SEARCH

SEARCH BY CITATION

Keywords:

  • amplified fragment length polymorphism;
  • branch support;
  • phylogenetic accuracy;
  • phylogenetics;
  • simulation

Abstract

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgments
  8. References
  9. Supporting Information

We examined the effect of increasing the number of sampled amplified fragment length polymorphism (AFLP) bands to reconstruct an accurate and well-supported AFLP-based phylogeny. In silico AFLP was performed using simulated DNA sequences evolving along balanced and unbalanced model trees with recent, uniform and ancient radiations and average branch lengths (from the most internal node to the tip) ranging from 0.02 to 0.05 substitutions per site. Trees were estimated by minimum evolution (ME) and maximum parsimony (MP) methods from both DNA sequences and virtual AFLP fingerprints. The comparison of the true tree with the estimated AFLP trees suggests that moderate numbers of AFLP bands are necessary to recover the correct topology with high bootstrap support values (i.e. >70%). Fewer numbers of bands are necessary for shorter tree lengths and for balanced than for unbalanced tree topologies. However, branch length estimation was rather unreliable and did not improve substantially after a certain number of bands were sampled. These results hold for different levels of genome coverage and number of taxa analysed. In silico AFLP using bacterial genomic DNA sequences recovered a well-supported tree topology that mirrored an empirical phylogeny based on a set of 31 orthologous gene sequences when as few as 263 AFLP bands were scored. These results suggest that AFLPs may be an efficient alternative to traditional DNA sequencing for accurate topology reconstruction of shallow trees when not very short ancestral branches exist.


Introduction

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgments
  8. References
  9. Supporting Information

Phylogenetic reconstruction requires choosing character sets that are appropriate for the problem under consideration based on their availability, start-up time, cost and expected efficacy (Grechko, 2002). In the last decade, the amplified fragment length polymorphism technique (AFLP) (Vos et al., 1995) has been extensively used as an efficient, versatile and relatively inexpensive source of genome-wide informative molecular characters for phylogenetic inference, particularly for nonmodel taxa for which no prior DNA sequence is available (Koopman, 2005; Meudt & Clarke, 2007). AFLP consists of a double restriction endonuclease digestion of total genomic DNA followed by the ligation of the fragment ends to oligonucleotide adapters. PCR primers annealing to each adapter and extended with additional selective nucleotides are used to amplify a subset of restriction fragments that are visualized and classified by length using gel or capillary electrophoresis. This results in highly reproducible DNA fingerprints that are usually recorded as a 1/0 band presence–absence matrix (Meudt & Clarke, 2007). Phylogenetic relationships are then inferred analysing the AFLP matrix directly or converting it into a distance matrix using dissimilarity measures (Koopman, 2005; Althoff et al., 2007).

Amplified fragment length polymorphism markers are appropriate for phylogenetic inference as long as sequence divergence is small, the topology of the underlying evolutionary tree is symmetric, and not very short ancestral branches exist (García-Pereira et al., 2010). As a multilocus method, AFLPs have the benefit of integrating phylogenetic signals from loci distributed throughout the genome, reducing the degree to which lineage sorting, hybridization and reticulation are expected to obscure phylogenetic reconstruction (Koopman, 2005). Therefore, AFLPs can still be used in cases where other methods are not appropriate (Holland et al., 2008). Current evidence suggests that AFLPs largely behave as neutral characters (Bonin et al., 2007), and an AFLP clock has been recently proposed for shallow divergences (Kropf et al., 2009). However, recent theoretical studies indicate that a major drawback of this technique is the low information content of AFLP markers (Simmons et al., 2007; García-Pereira et al., 2010). This weakness seems to have much larger negative impact on tree reliability than other commonly invoked limitations of AFLP data sets, such as the occurrence of size homoplasy due to the lack of homology of comigrating fragments (García-Pereira et al., 2010), or the dominant nature of AFLP characters (Simmons et al., 2007). This implies that to achieve the same level of accuracy in phylogenetic estimation, studies based on AFLPs require more characters than those employing codominant markers (Bensch & Åkesson, 2005).

Determining how many characters are required to eventually recover an accurate and well-supported phylogeny is a key issue in modern molecular phylogenetics (Wortley et al., 2005). Ideally, phylogenetic hypotheses should be constructed by utilizing the maximum number of characters, as there is no theoretical upper limit where extra sampling effort is worthless (Bonin et al., 2007). In practice, however, large volumes of data are expensive to collect, and the true degree of genome coverage remains difficult to evaluate without information on both the genome size and the localization of AFLP markers (Caballero & Quesada, 2010). Therefore, for some trees, phylogenetic methods fail even with massive data (Felsenstein, 1978; Kim, 1996). Alternatively, robust trees can sometimes be recovered from moderate amounts of data (Spinks et al., 2009). Far less clear, however, is how much data are required.

A limited number of attempts have been made to estimate the acceptable number of dominant markers to consider in particular situations to reconstruct a reliable tree. Albertson et al. (1999) used an empirical approach to determine the number of characters needed to obtain an accurate and well-supported AFLP tree in Lake Malawi cichlids. It was found that the 26 nodes of the inferred tree were resolved with 700 informative characters, but mean bootstrap increased from 75% to 90% when up to 1205 informative characters were scored (Albertson et al., 1999). Similarly, Hansen et al. (1999) found that 500 of a total of 11 309 characters scored were enough to obtain the best topology of an AFLP tree of sugarbeet species with a bootstrap support ≥99% for all nodes. Hollingsworth & Ennos (2004) found that 250 dominant characters were required to correctly cluster individuals at low levels of population differentiation, but the use of fewer markers (e.g. 50) was unable to resolve the tree topology even at much higher levels of population differentiation. Thus, there is usually a range in the number of dominant markers below which sampling variance is too large to provide an accurate phylogenetic tree, although it is still a matter of debate where this limit is (Bensch & Åkesson, 2005; Bonin et al., 2007). Conversely, sampling above this range does not necessarily increase the accuracy but may add some noise to the data (Hollingsworth & Ennos, 2004). However, there has not been yet any systematic and extensive exploration of the effect on phylogenetic accuracy of adding AFLP bands to a particular data set under different evolutionary scenarios. Experimental approaches aimed at assessing this question are, however, resource intensive and hampered by the fact that we rarely know the true tree.

Simulation studies are essential for identifying situations where different data matrices or different methods of phylogenetic inference excel or perform poorly (Felsenstein, 1978; Hall, 2005; Cantarel et al., 2006). Within this context, two components of performance are the following: (i) how the inferred tree matches or conflicts with the true tree and (ii) the branch support assigned to both correctly and incorrectly resolved clades. Here, we investigate one potential strategy for improving the performance of AFLP-based phylogenetic inference. Using in silico AFLP fingerprints, we assess the most likely approach to yield a resolved, accurate and well-supported phylogeny and how many AFLP bands would be needed for this. Because the probability of discordance between the estimated and the true tree also varies with tree shape and radiation model (García-Pereira et al., 2010), we conduct simulations using asymmetric and symmetric tree topologies under different radiation and divergence scenarios. Finally, empirical data are used to validate the applicability of in silico simulation settings for inferring the phylogeny of bacteria species of the genus Streptococcus with maximum accuracy, resolution and support.

Methods

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgments
  8. References
  9. Supporting Information

Simulation of DNA sequence evolution

Alignments of DNA sequences were generated with the software Seq-Gen (Rambaut & Grassly, 1997) using the Jukes & Cantor (1969) substitution model. Simulations with Seq-Gen were performed along rooted phylogenetic trees (hereafter referred as reference trees) with symmetric and asymmetric topologies encompassing 16 and 8 ingroup (+1 outgroup) taxa, respectively; to preserve the distribution of internode distances, the asymmetric trees had fewer taxa. Branch lengths were specified using three radiation models: uniform, recent and ancient. In uniform radiation, daughter branches were as long as the parent branch. In recent radiation, daughter branches were half the evolutionary distance of the parent branch, whereas for ancient radiation, daughter branches were twice as long as the parent branch. The minimum and maximum lengths from the most internal node to the tips were set to 0.02 and 0.05 substitutions per site, respectively. This is within the range of divergences for which AFLP data are informative enough for a reliable phylogenetic reconstruction (García-Pereira et al., 2010). For most simulations, sequence lengths were of about 3 Mb (see in silico AFLP below). For some data sets, the length of simulated sequences was increased more than two orders of magnitude (up to 800 Mb) to test the effect of genomic coverage on phylogenetic accuracy.

In silico AFLP

A computer program written in C (Caballero & Quesada, 2010) was used to simulate the cutting of the generated DNA sequences with restriction enzymes EcoRI and MseI, which are the typical enzymes used in AFLP studies. Only fragment sizes between 40 and 440 nucleotides were considered in the subsequent analyses. These correspond to polymerase chain reaction fragments between 72 and 472 bp long when the typical EcoRI and MseI adaptors are added, as these are typical boundary fragment lengths in empirical AFLP studies (Vos et al., 1995). A combination of selective nucleotides adjacent to the restriction sites was used to simulate the selective amplification of restriction fragments. As each additional selective nucleotide reduces the number of amplified fragments four times, one and three selective nucleotides were used at each end for small (3 Mb) and large (800 Mb) genomes, respectively, to ensure 100 AFLP bands per AFLP profile, as recommended for empirical studies (Vos et al., 1995; Caballero & Quesada, 2010). When necessary, the size of simulated genomes evolving under a specific model tree was adjusted to ensure a total of 100 bands per AFLP profile. The number of AFLP bands per data matrix was increased as required by generating additional AFLP profiles with different combinations of selective primers. Bands were scored as presence/absence characters. We note that although new technologies using capillary electrophoresis visualize fragments as peaks rather than as bands, they are also scored as present or absent as in our simulations (Holland et al., 2008).

Reconstruction accuracy

Phylogenies were estimated with PAUP*4 (Swofford, 2003) using the minimum evolution (ME) and maximum parsimony (MP) methods under the same settings that in the study of García-Pereira et al. (2010). ME and MP were conducted using heuristic searches (10 replicates) with random addition using TBR swapping. Estimated unrooted AFLP trees were compared with their corresponding unrooted reference trees by the program Ktreedist (Soria-Carrasco et al., 2007) after pruning the outgroups. This program computes a K-score that measures overall differences in the relative branch length and topology of two phylogenetic trees and allows the comparison between trees differing in divergence scales (Soria-Carrasco et al., 2007). Topological differences between reference and estimated trees were assessed using Robinson–Foulds (R-F) distance (Robinson & Foulds, 1981). High K-scores or R-F distances indicate a poor match between the estimated AFLP-based tree and the reference tree. A total of 50 replicates were run per simulation. An average K-score and R-F distance was computed for each set of replicates.

Branch support

Branch support was determined running 1000 bootstrap replicates per simulation (Felsenstein, 1985). Four different bootstrap consensus trees were estimated with minimum cut-offs of 63%, 86%, 95% and 98%. The averaged overall success of resolution (Simmons & Webb, 2006) was calculated as a weighted average of the four R-F distances resulting from the comparison of the respective consensus trees with the reference tree. As such, clades with a support of 100–98%, 97–95% and 94–86% are given four, three and two times, respectively, the weight as clades with 85–63% support (Simmons & Webb, 2006).

Streptococcus phylogeny

For purposes of comparing in silico AFLP results with high-throughput phylogenetic generation, we retrieved from the interactive Tree Of Life tool (iTOL) (Letunic & Bork, 2006) a tree of Streptococcus bacteria based on a set of 31 orthologous gene sequences (Ciccarelli et al., 2006). We used the following bacterial species (and strains): S. pneumoniae (R6, TIGR4), S. agalactiae (III, V), S. pyogenes (M1, MGAS8232, MGAS315, SSI-1) and S. mutans. Whole-genome bacterial sequences were retrieved for these species and strains using the iTOL web server. The estimated tree based on the set of 31 genes was used as the reference tree for comparison with the in silico AFLP trees based on the whole bacterial genomes.

Results

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgments
  8. References
  9. Supporting Information

Effect of band number on phylogeny accuracy

Figure 1 displays the relationship between number of AFLP bands and tree accuracy for phylogenetic reconstructions based on characters that were simulated onto: (i) two tree topologies (symmetric and asymmetric), (ii) three different radiation models (ancestral, uniform and recent) and (iii) two evolutionary divergences (0.02 and 0.05). Using simulated data sets, accuracy increased with the number of bands for all evolutionary scenarios examined. Tree accuracy displayed a rapid improvement that was subsequently slowed down as an increased number of bands were analysed (Fig. 1). As expected, more closely related sequences (divergence 0.02) produced more accurate trees. For a given number of bands, phylogenies from ancestral radiations were more difficult to reconstruct accurately than those from uniform radiations, which, in turn, were more difficult to reconstruct than recent radiations. However, high resolution to recover the correct tree topology (R-F distance) with a performance similar to that of DNA-based trees was achieved using only about 300 bands for recent and uniform radiations (Fig. 1a). When branch length is added to quantify the success of phylogenetic inference (K-score), AFLP-based trees were rather unreliable across the entire set of band numbers analysed and did not improve substantially once the amount of about 500 bands was reached (Fig. 1b). Consequently, unreliable branch length estimation appears to play a major role on this distinctive pattern of the K-score. Results from the ME and MP simulations were qualitatively very similar (Figs 1 and S1), indicating no substantial differences between distinct reconstruction methods within the assayed range of sequence divergences, as noted previously (García-Pereira et al., 2010). Because the ME and MP methods performed almost identically for all subsequent simulations, the ME method alone is presented hereafter.

image

Figure 1.  Relationship between number of sampled bands and tree accuracy. Trees were estimated using the ME method. Simulations were conducted using symmetric and asymmetric reference trees with ancestral, uniform and recent radiation. Tree accuracy was calculated comparing the estimated tree with the reference tree using (a) R-F distances and (b) K-scores. High R-F distances or K-scores indicate a poor match with the reference tree. Divergence values (0.02 and 0.05) represent the divergence in substitutions/position between the most internal node and the tip in the reference tree. Accuracy values for AFLP-based trees appear as circles. Triangles indicate the accuracy of the single DNA-based tree estimated from a 10-Kb simulated sequence.

Download figure to PowerPoint

We investigated the effect of doubling the number of taxa (from 16 to 32) on the accuracy of phylogenetic inference (Fig. S2). Increasing taxon number increases the number of possible trees and reduces internode distances for a given tree length, making it less likely to recover the correct tree (Grievink et al., 2010). This effect can be clearly visualized for ancestral radiations, which require more than 1000 bands to get a reliable tree (Fig. S2). However, it is still possible to recover the correct tree topology using only a relatively moderate number of bands (approximately 500) for uniform and recent radiations (Fig. S2).

Branch support

Figure 2 displays the averaged success of resolution on the optimal tree (or strict consensus), a measure of both tree topology and branch support performance against the number of bands. Our simulations showed that an increase in band number also results in an overall increase in branch support. This effect was most marked for shorter trees. However, an increase in the number of sampled bands did not necessarily equate to a linear increase in support values. Symmetric trees always reached higher overall levels of support with fewer data than did asymmetric trees, which generally showed a poor performance. Specifically, bootstrap support values ≥98% for all nodes were eventually reached in some data sets under symmetric tree topologies, but not under asymmetric topologies even when up to 1000 bands were sampled. Moderate amounts of data (400–500 bands) are enough in symmetric trees to recover most nodes with relatively high bootstrap values (i.e. ≥70%). Note that the averaged success of resolution yields moderate RF values when the number of bands allows accurate consensus topologies at intermediate but not at high bootstrap cut-off levels, as clades with 70% support are assigned one-quarter the weight of those with 98% support (Fig. S3).

image

Figure 2.  Relationship between number of sampled bands and the averaged success of resolution on the optimal tree (or strict consensus), a measure of both tree topology and branch support performance (Simmons & Webb, 2006). R-F distances were calculated comparing the estimated average consensus tree with the reference tree. High R-F distances indicate a poor match with the reference tree. Divergence values (0.02 and 0.05) represent the divergence in substitutions/position between the most internal node and the tip in the reference tree. Triangles indicate the accuracy of the single DNA-based tree estimated from a 10-Kb simulated sequence.

Download figure to PowerPoint

Tree visualization

To better visualize the relationship between tree accuracy and branch support, we plotted the reference tree and the DNA- and AFLP-based trees inferred from one of the replicate data sets. Figure 3 shows the results corresponding to a symmetric tree with ancient radiation, the most difficult radiation scenario to reconstruct accurately because of the short internal branch lengths. A total of 300 or fewer AFLP bands were clearly insufficient to recover accurate and well-supported relationships among all haplotypes, although some clades could be correctly reconstructed. The correct tree topology could be recovered at a high bootstrap support when sampling 500 bands for low divergences (0.02), but up to 1000 bands for larger divergences (0.05). External nodes displayed higher bootstrap support than deeper nodes, and nodes achieving 90% bootstrap were more prone to increasing support when increasing the number of sampled bands. The performance of AFLP-based trees for asymmetric trees (Fig. 4) was slightly better under the less difficult to reconstruct uniform radiation model, but branch support was drastically reduced even at small divergences. Seq-Gen introduces random variation into the lengths of the trees it generates with respect to the reference tree, which can be visualized when only one replicate is tested. However, for a given evolutionary model, low variances among replicates for K-score (average standard error = 0.0008) and R-F distance (average standard error = 0.389) indicated a strong correlation between the reference tree and simulated trees. Nevertheless, branch length estimation in AFLP-based trees was rather inaccurate under any topology and radiation model when compared with the DNA-based tree even when sampling a large number of bands (Figs 3 and 4).

image

Figure 3.  Phylogenetic reconstructions obtained with different numbers of sampled bands for symmetric trees with ancestral radiation. The divergence between the most internal node and the tip in the reference tree is (a) 0.02 or (b) 0.05 substitutions/position. ME trees estimated from one randomly chosen replicate data set. Numbers above branches are bootstrap values >50%.

Download figure to PowerPoint

image

Figure 4.  Phylogenetic reconstructions obtained with different numbers of sampled bands for asymmetric trees with uniform radiation. (a) The divergence between the most internal node and the tip in the reference tree is (a) 0.02 or (b) 0.05 substitutions/position. ME trees estimated from one randomly chosen replicate data set. Numbers above branches are bootstrap values >50%.

Download figure to PowerPoint

Effect of genome size on tree accuracy

Simulations presented so far assumed a genome size (approximately 3 Mb) typical of Archaea and bacteria. In order to assess whether genome size has an effect on the reliability of phylogenetic reconstruction, AFLP-based trees were reconstructed on sets of simulated sequences of a length typically found in eukaryotic genomes. Simulated genome sizes were increased up to 800 Mb, which represents a genome size similar to that found in the milkweed Asclepias syriaca, but much larger than those from other well-known model species such as Arabidopsis thaliana (120 Mb) or Anopheles gambiae (220 Mb). Supporting information Figures S4 and S5 display all measures of tree accuracy for the same evolutionary scenarios and number of AFLP bands displayed in Fig. 1. It is shown that the reliability of reconstructed trees is virtually the same for a given number of sampled bands, despite 100-fold differences in genome size. Only symmetric trees with ancient radiation recovered slightly but significantly worse tree topologies for large genomes when evolutionary divergence was large. This discrepancy, however, was not mirrored for the K-score, in which both tree topology and branch length accuracy are taken into account.

Practical example: streptococcus phylogeny

An important question raised by our results concerns their relevance to empirical data. To examine this question, in silico AFLP was performed on whole-genome bacterial sequences of Streptococcus spp. retrieved from the iTOL web server. The reference phylogeny is a gene-based reference tree obtained using 31 gene sequences (Fig. 5a). Overall, the in silico AFLP phylogeny of Streptococcus bacteria was fairly well supported when as few as 263 AFLP bands were sampled (cf. Fig. 5a–d). Relationships between Streptococcus strains were the same as in the empirical phylogeny, except S. pyogenes SSI-1. To examine the origin of this discrepancy, we simulated entire bacterial genomes of the same size and GC content than the assayed strains using the gene-based phylogeny as the reference tree. The DNA- and AFLP-based phylogenetic relationships estimated from these simulated data (Fig. 5e, f respectively) were well supported and mirrored those displayed by the gene-based reference tree (Fig. 5a). Thus, while problematic, the relative position of S. pyogenes SSI-1 when using real data does not appear to be an artefact of in silico AFLP.

image

Figure 5. Streptococcus phylogeny. (a) Empirical reference phylogeny based on a set of 31 orthologous genes. (b–d) AFLP-based ME trees estimated from whole bacterial genomes using different combinations (comb) of selective nucleotides. (e–f) ME trees estimated from bacterial genomes simulated along the reference model tree using (e) whole DNA sequences (2 Mb long) and (f) AFLPs (16 primer combinations; 994 bands). Numbers above branches are bootstrap values >50%.

Download figure to PowerPoint

Discussion

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgments
  8. References
  9. Supporting Information

Only a few empirical studies have addressed questions about the relationship between AFLP band sampling effort and phylogenetic resolution (Albertson et al., 1999; Hansen et al., 1999). This study contributes to estimating how much data would be required to improve the inference of an AFLP phylogeny in a reliable and efficient manner. The question was investigated using in silico AFLP on empirical and simulated data sets that evolved along model trees. Although these trees may not be a true representation of evolutionary relationships between real organisms, they are most likely a realistic approximation to the type of AFLP data found in many empirical studies. Our results directly pertain to the ongoing debate regarding the appropriateness of AFLP characters for phylogenetic reconstruction (Hollingsworth & Ennos, 2004; Kosman & Leonard, 2005; Althoff et al., 2007; Meudt & Clarke, 2007; Dasmahapatra et al., 2009).

All measures of accuracy and support showed a rapid improvement that was subsequently slowed down as an increased number of bands were analysed (Figs 1 and 2). At least two factors explain this plateau in resolution and support. First, as clades are resolved, a decreased number of clades remain unresolved. Consequently, the total number of newly resolved clades decreases for each further increment in band number (De Queiroz et al., 2002). Second, once a certain number of bands are sampled, only very difficult nodes will remain to be resolved. This will make more unlikely that any moderate increment in band number can resolve many of these recalcitrant clades (Wortley et al., 2005). In general, our results suggest that the likelihood of reaching such a plateau is strongly influenced by tree topology and radiation model. This is shown in Figs 1 and 2: in terms of tree topology, the data sets that provided greatest improvements in support, resolution and accuracy were those corresponding to symmetric trees. Alternatively, the existence of deep short internode distances and unbalanced trees facilitated the generation of plateaus in which no further substantial improvement was achieved despite increasing band sampling. The results obtained in the present study were not affected by the tree construction method used (Figs 1 and S1). A remarkable finding of our simulations is that branch length estimation was rather unreliable and did not improve substantially after a moderate number of bands were sampled (Fig. 1). This finding has important implications for current practices in AFLP phylogenetic reconstruction. A strict interpretation of our data suggests that the use of AFLP characters should be restricted to studies mostly aimed at determining a correct tree topology. Therefore, AFLPs must not be the marker of choice for accurate molecular dating (sensuKropf et al., 2009).

The results obtained in the present study are not linked to (or influenced by) the length of simulated sequences. All else being equal, reconstructed trees based on an identical number of bands generally yielded similar accuracies for data sets differing in genome size (Figs S4 and S5). Only small genomes outperformed larger genomes for ancient symmetric trees at large divergences (Fig. S4). This discrepancy is the result of using a different number of selective nucleotides rather than differences in genome size. As AFLP bands are scored as being present or absent, losses of a band may occur by changes in the restriction site and/or constituent nucleotides annealing with the selective nucleotides. Thus, shared band absences among taxa are more prone to homoplasy that shared presences because of the multiple ways in which a fragment can be lost (Simmons et al., 2007). As only one selective nucleotide was used for small genomes, but three selective nucleotides were used for large genomes, the incidence of homoplasy due to shared band losses is expected to be higher for the latter data sets. Larger divergences in ancient radiations and larger number of taxa in simulated symmetric than asymmetric trees would also provide more opportunities for mutational changes to occur, generating additional mismatches with selective nucleotides that would generate more band losses. Thus, the higher levels of homoplasy in ancient symmetric data sets produced when using a larger number of selective nucleotides would lead to a comparatively less reliable phylogenetic reconstruction, as observed (Fig. S4). Nevertheless, the impact of homoplasy on phylogenetic accuracy was rather small compared with that due to the number of sampled bands, as the low information content of AFLP characters is a major weakness for phylogenetic reconstruction (García-Pereira et al., 2010). Hence, all conclusions from our study apply to any organism, regardless of genome size.

The exact number of bands required to accurately infer the phylogeny depends upon the levels of accuracy, support and resolution demanded. It is unreasonable to expect any phylogenetic analysis to provide 100% accuracy and branch support (Wortley et al., 2005). However, good topology accuracy and support can probably be achieved with relatively small amounts of AFLP bands. For example, with 500 sampled bands, 100% of nodes were recovered with bootstrap support values of ≥86% for symmetric uniform trees (Fig. S3). With 600 bands, 100% of nodes were recovered with support values of ≥95%; with 700 bands, each node increased up to 100%. Thus, all else being equal, there was relatively little gain in overall number of supported nodes after the first 500 bands. Accordingly, band numbers in the range 500–600 would be informative enough to yield accurate and well-supported nodes (i.e. ≥70%) for most symmetric trees displayed in Fig. 2. Note that bootstrap provides very conservative estimations of the probability of correctly inferring phylogenies when conditions are favourable for phylogenetic analysis, as those simulated here (i.e. equal rates of change and symmetric topologies; Zharkikh & Li, 1992; Hillis & Bull, 1993). Thus, for symmetric trees, a bootstrap equal to or higher than 70% corresponds to a probability of 95% or more that the corresponding clade is real (Hillis & Bull, 1993). On the opposite side, less than approximately 300 bands would be usually insufficient to recover well-supported relationships for the majority of data sets simulated in this study. This value represents a lower bound for data sets with a simple divergence history; AFLP data sets with a similar divergence but unbalanced topologies and short ancestral branches would produce much worse supported trees (Fig. 2).

Our results are consistent with empirical studies reported by Albertson et al. (1999) and Hansen et al. (1999), showing that large AFLP data sets have better chance of correctly resolving phylogenies than do small data sets. However, for comparative purposes, the number of phylogenetically informative bands rather than the total number of sampled bands must be considered. In our simulations, the proportion of informative bands out of the total of bands sampled ranged from 87% for the shallowest phylogenies to 100% for the most divergent data sets. This reduces our lower threshold for accurate phylogenetic reconstruction up to approximately 261 informative bands. From a literature review (period 2005–2010) of 51 empirical phylogenetic studies on plants, animals and bacteria, we found that the number of informative AFLP bands scored ranged between 62 (Gong et al., 2008) and 2365 (Meudt et al., 2009). If we take the median value, 418, as the more common number of scored bands, we obtain in our simulations that this number is within of what it is generally acceptable for an accurate tree topology reconstruction when 16 or less taxa are sampled (Fig. 1). However, 24 (47%) of these studies sampled more than 16 taxa (20–59 taxa), and the majority (90%) displayed higher evolutionary divergences (0.10–1.06 substitutions per site) than those simulated here (Figs 1 and S2). Remarkably, 14 studies (27%) using 16 or less taxa were based on a number of informative bands below our threshold of approximately 261 (e.g. Zerega et al., 2005; Lambertini et al., 2006; Giraud et al., 2007; Schmidt-Lebuhn, 2007; Gong et al., 2008; Koblmuller et al., 2010; Shaheen et al., 2010; Arrigo et al., 2011). Hence, accurate phylogenetic reconstruction seems to remain an issue in many phylogenetic studies using AFLPs.

Generally, our in silico simulations using real bacterial genomes were in line with our a priori expectations in that a resolved and well-supported tree topology can be recovered with a moderate number of AFLP bands (Fig. 5). The reconstructed tree topology was consistent with the empirical phylogeny based on a set of 31 orthologous gene sequences, while branch lengths were rather unreliable even when up to 1166 bands were sampled, as predicted by our simulations (Fig. 1). Both data sets agreed on the monophyly of strains from the same bacterial species, but not on some relationships between strains within species. Specifically, direct comparison of our in silico AFLP tree with the empirical tree was complicated by differences in the placement of S. pyogenes SSI-1 strain (Fig. 5). A phylogeny based on the set of 1462 orthologous gene sequences among these same S. pyogenes strains strongly supported the pattern displayed by the empirical phylogeny (ATGC database; Novichkov et al., 2009). Further, a more detailed examination of S. pyogenes genomes revealed that the strain SSI-1 contains less noncoding DNA (1–2%) and three putative horizontally transferred genes (IMG database; Markowitz et al., 2009). A comparison between bacterial strains reveals that up to 301 S. pyogenes SSI-1 coding genes (15% of the total) are not orthologous with the closely related M1 strain (BioMart tool; EMBL-EBI database). These findings suggest that, as AFLPs screen entire bacterial genomes, the observed discrepancy for the AFLP phylogeny was most likely the result of a combination of ancient gene capture and gene loss events. This example illustrates that AFLPs may be an efficient and complementary tool to traditional DNA sequencing for assessing bacterial genomics.

Because simulation studies are performed under well-defined conditions, our results almost certainly overestimate our ability to reconstruct accurate phylogenies from empirical data. Uncertainties not included in our simulations such as a poor fragment mobility resolution (Pompanon et al., 2005) or incorrect scoring of bands (Holland et al., 2008) may have an important impact on both the information content and phylogenetic reconstruction of AFLP data sets. Thus, our estimates of reconstruction accuracy and support must be seen as best case studies. Despite these limitations inherent to any simulation study, our results argue strongly for the insights that can be gained from AFLP markers. Recent innovative developments for the quasi-codominant treatment of AFLPs for estimating population genetic parameters (Foll et al., 2010) have opened promising possibilities for further developments aimed at the codominat treatment of phylogenetically informative AFLP characters. Thus, despite the genomic era having brought about massive advances for the high-throughput generation of DNA sequence data, the development of approaches that make better use of AFLP markers will probably allow AFLPs to continue having a relevant role for testing important phylogenetic hypotheses in many model and nonmodel organisms.

Acknowledgments

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgments
  8. References
  9. Supporting Information

We are grateful to Simon Whelan for useful discussions and Mark P. Simmons and an anonymous reviewer for comments on the manuscript. This work was funded by the Ministerio de Ciencia e Innovación and Fondos Feder (CGL2009-13278-C02) and Xunta de Galicia (Grupos de Referencia Competitiva, 2010/80).

References

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgments
  8. References
  9. Supporting Information

Supporting Information

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. Acknowledgments
  8. References
  9. Supporting Information

Figure S1 Relationship between number of sampled bands and tree accuracy.

Figure S2 Relationship between number of sampled bands and tree accuracy when doubling the number of taxa.

Figure S3 Relationship between number of sampled bands and the averaged success of resolution on the optimal tree (or strict consensus), a measure of both tree topology and branch support performance (Simmons and Webb 2006).

Figure S4 Relationship between number of sampled bands and tree accuracy for genome sizes differing in two orders of magnitude: symmetric trees.

Figure S5 Relationship between number of sampled bands and tree accuracy for genome sizes differing in two orders of magnitude: asymmetric trees. Simulations were conducted using reference trees with ancestral, uniform and recent radiation.

As a service to our authors and readers, this journal provides supporting information supplied by the authors. Such materials are peer-reviewed and may be re-organized for online delivery, but are not copy-edited or typeset. Technical support issues arising from supporting information (other than missing files) should be addressed to the authors.

FilenameFormatSizeDescription
JEB_2361_sm_FigureS1.tif1272KSupporting info item
JEB_2361_sm_FigureS2.tif719KSupporting info item
JEB_2361_sm_FigureS3.tif401KSupporting info item
JEB_2361_sm_FigureS4.tif1169KSupporting info item
JEB_2361_sm_FigureS5.tif1180KSupporting info item

Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.