AFLPs support deep relationships among darters (Percidae: Etheostomatinae) consistent with morphological hypotheses

Abstract

Recent attention has focused on the efficacy of amplified fragment length polymorphisms (AFLPs) for resolving deep evolutionary relationships. Here we show that AFLPs provide resolution of deep relationships within the family Percidae that are more consistent with previous morphological hypotheses than are relationships proposed by previous molecular analyses. Despite in silico predictions, we were able to resolve relatively ancient divergences, estimated at >25 MA. We show that the most distantly related species share the fewest fragments, but suggest that large data sets and extensive taxon sampling are sufficient to overcome this obstacle of the AFLP technique for deep divergences. We compare genetic distances estimated from mitochondrial DNA with those from AFLPs and contrast traditional PAUP^* Nei–Li AFLP genetic distances with a recently proposed method utilizing the Dice equation with constraining nucleotides.

Introduction

Mitochondrial DNA (mtDNA) and nuclear introns are the most widely used genetic markers for phylogenetic analysis. However, discordance between topologies from nuclear and mitochondrial loci are common because of genes mutating at varying rates and being subjected to different stochastic processes (Maddison, 1997; Mendelson and Simons, 2006). Utilizing several loci and/or morphological characters have led to more robust topologies with greater statistical support, but, in many analyses, relationships utilizing multiple markers still remain unresolved or have low statistical support (Keck and Near, 2008; Ayache and Near, 2009). Examples where hybridization, introgression and saturation have confounded phylogenetic analysis are common in the literature (Mendelson and Simons, 2006; Keck and Near, 2010). Furthermore, for non-model organisms without a sequenced genome, primers for introns or other markers may not be available. Therefore, researchers may seek an alternative to using nucleotide sequences of a small number of loci for determining species relationships.

The amplified fragment length polymorphism (AFLP) technique (Vos et al., 1995) is becoming a more common method of phylogenetic analysis, as it samples a large number of loci throughout the genome, including mitochondrial and nuclear DNA. AFLPs use a total digest of genomic DNA by two restriction endonucleases, typically a rare cutter (six or more bases) and a common cutter (four bases) or two rare cutters. Theoretically, nucleotide polymorphisms that accrue over time between diverging species will yield different fragment lengths in the genomic digest. These different sized fragments are amplified by PCR, separated by gel electrophoresis and visualized by fluorescence. The pattern of fragments (bands on the gel) is used to determine patterns of speciation and to reconstruct evolutionary histories. Because AFLPs utilize the entire genome, they are able to detect rare and recent substitutions as well as retained alleles for hundreds or thousands of loci throughout the genome.

Although AFLPs underscore the importance of using a genome-wide set of genetic markers that includes both nuclear and mtDNA, their use is not without concerns. One main challenge of the AFLP technique in phylogenetics is that comigrating (same-size) fragments are assumed to be homologous, when in fact they may be homoplasious (Althoff et al., 2007; Garcia-Pereira et al., 2010). Homoplasy is not unique to AFLPs and presents a concern for all types of phylogenetic reconstruction. The likelihood of homoplasy because of saturation increases over time and can lead to hypotheses of close relationships in species that are distantly related. Using AFLPs may mitigate this problem; because different loci are subject to different stochastic processes and mutate at varying rates, utilizing the entire genome may capture phylogenetic signal that outweighs the noise of stochastic saturation.

Nonetheless, the probability of comigrating fragments being homologous decreases with species divergence, leading some to question the usefulness of AFLPs for deeper relationships (Althoff et al., 2007; Garcia-Pereira et al., 2010). A recent in silico study suggested that the utility of AFLPs breaks down at divergences >0.05 substitutions per position (from the most internal node to the tip of the tree), with all comigrating bands being non-homologous at a divergence (distance) of 0.4 substitutions per position (Garcia-Pereira et al., 2010). In another in silico study of genomic data from Drosophila, Althoff et al. (2007) reported that only 14% (2/14) of comigrating fragments from the D. simulans and D. yakuba genomes (estimated to have diverged 6 MA) were homologous. Graves (2009), however, questioned the protocol of this study, citing a robust AFLP phylogeny of pinnipeds (Dasmahapatra et al., 2009) in which authors eliminated bands that appeared to be non-homologous across all species (by only scoring those bands that amplified consistently with minimal size variation) and excluded smaller fragments. Using these methods, authors recovered moderate to strong phylogenetic signal at a divergence of ∼15 MA (Higdon et al., 2007; Dasmahapatra et al., 2009).

In this study, we investigate the use of AFLP analysis to resolve deeper divergences in distantly related species of darters (Teleostei, subfamily Etheostominae). Darters are members of the family Percidae usually placed in four genera: Ammocrypta, Crystallaria, Etheostoma and Percina. They are one of the most speciose groups of freshwater fishes in North America with over 200 species. Based on analyses of the mitochondrial and nuclear DNA sequences calibrated to the absolute age of centrarchid fossils, darters are currently thought to have diverged from a most recent common ancestor (MRCA) around 25 million years ago (95% highest posterior density: 20.5–33.2 MA) (Hollingsworth and Near, 2009). Previous research using mtDNA and nuclear introns has not been able to fully resolve the darter phylogeny, and different genetic markers have yielded phylogenetic and morphological discordance; for example, the phylogenetic position of Ammocrypta, and the relationship of Etheostoma zonale to snubnose darters (Wood and Mayden, 1997; Song et al., 1998; Sloss et al., 2004; Keck and Near, 2008; Mendelson and Wong, 2010). Many sympatric darter species have been found to share haplotypes because of hybridization and mitochondrial replacement, confounding phylogenetic analyses (Ray et al., 2008; Bossu and Near, 2009; Keck and Near, 2010).

Here we present an AFLP analysis of 69 darter species and show that most hypothesized relationships are consistent with previous morphological evidence. We propose that extensive taxon sampling and large numbers of AFLP fragments are sufficient to overcome the increased proportion of non-homologous comigrating bands in distantly related taxa. In addition, we propose an AFLP molecular clock based on previous studies that used sequence data calibrated to the absolute age of fossils of centrarchid fishes (Near and Benard, 2004; Near and Keck, 2005; Hollingsworth and Near, 2009). Finally, we show that the AFLP distance measures described by Mougel et al. (2002) may provide a more precise estimate of genetic distance than the more traditionally used Nei–Li equation.

Materials and methods

Taxonomic sampling and DNA extraction

Sixty-nine darter species representing the 4 genera and 25 subgenera recognized in the most recent classification (Page, 2000) were used in this study (Table 1). Specimen identity and locality for each species in this study is presented in Supplementary Table 1. Two Sander (S. canadense and S. vitreum) and two Perca (P. flavescens and P. fluviatilis) species were used as outgroups to root the phylogeny. Sander and Perca are members of the subfamily Percinae (Song et al., 1998). Genomic DNA from a single individual representing each species was extracted from tissue specimens using the QIAGEN DNeasy Tissue Kit (QIAGEN Inc., Valencia, CA, USA) and quantified using a NanoDrop ND-1000 spectrophotometer (Thermo Scientific, Wilmington, DE, USA).

Table 1 Classification of Etheostominae from Page (2000), with recent changes in names of species

AFLP protocol

The AFLP procedure was similar to that described by Vos et al. (1995). Genomic DNA (250 ng) was digested with EcoR1 and Pst1 restriction endonucleases. Double-stranded adaptors were ligated onto the sticky ends of the genomic digest using a T4 DNA ligase (New England Biolabs, Ipswich, MA, USA). The product was diluted 1:10 and then subjected to two rounds of increasingly selective PCR amplification. In the first round, 2 μl of the diluted genomic enzyme digest and 1 μl of each primer were combined with 0.1 μl Taq DNA polymerase and 5.9 μl mastermix consisting of 0.05 μl 100 mM dNTPs (Fermentas, Glen Burnie, MD, USA), 0.30 μl 50 mM MgCl₂ (Invitrogen, Carlsbad, CA, USA), 4.55 μl dH₂O and 1 μl 10 × PCR reaction buffer (Invitrogen). A single selective nucleotide (adenine) was added to the end of the primer that was complementary to the adaptors ligated to the genomic digest (EcoR1: 5′-GACTGCGTACCAATTC-A; Pst1: 5′-GACTGCGTACATGCAG-A). A non-hot start Taq DNA polymerase (Go-Taq, Promega Corp., Madison, WI, USA) was used in the first selective amplification. In the second amplification, 2 μl of a 1:40 dilution of the first amplification was used as a template with 1 μl of each primer, 5.9 μl mastermix and 0.1 μl hot-start Taq DNA polymerase (Invitrogen, Fermentas or Qiagen). Primers in the second amplification had two additional selective nucleotides (A+GG, AG, AA, GA or AT), and the EcoR1 primer was fluorescently labeled (FAM, Invitrogen) for automated fragment analysis. A total of 20 primer pairs were used: EcoR1+AGG with Pst1 AAA, AAG, AGG, AGA: EcoR1+AAG with Pst1 AAA, AAG, AGG, AGA: EcoR1+AAA with Pst1 AAA, AAG, AGG, AGA: EcoR1+AGA with Pst1 AAA, AAG, AGG, AGA: EcoR1+AAT with Pst1 AAA, AAG, AGG, AGA.

One μl of the second round of amplified genomic digest was added to 9 μl Hi–Di Formamide (Applied Biosystems, Carlsbad, CA, USA), and was sent to Yale's DNA Analysis Facility for automated fragment analysis using a GeneScan LIZ 500 size standard (Applied Biosystems) on an ABIPrism 3730 DNA analyzer (Applied Biosystems). Fragments were analyzed using GeneMapper v.3.7 (Perkin-Elmer Corp., Waltham, MA, USA).

Phylogenetic analysis

Individuals were scored for the presence or absence of same sized fragments, or bands, using 1 to indicate the presence and 0 to show the absence of a band. GeneMapper panels were created for each primer pair for fragments between 75.0 and 450.0 bp using bin widths 0.4 bp wide. The threshold intensity for scoring a peak as ‘present’ varied among primer pairs between 400 and 3000. One challenge of working with AFLPs is that different primer pairs vary in their overall intensity, including the intensity of background noise, such that a single threshold cannot be applied across all primer pairs. Thresholds were therefore chosen to maximize phylogenetic signal and to minimize background noise. The threshold yielding the greatest phylogenetic signal (average bootstrap statistical support across nodes) for each primer pair was identified and applied individually by re-estimating the tree in PAUP^* with the addition of each new primer pair. Automated peak scoring was modified manually by the removal of shoulders (stutter peaks) by a single individual who was unfamiliar with the taxon groups or labels to avoid systematic bias.

The resulting presence–absence matrix composed of 8056 characters was used to estimate phylogenetic relationships using distance and Bayesian methods. The matrix was translated into a distance matrix using the Dice distance equation with a Jukes–Cantor correction and 18 constraining nucleotides as described by Mougel et al. (2002) using DistAFLP (available from http://pbil.univ-lyon1.fr/ADE-4/microb). The Dice distance equation with constraining nucleotides is as follows:

where n_xy is the number of fragments shared between individuals, Δ_xy is the number of fragments found only in individual x or y and r is the number of constraining nucleotides. This equation assumes that shared alleles between two species were present in a common ancestor halfway between them. The Jukes–Cantor correction, $\widehat{t}=\text{-}(3/4)\text{ ln[1-(4/3)}{d}_{D_{xy}}\text{]}$, accounts for unobserved substitutions because it assumes equal rates of substitutions between all base pairs. The number of shared fragments between individuals for each pairwise comparison was determined manually in Microsoft Excel. We calculated the number of constraining nucleotides based on the number of nucleotides involved in restriction and selective PCR: r=E+P+N, where E is the number of nucleotides at the restriction site of EcoR1, P is the number of nucleotides at the restriction site of Pst1 and N is the number of selective nucleotides (r=6+6+6). These nucleotides constrain the results in that homologous fragments can be lost by a mutation at any one of the nucleotides involved in restriction or selective PCR; the greater the number of constraining nucleotides, the less likely a shared absence will be homologous. The distance matrix produced by DistAFLP with 1000 bootstrap replicates was analyzed with the Neighbor and Consense executables in PHYLIP Phylogeny Inference Package (Felsenstein, 1993) for neighbor-joining phylogenetic analysis.

The Nei–Li distance equation in PAUP^* (Swofford, 2000) was also used to calculate Nei–Li genetic distances that traditionally have been used in AFLP phylogenetic analyses. PAUP^* Nei–Li distances are based on a six base pair restriction site (r=6) using the following equation from Nei and Li (1979):

where Ŝ is an estimation of the proportion of ancestral restriction sites that remain unchanged in both individuals (Ŝ=2n_xy/[x_x+n_y]). Bootstrap analysis was performed in PAUP^* using 1000 replicates and a neighbor-joining algorithm.

Finally, we conducted a Bayesian phylogenetic analysis in MrBayes (Huelsenbeck and Ronquist, 2001; Ronquist and Huelsenbeck, 2003) using the restriction site binary model designed for restriction fragment data (for example, AFLPs) with the option ‘noabsencesites’ to compensate for bias introduced by the AFLP technique (characters that are absent in all taxa cannot be observed). Although Bayesian methods for AFLP markers have been developed, they are not computationally practical at this time. The analysis ran for 5 960 000 generations (convergence diagnostic <0.01). Burn-in was determined to be 25% of the generations in all analyses using the software Tracer (available at http://tree.bio.ed.ac.uk/software/tracer) to determine the stationarity of all parameters. This large burn-in value was consistent with the observation that very divergent sequences take longer to converge.

AFLP molecular clock

We inferred the age of three nodes in our phylogeny based on estimates in the literature calibrated to the absolute age of fossils of centrarchid fishes (Near and Benard, 2004; Near and Keck, 2005). Etheostoma tippecanoe is estimated to have diverged from other species of the subgenus Nothonotus ∼13.67±2.86 MA (Near and Keck, 2005), E. jordani is estimated to have diverged from other Nothonotus ∼11.85±2.35 MA (Near and Keck, 2005) and Percina caprodes and P. austroperca are estimated to have diverged ∼0.78±0.38 MA (Near and Benard, 2004). To determine the AFLP genetic distances of E. jordani and E. tippecanoe from other species of Nothonotus, we used the average genetic distance of the two species from each of the remaining members of the Nothonotus group identified by our phylogeny. The three node age estimates were plotted as a function of the AFLP Dice genetic distance with a Jukes–Cantor correction and 18 constraining nucleotides (‘DiceJC’) and against PAUP^* Nei–Li genetic distances in Excel to test for linearity (Etheostoma barbouri and E. smithi are estimated to have diverged around 16.3 MA (Hollingsworth and Near, 2009); however, this node age was excluded from the analysis because of the incongruence between mitochondrial and AFLP genetic distances).

The slope of the linear regression of AFLP genetic distance on the three node age estimates was used to determine the rate of AFLP divergence (as in Kropf et al., 2009). The AFLP mutation rate was calculated from the divergence rate using the equation D=2 μt’+4N_ancμ, where D is the divergence rate, μ is the mutation rate, N_anc is the ancestral population size and t’ is time expressed in generations after divergence (Kropf et al., 2009). If we assume the ancestral population size was very small for each speciation event, 4N_ancμ becomes negligible compared with the size of D, and the equation reduces to D=2 μt′. Divergence times for each species pair were then extrapolated using the equation of the linear regression of AFLP distance on node age.

AFLP and mitochondrial distance analysis

Cytochrome b (cyt b) sequences were obtained for 66 species from Genbank (http://www.ncbi.nlm.nih.gov/genbank), and jModelTest (Posada, 2008) was used to determine the parameters for the best fitting model of evolution. These parameters were implemented in PAUP^* to estimate maximum likelihood (ML) genetic distances based on cyt b. Microsoft XLSTAT was used to perform a Mantel test to determine the significance of the correlation between ML cyt b and AFLP genetic distances.

Performance of AFLP at deeper divergences

Relative node ages of all internal nodes were determined using the DiceJC tree in the software Figtree (University of Edinburgh, v1.3.1). To estimate the performance of AFLPs across a range of divergence, relative node ages were compared with PAUP^* bootstrap values and Bayesian posterior probabilities using Microsoft Excel. Microsoft Excel was also used to determine the correlation and statistical significance for each comparison.

Results

AFLP phylogenetic analysis

A total of 8056 characters were used in this analysis, of which 5942 (74%) were parsimony informative. The number of shared fragments between species ranged from a maximum of 235 to a minimum of 10 (average=42±0.360). A neighbor-joining cladogram based on DiceJC genetic distance with bootstrap support >50% for both DiceJC and PAUP^* Nei–Li genetic distances is shown in Figure 1. The Bayesian tree with posterior probability >0.5 is shown in Figure 2. Percina, Ammocrypta and Etheostoma are monophyletic: Percina 100% DiceJC bootstrap support (0.99 Bayesian posterior probability), Ammocrypta 100% (1.0), Etheostoma 52% or 64% (0.84) excluding Etheostoma cinereum. E. cinereum, the only member of the subgenus Allohistium, was recovered as sister to all other Etheostoma species with moderate bootstrap support (52%). However, Bayesian analysis (Figure 2) yielded a polytomy between Percina, E. cinereum and the remaining species of the genus Etheostoma.

Figure 1

Neighbor-joining tree depicting hypothesized relationships among darters (Etheostomatinae), based on 8056 AFLP characters. Perca and Sander used as outgroups. Bootstrap support >50% (1000 replicates) for DiceJC genetic distances (left) and PAUP^* Nei–Li genetic distances (right) are shown to the left of each node.

Figure 2

Bayesian Tree showing relationships among darters (Etheostomatinae). Posterior probabilities >0.50 shown and are based on 5 960 000 generations with a 25% burn-in.

One difference was observed between the topologies of the DiceJC and PAUP^* Nei–Li distance-based trees. The tree based on DiceJC distances recovered E. radiosum as sister to a group consisting of E. bison, E. lawrencei, E. caeruleum, E. swaini, E. luteovinctum and E. exile with strong bootstrap support (Figure 1). As in the Bayesian tree (Figure 2), the tree based on PAUP^* Nei–Li distances recovered E. radiosum as sister to E. bison and E. lawrencei, with moderate bootstrap support (50%).

AFLP molecular clock

We indirectly calibrated the AFLP molecular clock using three estimated node ages from mtDNA and nuclear intron analyses that were calibrated to the absolute age of centrarchid fossils (Near and Benard, 2004; Near and Keck, 2005). We found a strong positive linear correlation with respect to time for both DiceJC and PAUP^* Nei–Li distances (r=0.997; 0.998)(Figure 3). However, because of the limited calibrations to centrarchid fossils thus far, and because darter fossils are rare, we point out that the relationship between time and genetic distance may not be linear, particularly at node ages <1–2 million years. Ho et al. (2005) show that estimates of mutation rates are higher for recent divergences (<1–2 MA), which could lead to an overestimation of recent divergence times. Calibrations with additional fossils are needed to assess the utility of our molecular clock estimates.

Figure 3

The estimated age of nodes of the MRCA as a function of the AFLP pairwise genetic distances for three species-pairs of darters. Node ages and genetic distances were from Etheostoma tippecanoe from other Nothonotus darters (E. jordani, E. bellum, E. rufilineatum, E. aquali), E. jordani from other Nothonotus darters (E. bellum, E. rufilineatum, E. aquali) and Percina caprodes and P. austroperca (Near and Benard, 2004; Near and Keck, 2005). A full color version of this figure is available at the Heredity journal online.

We estimated an AFLP divergence rate of ∼0.0036±0.0024 nucleotide substitutions per site per million years using DiceJC genetic distances, or 0.0111±0.0068 using PAUP^* Nei–Li distances. The estimated AFLP mutation rate, calculated from the divergence rate and assuming darters have one generation per year (Page, 1983), was found to be between 6 × 10⁻¹⁰ and 3 × 10⁻⁹ or between 2 × 10⁻⁹ and 9 × 10⁻⁹ nucleotide substitutions per site per generation using the DiceJC and the Nei–Li divergence rates, respectively. This mutation rate is an approximation because of the uncertainty in generation time (for example, although rare, some species do not mature until 2 years of age, but then reproduce annually) and ancestral population size. A maximum DiceJC distance of 0.208 nucleotide substitutions per site was found in our study between P. fluviatilis (an outgroup) and P. caprodes, which would translate to ∼48.9 MA using our calibrated clock. The maximum PAUP^* Nei–Li distance for this species pair was 0.638, which also translates to a MRCA ∼48.9 MA. The maximum distance within the darter ingroup was 0.177/0.540 between E. hopkinsi and C. asprella, which translates to a node age of ∼40 MA (40.3 MA DiceJC and 40.1 MA PAUP^* Nei–Li). A strong positive correlation was found between DiceJC and PAUP^* Nei–Li node ages for each interspecific pairwise species comparison (r=0.9999, P<0001). We used our calibrated clock to estimate divergence times for all pairwise comparisons in our phylogeny (data not shown), and these node ages were compared with previously published divergence estimates based on mitochondrial and nuclear sequences.

AFLP and mitochondrial distance analysis

Supplementary Figure 1 shows the relationship between the number of shared AFLP bands and genetic distance, based on two methods of calculating genetic distance from restriction markers (DiceJC and PAUP^* Nei–Li). DiceJC and PAUP^* Nei–Li distances are strongly correlated with each other (r=1.00, P<0.001), and both increase logarithmically as the number of shared bands decrease; however, the means of the two genetic distances are significantly different (DiceJC: 0.131±0.0004, PAUP^* Nei–Li: 0.398±0.001; t=312, P<0001). PAUP^* Nei–Li distances also exhibit a broader range (0.082–0.638) and much more scatter about the regression line than the DiceJC estimates (0.027–0.208) (residual sum of squares: PAUP^* Nei–Li=1.77, DiceJC=0.183).

The pairwise comparison of DiceJC and PAUP^* Nei–Li genetic distances with ML cyt b genetic distances resulted in a strong positive correlation (DiceJC: r=0.718; PAUP^* Nei–Li: r=0.717; Mantel Test: P<0.0001) (Figure 4). Neither distance appears to be saturating.

Figure 4

Pairwise interspecific AFLP genetic distances as a function of ML cyt b genetic distances for 66 Percidae species. AFLP genetic distances were calculated with PAUP^* Nei–Li and the Dice equation with a Jukes–Cantor correction using 18 constraining nucleotides (Mougel et al., 2002). Correlation, PAUP^* Nei–Li r=0.717, DiceJC r=0.718, Mantel Test, P<0.0001. A full color version of this figure is available at the Heredity journal online.

Performance of AFLPs at deeper divergences

A statistically significant negative correlation was found between bootstrap statistical support and relative node age as estimated in FigTree (bootstrap: r=−0.526, P=3.76 × 10⁻⁵) (Figure 5), indicating that older nodes had overall less bootstrap support than more recent nodes. That correlation remained statistically significant after the removal of two extreme data points (r=−0.400, P=0.003). Posterior probabilities were not significantly correlated with relative node ages (r=−0.212, P=0.123), except when two extreme data points were removed (r=−0.374, P=0.006).

Figure 5

Relationship between relative node age (as estimated in Figtree) and statistical support >50% (0.50). Bootstrap r=−0.526, P<0.0001. Posterior probability r=−0.212, P=0.123. Large blue square indicates two data points at the same location. A full color version of this figure is available at the Heredity journal online.

Discussion

In this study we used AFLPs to infer relationships among mostly distantly related darter species representing four genera and 25 subgenera. These appear to be the deepest divergences analyzed using AFLPs to date (Graves, 2009; Dasmahapatra et al., 2009). We obtained a relatively robust tree with species of accepted genera and long-recognized subgenera clustering with statistical support. Relationships hypothesized from AFLPs agree with morphological hypotheses (for example, Bailey and Gosline, 1955; Page, 2000) to a greater extent than previous molecular analyses of nuclear and mitochondrial markers.

Phylogenetic relationships

The morphology-based genera first proposed in Bailey et al. (1954) and Bailey and Gosline (1955) have been almost universally accepted by researchers investigating evolutionary relationships among darters. One exception was Hubbs (1959), who continued to recognize Percina only for the morphologically distinctive logperches and Hadropterus for other species assigned by Bailey to Percina. In addition, several authors have recognized Crystallaria as a genus rather than as a subgenus of Ammocrypta (Simons, 1991; Page, 2000; Wood and Raley, 2000) and, most recently, Nothonotus as been recognized as a genus rather than as a subgenus of Etheostoma (Near and Keck, 2005). The AFLP data strongly support the monophyly of Percina with the logperches deeply embedded within the genus. Moreover, the 46 species of Percina share a morphological synapomorphy, the presence of scutes on the breasts of males (Bailey et al., 1954; Page, 1974).

The monophyly of Ammocrypta (with Crystallaria as a subgenus) or of Ammocrypta+Crystallaria has been nearly universally accepted since first recognized by Bailey et al. (1954), except by Simons (1992) who hypothesized Ammocrypta to be nested within Etheostoma. Our results strongly support the sister-group relationship of Ammocrypta and Crystallaria, and hypothesize this group to be sister to all other darters and not embedded within Etheostoma. The group containing the eight species of Ammocrypta and Crystallaria share a derived, exceptionally slender and translucent body. Recognition of Crystallaria as a genus is arbitrary but generally accepted because of the morphological distinctiveness of Crystallaria from species of Ammocrypta (Simons, 1991).

Although Etheostoma has never been diagnosed satisfactorily as a genus with morphological synapomorphies, it has been nearly universally accepted as monophyletic. The exception was Near and Keck (2005), who treated Nothonotus as a genus based on mtDNA sequence data in Song et al. (1998) and Sloss et al. (2004). However, Song et al. (1998) concluded that its phylogenetic position ‘cannot be resolved with the cyt b dataset, as evidenced by the lack of bootstrap and decay support.’ Sloss et al. (2004) found the position of Nothonotus to vary greatly depending on the type of analysis. Although these studies suggest that Nothonotus may be as divergent from other darters as the recognized genera are from one another, they neither identify the closest relative to Nothonotus nor indicate that Etheostoma with Nothonotus included creates a paraphyletic genus.

Our data show Etheostoma including Nothonotus to form a monophyletic group. The species of Etheostoma most distantly related to all other Etheostoma in our analysis is E. cinereum, a species that previously has been suggested to be the most divergent in the genus (Page and Whitt, 1973). Sloss et al. (2004) found E. cinereum to be the sister to the remaining Etheostoma, embedded in Etheostoma, or sister to Ammocrypta+Crystallaria, depending on the analysis. Our results indicate that recognizing Nothonotus as a genus and not Allohistium for E. cinereum creates a paraphyletic Etheostoma.

In addition to supporting the monophyly of long-established genera, AFLPs hypothesize a sister relationship between Percina and Etheostoma. Although this has been suggested previously in some analyses (for example, in a ML tree of mtDNA data by Sloss et al., 2004), most workers have treated Percina as sister to the remaining genera, leaving Ammocrypta as the sister to Etheostoma (Kuehne and Barbour, 1983; Page, 1983). Other interesting relationships, some previously not hypothesized, among species and subgenera are suggested by the AFLP data, but await discussion pending the analysis of data on more species.

AFLP molecular clock

Based on our AFLP molecular clock, we estimate the MRCA of the subfamily Etheostominae (‘darters’) to have existed ∼40 MA. Our estimated age of the MRCA of Percina and the Etheostoma subgenus, Catonotus, of ∼28 MA are similar to Hollingsworth and Near's (2009) estimated age of the MRCA of ∼25 MA (95% highest posterior density: 20.5–33.2 MA). Likewise, Hollingsworth and Near (2009) estimated the barcheek darter group to have diverged from the fantail darter group (subgenus Catonotus) ∼17 MA (95% highest posterior density: ∼13–22). Our clock estimates the age of this node to be ∼16 MA.

A notable exception to previous estimates is that E. barbouri and E. smithi are much more closely related (age of node ∼7.0 MA) in our analysis than Hollingsworth and Near's (2009) estimate based on a multigene phylogeny using mitochondrial and nuclear loci (16.3 MA, 95% highest posterior density: 12.4–20.5). E. barbouri and E. smithi have a smaller genetic distance based on AFLP (nuclear) DNA than that based on mitochondrial cyt b nucleotide sequence (Figure 4). We interpret this cytonuclear discordance as evidence that one or the other of this species pair experienced mitochondrial replacement at some time in the past, most likely E. barbouri. Previous phylogenetic analyses showed that two mtDNA loci (cyt b, ND2) place E. barbouri outside the barcheek darter group whereas, AFLP, S7 nuclear intron and morphological data all recover E. barbouri as a barcheek darter (Page et al., 2003; Mendelson and Simons, 2006). Interpretations of node age for E. barbouri and E. smithi are likely to be more reliable if based on nuclear rather than mtDNA distances.

AFLP and mitochondrial distance analysis

Restriction-site distance methods can be used with AFLP fragments because AFLP bands show a similar statistical behavior to restriction site data (Mougel et al., 2002). Genetic distances estimated by two restriction-site distance metrics, DiceJC and PAUP^* Nei–Li, are strongly correlated (r=1.00); however, PAUP^* Nei–Li distances yield a greater amount of scatter around the regression line of distance on shared bands as compared with DiceJC distances (Supplementary Figure 1). This scatter suggests that PAUP^* generated Nei–Li distances may not as accurately reflect genetic relationships, and perhaps suggests that, as shown by Mougel et al. (2002), constraining nucleotides are important in calculations of genomic divergence. Importantly, however, neither the topology (with the exception of one species, E. radiosum) nor the statistical support of nodes in the neighbor-joining tree based on PAUP^* Nei–Li distances were noticeably different from the DiceJC tree.

Several studies have suggested that the homology of comigrating bands should decrease with increasing genetic divergence. Garcia-Pereira et al. (2010) estimate that the utility of AFLPs breaks down at divergences >0.05 substitutions per position using PAUP^* Nei–Li distances, with all comigrating bands being non-homologous at a divergence of 0.4 substitutions per position. Althoff et al. (2007) also predict that the number of comigrating bands containing homologous fragments decreases with increasing genetic distance. If this were the case in our study, we would expect to see a greater amount of scatter about the regression line of AFLP distances on ML cyt b distances at larger distances. As cyt b distances are derived from truly homologous loci, a greater amount of scatter at larger distances would indicate a greater proportion of homoplasious AFLP bands. As in Dasmahapatra et al., 2009, we show a strong positive relationship between PAUP^* Nei–Li distances and ML cyt b distances between 0.082 and 0.638 substitutions per position, with a seemingly negligible increase in scatter about the regression line at larger distances, especially for DiceJC distances (Figure 4). This pattern suggests the phylogenetic signal-to-noise ratio is approximately constant through this level of divergence. Furthermore, because we are recovering moderate to strong phylogenetic signal at these deeper divergences, we suggest that a sufficient number of comigrating bands are indeed homologous at this level of divergence, and that perhaps large datasets and increased taxon sampling are sufficient to overcome noise created by homoplasious bands.

Performance of AFLP at deeper divergences

Our analysis of statistical support across increasing relative node ages indicates a significant effect of node age on bootstrap support but not on Bayesian posterior probability, unless two extreme data points are removed. This pattern suggests that statistical support for the deepest divergences relies on a few highly informative characters. Because the bootstrap method samples with replacement, a proportion of bootstrap replicates may fail to sample the few important characters that define older nodes. However, because both distance and Bayesian analyses yield moderate to strong phylogenetic signal for deeper relationships, and because posterior probabilities, which sample every character, do not show a significant correlation between node age and statistical support (Figure 5), we suggest that AFLPs are suitable for evaluating relationships among deeply divergent taxa.

We further suggest that extensive taxon sampling and large datasets (that is, a large number of AFLP fragments) may be necessary to overcome noise created by homoplasious fragments. As divergence time increases, species share fewer alleles (fragments), and the shared derived characters that once defined internal relationships are reduced in frequency. Sampling only a small fraction of the extant species or using a small number of characters could fail to detect these rare alleles shared between distantly related taxa. The greater the fraction of extant species (or individuals) sampled for a given taxonomic group, and the larger the AFLP dataset, the more likely two taxa would share common alleles (fragments) to link deeper relationships. Capturing these ‘intermediate’ species should therefore allow us to better detect internal relationships and reconstruct evolutionary histories. Most importantly, however, our study indicates that AFLPs can be used to resolve deeper divergences and determine species relationships at multiple levels of divergence.