Introduction

Studying spatial partitioning of genetic variation is viewed as a robust method for inferring gene flow among discrete populations, conditional on taking into account all processes shaping the present genetic structure (Hewitt, 2000). For coral reef fishes, contrasting results have been found relative to the relationship between present genetic structure and larval duration (Waples, 1987; Doherty et al., 1995; Shulman and Bermingham, 1995; Rocha et al., 2002; Bay et al., 2006), reproductive strategy (Shulman and Bermingham, 1995; Planes et al., 1997), species swimming abilities (Waples, 1987; Riginos and Victor, 2001; Purcell et al., 2006) or geographic distances (Planes et al., 1993, 1996; Planes and Fauvelot, 2002; Purcell et al., 2006). Fauvelot and Planes (2002) have shown, using allozymes, that the observed significant structuring of genetic variation among populations of eight coral reef fishes in a fragmented island system was not correlated with geographic distances at the studied spatial scale (ca 1500 km) nor associated with species reproductive strategy and larval duration. On average, only a third of the polymorphic loci contributed to the overall genetic differentiation. Conversely, the analysis of the genetic variation at the mitochondrial DNA (mtDNA) control region sequences in a subset of individuals revealed genetically homogeneous populations in all seven species under study (Fauvelot et al., 2003).

Such puzzling discordant patterns of genetic variation obtained at different loci are increasingly described (Karl and Avise, 1992; Pogson et al., 1995; Lemaire et al., 2000; De Innocentiis et al., 2001; Dufresne et al., 2002; Arnaud-Haond et al., 2003; Gomez-Uchida et al., 2003; Nielsen et al., 2006) and lead to question the interpretation made of genetic differentiation (FST) estimates in terms of effective dispersal. A key issue is that, under neutrality, all loci should theoretically be affected similarly by drift and migration (Lewontin and Krakauer, 1973; Slatkin, 1987), so that FST estimates should be homogeneous among loci, and multilocus FST estimates may be used to infer gene flow among populations. Yet, erroneous gene flow estimates may result from the interpretation of allelic distribution over multiple loci for two main reasons. One lies in the underlying high variance of the migration/drift process itself and its sensitivity to departure from equilibrium (Whitlock and McCauley, 1999). The other resides in selection or other non-neutral mechanisms (acting directly or through hitch-hiking) differentially affecting loci (Cavalli-Sforza, 1966; Slatkin, 1987). Indeed, disruptive selection will result in higher FST estimates than expected under neutrality, inflating the overall genetic differentiation. On the other hand, loci experiencing balancing selection will have allele frequencies more similar than expected under neutrality, reducing the FST estimates and further giving the false view of high gene flow.

In order to properly infer gene flow among populations, it appears therefore essential to (1) estimate genetic differentiation based on reasonable numbers of individuals (50–100 individuals; Ruzzante, 1998), loci (15–20 loci; Nielsen et al., 2006) and alleles (either few loci with many alleles or many loci with a few alleles; Kalinowski, 2002) and (2) discard loci behaving non-neutrally before the estimation of genetic differentiation. Several neutrality tests are available in the current literature that may allow detection of selected loci at ecological timescales (reviews in Guinand et al., 2004; Beaumont, 2005). Among these tests, those derived from the Lewontin–Krakauer method (that is, based on multilocus comparisons and the identification of FST outliers) appear robust when allele distributions are available at multiple loci (Beaumont and Nichols, 1996; Beaumont and Balding, 2004; Beaumont, 2005).

The goal of the present study is, for five species of coral reef fish, to infer gene flow among insular populations sampled within a 60–600 km spatial range and to test for the influence of reproductive strategies. Considering the problem of neutrality among loci, the theoretical neutral variance of single locus FST estimates in our insular system was first investigated in order to identify (and exclude from calculation) outlier loci before computing gene flow. For this purpose, the number of loci was increased to improve the power of the test. We reanalysed all previously assayed individuals from five species with a new set of loci with potentially higher mutation rates and gene diversities than allozymes, that is, intron length polymorphisms. Unlike exons, the nucleotide sequences of introns are little constrained so that introns accumulate mutations rapidly. Therefore, introns often display a high genetic variability (Palumbi, 1995), which make introns being targeted as markers of population variation and subdivision (Lessa, 1992; Borsa, 2003; Hassan et al., 2003; Friesen et al., 2005; Rohfritsch and Borsa, 2005; Berrebi et al., 2006). Exon-primed, intron-crossing polymerase chain reaction (EPIC-PCR) amplification is an appreciable strategy for finding DNA polymorphism in several species of different families (Hassan et al., 2002) and an alternative to the development of microsatellite loci. Length polymorphism at 11 intron loci was assessed in five coral reef fish species and the corresponding new data set was further added to the previous allozymes data set (Fauvelot and Planes, 2002) to conduct global analyses. The comparison of the distribution of genetic variation obtained from different markers using the same set of individuals from related species collected in the same area also provide a solid framework to test among mutation, migration and genetic drift as the main forces driving the distribution of genetic polymorphisms.

Materials and methods

Sampling

Sampling was carried out as described in Fauvelot and Planes (2002). Briefly, five different islands were sampled across French Polynesia based on a standard sampling effort of 50 individuals per island per species. Because the eight species selected for the original study were not all present in the five islands, and in order to compare effective dispersal among species at similar scales, we restricted further analysis to five species inhabiting four common islands: Rangiroa, Takapoto, Tetiaroa and Moorea (Figure 1). Geographic distances between islands ranged from 60 km (between Moorea and Tetiaroa) to 566 km (between Moorea and Takapoto). Chaetodon citrinellus is a butterflyfish strictly inhabiting lagoons while C. quadrimaculatus and Forcipiger flavissimus are both outer reef butterflyfishes. Dascyllus aruanus and Pomacentrus pavo are two damselfishes strictly inhabiting lagoons. Fishes were collected, tissue stored and DNA extracted as described in Fauvelot et al. (2003). A total of 952 individuals from five species (Table 1) were used for further genetic screening.

Figure 1
figure 1

Location of sampling sites of the five fish species in French Polynesia (South Pacific Ocean). Details of sites are reported in Fauvelot and Planes (2002) and sampling effort is indicated in Table 1.

Full size image
Table 1 Number of individuals scored at enzymatic and introns loci, indicated per species and per site
Full size table

Genetic analysis – EPIC-PCR

Eleven introns loci were tested for length polymorphism for each species: intron 7 of creatine kinase; intron 2 of glyceraldehyde-3-phosphate dehydrogenase; introns 2 and 5 of growth hormone; introns 1, 2, 4 and 5 of aldolase B; introns 1, 2 and 3 of gonadotropin-releasing hormone (Hassan et al., 2002). PCR amplification and electrophoresis follow protocols described in Hassan et al. (2002). Briefly, amplified PCRs (with forward primer labelled radioactively with 33P) were loaded on a 6% polyacrylamide gel in TBE buffer. The gels were run for 3 h at 50 W. After drying, the gels were exposed against films of Biomax autoradiograph. Individual genotypes were directly scored from the autoradiograph films with alleles being identified according to their size relative to the most common allele, as for allozymes. To ensure scoring accuracy and consistency across gels and populations, a locus-specific size standard from a pool of individuals with known genotypes was run at both sides of each gel. With this method, we were able to segregate 1 bp length difference bands. Individuals for which no PCR products were visualized on gel were further re-amplified in a new PCR using a 2°C lower annealing temperature. Results of the amplification successes across species are presented in Hassan et al. (2002) and detailed PCR conditions for each species and locus are available from the authors.

Data analysis

The allozymes data set was reanalysed for genetic divergences in the overall populations following the removal of populations from the data set (see sampling). Allele frequencies, number of alleles per locus (NA), observed heterozygosity (HO) and unbiased gene diversity (HE; Nei, 1987) were calculated within each population and over all populations for both allozymes and introns. Linkage disequilibrium between loci was tested using the procedure of Black and Krafsur (1985). Deviation from Hardy–Weinberg expectations (HWE) at introns loci was measured using Weir and Cockerham (1984) inbreeding coefficient f (equivalent to Wright's FIS), and locus conformance to HWE was assessed by permuting alleles within populations (1000 permutations). The presence of null alleles was tested using the program MICRO-CHECKER (Van Oosterhout et al., 2004). Genetic divergences were estimated using the Weir and Cockerham (1984) estimator. For each species, overall single locus FST values were estimated by jackknifing over populations and overall global FST estimates were obtained by jackknifing over loci. Unless specified, all computations were conducted using GENETIX (Belkhir et al., 2002).

The theoretical distribution of single locus FST estimates expected under neutrality was computed through simulations and outlier loci identified using Beaumont and Nichols’ (1996) test. The principle of this test is to estimate, conditional on gene diversity, the range of FST values expected under the combined effect of migration and genetic drift in a model approaching the studied system, and to compare the observed FST at this given locus to the simulated empirical distribution. The models for computer simulations were chosen as to fit at maximum the conditions of the studied system. Indeed, we chose the infinite allele model of evolution (Kimura and Crow, 1964), and the infinite island model of Wright (1943) as this is the best-suited available model for fragmented insular environments such as Pacific reefs. It should be noted that, for heterozygosities lower than 0.8 (which is our case), stepwise and infinite-allele models lead to similar simulated distributions of FST (Beaumont and Nichols, 1996), so that the choice of the mutation model has little impact on the simulated empirical distribution.

Coalescent simulations were performed to generate data sets with a distribution of FST close to the empirical distribution. Simulations were conducted assuming four sampled populations, with a metapopulation size of 100 populations (the approximate number of islands in French Polynesia), sample sizes of 50 individuals and an expected FST equal to the median estimate of FST calculated for all loci. Under the infinite allele mutation model, 20 000 HE and FST paired values were generated and the P-values for each locus were then calculated using the procedure described in Beaumont and Nichols (1996). Loci with unusually low or high FST estimates (that is, out of the 95% distribution of simulated FST), conditional on heterozygosity, were regarded as outlier loci. We used the currently distributed version of the test, namely Fdist2 (http://www.rubic.reading.ac.uk/~mab/software/fdist2.zip), described in Beaumont and Balding (2004). Because the test conducts simulations based on initial expected FST value set up by the user, this value was first fixed equal to the global FST (estimated over all loci). However, because this median value may be biased owing to the presence of an outlier locus, we systematically conducted again the simulations with a new global FST estimated omitting the outlier locus (or loci), and so on until no more loci were identified as outliers.

Overall populations and pairwise FST were estimated for each species over all nuclear loci (allozymes and introns) once removing outliers detected by the Beaumont and Nichols’ (1996) test. Significance of pairwise genetic divergence was tested using a permutation procedure (1000 permutations) in GENETIX (Belkhir et al., 2002). Mantel tests were performed to test for a correlation between geographical distance and multilocus FST (approach of Hutchison and Templeton, 1999). For each species, Mantel coefficient Z was computed and tested for significance through a permutation procedure with 100 permutations, under the null hypothesis that both matrices are not linearly related. To evaluate the level of correlation between genetic patterns exhibited by the different species, matrices of pairwise FST were compared with a Mantel test: Pearson correlation coefficient r was computed for each pair of species and its significance tested using a permutation procedure with 100 permutations. All tests were performed in GENETIX using the Mantel option.

Results

Out of the 11 intron primer pairs tested, five loci (AldoB1, AldoB5, CK7, GnRH3-3 and Gpd2) could not be further used for the five species because they lead to multibanded (several loci amplified), monomorphic or not clearly interpretable patterns (Table 2). The remaining six loci (AldoB2, AldoB4, GnRH3-1, GnRH3-2, GH2 and GH2) provided single locus amplifications with clearly interpretable length polymorphism, and finally, 2–3 polymorphic intron per species could be analysed (Table 2). Over all populations, the number of alleles per locus ranged from 2 to 12 (for GH2 over all F. flavissimus-sampled populations). Gene diversity among populations in each species was more or less even and variations in gene diversity were locus specific, such as that observed at enzymatic loci (Fauvelot and Planes, 2002), with no congruent patterns among species. Over all four populations, gene diversity at intron loci reached 0.726 for F. flavissimus at GH2, and 0.501 at enzymatic loci for F. flavissimus at Est1 (Table 3).

Table 2 Amplifications results of the 11 tested
Full size table
Table 3 Gene diversity (HE) over all four sampled populations and FST estimates for each scored locus (allozymes and introns) per species
Full size table

After standard Bonferroni correction of the 5% significance level, a single locus significantly deviated from HWE; GnRH3-1 in C. citrinellus populations. The cause of such departures is likely to be the presence of null alleles because the deviations resulted from significant heterozygote deficiencies, observed in all populations in this locus only. The occurrence of null alleles was further confirmed using MICRO-CHECKER and the average null allele frequency at GnRH3-1 locus across populations was 0.165 (estimator of Brookfield, 1996). However, this locus was maintained for further analysis, and genotypes were not adjusted according to nulls to allow for multilocus analysis.

Significant (P<0.01) linkage disequilibria were found between introns originating from the same gene (that is, GnRH3-1 and GnRH3-2 for C. citrinellus and C. quadrimaculatus; between AldoB2 and AldoB4 for D. aruanus). For C. citrinellus, the association was significant between the two most common alleles at each locus, and between rare alleles at GnRH3-1 with two of the four alleles at GnRH3-2. For C. quadrimaculatus and D. aruanus, the linkage involved rare alleles from one locus significantly associated with common alleles from the other locus. However, all loci were maintained for further analysis.

For each species, over all four populations, a wide range of single locus FST was observed among nuclear loci (allozymes and introns), from negative values up to 0.198 at ADA-1 for D. aruanus. Multilocus FST estimates ranged from 0.008 for C. citrinellus to 0.026 for D. aruanus (Table 3). Discrepancies among classes of markers were observed because only at enzymatic loci, global FST estimates ranged from 0.012 for C. citrinellus to 0.031 for F. flavissimus and D. aruanus, while when looking at only introns, global FST estimates were zero for all species except one, F. flavissimus (FST=0.008, P>0.05). However, these discrepancies were not marker related but locus related because single locus FST estimates at introns were similar to those of most of allozymes within species (Table 3).

Beaumont and Nichols’ (1996) test revealed that three loci demonstrated significantly higher FST values than expected under the null hypothesis of neutrality: ADA-1 in D. aruanus (P=0.991), and ADA and GDA in F. flavissimus (P>0.999 and P=0.953, respectively) (Figure 2). One additional locus, AldoB4 in P. pavo, showed significantly lower FST values (P=0.037) than expected under neutrality. Recalculating the global FST for each species without the outliers loci did not induce further rejection of neutrality for any other locus.

Figure 2
figure 2

Summary of the results of the Beaumont and Nichols’ (1996) neutrality test. For each locus, estimate of FST is plotted against its corresponding transformed P-value (expressed as logit(2∣P–0.5∣) as in Beaumont and Balding, 2004), corresponding to the rejection of the null hypothesis of neutrality. The vertical dotted line bar shows the critical transformed P-values used for identifying outlier loci at the 5% confidence level. Each outlier locus is denoted as follows: the two first letters refer to the species (Da, Dascyllus aruanus; Ff, Forcipiger flavissimus; Pp, Pomacentrus pavo), followed by the locus name.

Full size image

Excluding outlier loci detected by Beaumont and Nichols’ (1996) test, global FST were significant for all five species and ranged from 0.008 (P=0.01) for C. citrinellus to 0.019 (P=0.02) for D. aruanus (Table 4). Pairwise FST estimated from all nuclear loci excluding outliers ranged from 0 to 0.045 (between Rangiroa and Moorea for P. pavo; Table 4). None of the five species expressed significant correlation between genetic divergences and geographic distances between populations (all P>0.05). The comparison of pairwise FST matrices among the five species revealed that two species showed a significant correlation of their patterns of genetic structuring throughout the four sampled islands: C. quadrimaculatus and D. aruanus (r=0.843, P=0.03; see Table 4).

Table 4 Pairwise genetic differentiation (FST) estimated among islands excluding outlier loci
Full size table

Discussion

Discrepancies in population differentiation estimates among genetic markers were found in all five species. While intron length polymorphisms (in the present analysis) and sequences of the mtDNA control region (in Fauvelot et al., 2003) revealed homogeneous gene frequency over the study area, allozymes showed, over all loci, two- to fivefold higher values of global differentiation estimates. When looking at FST estimates at nuclear loci, we found high variance among loci within each species, with specific allozymes showing much higher FST values than introns and other enzymatic loci.

Neutral single locus FST variance and assumptions of neutrality tests

Identifying the theoretical variance of FST estimates is a challenging task. Under neutrality, a large variance of FST estimates is observed among loci because of the stochasticity of genetic drift (Lewontin and Krakauer, 1973; Beaumont and Nichols, 1996; Baer, 1999) but other factors may also inflate such variance. While different distributions of FST may be expected from genetic markers that differ in the mode of inheritance (Buonaccorsi et al., 2001), FST is theoretically insensitive to variation in mutation rate as long as mutation remains low compared to migration (Slatkin, 1995; but see Hedrick 1999). Meanwhile, the distribution of FST as a function of heterozygosity appears to be not strongly affected by the mutation rates typical of many nuclear genetic markers at mutation/drift equilibrium (Beaumont and Nichols, 1996). Discrepancies among loci may also originate from spurious typing errors and/or the presence of null alleles. We systematically reran several individuals from different gels side by side on new gels to minimize scoring errors. However, technical artefacts during allozymes migration and/or revelation cannot be totally discarded in generating scoring errors, and the occurrence of null alleles on specific gels may also increase variance in allele frequencies among gels/populations and inflate FST at particular loci.

Regarding the interpretation of Beaumont and Nichols’ (1996) test, the problem of identifying statistically significant departures from neutrality is complicated by confounding effects on the simulated distribution itself. Indeed, the 95–99% confidence interval of the simulated distribution relies on the assumptions of the chosen population structure model, which may be violated. Even though in large populations (such as the ones we are dealing with) migration model has little effect on the distribution of expected FST (Beaumont and Nichols, 1996), one could ask if migration/drift equilibrium is attained, as assumed in neutrality models. In natural populations, equilibrium is reached after roughly 1/m generations (with m being the migration rate; Slatkin, 1985a, 1985b) so that the higher the migration rate, the faster the equilibrium reached (Hardy and Vekemans, 1999). For our coral reef fishes, considering the high observed migration rates and a 2–3 years generation time, we can regard these populations as being at migration/drift equilibrium.

Nevertheless, two important caveats remain using this approach to detect outliers. The first one is that balancing selection is undetectable when the 95–99% confidence intervals of the expected distribution overlap zero (Beaumont and Balding, 2004). The second one is that simulations under neutrality are based on a specified global genetic differentiation supposed to be the ‘true’ FST expected overall loci (usually set equal to the observed genetic divergence over all loci for each species). However, this estimate may not be accurate if (1) a limited number of analysed loci is used and (2) outlier loci are present in the data set. One way to minimize the latter is to remove outliers and redo the test until no more outliers are detected. However, as marine organisms often exhibit high gene flow between populations, it may not be possible to detect loci under balancing selection because the confidence intervals of the expected distribution overlap zero (see Hoffman et al., 2006).

Finally, for all the reasons mentioned, outcomes of Beaumont and Nichols’ (1996) test regarding outlier loci being under selection should be interpreted with caution. In our study, the two allozyme loci identified as outliers in F. flavissimus and D. aruanus are both deaminases (ADA, adenosine deaminase; GDA, guanine deaminase). These loci may be under disruptive selection, but we cannot reject the hypothesis that these high FST estimates may be caused by mis-scoring of individuals. Contrary to ADA and GDA, AldoB4 in P. pavo exhibits an FST estimate significantly smaller to what was expected under neutrality. However, the observed FST estimate at this locus is negative (FST=−0.015), and because FDIST estimates corresponding P-values as the probability of getting values as small as or smaller than the simulated data points, it is clearly impossible to reveal balancing selection when the lower confidence limit for FST falls below zero. Finally, although it is difficult to conclude that these particular loci are driven by selection, all outlier loci were further excluded before inferring gene flow.

Gene flow in coral reef fishes

Including introns and removing outliers, global FST estimates are not significantly different from those based only on allozymes presented in Fauvelot and Planes (2002) (paired t-test, P=0.49). Geographic/genetic distance correlations for the five species remained insignificant, and they were still no significant correlative patterns of genetic structuring among species sharing similar reproductive strategy. Indeed, the only two species that show correlative pairwise FST matrix among islands differed in their reproductive strategies, C. quadrimaculatus releasing pelagic eggs and D. aruanus laying benthic eggs. However, it should be noted that these results may suffer from a low statistical power owing to the small number (four) of islands/populations compared, critical for the Mantel test.

While being significant, observed levels of genetic structuring for the five species are low (range 0.01–0.02), especially when compared to coral reef fishes lacking pelagic larval phase such as Acanthochromis polyacanthus and Pterapogon kauderni (Doherty et al., 1994; Planes et al., 2001; Hoffman et al., 2005; Bay et al., 2006). However, they are in the range described for related species with similar life histories and over comparable spatial scales (less than 1000 km) in the Pacific (Doherty et al., 1995; Riginos and Victor, 2001; Planes, 2002; Bay et al., 2006). Such rather high and chaotic levels of gene flow among islands accommodate Wright's island model in which each individual issued from one population is equally likely to move to any other population (Wright, 1931). This model may explain the lack of genetic structuring observed at mtDNA sequences. Indeed, such subtle genetic divergences may not be detected in high dispersive capability species using such molecular markers assessed on only 10 individuals per population (Fauvelot et al., 2003).

An interesting result is the significant pairwise divergences revealed for all pairwise comparisons for D. aruanus (see Table 4), in contrast to what was reported previously for this species over similar spatial scales in 1990. Indeed, Planes et al. (1993) described genetically homogeneous populations across Polynesia (minor genetic differences were detected only when comparing remote archipelagos). However, Planes et al. (1993) used a set of 12 polymorphic allozymes from which 9 were in common with our survey (we used seven additional polymorphic loci). Retaining the nine loci common to both survey (that is, AAT-2, Est1, GPD-1, GPD-2, GPI-1, GPI-2, IDH-1, IDH-2 and MDH-2), the global FST estimate over the four sampled islands decreased from 0.019 (P=0.002) to 0.009 (P=0.017), much more comparable to what was reported among archipelagos (FST=0.0076; Planes et al., 1993). This shows that global estimates of gene flow among populations may vary depending of the sets of loci used (in this particular case, by a factor of 2), and that, in order to avoid a bias in global FST estimates, a reasonable number of analysed loci should be considered (15–20; Nielsen et al., 2006; but see also Kalinowski, 2002 regarding the number of alleles).

In conclusion, our study shows that neutrality was rejected at few loci within species, although we could not conclude these loci being under selection. Our study demonstrates how cautious one has to be when trying to interpret present-day genetic structure in terms of gene flow, especially when using a limited number of loci and/or different sets of loci. Indeed, selection may influence particular loci, but as well, the expected neutral range of single locus FST is wide for loci with low heterozygosities, which can lead to erroneous estimates of genetic differentiation when a limited number of loci are used.