conditions, we might expect stabilizing selection to act on the traits mediating the interaction favoring intermediate levels of pericarp thickness and rostrum length. In this section, we integrate selective constraints into the model of weevil-Camellia coevolution developed in the previous section and explore how the inclusion of constraint alters our expectations for coevolution. Stabilizing selection can be readily incorporated into a quantitative genetic model by assuming the fitness of an individual decreases as its phenotype becomes more distant from some optimum value, θ. Often, it is assumed that stabilizing selection is of a Gaussian form such that fitness is given by: 𝑊 = exp[−𝛾(𝑧 − 𝜃)2 ] (13) where 𝛾 measures the strength of stabilizing selection, z is the phenotype of an individual, and θ is the phenotypic optimum (Figure 6). If we assume that stabilizing selection and species interactions influence fitness independently, individual fitness can be found by multiplying the fitness consequences of constraint (19) by the fitness consequences of biotic interactions (3). For our scenario of coevolution between the weevil and Camellia, this yields the following expressions for the fitness consequences of an encounter between an individual weevil with rostrum length x and an individual Camellia with pericarp thickness y: 𝑊𝑋 = exp[−𝛾𝑋 (𝑧𝑋 − 𝜃𝑋 )2 ](1 + 𝑠𝑋 𝑃(𝑥, 𝑦)) (14a) 𝑊𝑌 = exp[−𝛾𝑌 (𝑧𝑌 − 𝜃𝑌 )2 ](1 − 𝑠𝑌 𝑃(𝑥, 𝑦)) (14b) As was the case in the absence of constraint, our goal now is to calculate the population mean fitness of both weevil and Camellia so that coevolutionary change can be predicted using the classical quantitative genetics equation (1). Just as we did previously for our model without constraint, we can calculate population mean fitness by integrating expressions (20) twice: once over the frequency distribution of phenotypes in the interacting species, and once over the frequency distribution of phenotypes in the focal species. Our approach to evaluating these integrals will rely on our previous assumption that α is small and a new assumption that 𝛾 is also small. Biologically, these assumptions imply that the outcome of interactions between individuals does not depend too strongly on their phenotypes, and that stabilizing selection is relatively weak. Mathematically, these assumptions allow us to approximate (20) by a first order Taylor expansion near the points α = 0 and 𝛾 = 0. The result of this mathematical manipulation is the following approximate expressions for population mean fitness in weevil and Camellia: ̅𝑋 ≈ 1 + 𝑠𝑋 + ∫ ∫ (𝛼𝑠𝑋 (𝑥 − 𝑦) − 2𝛾𝑋 (2 + 𝑠𝑋 )(𝑥 − 𝜃𝑋 )2 ) 𝜑(𝑥)𝜑(𝑦)𝑑𝑥 𝑑𝑦 𝑊 2 4 =1+ 𝑠𝑋 𝛼𝑠𝑋 𝑠𝑋 (𝑥̅ − 𝑦̅) − 𝛾𝑋 (1 + ) ((𝑥̅ − 𝜃𝑋 )2 + 𝑉𝑋 ) + 2 4 2 ̅𝑌 ≈ 1 − 𝑠𝑌 + ∫ ∫ (𝛼𝑠𝑌 (𝑦 − 𝑥) − 2𝛾𝑌 (2 − 𝑠𝑌 )(𝑦 − 𝜃𝑌 )2 ) 𝜑(𝑥)𝜑(𝑦)𝑑𝑥 𝑑𝑦 𝑊 2 4 =1− (15a) (15b) 𝑠𝑌 𝛼𝑠𝑌 𝑠𝑌 (𝑥̅ − 𝑦̅) − 𝛾𝑋 (1 − ) ((𝑦̅ − 𝜃𝑌 )2 + 𝑉𝑌 ) − 2 4 2 ∆𝑥̅ ≈ 𝑆𝑋 𝐺𝑋 + 2𝛾𝑋 𝐺𝑋 (𝜃𝑋 − 𝑥̅ ) (16a) ∆𝑦̅ ≈ 𝑆𝑌 𝐺𝑌 + 2𝛾𝑌 𝐺𝑌 (𝜃𝑌 − 𝑦̅) (16b) where 𝑆𝑋 = 𝛼 𝑠𝑋 2(2+𝑠𝑋 ) and 𝑆𝑌 = 𝛼 𝑠𝑌 2(2−𝑠𝑌 ) . Although equations (22) are more complicated than those we studied in the absence of constraint, it remains possible to solve them directly and thus characterize the full time course of coevolution. In the presence of constraint, however, we cannot reduce this pair of equations to a single recursion for the variable 𝛿 = 𝑦̅ − 𝑥̅ as we did previously. Instead, we must study how the mean phenotype itself evolves in each species, which results in the following solutions for the population mean phenotypes of the coevolving species as a function of time, t: 𝑥̅ (𝑡) = 𝑦̅(𝑡) = 𝑆𝑋 +2𝛾𝑋 𝜃𝑋 −(1−2𝛾𝑋 𝐺𝑋 )𝑡 (𝑆𝑋 +2𝛾𝑋 (𝜃𝑋 −𝑥̅0 )) 2𝛾𝑋 𝑆𝑌 +2𝛾𝑌 𝜃𝑌 −(1−2𝛾𝑌 𝐺𝑌 )𝑡 (𝑆𝑌 +2𝛾𝑌 (𝜃𝑌 −𝑦̅0 )) 2𝛾𝑌 (17a) (17b) where 𝑥̅0 and 𝑦̅0 are the initial population mean phenotypes for weevil rostrum length and Camellia pericarp thickness, respectively. Inspection of equation (23) reveals that the coevolutionary dynamics of this system are very simple. Specifically, as evolutionary time progresses (i.e., increasing t), the term (1 − 2𝛾𝐺)𝑡 becomes smaller and smaller, ultimately approaching zero as 𝑡 → ∞. This occurs because the product 2𝛾𝐺 must be less than one due to our assumption of weak stabilizing selection, such that the quantity (1 − 2𝛾𝐺) is always between 0 and 1 and thus decays when raised to a high power, t. Consequently, as time proceeds, population mean phenotypes of the two species inevitably evolve to a unique equilibrium: 𝑆 𝑥̂ = 𝜃𝑋 + 2𝛾𝑋 (18a) 𝑋 𝑆 𝑦̂ = 𝜃𝑌 + 2𝛾𝑌 (18b) 𝑌 At this equilibrium, each species exceeds the optimal phenotype favored by stabilizing selection by an amount which depends on the strength of coevolutionary selection, 𝑆, relative to the strength of stabilizing selection, 𝛾. The equilibrium (24) provides us with the opportunity to understand how selective constraints influence the dynamics and outcome of coevolution between weevil and Camellia. Perhaps the single most important insight is simply that an equilibrium is now possible, and, in fact, inevitable. This is in stark contrast to our expectations for the coevolutionary process in the absence of constraint, which led us to believe that Camellia and Weevil would become locked in a coevolutionary arms race where trait means escalate in perpetuity. Furthermore, whereas this perpetual arms-race was inevitably won by the species with the greatest response to selection, we now see that the outcome of the coevolutionary process depends on relative levels of constraint in the interacting species. Specifically, at equilibrium, the phenotypic distance between mean rostrum length in weevils and mean pericarp thickness in Camellias, 𝛿 = 𝑦̅ − 𝑥̅ , is given by: 𝑆 𝑆 𝛿̂ = 𝜃𝑌 − 𝜃𝑋 + 2𝛾𝑌 − 2𝛾𝑋 𝑌 𝑋 (19) This quantity will be largest when the selective constraints acting on weevil rostrums (𝜃𝑌 , 𝛾𝑌 ) are smaller than the selective constraints acting on Camellia pericarps (𝜃𝑋 , 𝛾𝑋 ), and the strength of coevolutionary selection acting on weevil rostrum length (𝑆𝑌 ) exceeds that acting on Camellia pericarp thickness (𝑆𝑋 ). The relationship between selective constraints, coevolutionary selection, and the equilibrium difference between weevil and Camellia traits is illustrated in Figure 7. Although we have explored only how constraints influence coevolution mediated by a mechanism of phenotype differences, constraints also play an important role in interactions mediated by phenotype matching {, #946;Gavrilets, 1997 #77}. RESULT 25 HAS IMPORTANT IMPLICATIONS FOR THE GEOGRAPHIC STRUCTURE OF INTERACTIONS WHICH MAY SHED SOME LIGHT ON AVAILABLE DATA. FOR INSTANCE, IN THE CAMELLIA SYSTEM, A GMTC IS PRESENT. 25 SHOWS THIS COULD EASILY RESULT FROM SPATIALLY VARIABLE ABIOTIC OPTIMA, OR COSTS… Generalization 2: Interactions with a different mechanistic basis the interaction between the nest parasitic cuckoo Cuculus canorus and its reed warbler host Acrocephalus scirpaceus appears to depend upon egg coloration in both cuckoo and host {Aviles, 2012 #2666}. One of the most interesting aspects of this interaction is the large amount of spatial variation which exists among populations for rates of parasitism, rates of parasite egg rejection, and the degree to which the egg coloration of the parasite matches that of its host (Aviles et al. 2012). the coloration of cuckoo and warbler eggs match. If the cuckoo lays eggs which closely match the coloration of the warbler host, the warbler may fail to recognize and reject these eggs and raise them as its own. Alternatively, if the cuckoo lays eggs which differ appreciably in coloration from those of the warbler, the warbler may recognize them and eject them from the nest. Thus the probability that a cuckoo with egg coloration x succeeds when it encounters a warbler with egg coloration y could be reasonably modeled using the following function: 𝑃(𝑥, 𝑦) = 𝐸𝑥𝑝[−𝛼(𝑥 − 𝑦)2 ] (20) where the parameter α determines how perceptive the warbler is to differences in egg coloration (FIGURE 1). Although equation (2) captures an important component of fitness — the probability that a cuckoo succeeds or fails to have its eggs reared by the warbler —the impact cuckoo success (or failure) has on the fitness of cuckoo and warbler is also critical. This additional component of fitness can be easily modeled by assuming that successful cuckoo parasitism increases cuckoo fitness by some amount sX but reduces warbler fitness by some amount sY. Putting these two fitness components together yields expressions for the fitness consequences of an encounter between an individual cuckoo with egg coloration x and an individual warbler with egg coloration y: 𝑊𝑋 = 1 + 𝑠𝑋 𝑃(𝑥, 𝑦) (21a) 𝑊𝑌 = 1 − 𝑠𝑌 𝑃(𝑥, 𝑦) (21b) We now have simple expressions for the fitness outcome of encounters between individual cuckoos and warblers at our disposal. What we really need in order to capitalize on the classical framework of quantitative genetics, however, is expressions for the population mean fitness of both cuckoo and warbler. Population mean fitness is nothing more than the average or expected value of fitness within a given population and can thus be calculated by integrating over all possible values of fitness multiplied by the frequency with which they occur (i.e., a weighted average). As long as we are willing to assume cuckoos and warblers encounter one another at random with respect to egg coloration, calculating population mean fitness is particularly straightforward and can be accomplished in two discrete steps. First, calculate the expected fitness of particular phenotypes within the focal species by integrating over the phenotype distribution of the interacting species: 𝐸[𝑊𝑋 (𝑥)] = ∫ 𝑊𝑋 𝜑𝑦 𝑑𝑦 = 1 + 𝑠𝑋 ∫ 𝑃(𝑥, 𝑦)𝜑(𝑦)𝑑𝑦 (22a) 𝐸[𝑊𝑌 (𝑦)] = ∫ 𝑊𝑌 𝜑𝑥 𝑑𝑥 = 1 − 𝑠𝑌 ∫ 𝑃(𝑥, 𝑦)𝜑(𝑥)𝑑𝑥 (22b) where 𝜑(𝑥) is the frequency of cuckoos with phenotype x and 𝜑(𝑦) is the frequency of warblers with phenotype y. Next, calculate the expected value of (4) over the phenotype distribution of the focal species: ̅𝑋 = ∫ 𝐸[𝑊𝑋 (𝑥)]𝜑𝑥 𝑑𝑥 = 1 + 𝑠𝑋 ∫ ∫ 𝑃(𝑥, 𝑦)𝜑(𝑥)𝜑(𝑦)𝑑𝑥 𝑑𝑦 𝑊 (23a) ̅𝑌 = ∫ 𝐸[𝑊𝑌 (𝑦)]𝜑𝑦 𝑑𝑦 = 1 − 𝑠𝑌 ∫ ∫ 𝑃(𝑥, 𝑦)𝜑(𝑥)𝜑(𝑦)𝑑𝑥 𝑑𝑦 𝑊 (23b) Although equations (5) look a bit frightening, remember that the double integral in each is nothing more than two weighted averages: one taken over the phenotype distribution of the interacting species and the other over the phenotype distribution of the focal species. Our challenge now, is to evaluate the integrals in (5). Remarkably, this integration is not as challenging as you might think, and does not even require reaching for a calculus text. As long as we are willing to make a simplifying assumption, the integrals boil down to nothing more than an application of basic statistical rules for taking expectations of random variables. One way to make progress, and the one I personally favor, is to assume the probability of a successful interaction, 𝑃(𝑥, 𝑦), does not depend too strongly on the phenotypes of the interacting individuals. This, of course, begs the obvious question: what do you mean by “too strongly”? Although this question can be answered in several ways, the essence of all answers is that it must be possible to approximate the Gaussian function (2) with the quadratic function: 𝑃(𝑥, 𝑦) ≈ 1 − 𝛼(𝑥 − 𝑦)2 (24) while maintaining a sufficient level of accuracy. Formally, the quadratic approximation (6) can be derived by taking a first order Taylor Expansion of 𝑃(𝑥, 𝑦) around the point α = 0. This approximation ignores all terms of order α2 and higher which will be negligible for small values of α (Figure 2). Figure 2 makes clear that this approximation is quite accurate as long as α is small and the phenotypes of the two interacting individuals are not too far apart. Remembering the latter is very important when analyzing coevolutionary models which may generate sustained evolutionary change and drive the traits of the interacting species outside the range over which the approximation is valid! Our task now is to use the quadratic approximation (6) to facilitate evaluation of the integrals in (5). Substituting (6) into (5) shows that the integral can now be written as: ̅𝑋 ≈ 1 + 𝑠𝑋 ∫ ∫(1 − 𝛼(𝑥 2 − 2𝑥𝑦 + 𝑦 2 ))𝜑(𝑥)𝜑(𝑦)𝑑𝑥 𝑑𝑦 𝑊 (25a) ̅𝑌 ≈ 1 − 𝑠𝑌 ∫ ∫(1 − 𝛼(𝑥 2 − 2𝑥𝑦 + 𝑦 2 ))𝜑(𝑥)𝜑(𝑦)𝑑𝑥 𝑑𝑦 𝑊 (25b) where each integral is simply taking the expectation of the approximate fitness function over one of the random variables (either x or y). Taking expectations with respect to the non-focal species yields: ̅̅̅2 − 2𝑥̅ 𝑦 + 𝑦 2 )) 𝜑(𝑥)𝑑𝑥 ̅𝑋 ≈ 1 + 𝑠𝑋 ∫ (1 − 𝛼(𝑥 𝑊 (25a) ̅𝑌 ≈ 1 − 𝑠𝑌 ∫ (1 − 𝛼(𝑥 2 − 2𝑥𝑦̅ + ̅̅̅ 𝑊 𝑦 2 )) 𝜑(𝑦)𝑑𝑦 (25b) where overbars indicate expectations. Next, taking expectations over the focal species yields: ̅̅̅2 − 2𝑥̅ 𝑦̅ + ̅̅̅ ̅𝑋 ≈ 1 + 𝑠𝑋 (1 − 𝛼(𝑥 𝑊 𝑦 2 )) (26a) ̅̅̅2 − 2𝑥̅ 𝑦̅ + ̅̅̅ ̅𝑌 ≈ 1 − 𝑠𝑌 (1 − 𝛼(𝑥 𝑊 𝑦 2 )) (26b) with overbars again indicating expectations. The last step in calculating mean fitness is to remember a basic identity from statistics: 𝑉[𝑥] = 𝐸[𝑥 2 ] − 𝐸[𝑥]2 . Using this identity to replace the terms ̅̅̅ 𝑥 2 and ̅̅̅ 𝑦2 and performing a small bit of algebra yields the following more insightful and useful expressions for mean fitness: MANY FOUND THE STRETCH ABOVE CONFUSING… ̅𝑋 ≈ 1 + 𝑠𝑋 (1 − 𝛼(𝑥̅ − 𝑦̅)2 − 𝛼(𝑉𝑥 + 𝑉𝑦 )) 𝑊 (27a) ̅𝑌 ≈ 1 − 𝑠𝑌 (1 − 𝛼(𝑥̅ − 𝑦̅)2 − 𝛼(𝑉𝑥 + 𝑉𝑦 )) 𝑊 (27b) where Vx and Vy measure the phenotypic variance for trait x and trait y, respectively. Before proceeding to use mean fitness to predict coevolutionary change, it is worth taking a moment to think a bit about what these simple expressions tell us about the way in which cuckoo and warbler influence each other’s population mean fitness. Although crude, studying population mean fitness provides a tie to population growth and insight into how the phenotype distributions of the interacting species might impact their ecology. Two realizations are key: First, the population mean fitness of cuckoo and warbler depends on the extent to which their population mean egg colorations match. The closer the average egg coloration of the two species, the lower the fitness of the warbler and the greater the fitness of the cuckoo. This result is quite intuitive and suggests that close phenotype matching may have negative impacts on warbler population size but positive impacts on cuckoo population size. Second, and much less intuitively, the population mean fitness of cuckoo and warbler depends on levels of phenotypic variation in egg coloration in the two species. The more phenotypically variable the egg coloration of the two species, the greater the population mean fitness of the warbler and the lesser the population mean fitness of the cuckoo. The reason for this interesting result is a fundamental difference in the form of selection coevolution imposes on the two species. Specifically, when coevolutionary interactions are mediated by a mechanism of phenotype matching like that which defines the interaction between cuckoo and warbler, the species which does not want to match (in this case the warbler) experiences disruptive selection whereas the species which does want to match (in this case the cuckoo) experiences stabilizing selection {Nuismer, 2005 #1454;Kopp, 2006 #1643}. Consequently, phenotypic variance is “good” for the species which does not want to match (warbler) but “bad” for the species which wants to match (cuckoo). This qualitative difference in the form of selection acting on the two species has interesting consequences for levels of standing genetic variation in the two species {Nuismer, 2005 #1454;Kopp, 2006 #1643}, patterns of diversification across space {Yoder, 2010 #2309}, and potentially even the propensity for sympatric speciation (REFS). With expressions for population mean fitness of cuckoo and warbler in hand, we can now make the final step toward predicting the dynamics of coevolution between these species. Specifically, we need only substitute (10) into the classical quantitative genetics expression (1), evaluate the partial derivative, and ignore all higher order terms in α, which are negligible under our assumption that α is small. The result is a set of very simple equations predicting how cuckoo and warbler egg coloration will coevolve: ∆𝑥̅ ≈ 𝑆𝑋 𝐺𝑋 (𝑦̅ − 𝑥̅ ) (28a) ∆𝑦̅ ≈ 𝑆𝑌 𝐺𝑌 (𝑦̅ − 𝑥̅ ) (28b) 𝑠 𝑠 𝑌 Where 𝑆𝑋 = 2𝛼 1+𝑠𝑋 and 𝑆𝑌 = 2𝛼 1−𝑠 measure the strength of selection for and against egg coloration 𝑋 𝑌 matching in cuckoo and warbler, respectively, and GX and GY measure the additive genetic variance for cuckoo and warbler egg coloration, respectively. The pair of recursion equations defined by (11) are much simpler than those which we tackled in the previous chapter for scenarios where coevolution was mediated by a single gene of major effect. The equations are so simple, in fact, that it is possible to reduce them to a single equation which allows us to answer the biological questions posed at the beginning of this chapter. Specifically, if our goal is to understand the degree to which egg coloration of cuckoo and warbler matches, we need only study how the difference between the mean phenotype of the two species itself evolves. Defining a new variable 𝛿 = 𝑦̅ − 𝑥̅ and replacing all occurrences of 𝑦̅ − 𝑥̅ in (11) with δ, yields an expression for the change in δ which occurs over a single generation: ∆𝛿 = ∆𝑦̅ − ∆𝑥̅ = 𝑆𝑌 𝐺𝑌 (𝛿) − 𝑆𝑋 𝐺𝑋 (𝛿) = (𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 )𝛿 = 𝐾𝛿 (29) Because this recursion equation is so simple, we have the luxury of doing something we can almost never do in coevolutionary theory: deriving a general time dependent solution. This general solution can be found by following the logical progression of 𝛿𝑡 where the subscript t indicates time, or the current generation: 𝛿1 = 𝛿0 + 𝐾𝛿0 = (1 + 𝐾)𝛿0 𝛿2 = 𝛿1 + 𝐾𝛿1 = 𝛿0 + 𝐾𝛿0 + 𝐾(𝛿0 + 𝐾𝛿0 ) = (1 + 𝐾)2 𝛿0 𝛿3 = 𝛿2 + 𝐾𝛿2 = 𝛿0 + 𝐾𝛿0 + 𝐾(𝛿0 + 𝐾𝛿0 ) + 𝐾(𝛿0 + 𝐾𝛿0 + 𝐾(𝛿0 + 𝐾𝛿0 )) = (1 + 𝐾)3 𝛿0 ⋮ 𝛿𝑡 = (1 + 𝐾)𝑡 𝛿0 = (1 + 𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 )𝑡 𝛿0 (30) Ailene points out K is not obviously bounded on -1<K<inf Inspection of (13) shows that coevolutionary dynamics are extremely simple, with the outcome depending only on the sign of the quantity 𝐾 = 𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 , which must take values -1 < K < 1 because of our earlier assumption that α is small and G not too large. If this quantity is positive, the absolute distance between the population mean phenotypes of the two species increases over time, indicating that the warbler is winning the coevolutionary race and leaving the cuckoo behind (Figure 3). Unless the cuckoo is able to switch hosts, such a scenario would ultimately result in the extinction of the cuckoo since it would no longer be able to have its young reared by the warbler. In contrast, if the quantity K is negative, the absolute distance between average cuckoo and warbler egg coloration decreases over time until an equilibrium is reached where δ = 0. At this equilibrium, the cuckoo perfectly matches the coloration of its warbler host (Figure 3), such that the warbler suffers a chronic negative fitness impact from cuckoo parasitism. The simple analysis in the previous paragraph suggests the outcome of coevolution between cuckoo and warbler rests entirely on the sign of the quantity 𝐾 = 𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 . What is it about this quantity that makes it so important to the outcome of coevolution? Biologically speaking, this quantity is of central importance because it measures the relative rate of evolution in the two species for any specific distance, δ, between their mean phenotypes. For example, if the current difference in egg coloration between cuckoo and warbler is 6nm, the response to selection for the warbler is 6nm(𝑆𝑌 𝐺𝑌 ) and the response to selection in the cuckoo is 6nm(𝑆𝑋 𝐺𝑋 ). Consequently, if 𝑆𝑌 𝐺𝑌 exceeds 𝑆𝑋 𝐺𝑋 the warbler’s response to selection is greater than the cuckoo’s and the warbler will pull ahead in the coevolutionary race (Figure 3a). If, the converse is true, however, the cuckoo’s response to selection exceeds that of the warbler and it will capture the warbler in phenotype space (Figure 3b). Based on this analysis of our model we can begin to answer the questions we posed at the beginning of this chapter using what we have learned from our mathematical analysis. Generalization #3: Other forms of ecological interaction There is no doubt that antagonistic interactions such as those we have considered between cuckoos and warblers or weevils and camellias are taxonomically diverse and ecologically important. At the same time, however, antagonism is just one form of ecological interaction. For instance, competitive interactions among species are thought to structure biological communities, shape probabilities of invasion, and influence species ranges. Similarly, mutualistic interactions play a key role in the function of biological communities, facilitating seed-dispersal, pollination, and nitrogen fixation. Clearly, any theory of coevolution must accommodate these diverse and ecologically important forms of interaction. Although developing a general mathematical treatment for interactions of any ecological form is beyond the scope of this chapter, the general approach and several key insights can be illustrated by developing and analyzing a model for a specific biological interaction. Seed dispersal mutualisms are widespread and of significant ecological importance. For instance dispersal of the wild nutmeg, Virola sebifera, by chestnut mandibled toucans, Ramphastos swainsonii, plays an important role in nutmeg recruitment and range expansion within the Amazon. At the same time, nutmeg provides an important food source for toucan populations within this region {Howe, 1981 #2678;Howe, 1981 #2679}. Assuming the rate at which successful consumption and dispersal occur depends on the size of nutmeg seeds and the depth of toucan beaks, how might these traits coevolve within this mutualistic interaction? Within many seed-dispersal mutualisms, rates of successful consumption and dispersal increase as beak depth increases relative to seed size (REF). Consequently, an appropriate functional form for the interaction between nutmeg seed size, x, and toucan bill depth, y, would be very similar to the function we used previously to study coevolution between weevils and camellias: 1 𝑃(𝑥, 𝑦) = 1+𝐸𝑥𝑝[𝛼(𝑥−𝑦)] (31) where 𝛼 now measures the extent to which differences in beak depth and seed width influence the probability of consumption and dispersal (FIGURE 8). When a toucan succeeds in consuming and dispersing a nutmeg seed, the mutualistic nature of the interaction suggests that both individuals experience increased fitness: 𝑊𝑋 = 1 + 𝑠𝑋 𝑃(𝑥, 𝑦) (32a) 𝑊𝑌 = 1 + 𝑠𝑌 𝑃(𝑥, 𝑦) (32b) where 𝑠𝑋 and 𝑠𝑌 are the fitness benefits to a nutmeg of being dispersed, and to a toucan of consuming a nutmeg seed, respectively. Although the interaction between nutmeg and toucan shapes the fitness of the interacting species, we might expect that other forces contribute to fitness and constrain evolution in response to species interactions. For instance, it seems likely that nutmeg seeds which are very small or very large would have lower fitness for reasons extrinsic to the species interaction (e.g., life-history trade-offs, minimal provisioning, etc.). Similarly, it seems likely that very small or very large beak sizes would be detrimental to Toucan fitness for reasons unrelated to interactions with wild nutmeg (e.g., flight performance, ability to utilize alternative food sources, sexual selection). Consequently, it seems prudent to include stabilizing selection acting on the traits mediating the interaction as we did in the previous section such that the fitness consequences of an encounter between an individual nutmeg with seed size x and an individual Toucan with beak depth y are given by: 𝑊𝑋 = exp[−𝛾𝑋 (𝑧𝑋 − 𝜃𝑋 )2 ](1 + 𝑠𝑋 𝑃(𝑥, 𝑦)) (33a) 𝑊𝑌 = exp[−𝛾𝑌 (𝑧𝑌 − 𝜃𝑌 )2 ](1 + 𝑠𝑌 𝑃(𝑥, 𝑦)) (33b) where 𝛾𝑋 and 𝛾𝑌 are the strengths of stabilizing selection acting on seed size and beak depth respectively. Taking a step back and looking at equations (28), it becomes apparent that they are structurally identical to the equations we used to study the interaction between weevil and Camellia. The only difference being that now both species benefit from the interaction such that the sign preceding the parameter 𝑆𝑋 in (28a) is now positive rather than negative. Consequently, coevolution between nutmeg and Toucan must be described by an identical set of recursion equations differing only in the sign of the compound parameter, 𝑆𝑋 : ∆𝑥̅ ≈ −𝑆𝑋 𝐺𝑋 + 2𝛾𝑋 𝐺𝑋 (𝜃𝑋 − 𝑥̅ ) (34a) ∆𝑦̅ ≈ 𝑆𝑌 𝐺𝑌 + 2𝛾𝑌 𝐺𝑌 (𝜃𝑌 − 𝑦̅) 𝑠 (34b) 𝑠 𝑋 where 𝑆𝑋 = 𝛼 2(2+𝑠 𝑋) 𝑌 and 𝑆𝑌 = 𝛼 2(2+𝑠 . As was the case for equations (22) describing ) 𝑌 coevolutionary change between Weevil and Camellia, solving these recursion equations directly shows that coevolution between nutmeg and toucan always leads to a unique equilibrium where population mean seed size and beak depth are given by: 𝑆 𝑥̂ = 𝜃𝑋 − 2𝛾𝑋 𝑋 𝑆 𝑦̂ = 𝜃𝑌 + 2𝛾𝑌 𝑌 (35a) (35b) As expected, these expressions differ from the equilibrium solutions for antagonistic interactions only in the sign of the parameter, 𝑆𝑋 , which is now negative rather than positive. Our equilibrium solutions for population mean trait values in nutmeg and toucan provide two insights into mutualistic coevolution. First, we expect coevolution to drive toucan beak depths to values greater than the optimal value favored by stabilizing selection and nutmeg seed sizes to values smaller than those favored by stabilizing selection. This result makes sense given our original assumption that toucans with greater beak depth can consume a greater range of nutmeg seed sizes and nutmegs with smaller seeds can be consumed and dispersed by a greater range of toucan beak sizes (i.e., Figure 8). Second, in contrast to antagonistic interactions, both nutmeg and toucan evolve in a way that increases the rate of successful interactions between species (FIGURE 9). The reason for this result, which is actually quite general across mutualistic interactions, is that the fitness interests of both species are aligned such that both evolve to enhance the efficacy of the interaction. Conclusions and Synthesis HERE PRESENT A GENERALIZED MODEL WITH TAYLOR SERIES GENERAL FUNCTIONS… References Figure Legends Figure 1. The interaction function 𝑃(𝑥, 𝑦) for three different values of the parameter α. Larger values of α cause the probability of successful nest parasitism to fall off more rapidly with differences in cuckoo and warbler egg coloration. Figure 2. The exact interaction function 𝑃(𝑥, 𝑦) (solid line), and its quadratic approximation (dashed line), for three different values of the parameter α. As the value of α increases, the range of phenotypes over which the quadratic approximation remains accurate decreases. Figure 3. The coevolution of the difference between population mean egg coloration in warbler and cuckoo, 𝛿 = 𝑦̅ − 𝑥̅ , over 200 generations. The solid line shows a case where the composite parameter 𝐾 = 𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 > 0, indicating the warbler has a greater response to selection. In this case, the mean egg coloration of the warbler population increases more rapidly than that of the cuckoo population. The dashed line shows a case where the composite parameter 𝐾 = 𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 < 0, indicating the cuckoo has a greater response to selection. In this case, the egg coloration of the cuckoo increases more rapidly than that of the warbler population, resulting in a matching equilibrium where the egg coloration of the two species is identical. Figure 4. The interaction function 𝑃(𝑥, 𝑦) for three different values of the parameter α. Larger values of α cause the probability of successful weevil oviposition to decline more rapidly as pericarp thickness increases relative to rostrum length. Figure 5. The coevolution of the difference between population mean pericarp thickness in the Camellia and rostrum length in the weevil, 𝛿 = 𝑦̅ − 𝑥̅ , over 100 generations. The solid line shows a case where the composite parameter 𝐾 = 𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 > 0, indicating the Camellia has a greater response to selection. In this case, pericarp thickness increases more rapidly than rostrum length. The dashed line shows a case where the composite parameter 𝐾 = 𝑆𝑌 𝐺𝑌 − 𝑆𝑋 𝐺𝑋 < 0, indicating the weevil has a greater response to selection. In this case, rostrum length increases more rapidly than pericarp thickness. In both cases, coevolution generates sustained arms races where mean rostrum length and pericarp thickness increase in perpetuity. Figure 6. Gaussian stabilizing selection for three different values of the parameter γ. In all cases, the optimal phenotype is θ = 0. Figure 7. Coevolutionary dynamics of population mean trait values (left hand panels) and corresponding rates of interspecific interaction (right hand panels). In panels a and b, the weevil experiences only very weak selective constraints acting on rostrum length (γX = 0.01), whereas the Camellia pericarp is strongly constrained (γY = 0.06). As a consequence, coevolutionary selection is able to significantly exaggerate rostrum length but not pericarp thickness, thus increasing the average rate of parasitism by the weevil over time. In panels c and d, weevil rostrum length is strongly constrained (γX = 0.06) whereas Camellia pericarp experiences only weak constraints (γY = 0.01). As a consequence, coevolutionary selection is able to significantly exaggerate pericarp thickness but not rostrum length, thus decreasing the average rate of parasitism by the weevil over time. Figure 8. The interaction function 𝑃(𝑥, 𝑦) for three different values of the parameter α. Larger values of α cause the probability of successful consumption and dispersal of nutmeg seeds by the toucan to decline more rapidly as seed size increases relative to beak depth. Figure 9. Coevolutionary dynamics of population mean trait values (left hand panels) and corresponding rates of interspecific interaction (right hand panels). In panels a and b, the nutmeg experiences only very weak selective constraints acting on seed size (γX = 0.01), whereas Toucan beak depth is strongly constrained (γY = 0.06). As a consequence, coevolutionary selection is able to significantly reduce seed size but not beak depth. In panels c and d, nutmeg seed size is strongly constrained (γX = 0.06) whereas Toucan bill depth experiences only weak constraints (γY = 0.01). As a consequence, coevolutionary selection is able to significantly exaggerate bill depth but not significantly reduce seed size. In both cases, the rate at which toucans consume and disperse nutmeg seeds by increases over time because coevolutionary selection drives the mean phenotypes of both species in a direction which promotes the interaction.