chapter 6: natural selection on phenotypes and hartl – p. 6-1 from: conner, j. and d. hartl, a...

Conner and Hartl – p. 6-1

From: Conner, J. and D. Hartl, A Primer of Ecological Genetics. In prep. for Sinauer

Chapter 6: Natural selection on phenotypes

Natural selection and adaptation have been recurring themes throughout this book,from the very beginning of chapter 1. We discussed selection on genotypes (and discretephenotypes) in chapter 3, and now that we have a good understanding of the genetics ofcontinuously distributed traits we turn to selection on these common and ecologicallyimportant phenotypes. We discuss the very general and widely used regression-basedapproaches to measuring selection, and cover ways to identify the phenotypic traits thatare the direct targets of selection, as well as ways to determine the environmental agentsthat are causing selection. Identifying selective agents and targets is a powerful approachto understanding adaptation. Finally, we integrate this material with the concepts coveredin chapters 4 and 5 to show how short-term phenotypic evolution can be modeled andpredicted, and how this undertaking sheds light on constraints on adaptive evolution.Throughout the chapter the effects of genetic and phenotypic correlations among traitsare highlighted.

Evolution by natural selection has three parts (Figure 6.1; Endler 1986):

1. There is phenotypic variation for the trait of interest.2. There is some consistent relationship between this phenotypic variation and variation

in fitness.3. A significant proportion of the phenotypic variation is caused by additive genetic

variance, that is, the trait is heritable.

Numbers 1 and 2 represent selection on phenotypes, which occurs within ageneration and can be quantified using the selection differential (S) just as with artificialselection. Number 3 is heritability, which can be thought of as determining the amountof phenotypic change that is passed on to the next generation, because as with artificialselection, the amount of evolutionary change, R, is determined by the product h2S.Increases in either h2 or S produce a greater evolutionary change across generations.


Figure 6.1. Graphical depiction of the three elements of phenotypic evolution by natural selection. A.Histogram showing the frequency distribution for flower number in wild radish (data in A and B fromConner et al. 1996). Flower number is standardized to have a mean of zero and a standard deviation of one.B. The relationship between flower number and fitness. The spread of points along the X-axis showsexactly the same phenotypic variation as in part A. The fitted regression line is the estimated fitnessfunction; the slope of this line is the selection differential S. C. Hypothetical offspring-parent regressionfor flower number, showing that the trait is heritable.

6.1 The Chicago school approach to phenotypic evolution

An extremely fruitful approach to studying phenotypic evolution in nature is basedon the scheme depicted in Figure 6.1. This approach is sometimes referred to as the‘Chicago School’, because it was developed at the University of Chicago in the late1970s and early 1980s (Lande 1979; Lande and Arnold 1983). In this scheme, thebivariate plot shown in Figure 6.1B is the fundamental depiction of natural selection. Onthe X-axis is the phenotype, which can be any quantitative trait in the phenotypichierarchy, including physiology, morphology, behavior, or life-history traits. For

1. Phenotypic variation

2. Consistent relationship betweenfitness and phenotype

3. Some of the phenotypic variation is heritable

Parental flower number (standardized)

Offs

prin

g flo

wer

num

ber

-3 -2 -1 0 1 2

-1

0

1

A

B

C

Fitness function; slope = S

Offspring-parent regression;

slope = h2


measurements of selection it is very useful to standardize the phenotypic values by takingeach individual value and subtracting the population mean and dividing by the populationstandard deviation. This produces measures of selection that have useful statisticalproperties and are comparable across traits and organisms (e.g., Kingsolver et al. 2001).

On the Y-axis is fitness, which is a fundamental and very difficult concept, andtherefore a number of different definitions of fitness have been proposed, each withstrengths and weaknesses (Dawkins 1982, ch. 10; Endler 1986, ch. 2; deJong 1994). Inaddition to this conceptual difficulty, in practice fitness is very difficult to measure innatural populations. For the Chicago School methods, the best measure may be lifetimenumber of offspring produced, and this is the working definition that we will use in thischapter. This is certainly an excellent practical measure of fitness for organisms withnon-overlapping generations and stable population sizes, and is better than the fitnessestimates that have been used in most field studies of selection. For the actual analysisrelative fitness values are used, which are calculated by dividing the fitness of eachindividual by the average fitness of the population. This is different from how relativefitness was calculated for genotypes (Chapter 3), but the symbol (w) and the reason arethe same – selection operates through the fitnesses of individuals relative to otherindividuals in the population, not through absolute numbers of offspring produced.

The Chicago School methods estimate the strength of natural selection by regressingfitness on the phenotype. This provides a statistical estimate of the fitness function,which describes the relationship between fitness and the phenotype; this relationship isnatural selection. There are three basic types of phenotypic selection defined accordingto the shape of the fitness function. Directional selection is characterized by a linearfitness function (a straight line), and can be positive or negative depending on whetherfitness increases or decreases with increasing phenotypic value (Figure 6.1 B and Figure6.2 A respectively). Linear regression is used to fit a line through the data, and the slopeof this line measures the strength of selection. If standardized data are used, then thisslope equals the selection differential (S). This is the same measure as is used forartificial selection (Chapter 5), but is calculated in a different way because fitnesses innatural selection are continuously distributed (truncation selection is rare or nonexistentin nature). Recall that the selection differential is for phenotypic selection, and is not thesame as the selection coefficient (lower case s) that is used to measure genotypicselection (Chapter 3). The results of directional selection are very similar to truncationselection – if the trait is heritable, it will change the mean and may decrease the variance(Fig. 6.2B).


Figure 6.2 Hypothetical fitness functions (top graphs) and the effect of each kind of selection on thefrequency distribution of the population for that trait (bottom graphs). A and B show negative directionalselection, C and D stabilizing selection, and E and F disruptive selection. In the bottom graphs the graycurves are the population distribution before selection, and the black curves after selection.

Fitness functions are not always straight lines, so when the fitness function hascurvature quadratic regression is used to estimate the strength of selection:

Y = βX +γ2

X2 6.1

In this equation, β is the slope of the fitness function (equivalent to S withstandardized data) and γ measures the degree that the slope changes with increasing X. Inother words, γ estimates the amount of curvature in the fitness function, and is variouslycalled the non-linear, variance, or stabilizing/disruptive selection gradient (non-linearis preferable, and stabilizing/disruptive should be avoided). If β = 0 and γ is negative,then there is no overall trend in the data but the slope is constantly decreasing (Fig. 6.2C);this is called stabilizing selection. The key characteristic of stabilizing selection is thatthere is an intermediate optimum for fitness; in other words, fitness is highest at someintermediate fitness and is lower at the phenotypic extremes. Stabilizing selection byitself does not change the trait mean but does decrease trait variance (Fig. 6.2D).

A particularly clear pattern of stabilizing selection occurs on the size of galls formedby flies in the genus Eurosta on goldenrod plants (Weis and Gorman 1990; Fig. 6.3).Female flies lay their eggs inside goldenrod stems, and this causes the plant cells in the

Phenotypic trait

Fitn

ess

Fre

quen

cy

A

B

C

D

E

F


area to divide profusely, producing a large, hard ball inside which the larva feeds anddevelops. A parasitoid wasp often attacks these larvae, but only in galls that are smallenough in diameter that the ovipositor of the wasp can reach the fly larva. The fly larvaeare also eaten by birds, who are more attracted visually to the larger galls. Therefore, thetwo extreme gall sizes are each attacked at a higher frequency by different predators,reducing survivorship relative to the intermediate phenotypes.

Fig. 6.3 Stabilizing selection on gall size of Eurosta flies on goldenrod. Survival of the fly larvae wasthe measure of fitness, and the phenotypic trait was gall diameter measured to the nearest mm. Over 3500galls were measured and scored for survival to adulthood; the points in the graph represent the proportionthat survived at each gall size.

If β = 0 and γ is positive, then the slope is steadily increasing with X (Fig. 6.2E);when the fitness function has this shape the term disruptive selection is used. The keycharacteristic of disruptive selection is opposite to that for stabilizing selection -- there isan intermediate minimum for fitness, i.e., fitness is lowest at some intermediate fitnessand is higher at the phenotypic extremes. Like stabilizing selection, disruptive selectionby itself does not change the population mean, but it can increase the variance, at least inthe short term (Fig. 6.2F), opposite to the effect of stabilizing selection. These effects onvariance are why γ is called the variance selection gradient, and the sign of γ indicates theeffect on variance – negative is stabilizing selection, which decreases variance, andpositive is disruptive, which increases variance.

Disruptive selection is potentially important because it can maintain phenotypic andgenetic variation in the short term, and could result in adaptive differentiation and evenspeciation if the two phenotypic extremes become reproductively isolated. However,there is little strong evidence for disruptive selection in natural populations. Some of thebest evidence comes from bill size in finches, in which disruptive selection can occurthrough feeding on seeds from different species of plants. In both Darwin’s finches fromthe Galapagos and African finches, large billed birds have high fitness through theirability to crack abundant hard seeds and small-billed birds survive well on smaller, softer


seeds. Birds with intermediate bill sizes have the lowest fitness because they cannot useeither resource as efficiently as the phenotypic extremes (Schluter et al. 1985; Smith1990).

Often more complicated forms of selection can occur that do not exactly fit thesimple definitions of directional, stabilizing, or disruptive selection. Often the overalltrend of the fitness function is increasing or decreasing without an intermediate maximumor minimum as in directional selection, but it curves rather than being linear (Figure 6.4).In this case both β and γ can be significant, so a significant γ by itself is not goodevidence for stabilizing or disruptive selection, only that the fitness function curves.Standard regression techniques can sometimes be misleading when there is curvature inthe fitness function, so related curve-fitting techniques such as cubic splines and locally-weighted least-squares can be very useful in determining the shape of the fitness function(Schluter 1988).

Figure 6.4 Example of curved fitness function without an intermediate maximum or minimum infitness. The fitness measure is number of pollen grains removed from a flower during a pollinator visit (acomponent of male fitness), and the trait is the number of pollen grains produced by the flower. This issometimes called a saturating fitness function, because after a certain point (about 30,000 grains producedin this case) further increases in trait value no longer result in increased in fitness. The curve was fit usinglocally-weighted least squares; both the linear and quadratic terms in a regression analysis were significant.Modified from Conner et al. 1995.

6.2 Selective agents and targets

The regression-based and related techniques described above provide measures ofthe strength and pattern of selection, but by themselves tell us little about the causes ofthe selection. The fitness differences along the Y-axis in a fitness function plots (Figs.6.1 – 6.4) are caused by biotic and abiotic factors in the environment called selectiveagents. Examples of selective agents are the seeds eaten by Darwin’s and Africanfinches, the bird predators and wasp parasitoids of gall flies (Fig. 6.3), and the winterstorm that killed sparrows in the classic example of natural selection reported by Bumpus(1899). However, fitness differences do not always lead to selection, because thesedifferences may be random with respect to a given phenotypic trait. For example, the

-2

-1

0

1

10,000 30,000 50,000

Resid

ual #

pol

len

rem

oved

No. pollen produced


chance of encountering a mate, getting caught in a storm, or coming in contact with adisease may be random with respect to most or all phenotypic traits. Even if there is aconsistent relationship between fitness and some traits, there may be many others that donot affect fitness in a given generation. This is one reason why it is very important tofocus on specific phenotypic traits, called targets of selection, when thinking aboutselection. In our examples above bill size in finches and gall size in flies were thetargets. Much confusion results from thinking of the strength of selection on species orpopulations in general, without reference to a certain trait.

Therefore, when trying to understand the causes of selection, we need tosimultaneously consider both the X and Y axes in the fitness function, that is, specificphenotypic traits and the causes of fitness differences. It is very useful to think ofselection as an interaction between the phenotype with the environment, i.e., between thetargets of selection and the selective agents; it is this interaction that determines the shapeof the fitness function. The target of selection is a trait that helps the organism deal withthe selective agent, which is a challenge in the environment. In the examples of naturalselection above, the fitness functions for finch bill size are determined by the interactionof the beak dimensions found in a given population and the size and hardness of seeds inthe environment. Similarly, the fitness function for Eurosta gall size depends on thedistribution of gall sizes and the predators and parasitoids in that population’senvironment. If there is additive genetic variation for the selective target, it will evolveto be an adaptation to the selective agent.

This means that changes in either the target of selection or the selective agent canchange the fitness function. Figure 6.5 shows hypothetical fitness functions for a largerange of phenotypic values, greater than the range spanned by any one population(“overall” fitness functions). The pairs of vertical dotted and dashed lines denote thephenotypic ranges for three different populations. Note that the fitness function withineach population and thus the pattern of selection depends on the phenotypic mean andvariance of the population relative to the overall fitness function. In figure 6.5Apopulations 1 and 2 experience directional selection for an increase in the trait andpopulation 3 is undergoing weak stabilizing selection. Therefore, changes in the meanand variance of the target of selection (the phenotypic trait) can change selection on thattrait. If the trait is heritable, then the trait means in populations 1 and 2 will increaseacross generations (indicated by the arrows in the figure) until their distributions aresimilar to population 3. At this point the trait mean is at the optimum and the populationswill be at an equilibrium for this trait, because unless the trait distribution or the fitnessfunction changes, the trait mean will no longer evolve. This equilibrium is stable, becauseif something perturbs the trait mean away from the optimum, selection will return themean to the optimum.


Figure 6.5 Two hypothetical fitness functions, with the phenotypic positions of three populations mappedonto each. The pairs of dotted and dashed lines show the range of phenotypic values represented in each ofthe three populations. As shown, all have similar phenotypic variances but the means are 1<2<3. Arrowsshow the evolutionary changes expected if the trait is heritable. (Adapted from Endler 1986).

Figure 6.5B shows the effects of changes in the selective agent. When theenvironment changes, the position or shape of overall fitness function can change, whichcan also change the selection in each population without a change in the mean or varianceof the population. In this case, without any change in the trait distribution compared toFig. 6.5A, population 1 still experiences positive directional selection, population 2 nowis under weak stabilizing selection, and population 3 now has negative directionalselection. If the trait is heritable, both populations 1 and 3 will evolve in the direction ofpopulation 2.

Populations 1-3 in Figure 6.5 could represent subpopulations that are differentiatedfor the phenotypic trait of interest, or they could represent one population at differentpoints in time. Similarly, the two overall fitness functions in parts A and B could resultfrom different local environments, or environmental differences from season to season orfrom year to year. These points emphasize that fitness functions and the position ofpopulations on the overall fitness function can vary spatially and temporally. For anexample of temporally shifting fitness functions, we can turn again to one of the well-studied Darwin’s finches, Geospiza fortis (reviewed in Grant 1986; Grant and Grant2002). In drought years caused by La Niña events, large seeds predominate and there isdirectional selection for increased bill size. During wet El Niño years, small seeds are ingreater abundance and there is directional selection for smaller bills. Therefore, the size

A

B

3

Phenotypic trait

Fitn

ess

21


distribution of seeds in the environment determine the shape and position of the fitnessfunction, as in Fig. 6.5 above.

Note that both overall fitness functions in Figure 6.5 have an intermediate optimum.This seems like a reasonable assumption for most phenotypic traits, because it means thatthere is some trait value that is too small or too large to function optimally. It is not truefor fitness, and may not be true for fitness components like fecundity and lifespan.

We have been discussing the fitness function as an interaction between the selectiveagent and the target of selection, but estimating the fitness function by itself does notgenerally give much information about either the agents or targets. This is because wecan measure fitness without knowing the reasons for fitness differences amongindividuals with different phenotypes, and the phenotypic trait measured may not be thetrue target of selection, but the target is instead a trait correlated with the measured trait.The next two sections cover techniques and difficulties involved in identifying agents andtargets.

Multiple traits; direct and indirect selection

Direct selection is when there is a causal relationship between a phenotypic trait andfitness. The trait under direct selection is the target of selection and is the trait that ishelping the organism deal with the selective agent. Phenotypic correlations among traitscan cause indirect selection within a generation. This is illustrated in Figure 6.6, whichshow selection on horn and elytra (wing cover) length in male fungus beetles (Conner1988). Fig. 6.6 A and B show the fitness functions for each of these traits consideredseparately. These show almost identical fitness functions for the two traits, and thereforethe standardized selection differentials (S) are very similar.

Why is this true? There are two interconnected reasons. First, there is a strongphenotypic correlation between horn and elytra length (Fig. 6.6C), which means that ingeneral, a male with long horns also has long elytra. This means that the individuals arearranged in roughly the same order along the X-axis in both fitness functions (note thenumbered points in Fig. 6.6A and B). Second, each individual beetle has only onemeasure of fitness (lifetime number of females inseminated in this case), which meansthat they are arranged in exactly the same order along the Y-axis in the two fitnessfunction plots. Therefore, the two plots are not independent, but are essentiallymeasuring the same thing twice.

These univariate selection differentials, considering each trait separately, measurethe total selection acting on the trait, including both direct and indirect selection, but theyare close to worthless for determining the target of selection. This is because they do notseparate the direct from the indirect selection, so we have no way of knowing which ofthe two traits is causing the higher fitnesses. How do we determine if it is long horns orlong elytra that causes higher insemination success?


Fitn

ess

(life

time

no. i

nsem

inat

ions

)H

orn

leng

th (

mm

)

1

2

3

3

3

1

1

2

2

Elytra length (mm)

S = 0.49

S = 0.38

r = 0.91


Figure 6.6 Illustration of how phenotypic correlations cause indirect selection. A and B. Scatterplots withlinear fitness functions fit by simple regression for horn and elytra (wing cover) length. Standardizedselection differentials (S) are shown, which are the slopes of these lines after standardization. C.Scatterplot showing the strong phenotypic correlation between horn and elytra length. As an example, thedata points for three individual beetles are denoted on all three plots. Data from Conner 1988.

Selection gradients

To answer these questions, we can estimate selection gradients (β), which measuredirect selection on each trait after removing indirect selection from all other traits that arein the analysis. Instead of using simple regression, which estimates total selection as theselection differential (S), selection gradients are estimated using multiple regression. Inmultiple regression, a single dependent variable is regressed on multiple independentvariables simultaneously, and the effects of correlations among the independent variablesare controlled for statistically. The standard equation for multiple regression with kindependent variables is:

Y = b1X1 + b2X2 + b3X3 . . . bkXk 6.2

where the b terms represent the partial regression slopes of the relationships betweeneach X variable and Y, removing the effects of correlations among the X variables.Applying this same equation to selection gives:

Fitness = β1 trait 1 + β2 trait 2 + β3 trait 3 . . . βk trait k 6.3

where βk is the selection gradient, estimated as the partial regression slope. Therefore,each selection gradient measures the slope of relationship between fitness and that trait,just as with a selection differential, except that with gradients indirect selection isremoved by controlling for phenotypic correlations the other traits.

We can gain a better conceptual understanding of how this is done by taking anotherlook at the scatterplot depicting the correlation between horn and elytral lengths (Figure6.7) Here we have drawn representative vertical ellipses around three groups of beetles;within each group the beetles have the same elytra length but vary in horn length.Additional ellipses like this could be drawn. In multiple regression, this variance in hornlength (within each ellipse) that is independent of variance in elytra length can be used todetermine the slope of the relationship between fitness and horn length correcting for thisstrong correlation between horn and elytra length. Similarly, the horizontal ellipses showgroups of beetles in which there is variation in elytra length but little to no variation inhorn length; this variation can be used to estimate the selection gradients for elytralength.


Figure 6.7. Scatterplot of the correlation between horn and elytra length in fungus beetles (from Fig.6.6C), with added ellipses showing groups of beetles with variation in only one of the two traits. Suchvariation is used in multiple regression to estimate selection gradients, that is, direct selection correcting forphenotypic correlations among traits.

Examples of direct and indirect selection measures from fungus beetles andDarwin’s finches are given in Table 6.1. In both species live weight and two or threeother linear dimensions were measured. Just as we saw with two of the fungus beetletraits in Figure 6.6, all the selection differentials measuring total selection on the traits aresignificantly positive. This total selection includes strong indirect selection because thesetraits are highly positively correlated with each other. Using multiple regression toestimate direct selection, we see that only one or two traits are under positive directselection, horn length in beetles and weight and beak depth in finches. In fact, the directselection on beak width is negative, and there may also be negative direct selection onelytra length ( Conner 1988). The horns in male fungus beetles are used in fights overfemales, and larger-horned males win most fights, so this likely explains the directselection for and increase through insemination success (discussed further below).

In the finches there was direct selection for heavier birds with deeper, narrower bills.The selection was measured during a drought caused by La Niña, in which an importantfood item was woody fruits from the plant Tribulus cistoides. The finches crack theseeds by placing the side of their beaks across a corner of the fruit and twisting, and deepnarrow beaks are more efficient at performing this task.

Therefore, the significant positive total selection on elytra length and weight inbeetles, and on beak length and width in finches, was caused not by direct selection, butrather by strong positive indirect selection from the positively correlated traits that areunder positive direct selection (horn length in beetles and weight and beak depth infinches). The traits that are not under direct selection are not targets of selection, and are

Hor

n le

ngth

(m

m)

Elytra length (mm)

r = 0.91


therefore unlikely to be adaptations or to undergo current adaptive evolution. This is whyselection differentials are not useful for determining targets of selection or inferringadaptiveness of traits.

Total selection (S) Direct selection (β)

Fungus beetles:

Elytra 0.38** -0.33

Horn 0.49*** 0.94**

Weight 0.39** -0.16

Darwin’s finches:

Weight 0.62* 0.51*

Beak length 0.49* 0.17

Beak depth 0.60* 0.79*

Beak width 0.49* -0.47*

Table 6.1. Measures of total and direct selection in fungus beetles (Conner 1988) and Darwin's finches(Price et al. 1984, Table 2, 1976-77). Shown are standardized selection differentials (S) and gradients (β).All these traits are highly positively correlated with each other (r=0.69 to 0.91). Fitness measures werelifetime insemination success for the beetles and yearly survival for finches.

The contributions of direct and indirect selection to total selection are shown in thefollowing equation:

S1 = β1 + β2r12 + β3r13 + . . . 6.4

The total selection on trait 1 as measured by the standardized selection differential is thesum of the direct selection on that trait (β1) and the direct selection all correlated traitsweighted by the correlations between those traits and trait 1. Plugging in the numbers forelytra length in fungus beetles, where elytra length is trait 1, horn length is trait 2, andweight is trait 3:

Total Direct Indirect

Elytra Horn Weight

0.38 = -0.33 + 0.94 × 0.91 + -0.16 × 0.85

= -0.33 + 0.85 - 0.14


Note that the phenotypic correlation between horn and elytra lengths is 0.91 and thatbetween weight and elytra is 0.85. Here the negative contributions of both the directselection on elytra length (-0.33) and the indirect selection through weight (-0.16 × 0.85=-0.14) are outweighed by the much larger positive contribution made by indirect selectionthrough elytral length (0.94 × 0.91 = 0.85), resulting in positive total selection on elytralength.

Indirect selection is not the same as correlated responses across generations (lastchapter), but they are closely related. Indirect selection occurs within a generation and iscaused by phenotypic correlations, while correlated responses occur across generationsand are caused by genetic correlations Indirect selection is not as important as correlatedresponses are, because indirect selection does not affect evolutionary change. However,it can indicate possible correlated responses, if phenotypic correlations are similar to theunderlying genetic correlations. The main importance of indirect selection is that it beidentified and separated from direct selection so that the targets of selection can beidentified.

Selection gradients help determine the traits that direct selection is acting on andthereby identify the targets of selection. However, they can never prove which trait is thetarget of selection due to the problem of unmeasured traits. Multiple regression is astrictly observational, rather than experimental, approach, which means that onlynatural variation in the traits and environments are used. The phenotypic traits andfitnesses are measured on the organisms without any experimental manipulation. Theselection gradient analyses can remove indirect selection due to traits that are in theanalysis, but since it is impossible to measure or even identify all phenotypic traits, theremay be other traits that are actually the target of selection. Suppose that in the fungusbeetle study horn length was not measured, and the selection gradients were calculatedwith just elytra and weight. The analysis may have identified elytra length as the targetof selection, when the full analysis with three traits shows horn length to be the target.Similarly, it could be that horn length is not the true target, but it is correlated with anunmeasured fourth trait that is the actual target of selection. This seems unlikely in thiscase based on other knowledge about the function of horns in gaining access to females,but it cannot be ruled out. Knowledge of the biology of the organism is critical indeciding which traits should be measured and included in the selection gradient analyses.

Experimental Manipulation

While selection gradient analyses cannot prove which traits are the targets of directselection, they are an excellent place to start because they suggest which traits to focus onfor further study. In particular, a trait that does not have a significant selection gradient isnot likely to be a true target of selection and thus may not be worth further study. Toprove which are the targets, experimental manipulation of the phenotypic traits isnecessary. For example, the beetle horns could be cut off or extended and the fitnessconsequences of this manipulation measured in the natural population. If beetles areplaced randomly into the different groups for manipulations, then this eliminates theproblem of unmeasured correlated traits. With random assignment of beetles to thedifferent treatment groups (reduced vs. extended horns), these groups should not differ in


any other traits. Therefore, any differences in fitness between the groups can be reliablyattributed to the manipulated differences in horn length. If the extended horn group hashigher fitness, then this proves that horns are a target of selection.

Experimental manipulation was used to study the effects of tail length on matingsuccess in male long-tailed widowbirds, Euplectes progne (Andersson 1982). As theirname implies, males of this species have tails about half a meter long, and these tails areused in displays over their territories. The number of females nesting on the territories of36 males were recorded (this is the measure of fitness, as it likely helps determine thenumber of offspring sired by the male), and then these males were assigned randomly toone of four experimental treatments. The tails of one group were shortened to about 17cm in one group by cutting the tails and then gluing a short section back on. The rest ofthe cut tails were glued on another group of males who thus had their tails elongated by25 cm on average. There were also two controls; one group was just caught and bandedfor identification as all birds were, and the other had their tails cut and then the entire cuttail glued back on, to control for effects of cutting and gluing.

The results were dramatic (Figure 6.8) – the four groups had similar numbers ofnests on their territories before the manipulation, but afterward the elongated males hadfar more than the shortened males, with controls intermediate. This is strong evidencethat tail length is a target of selection, with females acting as the selective agent.

An important consideration with experimental manipulation is how much variationto create; the answer depends on the question being asked. If the goal is to determine ifthe trait is a target of current selection in a specific natural population, then themanipulation should not go beyond the trait variation present in that population. Thereason for this was shown in Figure 6.5 above – the phenotypic range expressed in anyenvironment can determine how selection is acting on that trait. If, however, the questionis more generally about how the trait interacts with the environment to affect fitness, thatis, whether or not it is an adaptation, then manipulations beyond the range of any givenpopulation are appropriate. These larger manipulations help to determine the overallfitness function as depicted in Figure 6.5. The widowbird study is an example of this.

While experimental manipulation is a powerful method to determine causation, itdoes have some disadvantages. Unlike selection gradient studies, only one trait can bestudied at a time. This can be a serious drawback, because measuring fitness is usuallydifficult. It can be difficult to do a proper control for the manipulation. In the case of thewidowbirds, cutting and then re-gluing the same feathers to maintain the same tail lengthwas an excellent control, because all the birds had their feathers cut and glued in the sameway. However, due to the complexity and integration of traits in living organisms it ishard to be sure that some other trait was not also being changed, and that trait is the targetof selection. In the widowbirds, it is possible that changing the length of the tail changedflight patterns or some other aspect of the male’s behavior that the females were actuallyresponding to. This seems unlikely in this particular case (Andersson showed thatshortened males actually displayed more often), but possible. Once again, detailedknowledge of the biology of the organism (e.g. display behaviors in the widowbirds) iscrucial.


Figure 6.8 Results from experimental manipulation of widowbird tail lengths. The measure of fitness isthe number of female nests per male. The top panel (a) shows the numbers before tails were manipulatedand the bottom (b) is after manipulation. Reprinted, with permission, from Andersson (1982).

A closely related problem is that the manipulation may not be producing abiomechanically equivalent structure. For example, there is some evidence that fungusbeetle horns have nerve cells and sensory hairs, so that cutting and re-gluing them as wasdone with the widowbird tails may fundamentally alter their function. Finally, there aremany traits for which experimental manipulation is extraordinarily difficult or perhapsimpossible. A good example may be the finch beak and weight traits in Table 6.1; it ishard to see how they could be manipulated without injuring the birds or altering manyother traits.

Therefore, the observational selection gradient and experimental manipulationapproaches are complementary. Only gradient analysis can produce a quantitativeestimate of the strength of natural selection in a population for use in predicting speedand direction of evolution (below). Only experimental manipulation can demonstratecausality. An excellent approach, rarely used, is to use a selection gradient approach firstto measure selection on a number of traits identified as potential adaptations fromknowledge of the biology of the organism, and then experimentally manipulate thosetraits with significant selection gradients to test for a causal relationship betweenphenotypic variation and fitness.


Correlational selection

From the above discussion of direct and indirect selection and the section oncorrelations among traits in the previous chapter it is clear that single traits cannot beconsidered in isolation. While the selection estimates for finches and fungus beetles inTable 6.1 show the effects of correlations on selection, we can also do the opposite andstudy the effects of selection on correlations. Correlational selection is when two traitsinteract to determine fitness, or in other words certain combinations of trait values havehigher fitness than other combinations. If both traits are heritable, correlational selectioncan cause evolutionary change in the correlation between the two traits.

So far the fitness functions we have been examining have been univariate (one traitonly; e.g. Figs. 6.1-6.5), but to understand correlational selection we need a three-dimensional phenotypic fitness surface (Figure 6.9). The fitness surface is alsosometimes called the phenotypic adaptive landscape, due to its close similarity toWright’s adaptive landscape. In both cases, the vertical (Z) axis is fitness, but inWright’s this was population mean fitness whereas in the phenotypic surface it isindividual fitness. In the version of Wright’s adaptive landscape that we discussed inchapter 3, the horizontal (X and Y) axes are allele frequencies at two different loci, whilein a phenotypic fitness surface they are the values of two different phenotypic traits.Therefore, either the allele frequencies or the mean and variance of the traits determine apopulation’s position on the landscape or surface respectively. Note that the individualsare spread out over the phenotypic surface, while the population mean in Wright’slandscape is one point determined by the allele frequencies. The overall shape ofWright’s landscape is determined by the environment and epistatic interactions betweenthe two loci; similarly, the shape of the phenotypic fitness surface is determined by theenvironment and functional interactions between the two phenotypic traits. The keydifferences are that each individual in the population can be placed on the phenotypicfitness surface depending on their phenotypes and individual fitness, while in Wright’ssurface the entire population is represented by one point as determined by the populationmean fitness and allele frequencies, which are a population level statistics.

Brodie (1992) measured selection on juvenile garter snakes. His measure of fitnesswas juvenile survival, which was measured using a mark-recapture study, where a groupof animals are marked and periodically recaptured to determine which are still alive.Brodie was mainly interested in two anti-predator traits, the degree of striping on thebody and the tendency for the snake to perform reversals, a rapid doubling-back whenbeing pursued by a predator. There were no significant linear or quadratic selectiongradients for the traits individually, but there was a significant correlational gradient.The correlational gradient is also estimated using multiple regression. Instead ofsquaring one of the traits as is done to get the quadratic gradient, the correlationalselection gradient for a pair of traits is the regression coefficient for the product of thetwo traits; this is called a cross-product.


Figure 6.9 Bivariate fitness surface for two antipredator traits in garter snakes. Positive scores for stripemean highly striped snakes, while negative scored snakes are blotchy. Positive scores for reversals aresnakes that reverse frequently. Fitness is on the vertical axis. Note that fitness depends on the combinationof the two traits – high values of both lead to low fitness, as does low values of both. The two adaptivepeaks are snakes that were highly striped and rarely reversed (back right of figure) and blotched snakes thatoften reversed (left foreground). Reprinted, with permission, from Brodie 1992.

The correlational selection gradient for the cross-product of stripes and reversals wasnegative, which selects for a negative correlation. Snakes that were highly striped (apositive stripe score in Fig. 6.9) had high fitness if they also had low (negative) scores forreversals (the fitness peak in the back right of Fig. 6.9), but low fitness if they were proneto reversals (the very low fitness valley at the rear left of the figure). Conversely,unstriped snakes had highest fitness if they also had a high tendency to reverse (front leftpeak in figure), but had low fitness if they did not reverse often (valley at the front right).Finally, note the ‘saddle’ in the figure, where both traits are at intermediate values; fitnessis intermediate also, not as high as the peaks but not as low as the wrong combination oftraits.

Therefore, very similar to the case with epistasis, the fitness of an individual with agiven phenotypic value for one trait depends on its value for the other trait. Selection inthis case does not act on traits individually, but on combinations of traits that arefunctionally integrated, that is, those that work together to affect organism function andfitness. In the garter snakes, the longitudinal stripes make speed of a straight-movingsnake appear to be slower than it actually is, so that predators misplace their strikes. Thereversals are often combined with a pause in movement, and the unstriped snakes areharder to see when motionless. Therefore, striped snakes that reverse and unstriped


snakes that move in a straight line are the easiest for predators to catch and therefore havelowest survival.

Correlational selection is very closely related to the epistatic selection that is at theheart of the shifting balance theory. The term correlational selection is best used whendiscussing selection on phenotypes, whereas epistatic selection refers to selection on theunderlying genotypes. Correlational selection can give rise to epistatic selection if thereare separate gene loci affecting the two phenotypic traits, and this epistatic selection cancreate linkage disequilibrium between these loci.

How do we study selective agents?

We have defined fitness as the lifetime number of offspring produced, and selectiongradients based on this fitness measure are likely to provide a good estimate of thestrength of selection. However, because so many different environmental factors affectlifetime offspring production, it is difficult to identify the selective agents from theseselection gradients. By dividing total lifetime fitness up into biologically meaningfulfitness components, we can obtain a better understanding of the environmental causes ofselection, that is, the selective agents.

Darwin was probably the first to break total fitness up into fitness components whenhe introduced the idea of sexual selection. Darwin recognized that many traits, such asthe showy plumage of many male birds, should decrease survivorship by making theanimals more vulnerable to predators. How then could these traits evolve by naturalselection? Darwin also recognized that these types of traits were often sexuallydimorphic, that is, much more pronounced in one sex, usually the males. In Darwin’sprocess of sexual selection, these traits evolve through differences in mating success,which can lead to differential offspring production. This is the process that leads to thelong tails of widowbirds. Therefore, in Darwin’s scheme there are two fitnesscomponents, survivorship and mating success.

Sexual selection is best thought of as a subset of natural selection, becausedifferential mating success is but one of several ways that fitness differences can occur.Some authors use natural selection in a narrower sense, to mean everything but sexualselection, whereas others call this non-sexual selection. Non-sexual selection is oftenbroken into mortality selection, caused by differences in lifespan among individuals withdifferent phenotypic values, and fecundity selection, caused by differences in offspringproduction. Obviously all these different terms can cause confusion, and often it isdifficult to decide where to draw the line between sexual and non-sexual (especiallyfecundity) selection.

An excellent way to avoid this confusing terminology and to focus on the real goalof identifying the selective agents is to define multiplicative fitness components that aremeasurable and biologically meaningful for the organism under study. By multiplicativewe mean they are defined so that the numerator of each component is the denominator ofthe next component (Table 6.2). This has two key advantages. First, it means that when


they are multiplied together they equal total fitness, because all but the last numerator(which is total fitness) cancel out. Second, it means that they are likely to be largelyindependent of each other, so that one can examine selection and selective agents at eachcomponent without it being confounded by earlier components. For example, in fungusbeetles the total number of females a male courts is determined by the male’s success infights with other males, as well as by his lifespan and attendance at the mating area. Ifwe measure selection using total lifetime number of females courted as the fitnessmeasure, we would not know if the selection was due to lifespan, attendance, or successin male competition. By dividing courtships by the total number of nights a male spent inthe mating area in his lifetime, however, the fitness component is number of femalecourted per night the male was in attendance, which is largely or entirely independent oflifespan or attendance. Therefore, we can isolate selection due to each component andmore easily identify the selective agents.

Table 6.2 Multiplicative fitness components for fungus beetles (Conner 1988). Lifespan is the number ofdays between the first and last sightings of each marked beetle, attendance is the number of nights the malewas seen on the surface of the fungi where mating occurs, courtships is the number of females courted inthe male’s life, and similarly copulation attempts and inseminations are the lifetime number of copulationsattempted and successful inseminations over the male’s lifetime. Therefore, the attendance/lifespancomponent is the proportion of days a male was alive that he was in the mating area, courtships/attendanceis the average number of females courted per day that the male was in attendance, and so on. The meanvalues are shown below each component; for example, males lived about 56 days on average, were in themating area two-thirds of their lives, courted one female per night in attendance, and so on. The bottomrow shows the selection gradients for horn length for each component and total fitness.

Table 6.2 shows the horn size selection gradients for each fitness component and fortotal fitness (lifetime inseminations). We discussed the latter gradient, 0.94, earlier in thechapter (Table 6.1). Then we said this was likely due to male competition for mates,because that is how the horns are used, but we did not have direct evidence. Theselection gradients for fitness components give us this evidence, because the largest andonly statistically significant gradient occurs in the copulation attempts per courtshipcomponent. Males often attempt to break up other male’s courtships before the courtingmale can attempt to copulate, but this is usually only successful if the attacker has longerhorns than the courting male. Therefore, the selection gradients fit with other knowledgeof the beetle’s biology, and strongly suggest that selective agent for horn size evolution isother conspecific males, and that horns are an adaptation to increase access to mates.

Measuring selection on components of fitness are an excellent first step inidentifying selective agents, but just as selection gradients cannot prove targets ofselection, additional evidence is usually needed to provide truly convincing evidence forselective agents. One way to provide additional evidence without altering the naturalpopulations is to measure selection gradients in different locations in which the putative

LifespanLifespan

Attendance LifetimeinseminationsAttendance

Courtships

Courtships

Copulationattempts

Attempts

Inseminations

0.26 0.03 0.19 0.30* 0.26 0.94**β

Mean 56.2 0.65 1.01 0.60 0.24 5.31


selective agent varies (Wade and Kalisz 1990). If the selection gradient varies acrosslocations in a way that is correlated with some environmental variable, then this isevidence that the environmental variable is the selective agent. This approach isillustrated in Figure 6.10. Here selection is measured in three hypothetical populationsthat differ for some environmental variable. The three fitness functions show that thestrength of selection increases as the value of the environmental variable increases; inother words, there is a positive relationship between the selection gradient and theenvironmental variable, which is evidence that this particular environmental variable is aselective agent.

Figure 6.10. Hypothetical illustration of determining the selective agent by measuring selection in severalpopulations that differ in an environmental variable (the putative selective agent). The three smaller plotsare the fitness functions within each population, and they are placed in the larger plot based on eachpopulation’s value for the selection gradient (slope of the fitness function) and the environmental variable.The positive relationship between these two is evidence that the environmental variable is a selective agent.

In a rare example of this approach, Stewart and Schoen (1987) measured selectionon a number of vegetative size and phenology (timing) traits in pale jewelweed,Impatiens pallida, at 24 locations within one population. They found that the strength ofselection on some of these traits increased in intensity with increases in light, soilnutrients and soil moisture.

The weakness of this approach in determining selective agents is directly analogousto the weakness of selection differentials in determining selective targets; neitheraccounts for correlations among independent variables. In figure 6.10, the positiverelationship between the strength of selection and the environmental variable couldactually be due a different environmental variable that is correlated with the one plotted.As with targets of selection, one solution is to perform multiple regression, that is, regress

Fitn

ess

Phenotype

Fitn

ess

Phenotype

Fitn

ess

Phenotype

Sel

ectio

n gr

adie

nt (β)

Environmental variable


the strength of selection in the different populations on multiple environmental variablesimultaneously. To our knowledge this has not been done.

Note that these are observational studies; natural populations are observed, andnothing is experimentally manipulated. Therefore, these types of studies cannot providedefinitive identifications of selective agents, for exactly the same reason as for targets ofselection – there could be unmeasured environmental factors correlated with themeasured factor that are actually the selective agents. Once again, manipulativeexperiments provide complementary evidence, and are necessary to prove whichenvironmental factor is the selective agent.

Dudley and Schmitt (1996) combined experimental manipulation of both theselective agent and target in a study of plant height as an adaptation to intraspecificcompetition. Working with a different species of jewelweed, Impatiens capensis, Dudleyand Schmitt manipulated plant height by exposing the plants to different ratios of red tofar red light, which is well known to produce differing levels of stem elongation. Theythen planted the resulting plants into natural populations at both low and high densities,thus experimentally manipulating this environmental variable. The selection gradientsfor height were positive at high density, indicating that taller plants were better able tocompete for light under crowded conditions, but negative at low density, indicating a costto growing tall when light is not the main limiting resource. This is a good example ofdensity dependent selection, because the direction of selection depended on the densityof the population. This is also excellent evidence for plant height as an adaptation forcrowding, because they have identified the selective agent and target throughexperimental manipulation. The different methods of identifying selective agents andtargets are summarized in Table 6.3.

To determine: First step Intermediate Definitive

Selective agents Multiplicative fitnesscomponents

Regress selection gradienton multiple environmental

variables

Manipulateenvironment

Targets of selection Selection gradients Manipulatephenotype

Table 6.3. Summary of the different methods of determining selective agents and targets discussed in thetext. The first step and intermediate columns are observational studies, and the definitive column isexperimental.

Box: Summary of bivariate relationships in ecological geneticsIn this chapter and the previous two we have introduced four different types of bivariateplots, so a review of these may be useful (Table 6.4). The key to keeping them straightand to understanding what each one means is to carefully consider what each axisrepresents (key to understanding any graph) and what each point in the plot represents.In the first three listed in Table 6.4, the axes represent phenotypic traits. If each pointrepresents means for a family, and the Y-axis is offspring mean and the X-axis is parents,then the plot represents additive genetic information. If the trait measured is the same in


parents and offspring then the slope of this relationship is the heritability, and if the traitsare different in parents and offspring then the covariance of the relationship is 1/2 theadditive genetic covariance. If instead of family means, each point represents oneindividual in the population, then the plots convey phenotypic information. If the traitson the X and Y axes are different, then the plot represents the phenotypic correlation. If atrait is on the X and fitness is on the Y-axis, then the plot is the fitness function.

Relationship: Measures: Y-variable:

X-variable: Each pointrepresents:

Parent-offspringregression slope

heritability offspringtrait A

parent traitA

family

Correlationbetween 2 traits

phenotypiccorrelation

trait A trait B individual

Parent-offspringcross covariance

1/2 geneticcovariance

offspringtrait A

parent traitB

family

Fitness function selection fitness phenotypictrait

individual

Table 6.4 Review of four different types of bivariate plots. The scatterplot at the top is a representativeexample of one possible relationship.End box

6.3 Predicting short-term phenotypic evolution (editor: any time ∆z appears in this section there should be a bar over the z as inequation 6.5)

We have now covered a number of related topics in the evolution of phenotypictraits by natural selection. In this chapter we discussed measuring the strength ofselection using selection differentials and gradients, and that the former includes indirectselection within generations caused by phenotypic correlations. In the previous chapterwe showed how the breeder’s equation, R = h2S, could be used to predict theevolutionary change across one generation; however, this equation is limited to one trait

Y

X


only and cannot take genetic correlations into account. Also in the last chapter wediscussed correlated responses to artificial selection caused by genetic correlations, but inthese experiments selection is typically applied to only one trait at a time, which is nothow selection acts in nature. Now we can combine all these elements in one generalequation that predicts evolution by natural selection for any number of traits:

∆z = Gβ 6.5

This is a multivariate version of the breeder’s equation from the previous chapter, R=h2 S, and like the breeder’s equation it is crucial for understanding phenotypic evolutionby selection. In both cases the term on the left of the equals sign is the change inphenotypic mean across one generation, and on the right are terms for additive geneticvariance and the strength of selection respectively. Because equation 6.5 is for multipletraits simultaneously, it also includes genetic covariances among traits. How is thisdone? Equation 6.5 is a matrix equation, in which each term actually represents a set ofrelated terms; this is shown in Table 6.5. Matrix algebra is a systematic method tosimplify math when a large number of terms are involved; computer spreadsheetprograms can perform simple matrix operations. Each of the three parts of Table 6.5represent exactly the same mathematical operation – the top is in compact matrix formthat is general for any number of traits, the middle is expanded matrix form for threephenotypic traits as an example, and the bottom is a system of three standard (non-matrix) equations for the three traits. The first and last terms of the matrix equation arecalled column vectors; they are nothing more than a list of variables (or the numbersthey represent), three in our three-trait example. In all cases the subscripts refer to thetrait in question, that is 1 for trait 1, 2 for trait 2, and so on. Thus, ∆z1 = R for trait 1, thatis, the change in mean across one generation, and β2 is the selection gradient measuringthe strength of selection on trait 2.

The G term is a square matrix, with the number of rows and columns equaling thenumber of traits. This is referred to as the G-matrix; many studies have been devoted toestimating the G-matrix and understanding its role in evolution. The diagonal of the G-matrix, in which the two subscripts are the same, represent the additive variances for thethree traits. That is, G11 = VA for trait 1, G22 = VA for trait 2 and so on. The off-diagonalelements, those with different subscripts, are the additive genetic covariances betweenpairs of traits. Thus G23 is the additive genetic covariance between traits 2 and 3. Notethat the G matrix is symmetrical, which means that each covariance appears twice, onceabove and once below the diagonal. Because the G-matrix consists of additive geneticvariances and covariances, it is estimated using the techniques described in the last twochapters, usually offspring-parent regression or nested half-sibling analysis.

∆z = Gβ

∆z 1∆z 2∆z 3

=

G11 G12 G13G12 G22 G23

G13 G23 G33

β1β2β3


∆z1 = G11β1 + G12β2 + G13β3

∆z2 = G12β1 + G22β2 + G23β3

∆z3 = G13β1 + G23β2 + G33β3

Table 6.5 The fundamental equation for predicting short-term phenotypic evolution, presented inthree equivalent ways. At the top is the compact matrix form that is general for any number of traits, themiddle is the expanded matrix form for three phenotypic traits as an example, and the bottom is a system ofthree standard (non-matrix) equations for the three traits. The terms in bold represent the response to directselection that leads to adaptive evolution, and the other two terms in each equation are the correlatedresponses to selection on correlated traits.

To solve for the predicted change across one generation (∆z) for each of the traits,each of the elements of the row of G adjacent to that ∆z term are multiplied by each ofthe selection gradients in order (spreadsheet and other computer programs have routinesfor matrix operations like this). The results of this process are the standard equationsshown in the bottom portion of Table 6.5. In boldface are the portions of evolutionarychange that are due to direct selection on the trait, that is, the additive genetic variance forthe trait times the selection gradient for that same trait; this represents adaptive evolutionover one generation. The other terms are correlated responses to selection on other traits.For example, the second term in the equation for ∆z1 is the selection gradient for trait 2times the additive genetic covariance between trait 1 and trait 2. These equations areanalogous to equation 6.4, in which there was direct selection and indirect selectioncaused by phenotypic correlations within generations. In equation 6.5, there is a directresponse to selection across generations, as well as correlated responses acrossgenerations due to genetic covariance. This equation ties much of quantitative geneticsand phenotypic evolution together in a unified way, similar to how F-statistics tied muchof single-locus population genetics together.

A rare example of evolutionary predictions in a natural population comes fromCampbell’s (1996) study of floral evolution in scarlet gilia, Ipomopsis aggregata (Table6.6). Campbell found three traits that were under direct selection – length and width ofthe floral corolla, and the proportion of time that the flowers had receptive stigmas(proportion pistillate). There was positive direct selection on all three traits and acorresponding prediction of an increase in the mean of all three in the next generation.Note, however, that the predicted changes for both corolla length and proportion pistillateare decreased by the negative covariance between these two traits. Thus, the adaptiveevolution of these traits is slowed or constrained by this negative covariance.


∆z G β

Corollalength

Corollawidth

Proportionpistillate

Corolla length 0.043 1.092 0.021 -0.039 0.05

Corolla width 0.035 0.021 0.025 0.004 1.22

Proportion pistillate 0.005 -0.039 0.004 0.002 0.96

Table 6.6. Predicted evolutionary change in the mean of three floral traits (∆z ), calculated as the product ofthe genetic variance/covariance matrix (G) and the vector of selection gradients (β). From Campbell(1996).

Exercise: Calculate ∆z for each of the three traits in Table 6.6 using the system ofequations at the bottom of Table 6.5. Make sure your answers match the values given inTable 6.6.

Equation 6.5 provides a general framework for understanding genetic constraints onevolutionary change. We can define an evolutionary constraint as any factor that slowsthe evolution of the most adaptive combination of traits; genetic constraints are a subsetof these (we discussed how drift and gene flow can constrain adaptive evolution inchapter 3) . The simplest genetic constraint is a lack of additive variance for the trait,because without variance there is nothing for selection to act upon. In terms of equation6.5, Gii, is close to zero, so the adaptive response to direct selection, GiiβI, is also close tozero regardless of the strength of selection. This means that even if there is strong directselection, adaptive evolution is slowed or constrained by the lack of additive geneticvariance. The low additive genetic variance for the proportion pistillate trait in Table 6.6(0.002) is part of the reason for the small predicted change (0.005).

The other source of genetic constraint comes from the other terms in the G-matrix,the genetic covariances. We have already seen how indirect selection within ageneration causes the total selection to be in the opposite direction as direct selection forelytra length in fungus beetles and beak width in Darwin’s finches (Table 6.1). In ananalogous way, correlated responses to selection can cause nonadaptive or evenmaladaptive evolution. Nonadaptive evolution is a change in a trait that is not underdirect selection (β ≈0), and maladaptive evolution is change in the opposite directionthan the direction that would increase adaptation of a given trait. In other words,maladaptive evolution occurs when ∆z and β have opposite signs.

Examples of both types of constraints can be found in a study of selection on bodysize traits in a laboratory population of flour beetles (Tribolium castaneum; Conner andVia 1992). The G matrix, selection gradient vector, and vector of predicted mean changeacross one generation are shown in Table 6.7. The total predicted change (∆z) is shown,and is broken down into the components due to direct selection and correlated responses(CR). There was strong directional selection for increased weight (β=0.26, weaker


selection for decreased width, and no selection on length. In spite of the strong selectionfor increased weight the predicted change in the mean was small, only 0.01 standarddeviations. The first reason was low additive variance for weight -- the heritability ofwas only 0.24, so that the predicted response to direct selection was only 0.05. The otherreason for the low predicted change is the negative selection on width, which had a strongpositive genetic covariance with weight, leading to a negative correlated response almostas strong as the direct response. Therefore, the adaptive evolution of increased weightwas constrained by a lack of additive genetic variation and by a genetic covariance withanother trait. The length and width show evidence for nonadaptive and maladaptiveevolution respectively, because both were predicted to increase due to their positivecorrelations with weight, in spite of the fact that the selection gradients indicate thatlength is at its optimum and width is above its optimum.

∆z G (x10-3) β

Total Direct CR Weight Length Width

Weight 0.01 0.05 -0.04 2.27 1.12 1.38 0.26

Length 0.03 0.00 0.03 1.12 0.41 0.45 0.00

Width 0.04 -0.03 0.07 1.38 0.45 0.46 -0.12

Table 6.7. Predicted changes across one generation for three body size traits in male flour beetles.The total predicted change (∆z ) is split into components due to direct selection on each trait and tocorrelated responses (CR) caused by selection on the other traits. Standardized values (in standarddeviation units) are shown for ∆z and selection (β) and unstandardized values are given for the G-matrix.All calculations were done on unstandardized values, however. Modified from Conner and Via (1992);data for females were eliminated for simplicity.

A graphical view of constraints caused by genetic correlations is given in Figure6.11. The elliptical cloud of points represents a hypothetical negative genetic correlationbetween flower size and number, which could be caused by trade-offs in allocation ofresources (Worley and Barrett 2000). The cross in the middle is the current bivariatemean, that this the mean for the two traits jointly in the population. The letters A-Crepresent three different possible evolutionary optima, that is, values for both traits thatmaximize fitness. The arrows show likely evolutionary trajectories, which are thephenotypic paths that the population would take in evolving to these optima. Becauseeach arrow represents a fixed amount of time (50 generations), long arrows denote rapidevolution and short arrows denote slow evolution. Evolution to the optimum marked byA (fewer, larger flowers) would be the fastest, because it is along the major (longest) axisof the ellipse, meaning that it is in the direction of most genetic variance in bivariatespace. Conversely, evolving to C (increased number and size of flowers) would beslowest, because that is the direction of least variation in bivariate space. Thus, evolutionto A is rapid but to C is constrained.


Evolution to point B is a little more complex. The optimum at B is increased flowersize with no change in flower number from the current mean shown by the cross. Sincethe population is far from the optimum for flower size, there is strong selection on flowersize but no directional selection on flower number. Therefore, the population evolvesincreased flower size in the direction of maximum genetic variance, that is, roughly alongthe major axis of the ellipse, as shown by the first arrow. This actually draws thepopulation away from the optimum flower number, thus creating maladaptation of flowernumber and selection to reduce flower number back to its optimum. As flower sizeapproaches its optimum and flower number is drawn further from its optimum, therelative strength of selection on these two traits switches, so that at some point selectionbecomes stronger on flower number. This is why the trajectory eventually turns thecorner and begins heading toward the optimum at B.

Figure 6.11 Hypothetical examples of evolutionary constraints caused by genetic correlation. Thegray points represent breeding values, so the elliptical cloud of points represents a negative geneticcorrelation between the two traits. The cross represents the current mean for both traits in the population,and the letters A-C are three evolutionary optima. The arrows represent evolutionary trajectories to thesethree optima; each arrow represents 50 generations.

The isoclines for fitness are not shown in figure 6.11 for simplicity, but note that inall three cases the population trajectories are climbing an adaptive peak, that is,population mean fitness is always increasing. This is even true in the trajectory towardspoint B; even though the initial evolution of flower number is maladaptive, the increasein fitness due to increased flower size is greater, leading to increased mean fitnessoverall. Therefore, genetic constraints cause a slowing of the ascent of a population to anadaptive peak, but even with constraints selection always increases population meanfitness. Also note that in all cases the arrows get shorter as the population approaches theoptimum. There are two causes for this. First, as the population gets closer to theadaptive peak selection becomes weaker, because the population is better adapted.Second, the selection depletes genetic variation for the traits, which also slows progresstoward the optimum.

Flower number

Flo

wer

siz

e (m

m)

20 30 40 50 60 70

7

8

9

6

CA B


To summarize constraints caused by genetic correlations (Table 6.8), if selection isacting in the same direction on two traits, there is a constraint only if the geneticcorrelation between the traits is negative. In fact, the response to selection can beaugmented if the genetic correlation is positive. Positive genetic correlations can cause aconstraint if selection is acting in different directions on the two traits, that is, if theselection gradients have opposite signs (e.g., weight and width in Table 6.7).

Signs of selection gradients

Same sign Opposite signs

Positive Augmented Constrained

Negative Constrained Augmented

Table 6.8. Summary of constraints due to a genetic correlation between a pair of traits. The speed ofevolution of the two traits can be either constrained or augmented depending on the sign of the correlationand the direction of selection on the two traits.

Readings questionsArnold 1987: (full references are below)

1. Why are genetic correlations important? How can they affect the evolution of traits?What is a correlated response to selection?

2. What are the two mechanisms that cause genetic correlations? Explain how eachcauses correlations.

3. Both parent-offspring regression and sib-analysis can be used to measure geneticcorrelations as well as heritabilities. What additional information is needed to use thesebreeding designs to estimate correlations over what is needed for heritabilities?

4. How can artificial selection experiments be used to reveal the presence of geneticcorrelations?

5. Explain what each of the terms in equation 9.1 represents.

6. In Fig. 9.7, why does the evolutionary trajectory marked b differ from the one markeda?

Berenbaum et al. 1986:

Sign

of

gene

tic


1. What breeding design (e.g. half-sibs, parent-offspring regression) did they use toestimate heritabilities and genetic correlations? What method to estimate selection?

2. For what traits were heritabilities, correlations, and selection measured? Are theseappropriate traits to address the questions posed in the Introduction? Are there othertraits that should have been included, in your opinion?

3. What is the likely cause of the significant negative selection differential for floweringdate?

4. How reliable are the estimates of genetic correlations given in Table 9? State yourevidence.

5. Table 9 shows a negative genetic correlation between pBERs and SPHs. Could thiscause a constraint, that is, a slowing of the evolution of the most adaptive combination ofthese traits (see also Table 5)?

6. On what do they base their conclusion that the evolutionary arms race between wildparsnip and the parsnip webworm has reached a temporary stalemate (p. 1227)? Is this areasonable conclusion to draw from the evidence presented in the paper?

Brodie 1992:

1. Define correlational selection. Why is it important in evolution?

2. Explain the hypothesis of correlational selection on stripedness and reversal behavior.In other words, why would we expect a functional relationship between stripes andreversal behavior?

3. What was Brodie's measure of fitness? Is this a good measure of fitness for thisstudy? Why or why not?

4. Do you think his choice of traits to study was reasonable? Was the rationale for eachtrait clear? Were there other traits that you think should have been included? If so, why?

5. Which traits were under significant directional selection? Which were undersignificant stabilizing/disruptive selection? Which pairs of traits were under significantcorrelational selection?

Campbell 1996:

1. Why is it important to measure heritabilities and genetic correlations in the field?

2. What breeding design (e.g. half-sibs, parent-offspring regression) did they use toestimate heritabilities and genetic correlations? What method to estimate selection?


3. For what traits were heritabilities, correlations, and selection measured? Are theseappropriate traits to address the questions posed in the Introduction? Are there othertraits that should have been included, in your opinion?

4. Which traits had significant heritabilities? Which pairs of traits were significantlygenetically correlated?

5. Give two reasons why the predicted response to selection (based on the multivariateequation; Table 8) is so much less for Proportion Pistillate than it is for Corolla Length,when the selection gradient for the former was so much greater than for the latter.

6. Why is there a range of predicted responses for Proportion pistillate even though thetwo estimates of the selection gradient for that trait are the same?Conner 1988:

1. What is direct and indirect selection? What are targets of selection and selectiveagents? What methods used in this paper help identify targets and agents of selection?

2. Explain how multiplicative fitness components (or episodes of selection) are definedso that they are independent of each other, and multiply to equal total fitness. What is thebiological meaning of each of the five components used in this paper?

3. Focusing on the differentials and gradients for total fitness of the cohort of 67 malesonly (the bottom lines in Tables 3 and 5),what are the similarities and differencesbetween the total selection on the three traits (selection differentials) and the directselection on the traits (selection gradients)? What are the reasons for the differencesbetween total and direct selection?

4. What are the main targets of selection identified by this study, and what are the likelyselective agents causing this selection? Why is it important to identify targets and agentsof selection?

Grafen 1988:

1. Explain the distinction between adaptation and selection in progress. Is thisdistinction always clear, or are there areas where they come together? Are the methodsof selection differentials and gradients developed by Lande and Arnold and used inConner (1988) and Brodie (1992) useful for studying adaptation? Selection in progress?

2. What are the advantages and disadvantages of using natural and artificial variation tostudy adaptation and selection in progress? Does selection gradient analysis help preventbeing mislead by Grafen’s three ‘malignant’ causes of natural variation?

What are the three problems in the interpretation of selection gradients identified byGrafen? How serious are they likely to be, in your opinion? Were any of them problemsfor the selection gradient analyses presented by Conner (1988) and Brodie (1992)?


4. What are the problems in measuring fitness identified by Grafen? Again, how seriousare they, and which of them apply to the Conner and Brodie papers?

Wade and Kalisz 1990:

1. From where does fitness arise, in their opinion?

2. What should be experimentally manipulated to determine the target of selection? Whatshould be manipulated to determine the selective agent? Recall the definitions of theseterms.

3. Focusing on the top three panels in Fig. 2, what causes the differences in slopes(selection gradients) across the three densities? What does this tell you about selectiveagents in this case?

ReferencesAndersson, M. 1982. Female choice selects for extreme tail length in a widowbird.

Nature 299:818-820.Arnold, S. J. 1987. Genetic correlation and the evolution of physiology. Pp. 189-215 in

M. E. Feder, A. F. Bennett, W. W. Burggren and R. B. Huey, eds. New Directionsin Ecological Physiology. Cambridge Univ. Press, Cambridge, UK.

Berenbaum, M. R., A. R. Zangerl, and J. K. Nitao. 1986. Constraints on chemicalcoevolution: Wild parsnips and the parsnip webworm. Evolution 40:1215-1228.

Brodie, E. D., III. 1992. Correlational selection for color pattern and antipredatorbehavior in the garter snake Thamnophis ordinoides. Evolution 46:1284-1298.

Bumpus, H. C. 1899. The elimination of the unfit as illustrated by the introducedsparrow, Passer domesticus. Biol. Lect. Woods Hole Mar. Biol. Sta. 6:209-226.

Campbell, D. R. 1996. Evolution of floral traits in a hermaphroditic plant: fieldmeasurements of heritabilities and genetic correlations. Evolution 50:1442-1453.

Conner, J. 1988. Field measurements of natural and sexual selection in the fungus beetle,Bolitotherus cornutus. Evolution 42:736-749.

Conner, J., and S. Via. 1992. Natural selection on body size in Tribolium: possiblegenetic constraints on adaptive evolution. Heredity 69:73-83.

Conner, J. K., R. Davis, and S. Rush. 1995. The effect of wild radish floral morphologyon pollination efficiency by four taxa of pollinators. Oecologia 104:234-245.

Conner, J. K., S. Rush, and P. Jennetten. 1996. Measurements of natural selection onfloral traits in wild radish (Raphanus raphanistrum). I. Selection throughlifetime female fitness. Evolution 50:1127-1136.

Dawkins, R. 1982. The Extended Phenotype. Oxford University Press, New York.deJong, G. 1994. The fitness of fitness concepts and the description of natural selection.

the Quarterly Review of biology 69:3-29.


Dudley, S. A., and J. Schmitt. 1996. Testing the adaptive plasticity hypothesis: Density-dependent selection on manipulated stem length in Impatiens capensis. Am. Nat.147:445-465.

Endler, J. A. 1986. Natural Selection in the Wild. Princeton Univ. Press, Princeton, NJ.Grafen, A. 1988. On the uses of data on lifetime reproductive success in T. H. Clutton-

Brock, ed. Reproductive Success. The University of Chicago Press, Chicago.Grant, P. R. 1986. Ecology and Evolution of Darwin's Finches. Princeton University

Press, Princeton, NJ.Grant, P. R., and B. R. Grant. 2002. Unpredictable evolution in a 30-year study of

Darwin's finches. Science 296:707-711.Kingsolver, J. G., H. E. Hoekstra, J. M. Hoekstra, D. Berrigan, S. N. Vignieri, C. E. Hill,

A. Hoang, P. Gibert, and P. Beerli. 2001. The strength of phenotypic selection innatural populations. Am. Nat. 157:245-261.

Lande, R. 1979. Quantitative genetic analysis of multivariate evolution, applied tobrain:body size allometry. Evolution 33:402-416.

Lande, R., and S. J. Arnold. 1983. The measurement of selection on correlated characters.Evolution 37:1210-1226.

Price, T. D., P. R. Grant, H. L. Gibbs, and P. T. Boag. 1984. Recurrent patterns of naturalselection in a population of Darwin's finches. Nature 309:787-789.

Schluter, D. 1988. Estimating the form of natural selection on a quantitative trait.Evolution 42:849-861.

Schluter, D., T. D. Price, and P. R. Grant. 1985. Ecological character displacement inDarwin's finches. Science 227:1056-1059.

Smith, T. B. 1990. Natural selection on bill characters in the two bill morphs of theAfrican finch Pyrenestes ostrinus. Evolution 44:832-842.

Stewart, S. C., and D. J. Schoen. 1987. Pattern of phenotypic viability and fecundityselection in a natural population of Impatiens pallida. Evolution 41:1290-1301.

Wade, M. J., and S. Kalisz. 1990. The causes of natural selection. Evolution 44:1947-1955.

Weis, A. E., and W. L. Gorman. 1990. Measuring selection on reaction norms: Anexploration of the Eurosta-solidago system. Evolution 44:820-831.

Worley, A. C., and C. H. Barrett. 2000. Evolution of floral display in Eichhorniapaniculata (Pontederiaceae): Direct and correlated responses to selection onflower size and number. Evolution 54:1533-1545.

chapter 6: natural selection on phenotypes and hartl – p. 6-1 from: conner, j. and d. hartl, a...

Documents