QST vs. FST: methodological issues
We did not find any evidence of selection for different optima in the populations studied. The divergence found for all traits can be explained by drift alone, as QST was never significantly different from FST, either calculated according to a full or a half-sib design. Importantly, these conclusions only hold if QST was calculated without any methodological bias. In practice, this is rarely the case and we will discuss below how several statistical or biological issues can influence our results and more generally the outcome of QST–FST experiments.
First, the number of populations studied is relatively low even though many families (30 per population) and individuals (total of 600) were used for the quantitative analyses. O'Hara & Merilä (2005) showed that the bias and the variance in the estimation of QST are especially large when few populations are studied. However, the bias and the variance in QST due to the number of populations decrease for low values of QST. Goudet & Büchi (2006) found similar results in a simulation study, with a marked decrease in the variation of QST for values around 10% and below, which is in the magnitude of our results.
A second point to consider is the type of statistical tool used to analyse quantitative traits. The most widely used method to extract components of genetic variance from phenotypic data is the classical hierarchical anova based on either the method of moments or REML analyses. These two frequentist approaches gave similar results. The outcomes from Bayesian analyses were more difficult to interpret because for two traits out of four, posterior distributions had odd shapes with CI for QST between 0 and 1. This problem can be explained by (i) low sample size for fecundity and (ii) heterogeneity of within-population variance for age at maturity (Waldmann et al., 2005). Furthermore, the shape of the posterior distributions was sensitive to the prior used (Gamma vs. Uniform) and this is still unclear which prior should be used for inferences on QST: Waldmann et al. (2005) found similar results using Gamma and Uniform distributions, while O'Hara & Merilä (2005) showed that the latter had better frequentist properties. We found that for heritability a Uniform distribution performed better than a Gamma, but for QST we observed the reverse. To further investigate this issue we made some simulations with a setting based on our sampling scheme (600 individuals distributed in four populations with 30 families each) and according to the method described in Goudet & Büchi (2006). For a trait with a purely additive basis and an expected value for FST of 0.18, we found a better estimate of QST using a Gamma distribution than a Uniform distribution. However, this was only true for the number of populations corresponding to our sampling design because with more populations the two priors performed similarly.
Importantly, the three statistical methods used gave similar results for size at maturity, which suggests that they perform equally well when the genetic variance is evenly distributed within the different populations and even when few populations are present. The results obtained with frequentist and Bayesian analyses have rarely been compared in quantitative genetics studies but recently O'Hara & Merilä (2005) found similar estimates for QST using REML and Bayesian methods, except a slight downward bias with REML for high QST values. Differences in variance between families (reflected by CVs and heritability values) among populations can result in very wide CIs for QST, even for a trait with reasonable sample size like age at maturity. CIs calculated by bootstrap over families appear narrower than Bayesian ones, which is in accordance with the results of O'Hara & Merilä (2005) who showed that the precision of CI for QST is highly variable depending on the method used, the bootstrap of families (according to dam or sire) leading to narrower CI than for instance, Bayesian CI. However, these authors also found a better coverage of Bayesian CI compared with other methods. For all traits, it is also worth considering that when expressed as evolvability (CV), the within-population variance is more homogeneous among populations than heritability. In the literature, values for within-population variances are rarely given (Waldmann et al., 2005); it is thus unclear how often these values could be heterogeneous.
A third methodological issue to consider is the fact that we started the experiment with G0 individuals supposed previously outcrossed in the field. There are two potential biases related to this approach. First, depending on whether these individuals mated with one or several partners, the progenies were then made of full-sibs or half-sibs. However, in both cases the conclusion of our study is similar: QST for survival and fecundity are not significantly different from zero and QST for age and size at maturity do not differ from FST. Secondly, it is also possible that some G0s may have selfed instead of outcrossed. It could be taken into account by using a non-null FIS value in the calculation of QST (Bonnin et al., 1996). With FIS values of 0.1 and 0.2, which represent selfing rate of 0.18 and 0.33, the conclusions of our study are not affected. Thus, even with a selfing rate as high as 30%, QST is never higher than FST for any trait. Also, as we detected a strong inbreeding depression for hatching rate, outcrossing is probably predominant over selfing in our study populations but a molecular analysis will be necessary to confirm this assumption. In addition, the nonlinear decrease of fitness as a function of inbreeding might indicate an influence of epistasis on the genetic architecture of hatching rate (Willis, 1993). However, as we do not know whether individuals from treatment T2 are full- or half-sibs, we cannot conclude on this point.
Another important issue is the potential influence of nongenetic maternal effects on the three traits that were measured on G1 individuals rather than on G2s. However, maternal effects are known to act more strongly on precocious traits like early growth and survival (Mousseau & Fox, 1998; Pakkasmaa et al., 2003), whereas all traits used here to infer QST were measured relatively late in the life cycle. Nonadditive genetic factors and common environment effects may also have influenced these traits as they were only discarded in the case of half-sib families. Common environment effects are supposed to be relatively low as relatives were kept isolated most of the time during the experiment. In contrast, dominance variance is known to be high in fitness-related traits and could have increased Vf and thus diminished QST. However, it is worth considering that even in an ideal situation where additive variance is perfectly estimated, theoretical studies have demonstrated that epistasis and dominance effects can lower QST (Whitlock, 1999; Lopez-Fanjul et al., 2003; Goudet & Büchi, 2006). This could potentially explain the very low QST we found for fecundity and longevity. In the case of fecundity, no conclusion can be drawn because of low sample size. For longevity, we found a clear inbreeding depression (δ = 0.27), which strongly suggests the action of dominance on the genetic architecture of this trait. However, in a recent simulation study, Goudet & Büchi (2006) found that dominance was likely to deflate QST relative to FST but the effect was only strong for high levels of structure (FST > 30%). As FST was 0.18 in the present study, dominance should not have greatly influenced the results. Nevertheless, it is quite clear that the QST–FST approach shows its limitations in situations where stabilizing selection is suspected but the action of dominance (and possibly other nonadditive factors) cannot be neglected (Toro & Caballero, 2005). In contrast, when QST > FST, the conclusion of the action of directional selection for different local optima is robust to the effect on nonadditive gene actions (Goudet & Büchi, 2006).
Absence of local adaptation
At this scale, we did not detect any evidence of local adaptation despite a high environmental heterogeneity between well-differentiated populations (FST = 0.18). However, demonstrating the occurrence of local adaptation is difficult because, as stated by Kawecki & Ebert (2004), this process can be ‘hindered by gene flow, confounded by genetic drift, opposed by natural selection due to temporal variability, and constrained by lack of genetic variation or by the genetic architecture of underlying traits’. Among these factors, gene flow and genetic drift are worth considering because they are likely to have a major influence on molecular and quantitative genetic structure in a floodplain context. For gene flow, this is illustrated by the striking differences among pairwise FST: 15% between BX1 and BX2, which are spatially the closest populations (500 m), while FST is <5% between PN and PL, which are 6700 m far from each other. All other pairwise comparisons are found between 15% and 25%. If one assumes that dispersal of freshwater snails is passive and occurs from upstream to downstream (Cellot & Bournaud, 1988; Cellot, 1996), these results suggest that the level of gene flow from PL to PN is higher than from PL to any other sites. This is consistent with the fact that PL is located upstream from the other sites and frequently connected to the river, while PN is rarely connected but strongly influenced by floods as reflected by the very low organic content from its sediments. In contrast, BX2 is frequently connected to the river but it is unlikely to receive migrants from the river (and thus from PL) because of its geographical setting (Citterio & Piégay, 2000). However, gene flow may occur from BX1 to BX2, BX1 being located upstream from BX2, but this must be very rare as they are rarely connected to each other (and there is no current between the two sites), and FST appears relatively high between these sites. Another feature of floodplain habitats is the temporary character of certain pools, which can be linked to high demographic fluctuations that increase the effects of genetic drift. For instance, BX2 has dried out for a total of 13 months between September 2000 and September 2003. The genetic diversity in this site is reduced compared with PN and PL, but it is similar to the one found in BX1 that rarely dried out during the same period. This suggests that other factors should be taken into account to explain the observed genetic diversity. For example, population size could be low in BX1, or few individuals may have founded this population, as BX1 is the most isolated of all sampled sites. The action of genetic drift and gene flow cannot be definitely disentangled in the present study but given the floodplain context, our results suggest that these forces strongly influence R. balthica populations by shaping their neutral genetic structure and potentially counteracting their local adaptation. Then, it would be interesting to investigate the temporal fluctuations in population size and genetic structure in order to measure the relative importance of genetic drift and gene flow in the evolution of these populations.