The well-known phenotypic diversity of male sexual displays, and the high levels of genetic variation reported for individual display traits have generated the expectation that male display traits, and consequently male mating success, are highly evolvable. It has not been shown however that selection for male mating success, exerted by female preferences in an unmanipulated population, results in evolutionary change. Here, we tested the expectation that male mating success is highly evolvable in Drosophila bunnanda using an experimental evolution approach. Female D. bunnanda exhibit a strong, consistent preference for a specific combination of male cuticular hydrocarbons (CHCs). We used female preference to select for male mating success by propagating replicate populations from either attractive or unattractive males over 10 generations. Neither the combination of CHCs under sexual selection (the sexual signal) nor male mating success itself evolved. The lack of a response to selection was consistent with previous quantitative genetic experiments in D. bunnanda that demonstrated the virtual absence of genetic variance in the combination of CHCs under sexual selection. Persistent directional selection, such as applied by female mate choice, may erode genetic variance, resulting in multitrait evolutionary limits.

Determining how much genetic variation in fitness is maintained in populations has been a major goal of evolutionary biology since the publication of Fisher's fundamental theorem of natural selection (Fisher 1930). Male fitness is primarily determined by mating success, which in many species is a consequence of the mating preferences of females for particular male display traits. The recurrent action of female preference is predicted to erode genetic variation in these male display traits and, consequently, in male mating success, leading to the paradox of the lek (Borgia 1979; Taylor and Williams 1982; Kokko et al. 2003). In contrast to this prediction, substantial genetic variance has been reported for male display traits (Pomiankowski and Moller 1995), leading to the development of theories explaining the maintenance of genetic variation in traits under persistent directional sexual selection (e.g., Pomiankowski and Moller 1995; Rowe and Houle 1996), and to the expectation that male display traits, and consequently male mating success, are highly evolvable.

In support of the expectation that the lek paradox is resolved (and male mating success is evolvable) there are numerous examples of male traits that exhibit heritable variation within populations (Pomiankowski and Moller 1995; Brooks and Endler 2001) or divergence among populations (Hughes and Leips 2006), including a heritable basis to metrics of male attractiveness or mating success (Hughes 1995; Drnevich et al. 2004; Drapeau et al. 2006; Etges et al. 2007; Rundle et al. 2007; Taylor et al. 2007). Recent studies in natural populations have also suggested there may be genetic variation in male fitness (McCleery et al. 2004; Qvarnstrom et al. 2006; Foerster et al. 2007). In addition, artificial selection experiments have been used to demonstrate the evolvability of individual aspects of male mating success, such as morphology (Wilkinson 1993) or behavior (Mackay et al. 2005; Fischer 2006; Meffert and Regan 2006).

However, there have been very few studies that have assessed the ability of female choice itself (rather than experimenter-applied artificial selection) to change the mean of the traits under sexual selection. Indeed, evidence from the only manipulative experiment of this type that we are aware of has indicated that selection on male attractiveness as assessed by females resulted in no evolutionary response of male mating success (Hall et al. 2004). A lack of an evolutionary response to sexual selection has also been observed in a field population of red deer, where male antler size, which is heritable and associated with male mating success, did not evolve over an extended period of observation (Kruuk et al. 2002). Furthermore, observational evidence of the evolution of sexual signals in natural populations is scarce (reviewed by Svensson and Gosden 2007), particularly compared to the observational evidence for natural selection (Endler 1986; Hendry and Kinnison 2001). Svensson and Gosden (2007), in their recent review of the evidence for sexual signal evolution in wild populations, noted that in the few studies in which sexual traits did evolve sexual selection was typically not the agent, rather changes in the natural selection regime were considered responsible.

The lack of observed evolutionary responses for sexual signals is particularly perplexing given that sexual selection tends to be stronger on average than natural selection in field populations (Kingsolver et al. 2001). The general paucity of manipulative evidence, and the apparent contradiction between current observational (Pomiankowski and Moller 1995; McCleery et al. 2004; Qvarnstrom et al. 2006; Foerster et al. 2007; Rundle et al. 2007) and manipulative or long-term evidence (Kruuk et al. 2002; Hall et al. 2004) makes it unclear how much evolutionary potential is typically maintained in male mating success within a single population.

One explanation for this apparent contradiction is that measures of genetic variance in single traits may not be reliable indicators of evolutionary potential in either field or laboratory populations (reviewed in Blows and Hoffmann 2005). Selection is unlikely to act on individual traits in isolation of other aspects of phenotype (as in artificial selection experiments), and sexual signals in particular are probably multimodal and, within a modality may have multiple components (Candolin 2003; Scheuber et al. 2004; Blows and Chenoweth 2006). Recent empirical studies, focused on multicomponent sexual signals, have attempted to estimate the additive genetic variation available in the direction of sexual selection (Blows et al. 2004; Hine et al. 2004; Van Homrigh et al. 2007). These studies, in contrast to estimates of genetic potential in individual traits, observed little genetic variation available in the direction that selection was acting through female preferences in the population.

There are two key limitations to statistical analyses that infer the level of genetic variance in a population from breeding designs. First, it is not possible to definitively demonstrate that genetic variance does not exist (or is at least very small) without large sample sizes (Mezey and Houle 2005; Blows 2007). Although statistical estimates from quantitative genetic experiments suggest additive genetic variation in complex sexual signals is lacking (Blows et al. 2004; Hine et al. 2004; Van Homrigh et al. 2007), practical limitations on our ability to estimate small genetic variances suggest manipulative experiments are required to confirm the inference that there is no genetic variance. Second, realized heritabilities often deviate substantially from the prediction made from quantitative genetic estimates of heritability in the base population (Falconer and Mackay 1996). In particular, this has been observed for traits closely associated with reproductive success, where asymmetrical responses are common (Frankham 1990; Hill and Caballero 1992), and such asymmetries may also be expected for male traits under sexual selection (Rowe and Houle 1996; Svensson and Gosden 2007).

Here, we manipulate male mating success during experimental evolution to test if female preference, as exercised by females themselves, is capable of driving the evolution of male traits. Drosophila bunnanda is a member of the D. serrata species group (Schiffer and McEvey 2006) in which females choose among potential mates based on male cuticular hydrocarbons (CHCs), which act as contact pheromones (Hine et al. 2002, 2004; Blows et al. 2004; Petfield et al. 2005). Female D. bunnanda show a mating preference for a combination of nine male cuticular hydrocarbons CHCs (Van Homrigh et al. 2007). Although individual display traits (CHCs) of D. bunnanda males exhibit genetic variance, as observed for male display traits in many other species, more detailed quantitative analyses revealed little genetic variance for the specific combination of the nine CHCs favored by females (Van Homrigh et al. 2007). At the same time, estimates of the strength of selection on the male traits are high (Van Homrigh et al. 2007). The median selection gradient estimated for the nine CHCs in D. bunnanda was 0.27, stronger than the median level of selection (0.18) found across all taxa reviewed by Kingsolver et al. (2001). The lack of genetic variance in the direction of sexual selection predicts that although female preference exerts strong selection it cannot drive further evolution of male display traits, or male mating success.

We tested these predictions using an experimental evolution approach in which only attractive males were allowed to sire offspring in some populations and only unattractive males were allowed to sire offspring in other populations. Control populations, in which an equal number of attractive and unattractive males were given the opportunity to contribute offspring, were also maintained throughout the 10 generations of experimental evolution. We assessed the response to selection using two different types of traits: attractiveness, assaying both the combination of male display traits (CHCs) known to be under sexual selection and mating latency, a commonly used index of male attractiveness in Drosophila (Taylor et al. 2007) and; male mating success itself, assayed in competitive trials. We show that neither the combination of male display traits under sexual selection or latency, nor male mating success itself evolved in response to our experimental manipulations, demonstrating an evolutionary limit to male mating success in this species.

Materials and Methods


This experiment was initiated from the mass-bred laboratory population of D. bunnanda described in Van Homrigh et al. (2007; see also McGuigan and Blows 2007). This population was founded from eight isofemale lines from Lake Placid, North Queensland, Australia. Nine selection lines were established from this population, three replicate lines for each of the three treatments: attractive, unattractive, and control. Selection lines were established, and selection was applied each generation, using binomial mate choice trials. The conditions of the mate choice trials exactly replicated those used in the experiment reported by Van Homrigh et al. (2007), which generated the prediction that male sexual signals could not evolve further, and led us to the manipulative test reported in this article. Flies from each selection line (including control lines), and from the nonselected stock population were sexed at emergence. Females from the stock population were held individually on 7 mL of standard laboratory media supplemented with baker's yeast (Rundle et al. 2005). Males from the selection lines were held five per vial under the same conditions. Virgins were aged to seven days, then two virgin males (from the same selection line but different holding vials) were introduced into a female's vial. Trials were conducted from 0800 h until 1200 h, and the order in which each selection treatment was assayed was randomized every generation. Occasionally (approximately once in 185 trials) a female failed to choose; males from these trials did not contribute to the next generation in any treatment. This mate choice protocol was followed every generation when selection was applied, and in the mate choice trials used to assess evolutionary response at the conclusion of the experiment.

All populations were maintained at 25°C with 12 h days at a census population size of 50 males and 50 females. Selection lines were initiated following 225 binomial mate choice trials (described above). Males were classified as attractive or unattractive depending on whether they succeeded in mounting the female. The three unattractive populations were each established by 50 unsuccessful males and 50 virgin females (i.e., females that were not exposed to males in mate choice trials); the three attractive populations were each started from 50 successful males and 50 virgin females (not from the mate choice trials); and the three control populations were initiated by 25 successful and 25 unsuccessful males, and 50 virgin females (not from the mate choice trials). Male and female flies were held together to mate and lay for three days, and then discarded. All populations were maintained in two bottles, each containing 25 males and 25 females, mixed each generation. The stock population was maintained at the same density (25 pairs per bottle) to provide the females used to assess male attractiveness every generation.

Selection was applied every generation as described above, with a virgin female from the nonselected stock (ancestral) population presented with two virgin males, both from the same selection line. Fifty such trials were conducted for each attractive and unattractive selection line each generation (300 total trials per generation), and 25 trials per line for the control populations (75 trials per generation). Only males rejected by the stock females were allowed to initiate the next generation in unattractive lines, only successful males initiated the next generation in the attractive lines, and for the control lines both the successful and unsuccessful males were used to found the next generation. Males were placed in the bottles with 50 virgin females from their own selection line to initiate each generation. Selection was applied each generation for 10 generations.

Applying selection to male mating success in this fashion was conducted for two reasons. First, it is known from previous experiments in D. bunnanda (Van Homrigh et al. 2007), and the closely related D. serrata (Hine et al. 2002; Blows et al. 2004; Hine et al. 2004), that females exert strong sexual selection on male CHCs under exactly the same conditions as selection is being applied here. Specifically, |β| estimated for the nine male CHCs in D. bunnanda that are under sexual selection ranged from 0.12 to 0.88 with a median value of 0.27 (Van Homrigh et al. 2007; Table 1). In a review of more than 2500 published estimates of |β| Kingsolver et al. (2001) reported the median value to be 0.18. Furthermore, Kingsolver et al. (2001) noted a negative association between sample size and the magnitude of β, indicating that when β was more accurately estimated it tended to be smaller. Specifically, when sample sizes were greater than 1000, estimates of |β| were mostly ≤ 0.1. The estimates of Van Homrigh et al. (2007) were made on a sample size of over 900 males suggesting that, under the controlled laboratory conditions used during these experiments, sexual selection in on D. bunnanda CHCs is strong relative to field estimates.

Table 1.  Sexual selection on male CHCs.
Trait Ancestral A Selection G10 Selection BG9
  1. Ancestral A: the vector of standardized selection gradients estimated in the nonselected ancestral population (Van Homrigh et al. 2007). Selection G10: the vector of standardized selection gradients estimated after 10 generations of selection. This selection gradient was estimated using females from the nonselected stock population to chose between a male from a selection treatment and a nonselected stock male. Selection BG9: the population regression slopes (across all treatments) estimated as the fixed effect in the random coefficient analysis (Model 4) of the preference functions of the females from the selection lines (after nine generations of selection) for the males from the nonselected stock population.

2-Me-C24 0.135 0.032 0.043
C25:1 (A)−0.876−1.051−0.858
C25:1 (B)−0.163 0.046−0.110
C25H48(B) 0.265 0.437 0.303
7,11-C27:2 0.827 0.803 0.731
C27:1 0.156 0.329 0.101
C27H50 (A)−0.478−0.534−0.204
2-Me-C28 0.294 0.248 0.134

Second, females were able to assess all aspects of male phenotype, mimicking the natural situation more fully than if we applied artificial selection on individual components of male mating success, such as individual male sexual signals (the CHCs themselves), mating speed (e.g., Mackay et al. 2005), or other metrics of attractiveness such as courtship latency (e.g., Taylor et al. 2007). This experimental design therefore allowed us to directly test whether strong sexual selection is capable of changing male trait means. Given the previous quantitative genetic experiment that indicated very little genetic variance was present in the multivariate direction that sexual selection is applied by females (Van Homrigh et al. 2007), our prediction is that it cannot.


We studied the same nine CHC traits as in Van Homrigh et al. (2007) to investigate the evolution of male attractiveness under the three experimental sexual selection regimes. Starting from the second generation in which selection was applied, 30 six-day-old virgin males were sampled for their CHCs each generation for six generations (generations two through seven). Selection was then maintained for a further three generations without sampling.

After 10 generations of selection (i.e., in Generation 11) we again sampled the CHCs of each selection line, this time within the context of the mate choice trials. Selection was applied throughout this experiment on male mating success itself, not on the CHCs that are known to affect male mating success. Therefore, it was important to assess the evolutionary response of male mating success, in addition to that of the sexual signal. From each of the nine selection lines we took 100 males and randomly combined them with a male from the nonselected stock (ancestral) population, and presented both males to a stock female. Stock males had their right wings clipped to allow identification (Van Homrigh et al. 2007). The time it took the stock female to choose a mate (mating latency), a common surrogate of attractiveness in Drosophila studies (Ritchie et al. 1998; Taylor et al. 2007), and the mating success of the male from the selection line was recorded. Selection line males (whether they gained a mating or not) were then sampled for CHCs. Stock males were discarded.


We first assessed whether our experimental procedure applied consistent selection during the experiment by allowing females from the stock population to choose every generation. To do this we analyzed the CHC and mating success data from Generation 11 (the 900 selection line males that were competed against 900 stock males) to estimate the vector of standardized linear selection gradients () using multiple regression (Lande and Arnold 1983). In this analysis of the 900 males from Generation 11 the nine assayed CHCs accounted for significant variation in male mating success (F9,880= 55.32, P < 0.0001, r2= 0.36). We estimated the vector correlation between the  estimated from this analysis (G10) and that reported for the ancestral population in Van Homrigh et al. (2007) (A), estimated in experiments conducted approximately 18 months (36 generations) prior to the estimate obtained in Generation 11 of this experiment. These two estimates of  were very similar (Table 1), with a vector correlation between the two vectors of selection gradients of 0.97. A visual impression of the consistency of the direction of sexual selection across generations is given in Figure 1, in which the linear equations for both estimates of  have been applied to the males measured at Generation 11. This result highlights the repeatability of our characterization of male attractiveness, and strongly supports our assumption that the same selection pressure was applied throughout the 10 generations of selection.

Figure 1.

The combination of CHCs under sexual selection. Male attractiveness was estimated by applying the vector of selection gradients (Table 1) estimated either in Generation 11 (G10: y-axis) or in the ancestral population (Van Homrigh et al. 2007) (A: x-axis) to the individual male CHC measures to generate an attractiveness score for each male assayed in Generation 11 (N= 880).

Because  represents the multivariate trait combination upon which sexual selection is applied most strongly, we were specifically interested in whether there was any phenotypic evolution in this trait combination. We chose to apply the linear equation of A from the ancestral population in all further analyses as this estimate was made on the same dataset as the estimate of genetic variation that supplies our prediction for the evolutionary response (Van Homrigh et al. 2007). We therefore applied the ancestral standardized vector A (Van Homrigh et al. 2007; Table 1) to each male's CHC measures to generate a univariate measure of attractiveness for each male. This univariate measure, called attractiveness, was then subjected to statistical analyses to determine the response to selection.

We assessed the evolutionary response of male attractiveness in several ways. First, we took a typical regression approach, regressing the line means of our trait of interest on generation (Falconer and Mackay 1996). Taking the CHC data from the 30 males sampled per selection line per generation we applied the ancestral A to generate attractiveness scores (as described above). We regressed line mean attractiveness on generation and tested whether the slope of each of the nine populations was significantly different from zero, which would provide evidence that the population had evolved in attractiveness. As a consequence of the potential for random genetic drift to contribute to short (and long) term trends in phenotypic evolution, it was important to consider how idiosyncratic the responses of the replicate selection lines within treatments were (Falconer and Mackay 1996). Therefore a t-test was used to determine if the average slope of the three replicate selection lines within treatment differed from zero. Rejecting this null hypothesis would indicate support for the alternative hypothesis that the selection regime applied by us had caused phenotypic evolution.

Second, several anonymous reviewers requested an analysis of these data (attractiveness scores for 30 males per line per treatment over 6 generations) to determine if the response slopes differed among selection treatments. To do this, we used random coefficient modeling (implemented in SAS version 9.01, SAS, Cary, NC) with the model


where the response variable (y) was the vector of attractiveness scores for all males from the jth selection line within the kth selection treatment. Two fixed effects were modeled, the intercept (α) and the population-wide regression slope G, which describes the mean change in attractiveness over generations across all animals in the experiment (Generation was the only fixed effect in the model). The random effects (δ(P)jk and δ(T)jk) model the departure from the population-wide regression slope (G) of each jth replicate Population within the kth selection treatment, and of the kth selection Treatment, respectively. The design matrices (Xjk, Z(P)jk, and Z(T)jk) relate the individual phenotypic information to the corresponding fixed or random effect.

This analysis determined whether: (1) there was a change in mean attractiveness during the experiment, indicated by significance of the fixed effect, Generation, and; (2) the different selection treatments varied in their change in attractiveness across generations, which would indicate an evolved response to the selection regime. Whether any of the variation in attractiveness across the six generations could be attributed to the selection treatment was determined by comparing the model fit when a single regression slope was fitted (i.e., δ(T)jk was not included in the model) versus when each of the three selection treatments were allowed their own slope (δ(T)jk was included in the model). Model fit was assessed using a chi-square test of the log-likelihood ratio, with one degree of freedom, and halving the resultant probability as necessary for a test of a variance component, which is constrained to be greater than zero (Self and Liang 1987; Littell et al. 1996). Separate intercepts were not fit for random effects as all populations were derived from a common ancestor at the beginning of the experiment.

Third, we used the data from Generation 11, where the large sample size (100 males per line) allowed us to determine the divergence in both attractiveness, and in other combinations of the CHCs. We assessed phenotypic divergence for both a univariate (attractiveness) and multivariate (CHC) analysis in SAS (version 9.1) using the mixed model


where D was the effect of sampling day (mate choice trials were conducted over two consecutive days: fixed effect), T was the effect of the three selection treatments (fixed), M was male mating success (fixed), where males were either successful or unsuccessful in mating trials, T×M was the interaction between treatment and mating success, and P was replicate selection population nested within treatment (random effect). Because of a lack of degrees of freedom, arising from the large number of traits (9) relative to the number of treatments (3) and replicate lines within treatment (3), it was not possible to test the multivariate main effect of treatment in the (nine CHC) multivariate analysis of variance (MANOVA). analysis of variance (ANOVAs) of individual CHCs did not suggest any CHC had responded to the selection regimes (main effect of treatment: F2,6 ranged from 0.17 to 1.80, all P > 0.2 without correction for multiple tests). The treatment by mating success interaction provides a sensitive indication of whether any combination of CHCs changed between successful and unsuccessful males across treatments, and this interaction term was used to assess the multivariate response to selection.

CHCs in male D. bunnanda account for only some of the variation in male mating success (Van Homrigh et al. 2007). We therefore analyzed another measure of male attractiveness, the time it took the stock female to choose between the stock male and a male from one of the selection treatments. Mating latency was natural log-transformed, and analyzed using ANOVA Model 2 (in its univariate form). We also analyzed male mating success itself to assess response to the female preference manipulation. Male mating success after 10 generations of selection was a binomial score, so a generalized mixed model (using a logit link function and binomial distribution) was applied using SAS (version 9.01) and the model:


where mating success was the response variable (y), and the sources of variance were as in Model 2 (Day [fixed], Treatment [fixed], and replicate Population within treatment [random]).


We were also interested in whether female preference had evolved in response to the selection applied to male attractiveness. No selection was applied to female preferences, but they may have evolved as a correlated response to the selection on male attractiveness. In Generation 10 we assessed female preference in the same binomial mate choice design that was used throughout this experiment. From each of the nine selection lines, we took 100 females and each of these females was presented with two males from the nonselected stock (ancestral) population. We again recorded mating latency, the time it took the female to make her choice, and we sampled both of the stock males for their CHCs. As with the assay of male mating success, this analysis does not confound male and female evolution because the male phenotypes that a female was allowed to choose between came from the stock population, which was not subjected to artificial selection during this experiment.

The time it took a female from a selection line to choose between the two stock males was analyzed in a similar manner to the mating latency of stock females presented with selection line males described above. Latency was natural log-transformed, and the main effect of selection treatment tested using ANOVA, based on Model 3 (the response in this case was latency, not mating success, and so Model 3 was run as a standard ANOVA without specifying a bionomical distribution of the response variable). A Student–Newman–Keuls multiple comparison test was applied.

Variation in female preference was analyzed using the random coefficient modeling approach developed in McGuigan et al. (2008). Applied to female preference functions, random coefficient models differ slightly from the analysis of the evolution of male attractiveness described above. Instead of modeling a single independent variable (Generation), analysis of female preferences involves modeling multiple independent variables (male CHCs). In the model:


for each female from each of the jth replicate selection lines within each of the kth selection treatments the response (y) is the male mating success score. Two fixed effects were modeled, the intercept (α) and the experiment-wide regression slope (B) for the variables represented in the design matrix (the male CHCs and sampling Day). The vector of slopes, B, generated in this analysis as fixed effects is equivalent to the directional selection gradient, β (McGuigan et al. 2008). The fixed effect, therefore, estimates the grand population mean linear female preference function across all females. The random effects (for replicate Population and for selection Treatment, respectively: δ(P)jk and δ(T)jk) are two variance–covariance matrices that describe the departure of the regression slope of the jth replicate population within the kth selection treatment, and of the kth selection treatment from the grand population mean slope described by B. The design matrices (Xjk, Z(P)jk, and Z(T)jk) again relate individual phenotypic information to the corresponding fixed or random effect. A chi-square test of the log-likelihood ratio was used to test the effect on model fit of fitting a single preference function for females from all selection treatments (δ(T)jk was left out of the model) compared to allowing variation in preference functions across the three selection treatments (δ(T)jk was included in the model) (Littell et al. 1996). All models were implemented in proc MIXED, SAS ver. 9.01.



Analysis of attractiveness during the first six generations of selection (Fig. 2) indicated no significant response of male attractiveness to the treatments we had imposed. One of the three populations selected for unattractive males displayed a significant change in mean attractiveness, but the average slope of the three replicates was not significantly different from zero (t-test; t2=−3.493, P= 0.073). Selection for attractive males did not change the mean attractiveness significantly in any of the three replicate populations, and the average slope of the three lines was not significantly different from zero (t2=−2.190, P= 0.160). One of the control populations displayed a significant change in mean attractiveness, but again the average slope of the three replicates was not significantly different from zero (t2=−2.219, P= 0.157). The random coefficient analysis supported the general observation of a slight decline in attractiveness during the experiment (the fixed effect of generation was significant: F1,2= 28.91, P= 0.0329, β=−0.035). However, there was no variation around this population mean slope that could be attributed to the selection treatments (χ2= 0.22, df = 1, P= 0.320). Therefore, we had no evidence that male CHC attractiveness evolved in response to the experimental selection regimes after six generations, or consequently that genetic variance for male CHC attractiveness was segregating in the ancestral population.

Figure 2.

The response in male attractiveness over the first six generations of selection (in phenotypic standard deviations). Thirty males were sampled per line per generation. Within each treatment, the solid line represents population 1, the dotted line is population 2, and the dashed line is population 3. Linear regression on the generation means were used to determine if a response had occurred in each population. Unattractive populations; b1=−0.175 ns, b2=−0.083 ns, b3=−0.251 (P= 0.005). Attractive populations; b1=−0.055 ns, b2=−0.103 ns, b3=−0.270 ns. Control populations; b1=−0.063 ns, b2=−0.191 (P= 0.030), b3=−0.048 ns.

After 10 generations of selection the change in mean attractiveness was again assessed. Multivariate analysis of variance (using Model 2) of the nine CHCs failed to reveal any combination of CHCs that had evolved significantly during the 10 generations of experimental sexual selection (mating success × treatment interaction; Wilks' lambda = 0.984, F18,1718= 0.74, P= 0.775). We explicitly asked whether the combination of CHCs known to be under sexual selection (i.e., the attractiveness score) had responded to the selection treatments. Successful males from the attractive populations had the highest mean attractiveness scores, and unsuccessful males from unattractive populations had the lowest mean attractiveness scores (Fig. 3), consistent with the directions in which selection was applied. However, this interaction was not significant (Model 2, mating success × treatment interaction; F2,867= 2.31, P= 0.100), and our experiment consequently provided no evidence for the evolution of the combination of male display traits known to be under sexual selection.

Figure 3.

The response of male CHC attractiveness after 10 generations of selection (N= 880). Males were classified as either unsuccessful (stock female mated with stock male) or successful (stock female mated with selection line male). Treatments left to right are unattractive (solid bar) (N= 164 and 135, unsuccessful and successful), attractive (dashed bar) (N= 147 and 153), and control (gray bar) (N= 142 and 158).

In generation 11 we also assayed our populations for changes in mating success. First, we report that males from the selection lines did not differ in mating success relative to the nonselected stock (ancestral) males that they were competed against: selection line males were successful 447 times and unsuccessful 453 times in 900 trials. We then asked whether males from the different treatments differed in their mating success. There was no difference in mating success across the three selection treatments (Model 3, main effect of treatment: F2,6= 1.87, P= 0.234). Mating success of males from unattractive populations was on average over 5% lower than control populations (Fig. 4A), but this difference was not statistically significant (Model 3, contrast between unattractive and control populations; F1,6=−3.38, P= 0.116). Similarly, there was a suggestion that our selection had affected some aspect of male attractiveness: the nonselected stock females took longer to make a mate choice decision when presented with a male from an unattractive population (Fig. 4B). However, this change was also not significant (Model 2, main effect of treatment; F2,6.028= 2.44, P= 0.168).

Figure 4.

Mating success and attractiveness after 10 generations of selection. (A) Variation in male mating success among selection populations (unsuccessful was scored as zero, successful as one: N= 900) and (B) time taken by a nonselected stock (ancestral) female to choose between a stock (ancestral) male and male from a selection population (N= 886).


The estimate of sexual selection on male CHCs across all treatments in Generation 10, BG9 was very similar to the estimates of  from the nonselected stock females from the other experiments (Table 1: vector correlation between the ancestral A and BG9 was 0.97), suggesting that female preferences had not evolved. Consistent with this interpretation, no variance in female preference was attributable to the selection treatments (Model 4: χ2= 5.7, df = 9, P= 0.770). However, females from the different selection lines did diverge in their mating latency (Model 3, main effect of treatment; F2,6.025= 5.81, P= 0.0395; Fig. 5). The Student–Newman–Keuls test indicated females from the lines selected to be more unattractive were significantly slower to make a mate choice decision than females from other selection lines (contrast: F1,6.025= 15.30, P < 0.001), with no difference in mating latency between females from the lines selected for more attractive males and the control lines selected to maintain the ancestral level of attractiveness (F1,6.025= 2.12, P= 0.1459).

Figure 5.

Mating latency, the time taken by a female from a selection line to choose between two males from the stock (ancestral) population (N= 882).


Ten generations of experimental evolution did not significantly change male display traits or male mating success, despite the strong sexual selection operating on male CHCs. This result is consistent with our prediction based on a previous experiment in which we were unable to detect any genetic variance for male traits in the direction of sexual selection (Van Homrigh et al. 2007). This lack of additive genetic variation in multivariate male sexual displays, coupled with our report of a lack of evolutionary response of that display to persistent experimental sexual selection, suggests the continual application of preference each generation depletes the genetic variance in the male traits, and prevents their further evolution. Mechanisms to increase the retention of additive genetic variation in display traits (e.g., Pomiankowski and Moller 1995; Rowe and Houle 1996) may prove sufficient for individual traits, but do not seem to maintain variance in the complex set of displays that are under sexual selection (Blows et al. 2004; Hine et al. 2004; Van Homrigh et al. 2007). This suggests that evolutionary limits may exist for traits experiencing persistent natural or sexual directional selection.

Although none of our analyses were able to demonstrate significant and consistent evolution in male attractiveness or male mating success in our treatments, trait means were typically in the predicted direction of change (e.g., lower attractiveness and lower mating success in the unattractive lines). We observed a nonsignificant increase in mating latency in the males from the unattractive selection lines (when nonselected stock females did the choosing), coupled with a significant increase in mating latency of females from the same unattractive selection lines. This result indicates that selecting for more unattractive males had a correlated response on female mating behavior, and suggests coevolution of mating behavior between the sexes. Although female mating latency has been associated with male attractiveness measures (Ritchie et al. 1999; Taylor et al. 2007) it is unclear what aspect of behavior evolved in our experiment.

It is possible that a larger experiment may have demonstrated some response to selection in male mating success. It is important to note however, that even if this were the case, the realized responses (and hence the level of genetic variance in the base population) would still be very small, particularly for increasing male attractiveness. There are at least two potential confounding forces that could have contributed to the lack of response we observed in male mating success. First, the opportunity remained for a second round of sexual selection within populations after the application of our treatments (during the three days males and female were held together). For the populations selected to be more attractive, this potential second round of sexual selection should have acted in the same direction as the treatment, increasing the intensity of selection and making the observation of no response more convincing. If females in the control populations persisted in rejecting the same males as were unsuccessful in the mate choice trials we would expect the controls to most closely resemble the attractive populations, and the difference between control and unattractive populations to increase. In the unattractive populations, the potential second round of selection would have opposed the first, slowing the evolution of unattractive male displays. In contrast to these expectations, unsuccessful populations displayed the greatest (although nonsignificant) response to selection over the first six generations (Fig. 2), but were not significantly different from the controls (Figs. 3 and 4).

Second, the general decline in mean attractiveness, including in the control populations, might suggest a low level of inbreeding depression. However, if alleles with detrimental effects on attractiveness existed in the ancestral population we would expect that inbreeding in combination with selection would have driven rapid fixation of such detrimental alleles in the populations selected to be unattractive. Again, such an interpretation is not consistent with the lack of a significant response to selection in the unattractive populations. The general decline in CHC attractiveness may reflect some environmental trend in husbandry, highlighting the importance of maintaining control populations in selection experiments (Falconer and Mackay 1996). Mating success of the nine selection lines was not lower than that of the nonselected stocks after 10 generations of selection, suggesting that the observed decline in CHC attractiveness scores may have been specific to the CHCs.

The dramatic phenotypic diversity of male sexual displays, the high genetic variances reported for individual display traits, and the requirement for females to receive indirect genetic benefits from exercising choice all support the expectation that male display traits are highly evolvable (Pomiankowski and Moller 1995; Rowe and Houle 1996). In contrast, long-term field studies in free-living organisms have provided evidence for a lack of evolutionary response in sexual traits over long periods, in the presence of both directional selection and heritable variation in single traits (Merila et al. 2001; Kruuk et al. 2002). Limits to evolutionary change are receiving increasing attention (Blows and Hoffmann 2005), particularly with regard to multivariate genetic constraints (Mezey and Houle 2005; Hine and Blows 2006; Foerster et al. 2007). Although single traits may exhibit substantial levels of genetic variation, persistent sexual (or natural) selection may deplete genetic variance in those trait combinations that respond to selection, supplying a general mechanism for evolutionary limits. In other words, genetic variation may be not lacking for sexual traits in general, but rather for the specific combinations of traits that are favored (Hine et al. 2004; Van Homrigh et al. 2007). This suggests that a change in the trait combination favored by selection may result in rapid phenotypic evolution, although the evidence for this is currently limited (Svensson and Gosden 2007). Thus, the dramatic phenotypic diversity of male sexual displays may depend on frequent changes in selective optima, generated by either a change in the environment or in female preference itself.

Associate Editor: K. Hughes


Financial support for this experiment was provided by the Australian Research Council. We thank D. Petfield, E. Hine, Y. Eiby, and M. Higgie for assistance with maintenance and sampling of selection lines; and L. Rowe, K. Hughes, and four anonymous reviewers for their detailed and constructive comments on previous drafts.