The maintenance of genetic variation in traits closely associated with fitness remains a key unresolved issue in evolutionary genetics. One important qualification on the observation of genetic variation in fitness-related traits is that such traits respond asymmetrically to selection, evolving to a greater extent in the direction of lower fitness. Here we test the hypothesis that standing genetic variation in fitness-related traits is principally maintained for unfit phenotypes. Male Drosophila bunnanda vary in mating success (the primary determinant of male fitness) due to female mate choice. We used competitive mating success to partitioning males into two groups: successful (high fitness) and unsuccessful (low fitness). Relative to successful males, unsuccessful males harbored considerably greater levels of additive genetic variation for sexual signaling traits. This genetic asymmetry was detected for a multivariate trait that we demonstrated was not directly under stabilizing sexual selection, leading us to conclude the trait was under apparent stabilizing selection. Consequently, our results suggest genetic variance might be biased toward low fitness even for traits that are not themselves the direct targets of selection. Simple metrics of genetic variance are unlikely to be adequate descriptors of the complex nature of the genetic basis of traits under selection.

Genetic variation in natural populations is epitomized by two basic observations. First, genetic variation in individual traits is ubiquitous, with very few exceptions to this rule (Lynch and Walsh 1998; Blows and Hoffmann 2005). Second, traits highly correlated with fitness tend to express most genetic variation (Houle 1992). As stabilizing and directional selection should eliminate genetic variation, both of these observations sit uncomfortably with the established pervasiveness and strength of selection in natural populations (Endler 1986; Conner 2001; Hoekstra et al. 2001; Kingsolver et al. 2001; Rieseberg et al. 2002; Hereford et al. 2004; Knapczyk and Conner 2007). Reconciling these apparently contradictory observations is a long-standing aim in evolutionary genetics (Barton and Turelli 1987; Bulmer 1989; Rowe and Houle 1996; Bürger 1998; Hill and Mbaga 1998; Turelli and Barton 2004; Zhang et al. 2004). However, there is currently no compelling explanation for the maintenance of the observed high levels of genetic variation in fitness traits given the established strength of selection (Johnson and Barton 2005).

It is possible that the paradox of strong selection and high genetic variance is an artifact of how we estimate these parameters (Johnson and Barton 2005). Although strong stabilizing selection has been reported for many traits (Kingsolver et al. 2001), it is unlikely that all such traits truly affect fitness. Apparent stabilizing selection, in which the trait does not itself affect fitness but is pleiotropically correlated with a trait that does, is likely to be common (Johnson and Barton 2005). Multivariate stabilizing selection appears to be much stronger on some trait combinations than others (Blows and Brooks 2003), supporting this view that only a subset of traits are under real stabilizing selection.

Similarly, genetic variance does not appear to be evenly distributed among all trait combinations in multivariate trait space (Kirkpatrick and Lofsvold 1992; Schluter 2000; Hansen et al. 2003; Mezey and Houle 2005; Hine and Blows 2006; McGuigan and Blows 2007). Furthermore, multivariate trait combinations under directional (Blows et al. 2004; Hine et al. 2004; Van Homrigh et al. 2007) or stabilizing (Hunt et al. 2007) selection have been demonstrated to be associated with low levels of genetic variance. Together, these observations suggest genetic variation is not as ubiquitous as it appears, and that measures of genetic variance in individual traits might not accurately reflect the availability of genetic variation to selection (Blows 2007; Hansen and Houle 2008; Kirkpatrick 2009).

Low estimates of genetic variance for specific trait combinations affecting fitness suggest that genetic variance itself is more responsive to selection than expected under the classic infinitesimal model of quantitative genetics (Bulmer 1971; Lande 1980). The extent to which genetic variance responds to selection on the traits depends on a number of genetic details that are difficult to establish empirically (Barton and Turelli 1987; Johnson and Barton 2005). Information from selection and quantitative genetic experiments can provide insight into the evolutionary dynamics of genetic variances without needing first to solve the empirical challenges of measuring specific details of the genetic architecture (Turelli 1988).

Perhaps the most well-supported and consistent empirical observation relevant to how genetic variances evolve is asymmetrical phenotypic responses to artificial selection (Frankham 1990; Hill and Caballero 1992; Falconer and Mackay 1996). In a review of 30 studies selecting on reproductive fitness Frankham (1990) reported that 24 studies observed a smaller response to selection for higher fitness than to selection for lower fitness. Several potential mechanisms have been proposed to underlie asymmetrical responses to divergent selection: drift, different selection differentials, inbreeding depression, maternal effects, scalar asymmetry, indirect selection, and genetic asymmetry (Frankham 1990; Hill and Caballero 1992; Falconer and Mackay 1996). This last mechanism, genetic asymmetry, which might be underlain by asymmetric allele frequency, genes of major effect, or directional dominance, is expected to be the most important (Frankham 1990; Hill and Caballero 1992; Falconer and Mackay 1996).

Fitness traits are, by definition, under strong selection, which might generate strongly asymmetrical allele frequencies through fixation of favorable alleles. Further, there is both a theoretical expectation (Fisher 1930; Orr 1998) and empirical evidence (Caballero et al. 1991; Mackay et al. 1994; Lynch et al. 1999; Eyre-Walker and Keightley 2007; Haag-Liautard et al. 2007) that most mutations are deleterious with respect to fitness. This mutation bias would ensure that the greatest proportion of low-frequency alleles would contribute to lower fitness. When the selection regime is altered to favor previously unfit phenotypes, increasing frequencies of either new mutations or alleles previously held at low frequency by selection could result in increasing heritability and cause a rapid response to selection (Hill and Caballero 1992; Falconer and Mackay 1996). The effect of asymmetry in allele frequencies may be further exacerbated by differences in allelic effect sizes (Frankham 1990; Hill and Caballero 1992; Falconer and Mackay 1996). Alleles of large effect contribute disproportionately to genetic variance, and have a greater probability of being deleterious (Fisher 1930; Orr 1998).

Genetic asymmetry provides an intuitive explanation for asymmetric responses to selection, which in turn suggests a failure of the infinitesimal model. However, asymmetric responses to selection do not directly inform as to whether standing genetic variation in natural populations is asymmetric (Frankham 1990; Hill and Caballero 1992). Mbaga and Hill (1997) found little evidence of asymmetry in standing genetic variation for body size in mice, a trait that responds asymmetrically to bidirectional selection (Falconer 1953, 1955, 1973). Asymmetry of standing additive genetic variation has been implicated through nonlinear parent–offspring regression in several studies (reviewed in Mbaga and Hill 1997), but the question of whether additive genetic variance in populations is principally maintained for low-fitness phenotypes has not been directly addressed.

Here, we directly estimate the standing genetic variance underlying high and low fitness. Male mating success is the primary determinant of male fitness in many species in which males contribute only genes to their offspring. Consequently, males that are successful in gaining matings, by definition, carry alleles with net beneficial effects on fitness whereas males unsuccessful in gaining matings carry alleles with detrimental effects on fitness. By exploiting an experimental system in which we can unequivocally distinguish between individuals that differ in fitness, we can determine whether standing genetic variation differs for high versus low-fitness groups.

Drosophila bunnanda is a rainforest endemic of northeastern Australia, and a recently described member of the D. serrata species group (Schiffer and McEvey 2006). In this species group, male cuticular hydrocarbons (CHCs) affect male fitness through their contribution to male mating success (Blows and Allan 1998; Blows 2002; Howard et al. 2003; Higgie and Blows 2007). We have previously reported strong directional sexual selection acting on male CHCs through female mate choice in both D. bunnanda (Van Homrigh et al. 2007; McGuigan et al. 2008a) and in its close relative D. serrata (Blows et al. 2004; Hine et al. 2004). Although individual CHCs are associated with additive genetic variation, the specific combination of CHCs favored by females is associated with very low levels of additive genetic variance (Blows et al. 2004; Hine et al. 2004; Van Homrigh et al. 2007). These studies suggest that genetic variance for CHCs has responded to the persistent directional selection applied through female mate choice, and further suggest the possibility that standing genetic variation for these fitness-related traits will be biased toward low-fitness phenotypes. The experimental coupling of mate choice trials with a large quantitative genetic breeding design in the study reported in Van Homrigh et al. (2007) provided us with the opportunity to explicitly test this hypothesis. In this paper, we determine that high fitness (successful in gaining a mating) and low-fitness (unsuccessful in gaining a mating) males differ in the genetic basis of their CHCs, and characterize these genetic differences.

Materials and Methods


The genetic design and sample collection has been described elsewhere (McGuigan and Blows 2007; Van Homrigh et al. 2007). Briefly, we used paternal half-sibs to estimate genetic parameters, mating each of 125 males to four females to generate a total of 500 paternal half-sib families. At emergence, we collected two sons per family (a total of 1000 males). The mating success (fitness) of each of these males was determined using a bivariate mate choice trial in which a female chose between a male from this paternal half-sib breeding design and a random male from the same population, of the same age, raised under the same conditions. The choosing females were also sourced from the paternal half-sib breeding design, but in any particular mate choice, trial was unrelated to the male from the paternal half-sib design (and the random male) (McGuigan et al. 2008b).

In the D. serrata species complex females actively control copulation by forcibly dislodging rejected males shortly after they mount and before they transfer sperm (Hoikkala et al. 2000). Thus, although some male–male interactions occur in the presence of females, mating success in these trials reflects female choice (Hine et al. 2002). The direction of selection applied to male traits by female choice in D. bunnanda is consistent across experimental conditions, and over multiple generations (McGuigan et al. 2008a).

Males from the paternal half-sib breeding design were clipped on the left wing whereas the random males had a small piece clipped from their right wing to allow discrimination. Paternal half-sib males were recorded as being successful or unsuccessful depending on whether they or the random male mated with the female. Importantly, males from the breeding design were rejected as frequently as they were accepted, leading to roughly equal numbers of successful and unsuccessful males. Mate choice trials were conducted over three successive days with all flies in any particular trial of the same age (6, 7, or 8 days old). At the conclusion of the mate choice trial, cuticular hydrocarbons were extracted in hexane. A gas chromatograph was used to analyze the CHC composition of each male, with peak area calculated for 10 peaks, the area of each peak divided by the total CHC content and then logcontrast scores calculated for nine CHCs using C25 as the divisor (Van Homrigh et al. 2007). Logcontrast scores were standardized (mean of zero, standard deviation of 1) prior to analysis.

In this experiment, males were assigned to fitness categories based on whether they were successful or unsuccessful in the binomial mate choice trials. From previous analysis of this data we know that CHCs are under strong directional sexual selection through female mate choice (Van Homrigh et al. 2007). Van Homrigh et al. (2007) used a multivariate regression analysis to estimate the sexual selection on CHCs (Lande and Arnold 1983). Of the 14 CHCs measured in D. bunnanda males (excluding C25, the divisor for logcontrasts), the best model of sexual selection contained the nine CHCs analyzed here (Table 1). This model accounted for 11% of the variation in male mating success. The selection gradients for six of the nine CHCs were significant, and furthermore five of the CHCs experienced selection stronger than the median magnitude of directional sexual selection gradients (0.16) found across all taxa by a recent comprehensive review (Kingsolver et al. 2001) (Table 1). Thus, CHCs are considered as fitness-related traits, and expected to exhibit genetic asymmetry.

Table 1.  The trait combination (vector) describing sexual selection (β), the first eigenvectors of I, and of G for unsuccessful (U), successful (S), and all males, and the additive genetic variance (VA) associated with vectors of G.
 β1imaxU gmaxS gmaxPooled gmax2
  1. 1The linear selection gradient reported in Table 1 of Van Homrigh et al. (2007).

  2. 2First eigenvector of the G reported in Table 2 of Van Homrigh et al. (2007).

VA 0.0212.9581.5761.248
2-Me-C24 0.1350.2790.0520.5300.187
C25H48(B) 0.2650.3530.3270.1170.088
7,11-C27:2 0.8270.2240.4210.1940.526
C27:1 0.1560.3140.2930.2360.282
2-Me-C28 0.2940.3110.1620.4260.233


We estimated the additive genetic (sire) covariance matrices (G) separately for the successful and unsuccessful males to compare the genetic basis of the nine CHCs. G was estimated using restricted maximum likelihood in the multivariate mixed model:


where sampling day (T) was a fixed effect, and variance-covariance matrices at the sire (S) and dam nested within sire (D) levels were random effects. Analyses were implemented using the MIXED procedure in SAS (version 9.1; SAS Institute Inc., Cary, NC). We used factor analytic modeling (Hine and Blows 2006) to determine statistical support for the estimated G.

There are several approaches for the statistical comparison of G-matrices that have been separately estimated from different populations (Steppan et al. 2002). These methods involve estimating G from each population, followed by the comparison of various aspects of the geometry of these covariance matrices (Phillips and Arnold 1999; Steppan et al. 2002; Arnold et al. 2008). Here, we take advantage of an unusual feature of our experimental design to adopt a different approach to comparing G. Because individual sires had sons that were sexually successful and sons that were sexually unsuccessful, we were able to use an approach analogous to the determination of a genotype by environment interaction using univariate mixed models (Lynch and Walsh 1998, p. 667). Under this approach, we tested the hypothesis that G differs between the two groups of males and then characterized these genetic differences, all within the same restricted maximum likelihood mixed model framework in which the G were estimated. Using this approach, we identified those trait combinations whose genetic bases differ most between the two groups of males, as opposed to searching for differences in the geometry of G-matrices that were estimated in separate analyses.

To test the hypothesis that the genetic basis of the CHC traits differed between the two fitness groups we analyzed all males simultaneously for the nine CHCs using the multivariate mixed model:


which differs from model (1) by the addition of mating success of each individual (M) as a fixed effect, and a variance–covariance matrix for the interaction between sire effect and the son's mating success (I) as a random effect. The covariance matrix I accounts for the difference in the genetic basis of the traits between the two groups of males (i.e., differences between G). Significant variance in the I matrix indicates that sire (genetic) effects on successful and unsuccessful sons differ. We used factor analytic modeling with log-likelihood ratio tests (Hine and Blows 2006) to demonstrate that the variance contained in I accounted for a significant amount of the variance in CHCs. That is, to test the hypothesis that nine CHC G estimated for sexually unsuccessful males differed from the nine CHC G estimated for sexually successful males. This approach has the further advantage that it allows us to generate a low-dimensional interpretation of observed genetic differences by estimating the statistically supported eigenvectors of the covariance matrix I.

As detailed previously, male mating success (fitness) in D. bunnanda is strongly associated with a specific linear combination of the nine CHCs (Van Homrigh et al. 2007; McGuigan et al. 2008a). We therefore determined whether this specific combination of CHCs associated with fitness was genetically differentiated between the two groups of males. Using the loadings of the vector of linear sexual selection gradients (β: Table 1) we calculated a CHC attractiveness score for each male, applying the coefficients in β to each male's standardized logcontrast CHC profile. The genetic variation in male CHC attractiveness was determined by applying model (1) separately for each group of males, and the hypothesis of a different genetic basis was tested using model (2).


Previous analysis of all males in this experiment (irrespective of mating success) using factor-analytical modeling at the sire level indicated statistical support for four dimensions (eigenvectors) of G, accounting for 95% of the genetic variance (Van Homrigh et al. 2007). Although the number of sires remained high when the two groups of males were analyzed separately (of the 125 sires in the full dataset, 121 and 122 contributed to unsuccessful and successful males, respectively), dividing the dataset into the two fitness groups halved the number of individual sons available for estimating each G (444 unsuccessful and 459 successful males). To determine whether splitting the dataset adversely affected our power to detect genetic variance we used factor analytic modeling of the sire-level covariances (i.e., of G). Four dimensions of G were supported in unsuccessful males (4 to 3 dimensions: χ2= 14.24, df = 6; P= 0.027), accounting for 97% of the genetic variance. In the successful males, three dimensions of G were supported (3 to 2 dimensions: χ2= 18.98, df = 7; P= 0.008), accounting for 90% of the genetic variance. These results indicate we had sufficient statistical power to estimate G separately in each group of males, but suggested less genetic variance (requiring greater statistical power to detect) in the successful males.

Visual inspection of the nine CHC G (Table 2) revealed that individual trait variances were greater for seven of the nine CHCs in the unsuccessful males. The total genetic variance (estimated as the sum of the variances (diagonals in Table 2), or as the trace of G (the sum of the eigenvalues; see Kirkpatrick 2009) for unsuccessful males (4.24) was 1.63 times greater than that for the successful males (2.61). These observations further suggest that additive genetic variance might differ between the two groups of males. Trait covariances were predominantly positive in both groups of males, with only six of the 72 covariances being negative. Although the covariances were not identical between G (Table 2), it is difficult to interpret differences in genetic covariation simply from visual inspection of the two matrices. We therefore tested whether the two G differed by applying factor analytic modeling of the sire by mating success interaction (I).

Table 2.  The genetic variance (bold, on the diagonal) covariance matrix (G) for the nine CHCs in unsuccessful and successful males.
  2-Me-C24 C25:1(A) C25:1(B) C25H48(B) 7,11-C27:2 C27:1 C27H50(A) 2-Me-C28 2-Me-C30
 Unsuccessful males
  C25:1(A) 0.1060.728       
  C25:1(B) 0.164 0.6180.605      
  C25H48(B) 0.095 0.482 0.3890.580     
  7,11-C27:2−0.039 0.625 0.505 0.2790.639    
  C27:1−0.194 0.440 0.265 0.229 0.4320.360   
  C27H50(A) 0.120 0.416 0.353 0.333 0.330 0.2310.401  
  2-Me-C28 0.051 0.223 0.157 0.083 0.187 0.106 0.1650.167 
  2-Me-C30 0.049 0.401 0.296 0.242 0.370 0.323 0.295 0.2820.454
 Successful males
  C25:1(A) 0.1830.144       
  C25:1(B) 0.323 0.2000.341      
  C25H48(B) 0.244−0.030 0.1020.196     
  7,11-C27:2 0.128 0.104 0.142−0.1620.171    
  C27:1 0.110 0.113 0.137−0.021 0.1080.162   
  C27H50(A) 0.365 0.082 0.173 0.069 0.060 0.0840.305  
  2-Me-C28 0.289 0.109 0.188 0.059 0.133 0.190 0.2370.355 
  2-Me-C30 0.124 0.104 0.106−0.040 0.131 0.210 0.131 0.3320.377

Two dimensions of I were statistically supported (difference between 2 and 1 dimensions: χ2= 30.34, df = 8; P < 0.001), demonstrating a difference in G between males of high versus low fitness. Most (88%) of the genetic differences between the two groups of males were attributable to a single multivariate trait combination (the first eigenvector of I, imax; Table 1). Because visual inspection of the two G suggested differences in total additive genetic variation, we estimated the additive genetic variance associated with imax in each group. This was done by projecting imax into the space of G using imaxTGimax (Kennedy et al. 1993). There was twice as much additive genetic variance for imax in the G estimated from unsuccessful (VA= 2.60) compared to the G for successful (VA= 1.33) males, supporting our hypothesis that more genetic variation was maintained for low-fitness phenotypes.

The vector imax was characterized by similar contributions from all CHCs, and described a very similar trait combination to that described by the first eigenvector of G (gmax) (Table 1; Fig. 1). The vector correlations between imax and the gmax of unsuccessful, successful and all males were, respectively, 0.93, 0.92, and 0.89. This strong similarity of imax and gmax suggested that the genetic difference between high- and low-fitness males was in the multivariate trait combination that was associated with the most additive genetic variance within each fitness group, gmax. To directly determine whether there was a difference in genetic variance in gmax between the two groups, we applied the coefficients of gmax (Table 1) to each male's CHC profile to generate univariate gmax scores, which were then analyzed using model (2) to test the hypothesis that these two groups of males differed in the genetic variance associated with gmax. Inclusion of the sire by mating success interaction term significantly improved model fit (χ2= 4.496, df = 1; P= 0.017, one-tailed test), indicating statistical support for a difference in the variance accounted for by gmax. In unsuccessful males, gmax was associated with 1.9 times more additive genetic variance than in successful males (Table 1; Fig. 2). Therefore, we conclude that the group of sexually unsuccessful males was associated with greater additive genetic variance for the multivariate trait gmax than was the group of sexually successful males. Thus, a difference in variance in gmax was the major difference between the two G, rather than any difference in CHC covariation.

Figure 1.

The relationship between individual imax scores (determined by applying the major eigenvector of I to each male's CHC profile) and gmax scores (determined by applying the major eigenvector of G (estimated for each fitness group separately) to each male's CHC profile) in sexually unsuccessful (black squares) and successful (gray circle) males.

Figure 2.

The distribution of breeding values for A) gmax scores and B) β scores. Sire breeding values were estimated for sexually successful and unsuccessful males separately using model (1).

For the combination of CHCs explicitly associated with fitness, the β attractiveness scores, there was very little genetic variation across the whole population (VA= 0.021: Van Homrigh et al. 2007). When estimated for each group of males separately, genetic variance was greatest in the direction of lower fitness (unsuccessful males VA= 0.022; successful males VA= 0.015; Fig. 2). However, there was no statistical support for a genetic difference in CHC attractiveness between the fitness groups (sire by mating success interaction in model 2: χ2= 0.00, df = 1; P= 0.500).

The genetic asymmetry in gmax between sexually successful and unsuccessful males is consistent with both mutational bias toward low fitness and selection toward high fitness. However, we could find no evidence that gmax is a trait that affects fitness. We applied the vector of trait loadings describing gmax in the whole population (Van Homrigh et al. 2007; Table 1) to the CHC scores of each male to generate univariate gmax scores. We subjected this univariate trait to second-order polynomial regression analysis (Lande and Arnold 1983) and found no evidence of linear (F1,901= 2.855, P= 0.091) or nonlinear (F1,901= 0.10, P= 0.7501) sexual selection. The sample size in this dataset is large (903 males), and we have considerable statistical power to detect selection (Kingsolver et al. 2001). In contrast to the nonsignificant selection on the univariate estimate of gmax, regression analysis of the univariate β CHC attractiveness scores indicated strong directional selection (F1,901= 149.580, P < 0.0001, b= 0.3773). Consistent with our interpretation that gmax is not a trait combination under sexual selection, the vector correlation between gmax and β (Table 1) was 0.02.


Understanding how genetic variance for fitness traits evolves is a central aim of evolutionary genetics (Barton and Turelli 1987; Rowe and Houle 1996; Johnson and Barton 2005; Zhang and Hill 2005). Here we take an underused approach to investigating the evolution of genetic variances, combining a quantitative genetic experiment with a fitness assay in D. bunnanda to test the hypothesis that standing genetic variation within a population was biased toward low-fitness phenotypes. The genetic basis of sexual signal traits (CHCs) in unfit males was characterized by significantly more additive genetic variation than for fit males. Substantial asymmetry in the distribution of standing genetic variance among groups of individuals within a single population is a novel empirical observation, and indicates that the distribution of breeding values for these traits does not conform to the standard normality assumption of quantitative genetics (Bulmer 1971; Lande 1980).


The large difference in genetic variance associated with gmax, coupled with the lack of difference in gmax trait mean between sexually successful and unsuccessful males, suggests that stabilizing selection (selection on the variance) is operating on this multivariate CHC trait. However, this is not stabilizing sexual selection, at least in the way that we measure it. Regression analyses provided no evidence that female preferences generated stabilizing sexual selection on gmax. We therefore suggest that the large difference in genetic variance in gmax has been generated by apparent stabilizing selection.

Apparent stabilizing selection occurs when selection acting on an unknown, genetically (pleiotropically) correlated trait is manifested as stabilizing selection on the analyzed trait (Roberston 1956; Hill and Keightley 1988; Barton 1990; Kondrashov and Turelli 1992). Not all traits can be under real stabilizing selection, affecting fitness independently, because population fitness would then be very low (Johnson and Barton 2005). Apparent (as opposed to true) stabilizing selection is therefore likely to be relatively common in the presence of pleiotropy between traits and fitness. The lack of evidence of direct sexual selection on gmax through female preference, and the fact that gmax and β are almost orthogonal, indicates that the indirect selection on gmax is not mediated through the sexual signal function of the CHCs.

Cuticular hydrocarbons are known to be involved in resource allocation tradeoffs with a number of functions in insects (Stanley-Samuelson and Nelson 1993), because the constituent lipids comprise a primary energy reserve. For example, in D. melanogaster a direct pleiotropic relationship exists between CHCs and female oocyte production (Wicker and Jallon 1994). In D. bunnanda, male CHCs are genetically correlated with wing size, a measure of body size, and condition (Van Homrigh et al. 2007). Competition among traits for the same lipid resources implies the potential for natural selection to operate indirectly on CHCs. Importantly, if selection on a component of fitness other than male mating success has lead to asymmetry of gmax due to pleiotropic correlations, males with low mating-success fitness must also have low fitness through this other, unknown, component of fitness. That is, for us to observe genetic asymmetry in gmax between the two groups that we defined through male mating success, aspects of fitness genetically correlated with gmax must sort males in the same way as mating success. When sexually selected traits evolve condition-dependent trait expression it is expected that these condition-dependent sexual signals will become pleiotropically related to other fitness components, which all depend on the same common pool of resources (Rowe and Houle 1996; see also Whitlock and Agrawal 2009).

The observation that variance in gmax is greater for unfit males suggests that additive genetic variation in male CHCs might predominantly be maintained by alleles with detrimental effects on fitness. Evidence that standing genetic variance is biased toward lower fitness in a nonmanipulated population supports the general empirical observation of asymmetric selection responses (Frankham 1990; Hill and Caballero 1992; Falconer and Mackay 1996). Furthermore, our results suggest that this bias can exist even for trait combinations that are not themselves the direct targets of selection. Genetic asymmetry might be due to directional dominance, asymmetrical allele frequencies or major genes (Frankham 1990; Hill and Caballero 1992; Falconer and Mackay 1996). For quantitative traits, the frequency of contributing alleles and the distribution of their phenotypic effects have proven surprisingly resistant to empirical estimation (Charlesworth et al. 2007; Macdonald and Long 2007; Kelly 2008). It therefore remains to be determined how genetic asymmetry is generated.


The CHCs of male D. bunnanda are under strong directional sexual selection through female mate choice (Van Homrigh et al. 2007; McGuigan et al. 2008a). Very low additive genetic variation remains for the specific combination of CHCs under sexual selection in the population as a whole (Van Homrigh et al. 2007) or, as we have shown here, specifically for low- or high-fitness males. A bidirectional artificial selection experiment verified this lack of genetic variation, with no evolutionary response to selection for either increased or decreased mating success (McGuigan et al. 2008a). These results indicate that the genetic variance for traits closely related to male mating success is low, but not demonstrably asymmetric. However, the low level of genetic variance in β impedes our ability to detect asymmetry, and this negative result should be interpreted cautiously given the lack of statistical power resulting from the very low variance.

Correlational (Hunt et al. 2007) and manipulative (Hall et al. 2004) evidence of very low genetic variance in multivariate male sexually selected traits is available from other taxa. These observations suggest sexual selection may typically be very effective at depleting genetic variance in the phenotypes that are the specific targets of female preference. It is important to note that the use of binomial mate choice trials to define fitness in our experiment is likely to have underestimated any bias in genetic variance in male attractiveness. Male mating success in many species is often more skewed than our binomial tests allowed, with few males gaining most matings (Andersson 1994). For example, in D. melanogaster, highly skewed male mating success is likely to contribute to the low effective population sizes that have been reported under laboratory rearing conditions (Briscoe et al. 1992).

The ability of females to assess male genetic quality is central to theories explaining the maintenance of female preference for male sexual traits (Pomiankowski et al. 1991; Rowe and Houle 1996; Reinhold 2002; Lorch et al. 2003), and the maintenance of sexual reproduction through the reduction of mutational load (Whitlock 2000; Agrawal 2001; Siller 2001). Maintenance of female mate preference by a good genes process in D. bunnanda may not depend on substantial genetic variance in the male sexual display, but rather on the ability of females to detect the phenotypic consequences of new mutations each generation (Whitlock 2000; Tomkins et al. 2004). This suggests the possibility that CHCs are a sensitive indicator of deleterious mutations, but that selection is very effective at removing deleterious alleles and fixing beneficial ones. If females can detect (and avoid) novel deleterious mutations through their effect on male displays, the frequency of detrimental alleles will be kept at low levels and therefore not contribute to population parameters such as the additive genetic variance. Further insights into the maintenance of female mate choice for indirect benefits may not depend on demonstrating additive genetic variation in male quality, but rather on demonstrating the ability of females to detect (and avoid) novel mutations (Radwan et al. 2004; Tomkins et al. 2004).


The contradictory observations of the ubiquity of genetic variation and pervasiveness of selection represent a major challenge to evolutionary genetics (Johnson and Barton 2005). One possible resolution is that these observations are misleading, and that most traits are not directly under selection, and therefore that much of the genetic variation present in a population is not subject to strong contemporary selection. Simple metrics of genetic variance are unlikely to be adequate descriptors due to the complex nature of the genetic basis of traits under selection (Rice 2008; Blows 2007; Hansen and Houle 2008; Kirkpatrick 2009). Much of theoretical and empirical quantitative genetics has relied on the assumption that the distribution of breeding values is symmetrical about the population mean. Although this assumption may hold for simple metric traits, the genetic basis of multivariate phenotypes closely associated with fitness may differ substantially from the expectation of symmetry. Interpreting metrics of genetic variance for single traits without consideration of genetically correlated traits or of the selection acting on those traits may therefore be of little value in understanding the behavior of genetic variation in natural populations. A more detailed knowledge of the consequences of persistent selection on the genetic variance is required if we are to understand how genetic variance, and evolutionary potential, is maintained in natural populations.

Associate Editor: J. Wolf


We thank Aneil Agrawal, Locke Rowe, Bruce Walsh, and anonymous Reviewers for comments on previous drafts. This manuscript benefited greatly from the constructive criticisms from Editor Nick Barton and Associate Editor Jason Wolf, as well as insightful comments from several anonymous reviewers. We gratefully acknowledge the work of A. Van Homrigh and M. Higgie in generating the CHC data. This research was supported by the Australian Research Council.