Health Articles

 

Msp315 988.1000

Evaluating the Relationship between Evolutionary Divergenceand Phylogenetic Accuracy in AFLP Data Sets Marı´a Jesu´s Garcı´a-Pereira, Armando Caballero,* and Humberto Quesada Departamento de Bioquı´mica, Gene´tica e Inmunologı´a, Facultad de Biologı´a, Universidade de Vigo, Vigo, Spain *Corresponding author: E-mail: [email protected].
Associate editor: Dan Graur Using in silico amplified fragment length polymorphism (AFLP) fingerprints, we explore the relationship between sequencesimilarity and phylogeny accuracy to test when, in terms of genetic divergence, the quality of AFLP data becomes too lowto be informative for a reliable phylogenetic reconstruction. We generated DNA sequences with known phylogenies using balanced and unbalanced trees with recent, uniform and ancient radiations, and average branch lengths (from the mostinternal node to the tip) ranging from 0.02 to 0.4 substitutions per site. The resulting sequences were used to emulate theAFLP procedure. Trees were estimated by maximum parsimony (MP), neighbor-joining (NJ), and minimum evolution (ME)methods from both DNA sequences and virtual AFLP fingerprints. The estimated trees were compared with the referencetrees using a score that measures overall differences in both topology and relative branch length. As expected, the accuracyof AFLP-based phylogenies decreased dramatically in the more divergent data sets. Above a divergence of approximately 0.05, AFLP-based phylogenies were largely inaccurate irrespective of the distinct topology, radiation model, or phylogeneticmethod used. This value represents an upper bound of expected tree accuracy for data sets with a simple divergencehistory; AFLP data sets with a similar divergence but with unbalanced topologies and short ancestral branches producedmuch less accurate trees. The lack of homology of AFLP bands quickly increases with divergence and reaches its maximumvalue (100%) at a divergence of only 0.4. Low guanine-cytosine (GC) contents increase the number of nonhomologousbands in AFLP data sets and lead to less reliable trees. However, the effect of the lack of band homology on tree accuracy issurprisingly small relative to the negative impact due to the low information content of AFLP characters. Tree-buildingmethods based on genetic distance displayed similar trends and outperformed parsimony at low but not at highdivergences. However, the impact of using alternative phylogenetic methods on tree accuracy was generally small relativeto the uncertainty arising from factors such as divergence, nonhomology of bands, or the low information content of AFLPcharacters. Nevertheless, our data suggest that under certain circumstances, AFLPs may be suitable to reconstruct deeperphylogenies than usually accepted.
Key words: AFLP, phylogenetic inference, homoplasy, computer simulation, accuracy, phylogeny.
2005) and Bayesian analysis (Brouat et al. 2004; Luoet al. 2007).
Amplified fragment length polymorphism (AFLP; Vos et al.
The appropriateness of AFLP data for phylogenic recon- 1995) is becoming an extensively used DNA fingerprinting struction is however compromised by several properties of method for many studies on plants, animals, and microor- AFLP markers that have led many authors to question their ganisms (Mueller and Wolfenbarger 1999; Meudt and utility as phylogenetic characters (e.g., Hollingsworth and Clarke 2007). The discovery that many AFLP data sets con- Ennos 2004; Kosman and Leonard 2005). One of the most tain phylogenetic signal (Giannasi et al. 2001; Despres et al.
2003; Koopman 2005) has stimulated its use as a source of important drawbacks of the AFLP technique is that comi- genetic information for phylogenetic inference, particularly grating bands of the same length may not be homologous among closely related genera or species (Meudt and Clarke with one another, a phenomenon termed ‘‘fragment size 2007). The technique is based on the selective amplification homoplasy'' (Vekemans et al. 2002). This implies that of DNA fragments from digested total genomic DNA, gen- within an individual, a band may contain fragments from erating reproducible fingerprints that are usually recorded different regions of the genome and that, in comparisons as a 1/0 band presence–absence binary matrix. It is partic- among taxa, fragments of equal size may also not come ularly suitable for nonmodel organisms for which no prior from the same locus (Gort et al. 2006; Althoff et al.
DNA sequence is available, producing a large number of 2007). In addition, the generally anonymous nature of genomewide informative markers per assay unit at a rela- bands may lead to the inadvertent presence of contami- tively low cost (Bensch and A ˚ kesson 2005). Phylogenetic nant DNA from parasites, symbionts, or hosts (Karp relationships are usually inferred from AFLP data convert- et al. 1996). The absence of bands is also susceptible to ing the binary matrix into a distance matrix using dissim- the lack of homology because they are treated as homol- ilarity measures, or using the binary matrix directly for ogous despite the multiple ways in which a fragment can be character-based methods such as parsimony (Koopman lost (Simmons et al. 2007).
The Author 2009. Published by Oxford University Press on behalf of the Society for Molecular Biology and Evolution. All rights reserved. For permissions, pleasee-mail: [email protected] Mol. Biol. Evol. 27(5):988–1000. 2010 Advance Access publication December 21, 2009 Phylogenetic Accuracy of AFLPs · doi:10.1093/molbev/msp315 Empirical and theoretical studies that have tested the Computer simulation provides an efficient alternative to homology of comigrating bands have demonstrated that experimental approaches because it can mimic the AFLP the lack of homology is a real and prevalent component technique producing virtual AFLP fingerprints that enable of AFLP data sets (e.g., Rouppe van der Voort et al.
the identification of fragments based on their length and 1997; El-Rabey et al. 2002; Althoff et al. 2007). The propor- DNA sequence (Rombauts et al. 2003). This approach al- tion of nonhomologous bands ranged from about 10% lows for the generation of data sets that evolve following within closely related genotypes (Rouppee van der Voort a given model tree, constructing phylogenetic trees from et al. 1997; Hansen et al. 1999) to as high as 100% for dis- that data set, and then comparing the inferred trees with tantly related taxa (O'Hanlon and Peakall 2000). Therefore, the model tree to determine the accuracy of phylogenetic nonhomology rapidly increases among AFLP profiles from estimation. In silico, AFLP has already been applied to increasingly divergent taxa (Althoff et al. 2007) and with quantify the effect of using presence–absence characters the number of bands per assay unit (Koopman and Gort in AFLP data relative to standard multistate characters 2004). The artefactual similarities among taxa originated by (Simmons et al. 2007), to test band homology (Koopman the lack of band homology may contribute to spurious and Gort 2004; Althoff et al. 2007), to select restriction en- phylogenetic relationships, and on average, this should zymes (Peters et al. 2001), or to assess the effect of homo- be more of a problem when the studied species are dis- plasy on the estimation of population genetic parameters tantly related (Bensch and A ˚ kesson 2005).
and the detection of natural selection in AFLP data sets Concerns other than lack of homology have also been (Caballero et al. 2008).
expressed that may potentially limit the usefulness of AFLP In this study, we take a bioinformatics approach to data for phylogenetic analyses (e.g., Bensch and A assess the accuracy of AFLPs for phylogenetic reconstruc- 2005; Koopman 2005; Meudt and Clarke 2007). One impor- tion. Using in silico AFLP fingerprints, we explore the rela- tant limitation is that AFLP bands are usually scored as tionship between sequence similarity and phylogeny dominant markers (Simmons et al. 2007). Therefore, bands accuracy over a wide range of tree topologies, phylogenetic that do not comigrate are not necessarily independent of methods, and evolutionary divergences to test when, in one another even though they are scored as different pres- terms of genetic divergence, the quality of AFLP data be- ence–absence characters (Koopman 2005). There is also an comes too low to be informative for a reliable phylogenetic asymmetry in the probability of losing or gaining fragments, because it is more likely destroy rather than generate newrestriction sites (Luo et al. 2007). These features would Materials and Methods increase the amount of stochastic noise in the data,making them less likely to recover the correct phylogenetic Simulation of DNA Data Sets relationships (Koopman 2005; Simmons et al. 2007).
Alignments of DNA sequences were generated with Seq- Given the popularity of AFLP markers for phylogenetic Gen (Rambaut and Grassly 1997), a program that evolves inference, it is surprising that relatively little has been done sequences along a specific tree. The input of the program is to investigate the actual reliability of AFLPs. Some authors a tree (hereafter referred to as the reference tree) in which have tried to determine the accuracy of AFLP-based trees by the user specifies its topology, number of sequences to be comparing them with trees obtained using an independent simulated, and the average length of the branches on the source of data (e.g., Marhold et al. 2004; Sullivan et al. 2004; tree. We used the Jukes and Cantor (1969) substitution ¨ller et al. 2007; Pellmyr et al. 2007; Dasmahapatra model in Seq-Gen, which assumes equal base frequencies et al. 2009). Koopman (2005) compared AFLP and internal and equal mutation rates. The length of the simulated transcriber spacer (ITS) data sets over a wide variety of taxa sequences was limited to 40,000 nt in order to get 90– (plant, fungi, and bacteria) and found that AFLP-based phy- 100 bands per AFLP profile, as recommended for logenies were largely consistent with those derived from se- experimental studies (Vos et al. 1995).
quence data. Approaches have also been developed aimed at Simulations with Seq-Gen were performed along phy- testing the significance of a given pairwise similarity among logenetic trees with two different topologies, a symmetric distinct AFLP fingerprints, enabling the removal of distantly tree and an asymmetric tree. These known trees (refer- related genotypes contributing to noise in AFLP data sets ence trees) were manually constructed. We sought a range (Koopman and Gort 2004). However, the comparison of of evolutionary divergences comparable with those likely fragment homology among taxa suggested that the phylo- to be encountered when estimating phylogenies from ex- genetic usefulness of AFLPs varies greatly depending on the perimental data sets. The minimum and maximum time since divergence and the specific genomic features of lengths from the most internal node to the tips were the compared taxa (Althoff et al. 2007). Although it is gen- 0.02 and 0.40 substitutions per position, respectively.
erally accepted that AFLPs may not provide an accurate es- The branch lengths of the trees in figure 1 were multiplied timate of species phylogeny when genetic divergence is too by factors of 0.2, 0.5, 1, 2, 3, and 4, so that we used in total 6 high, it is still a matter of debate where this limit is (Meudt different sets of evolutionary divergences. Saturation of and Clarke 2007). Experimental studies aimed at assessing simulated DNA data sets was tested by plotting pairwise this question are, however, resource intensive and hampered transition and transversion sequence differences against by the fact that we rarely know the true tree.
Garcı´a-Pereira et al. · doi:10.1093/molbev/msp315 enzymes EcoRI and MseI, which are the typical enzymesused in AFLP studies. The program searches for all the re-striction sites and returns all fragments that would resultfrom the digestions. The output is the list of fragmentssorted by their length, as it would appear in a real exper-iment. The in silico AFLP profiles were generated withoutselective nucleotides to obtain the highest possible numberof AFLP fragments. Only fragment sizes between 40 and 440nt were considered in the subsequent analyses, which cor-respond to polymerase chain reaction (PCR) fragments be-tween 72 and 472, because the typical EcoRI and MseIprimers contain a 16-bp long sequence when excludingthe selective nucleotides. This set of fragment sizes is equiv-alent to the range of fragment sizes normally scored inAFLP studies (Meudt and Clarke 2007). This yielded 90–100 bands per AFLP profile, as recommended for experi-mental data sets (Vos et al. 1995; Vekemans et al. 2002).
Data from different profiles were converted into a binary1/0 presence–absence matrix.
True Fragments versus Experimental BandsBecause all selected fragments were individually identified,this allowed us to count the total number of fragments(hereafter referred as ‘‘true fragments''), the number ofbands in a putative electrophoretic setting (hereafter re-ferred as ‘‘experimental bands''), and the proportion ofthese that are nonhomologous.
A comparison between the phylogenies inferred from true fragments and experimental bands allowed us to as-sess the impact of the lack of band homology on phyloge-netic accuracy in AFLP data sets. To this aim, two differentbinary matrices were generated from each simulated data FIG. 1. Reference tree topologies. Alignments of DNA sequences set, a matrix including the true fragments, and a matrix were generated with Seq-Gen along symmetric and asymmetric including the experimental bands.
trees with ancestral, uniform, and recent radiation. The scale bar, insubstitutions/position, corresponds to the trees with a divergence ofx1. All outgroup sequences were pruned for presentation.
Phylogenetic AnalysesAFLP-based phylogenies were estimated with PAUP* For each tree topology, branch lengths were specified (Swofford 2002) using the three most widely used methods using three radiation models: uniform, recent, and ancient.
in AFLP data sets (Koopman 2005): 1) neighbor-joining In uniform radiation, daughter branches were as long as the (NJ), 2) minimum evolution (ME), and 3) maximum par- parent branch. In recent radiation, daughter branches were simony (MP). For distance-based trees (NJ and ME), each half the evolutionary distance of the parent branch, binary file was used to make a distance matrix using the Nei whereas for ancient radiation, daughter branches were and Li (1979) method. This distance method uses shared twice as long as the parent branch. These trees had several presences and is less susceptible to lack of homology than short internal branches that made them difficult to resolve.
methods based on shared absences and presences (e.g., Eu- Thus, they are trees where the interplay between sequence clidean distance; Kosman and Leonard 2005). The charac- similarity and tree topology will differentially impact the ter-based MP method used the Wagner parsimony phylogenetic accuracy. An outgroup was included in the criterion directly on the binary matrix. Although this crite- reference trees used as input in Seq-Gen to root the trees.
rion assumes equal loss–gain probabilities of restriction This outgroup could be easily identified as the first se- sites, it better fits the AFLP data than methods assuming quence of the DNA alignment generated by SeqGen. For unequal probabilities such as Dollo parsimony (Koopman each tree created, 1,000 independent replicate data sets 2005). MP and ME were conducted using heuristic searches were generated by simulating evolutionary change along (10 replicates) with random addition using tree bisection and reconnection swapping. When several equally optimaltrees were found (for MP and ME), only the first one was Simulation of AFLP Fingerprints used. Given that the primary objective of this study is to A computer program written in C was used to simulate the assess the accuracy of both branch length and topology cutting of the generated DNA sequences with restriction estimation for AFLP-based trees, we were forced to select Phylogenetic Accuracy of AFLPs · doi:10.1093/molbev/msp315 a single optimal tree rather than working with a strict con-sensus to measure performance. NJ, ME, and MP trees werealso directly estimated from DNA sequences for the sake ofcomparison. This allowed us to quantify how much less ac-curate AFLP-based trees are with respect to the trees esti-mated from the corresponding source DNA sequences. Wedid not perform Bayesian trees because of the enormouscomputational time required for doing enough numberof replicates of all simulations performed.
Tree EvaluationEstimated trees were compared with their correspondingreference trees by the program Ktreedist (Soria-Carrascoet al. 2007), which takes into account both topologyand branch length information of a phylogenetic tree. Thisprogram computes a K-score that measures overall differ-ences in the relative branch length and topology of twophylogenetic trees after scaling one of the trees to havea global divergence as similar as possible to the other tree.
This allows for the comparison among trees differing inevolutionary rates and/or divergence scales (Soria-Carrascoet al. 2007). Trees compared with Ktreedist must be allrooted or unrooted. To this end, trees were saved with es-timated branch lengths, the outgroups were pruned, andunrooted reference and estimated trees were comparedwith Ktreedist. High K-scores indicate a poor match be-tween the estimated tree and the reference tree. An aver- FIG. 2. Relationship between sequence divergence and the pro- age K-score was computed for each set of 1,000 replicates, portion of nonhomologous bands. Simulations were conducted except for DNA-based trees, in which only 100 replicates using asymmetric (a) and symmetric (b) reference trees with recent, were used because of computational demands.
uniform, and ancestral radiation. Divergence values represent the For some data sets, we also compared reference and es- divergence in substitutions/position between the most internal timated trees using the symmetric difference or Robinson– node and the tip in the reference tree.
Foulds (R–F) distance (Robinson and Foulds 1981), whichonly takes into account the topology of a phylogenetic tree.
0.163 ± 0.0003). Thus, as the evolutionary divergence among This allowed us to evaluate the relative contribution of tree AFLP profiles increased, many of the bands became unique topology and branch length to tree accuracy.
to a taxon, only a few bands were shared among taxa, andan increasing amount of these became nonhomologous.
The proportion of nonhomologous bands was smaller in recent than in uniform radiations, which in turn had We generated phylogenies from two tree topologies (sym- a smaller overall number of nonhomologous bands than metric and asymmetric), three radiation models, three phy- did ancestral radiations. Therefore, the impact of the lack logenetic methods, and six evolutionary divergences. The of homology on recent radiations with respect to that symmetric and asymmetric tree topologies included 16 observed for ancient and uniform radiations was relatively and 8 taxa, respectively (to preserve the distribution of in- smaller in asymmetric than in symmetric trees.
ternode distances, the asymmetric trees had fewer taxa).
Figure 3 shows the quality of phylogenetic reconstruc- The comparison of the number of base changes due to tions in simulated data sets by comparing the reference transitions and transversions as a function of the diver- tree to the AFLP-based trees obtained with the three phy- gence among paired sequences fitted a linear regression logenetic methods. The two distance-based methods (NJ on each simulated data set (r2 5 0.99; P , 0.001), thus and ME) showed almost identical performance for any tree not suggesting saturation.
topology and evolutionary divergence; because the two The percentage of nonhomologous experimental bands distance-based methods performed almost identically, increased (i.e., homology decreased) with increasing evolu- the ME method alone is presented hereafter. The relative tionary divergence, reaching the value of 100% at a diver- efficiencies of MP with the distance-based methods in ob- gence of 0.4 (fig. 2). This increase is entirely explained by the taining the correct topology and in estimating the branch decreasing homology among comigrating bands from lengths were critically affected by the tree topology. For different taxa, because the proportion of bands containing asymmetric trees, the distance-based methods were more nonhomologous fragments within an individual remained ef- efficient than the character-based MP method (fig. 3a). The fectively invariable in all the simulated data sets (average 5 superiority of distance-based trees disappeared for the Garcı´a-Pereira et al. · doi:10.1093/molbev/msp315 FIG. 3. Relative accuracy of phylogenetic methods for reconstructing AFLP-based trees. Simulations were conducted using asymmetric (a) andsymmetric (b) reference trees with ancestral, uniform, and recent radiation. K-scores were calculated comparing the estimated tree with thetrue tree. High K-scores indicate a poor match with the reference tree. Divergence values represent the divergence in substitutions/positionbetween the most internal node and the tip in the reference tree. The data points of ME and NJ are almost completely overlapping. Trees wereinferred from experimental bands.
symmetric tree topology, for which MP and distance-based long sequence (Vos et al. 1995). This analysis, however, did methods had similar accuracy results, with MP slightly in- not have an appreciable effect on the reliability of the newly ferior in the ancestral radiation model at large evolutionary reconstructed AFLP trees, yielding nearly identical results divergences (.0.10) (fig. 3b).
to those displayed in figure 4.
The overall success of phylogenetic reconstruction based Because researchers are often interested on the accuracy on characters that were simulated on symmetrical and of tree topology, we evaluated this feature directly using R– asymmetrical trees using true fragments, experimental F topological distance. Figure 5 shows the average R–F bands, and aligned DNA sequences, are presented in distance to the reference tree for the same trees estimated figure 4. As expected, DNA sequence-based phylogenies in figure 4. In general, R–F distances displayed a similar pat- produced much more reliable estimates of the true phylog- tern to that shown by K-scores, with low evolutionary di- eny than AFLP-based phylogenies under all the radiation vergences producing more reliable trees. However, in terms models and evolutionary divergences. More closely related of topology, the likelihood of estimating the correct tree is sequences produced more accurate trees, and accuracy most dependent on whether radiation was ancient or re- steadily decreased with evolutionary divergence. Thus, cent. Remarkably, the likelihood of recovering the correct for a given number of taxa, the K-score increased as the tree topology was outstandingly higher for recent radia- average branch length increased, and this increase occurred tions, which displayed relatively low R–F distances even much more rapidly for AFLP-based trees than for DNA- at large evolutionary divergences. This pattern was not ob- based trees. For divergences below approximately 0.05, served for uniform and ancestral radiation models, in which phylogenies from ancestral radiations were more difficult the likelihood of recovering the correct topology rapidly to reconstruct accurately than those from uniform radia- decreased with evolutionary divergence. Consequently, un- tions, which, in turn, were more difficult to reconstruct reliable branch length estimation appears to play a major than recent radiations (not easily appreciated in fig. 4 be- role on the magnitude of K-score for recent radiations.
cause of the scale). In all cases, the accuracy of AFLP-based In contrast, inaccurate tree topology estimation has trees decreased dramatically at greater divergences, with no a key contribution to K-score for uniform and ancestral substantial differences in performance among the different radiation models. Because fragment homology is increased AFLP-based trees conducted on experimental bands had by excluding smaller fragments (Althoff et al. 2007), we a generally lower accuracy than AFLP-based trees con- tested the effect of fragment size on tree accuracy using ducted on true fragments (figs. 4 and 5). In some ME trees, only large fragments. We only used fragments longer than the reliability of trees inferred from experimental bands 100 nt, which correspond to PCR fragments of 132 bp, be- strikingly outperformed that of trees inferred from true cause the typical EcoRI and MseI adaptors contain a 16-bp fragments at large evolutionary divergences (.0.20) Phylogenetic Accuracy of AFLPs · doi:10.1093/molbev/msp315 FIG. 4. Relationship between sequence divergence and tree accuracy. Trees were estimated using both MP (a) and ME (b) methods. Simulationswere conducted using asymmetric and symmetric reference trees with ancestral, uniform, and recent radiation. K-scores were calculatedcomparing the estimated tree with the true tree. High K-scores indicate a poor match with the reference tree. Divergence values represent thedivergence in substitutions/position between the most internal node and the tip in the reference tree. AFLP-based trees were inferred fromtrue fragments (i.e., fragments individually identified irrespective of their size) and experimental bands (i.e., bands observed in a putativeelectrophoretic setting, which could include one or more fragments of a given size). NJ yielded similar results to those of ME and is notpresented.
despite the presumably negative impact of nonhomolo- Because the number of nonhomologous experimental gous bands (figs. 4b and 5b). This unexpected observation bands is a function of the number of fragments produced, is the consequence of an excessively large divergence in which in turn depends on the GC content (GC 5 0.5 in the some pairwise comparisons among AFLP profiles, for which results shown so far), we evaluated the impact of varying the Nei and Li distance is undefined. In these cases, the the GC content on the overall success for phylogenetic re- program PAUP*4 arbitrarily sets the values of these un- construction (fig. 7). For these simulations, we used the defined distances as twice the distance of the largest de- HKY model (Hasegawa et al. 1985) in Seq-Gen assuming fined distance in the distance matrix. To avoid the equal transition and transversion rates. The relationship undesirable effects of these undefined distances on tree between phylogenetic accuracy and genome GC content accuracy, we tried a value of 1.0 as the default upper yielded contrasting results depending on the set of markers bound divergence in the matrix generated by PAUP*4.
used to infer the phylogeny. There is a negative relationship This resulted in a consistently lower efficiency of between phylogenetic accuracy and GC content for AFLP- experimental bands than true fragments for inferring based trees inferred from experimental bands. In contrast, AFLP-based trees across the entire set of evolutionary this relationship turns out to be positive for AFLP-based divergences (fig. 6), as expected.
trees inferred from true fragments. The topology or Garcı´a-Pereira et al. · doi:10.1093/molbev/msp315 FIG. 5. Testing the effect of tree topology on phylogenetic accuracy. Average R–F topological distances to the reference tree for the same treesestimated in figure 4. High R–F scores indicate a poor match with the reference tree.
phylogenetic method used had a small effect on this rela- To better visualize the relationship between K-score and tionship, which was more apparent in asymmetric than in tree accuracy, we plotted the reference tree and the DNA- symmetric trees. No appreciable effect of GC content on and AFLP-based trees estimated from one of the replicate phylogenetic accuracy was observed for DNA-based trees.
data sets. Seq-Gen introduces random variation into the FIG. 6. Testing the effect of undefined distances on tree accuracy. A value of 1.0 was used as the default upper bound divergence in the distancematrix generated by PAUP*4. Trees are estimated using ME for the three cases displaying an anomalous pattern in Figure 4b. K-scores werecalculated comparing the estimated tree with the true tree. High K-scores indicate a poor match with the reference tree. Divergence valuesrepresent the divergence in substitutions/position between the most internal node and the tip in the reference tree. AFLP-based trees wereinferred from experimental bands and true fragments.
Phylogenetic Accuracy of AFLPs · doi:10.1093/molbev/msp315 FIG. 7. Relationship between tree accuracy and GC content. Simulations were conducted using asymmetric (a) and symmetric (b) referencetrees. The divergence between the most recent common ancestor of the ingroup and the tip in the reference tree was set to 0.1 substitutions/position. K-scores were calculated comparing the estimated tree using MP or ME with the reference tree. High K-scores indicate a poor matchwith the reference tree. To make fair comparisons among trees estimated from different GC contents, sequence lengths ranged from 20 to 40kb to get 90–100 experimental bands per AFLP profile.
lengths of the trees it generates with respect to the refer- As expected, the accuracy of AFLP-based phylogenies was ence tree that may obscure the evolutionary model when the greatest in the least divergent data sets (Althoff et al.
only one replicate is tested. However, for a given evolution- 2007). However, the efficiency of phylogenetic reconstruc- ary model, K-scores from independent replicates were tion decreased dramatically in more divergent data sets.
nearly identical to each other (standard error across Above a divergence of approximately 0.05, AFLP-based 1,000 replicates ranged from 0.00005 to 0.002), indicating phylogenies were largely inaccurate irrespective of the dis- a strong correlation between the reference tree and the tinct topologies, radiation models, or phylogenetic meth- resulting simulated trees. Figure 8a and b shows the results ods used (fig. 4). For these divergences, AFLP profiles corresponding to a symmetric topology under a uniform from different taxa shared so few fragments that trees in- radiation model. At small evolutionary divergences typical ferred from experimental bands were almost effectively of very closely related genotypes, haplotypes were correctly random with respect to the reference tree (fig. 8b). For di- grouped within the AFLP-based trees but estimated branch vergences below 0.05, generally the likelihood of estimating lengths were very inaccurate when compared with those an accurate AFLP-based tree was most dependent on tree obtained in the DNA-based tree (fig. 8a). However, topology or on whether radiation was ancient or recent. At AFLP-based trees produced vastly incorrect topologies this divergence level, simple symmetric topologies were re- and branch lengths for large evolutionary divergences typ- markably more likely to be accurately reconstructed than ically above the species rank (fig. 8b). The performance of the more complex asymmetric topologies (fig. 8a and c).
AFLP-based trees was drastically reduced for asymmetric Similarly, the efficiency in recovering the correct tree topologies even at small divergences (fig. 8c) was sensibly superior for recent than for uniform radiations,which in turn had a better performance than ancient ra- diations (figs. 4 and 5), as expected from their increasinglyshorter ancestral branches (Cantarel et al. 2006). Conse- Relationship between Sequence Divergence and quently, AFLP markers appear to be appropriate for phy- Phylogenetic Accuracy logenetic inference as long as divergence is small, the We have evaluated the limits of AFLP markers for phylo- topology of the reference tree is symmetric, and not very genetic reconstruction by assessing the accuracy of recov- short ancestral branches exist.
ering the correct tree for diverse data sets evolving under It is not feasible to precisely define the taxonomic rank different tree topologies, radiation models and evolution- that would correspond to the divergences assayed in this ary divergences. Our results demonstrate that phylogenetic study. Levels of genetic divergence between congeneric accuracy, as measured in both topology and branch length, species and confamilial genera vary among taxa as a result is seriously compromised by sequence divergence and, to a of differences in age, rate of molecular evolution, or biolog- lower extent, by tree topology and radiation pattern (fig. 4).
ical features (Johns and Avise 1998; Sites and Marshall Garcı´a-Pereira et al. · doi:10.1093/molbev/msp315 FIG. 8. Phylogenetic accuracy with respect to evolutionary divergence. (a) Symmetric trees. The divergence between the most recent commonancestor of the ingroup and the tip in the reference tree is of 0.02 substitutions/position. (b) Symmetric trees. The divergence between themost internal node and the tip in the reference tree is of 0.20 substitutions/position. (c) Asymmetric trees. The divergence between the mostinternal node and the tip in the reference tree is of 0.02 substitutions/position. MP trees estimated from one randomly chosen replicate dataset. The scale bar of MP trees corresponds to the number of steps.
2003). An extensive literature review revealed a rough to- and Peakall 2000; Bussel et al. 2005; Koopman 2005; Althoff pological congruence of AFLP and ITS trees (Koopman et al. 2007). For example, human and chimpanzee have ac- 2005), suggesting that AFLP markers are informative at cumulated a genomic divergence of 1.24% after they di- somewhat lower taxonomic levels than ITS sequences. If verged from each other 6 Ma (Chen and Li 2001). This we use congruence with ITS as a proxy for determining tax- implies that sequences evolving under the scenario repre- onomic rank, AFLP markers should be useful as a source of sented in figure 8a would produce reliable AFLP-based tree phylogenetic information at a level generally not exceeding topologies among species that diverged 19 Ma, provided an closely related congeneric species (Savelkoul et al. 1999) evolutionary rate similar to that of hominids.
and would be quite informative in determining phylogeo- Because simulation studies are performed under well- graphic patterns. Our data, however, suggest that under defined conditions, our results almost certainly overesti- certain circumstances, AFLPs may also be suitable to recon- mate our ability to reconstruct accurate phylogenies struct deeper phylogenies than usually accepted (O'Hanlon from empirical data. For example, a substantial source of Phylogenetic Accuracy of AFLPs · doi:10.1093/molbev/msp315 uncertainty in AFLP data not included in our simulations is simulations show this comparison and indicate that the that due to an insufficient fragment mobility resolution or lack of band of homology among taxa quickly increases an incorrect scoring of bands (Pompanon et al. 2005).
with divergence, thus rapidly compromising the phyloge- Holland et al. (2008) found that the choice of parameter netic usefulness of AFLP data sets (fig. 2). The upper limit settings in automated scoring software notoriously affects to fragment homology is reached at a divergence threshold phylogenetic resolution. Similarly, the data sets generated of only 0.4, in which the few bands that are still shared in this study were based on the simulation of random DNA among taxa are all nonhomologous. Consequently, sequences, but in fact DNA sequences of living organisms data sets containing highly diverged genotypes, typical follow specific nonrandom patterns (Karlin and Burge of ancient radiations, are consistently more affected by 1995). These compositional heterogeneities have an impor- the lack of band homology than data sets displaying more tant impact on the amount of nonhomologous bands, closely related genotypes, as those seen in recent radiations genomic coverage, and information content of AFLP data sets (Young et al. 1999; Peng et al. 2000; Campbell and One of the primary motivations of this study was to Bernatchez 2004; Caballero and Quesada 2010). Thus, quantify the impact of the lack of homology among AFLP our estimates of reconstruction accuracy must be seen bands on phylogenetic accuracy. To this end, we compared as best case values.
the relative accuracies of trees based on experimental We note that although our in silico simulations did not bands (susceptible to be nonhomologous) and trees based use selective nucleotides, in practice, selective nucleotides on the corresponding true fragments (where nonhomology are always employed in experimental AFLP studies. How- is nonexistent) (figs. 4 and 5). As expected, AFLP-based ever, the role of selective nucleotides is just to sample a ran- trees inferred from experimental bands had a generally dom subset of restriction fragments to get a reduced lower accuracy than AFLP-based trees inferred from true number of bands (90–100) per AFLP profile (Vos et al.
fragments. However, the difference in efficiency due to 1995). The number of bands containing nonhomologous the lack of band homology was surprisingly small compared comigrating fragments within an AFLP profile increases with the much more reliable trees obtained when using rapidly with the number of bands (Gort et al. 2006) and DNA sequences.
is independent of the genome size for a given GC content The dramatically lower performance of AFLP-based trees (Caballero and Quesada 2010). This implies that for a similar with respect to DNA-based trees is likely the consequence in silico setting, results with and without selective nucleo- of the much higher information content and lower nucle- tides would be similar provided that the number of bands otide sampling error of DNA sequences (Kumar and per AFLP profile is the same. This was corroborated by per- Gadagkar 2000). Indeed, increasing the number of AFLP forming additional in silico AFLP runs using one combina- markers would still result in a poor performance, as indi- tion of selective nucleotides (A for both restriction sites) cated by the more than 1,000 true fragments scored at a di- under different radiation models, tree topologies, and evo- vergence threshold of 0.4 (fig. 4). The difference in lutionary divergences. For fair comparisons, the genome performance between DNA sequences and AFLPs did size in these simulations was increased to 12 Mb in order not change substantially even when the length of DNA se- to get a number of bands (90–100) per AFLP profile similar quences used to construct sequence-based trees was set to to that observed without selective nucleotides. As ex- a number of nucleotides (700) similar to the number of pected, the accuracy of phylogenetic reconstruction using AFLP fragments scored across each evolutionary divergence selective nucleotides was very similar to that without selec- (not shown). This indicates that the poorer performance of tive nucleotides Supplementary AFLP-based trees is not the result of sampling a much lower Material online).
number of informative AFLP fragments than nucleotides.
Other commonly invoked limitations of AFLP data sets, Impact of the Lack of Band Homology on such as the asymmetry of loosing and gaining fragments, Phylogenetic Accuracy the dominant nature of AFLP characters, or the possible One basic assumption of AFLP-based phylogenies is that all nonindependence of fragments, do not appear to have comigrating bands are homologous or that only a very an effect as dramatic on phylogenetic accuracy as that seen small proportion of bands are nonhomologous (Koopman here (Koopman 2005; Althoff et al. 2007; Simmons et al.
2005). In the latter case, it is assumed that the phylogenetic 2007). Thus, the lower information content of AFLP data signal from homologous bands would overcome the phy- sets is a factor with a much higher negative impact on phy- logenetic noise from nonhomologous bands (Bussel et al.
logenetic accuracy than the lack of band homology itself.
2005). Several studies, however, have pointed out that the Consequently, even in experiments where all fragments lack of band homology could be more prevalent than orig- could be individually characterized by sequencing, it would inally thought and that it is especially problematic for dis- not be possible to substantially improve the reliability of tantly related taxa (O'Hanlon and Peakall 2000; Vekemans AFLP-based trees. This is a challenging finding with respect et al. 2002; Mechanda et al. 2004; Althoff et al. 2007; to the traditional view that the lack of band homology is Caballero and Quesada 2010). Nevertheless, because taxo- the most important drawback of AFLP markers regarding nomic rank is not comparable among taxa, it is difficult phylogenetic accuracy (O'Hanlon and Peakall 2000; Bussel to interpret the general relevance of these results. Our et al. 2005; Althoff et al. 2007).
Garcı´a-Pereira et al. · doi:10.1093/molbev/msp315 In silico AFLP studies have shown that the proportion of the utility of AFLP markers for phylogenetic reconstruction nonhomologous bands increases linearly with the total under certain evolutionary scenarios, despite the consis- number of fragments scored, which in turn rely on the tently poorer information content of AFLP data sets with GC content (Caballero and Quesada 2010). Here we show respect to DNA sequences.
that the lack of homology determines a distinct impact ofGC content on the accuracy of trees inferred from exper- Differences in Performance among Phylogenetic imental bands and true fragments (fig. 7). As GC content increases, EcoRI and MseI cuts are less frequent because of We did not find clear-cut differences in the performance of their AT-biased recognition motifs. This results in a lower the three phylogenetic methods commonly used in AFLP number of fragments and a lower probability of comigra- data sets. Distance methods (NJ and ME) had similar recon- tion (Vekemans et al. 2002), which increases the accuracy struction efficiencies, likely as a consequence of NJ being of trees inferred from experimental bands (fig. 7). In con- a special case of ME (Rzhetsky and Nei 1992). When tree trast, because true fragments are by definition always ho- topologies are symmetrical, all tree-building methods pro- mologous, increased GC content results in a decrease of duced similar trends, with distance methods performing the number of informative homologous characters. This better with some radiation models at large distances in turn leads to less reliable trees. These findings have (fig. 3b). Distance methods outperform MP when tree to- two significant implications. First, the proportion of ho- pologies are asymmetrical (fig. 3a). However, undefined dis- mologous bands and the concomitant phylogenetic ac- tances may seriously decrease the relative efficiency of curacy depends on a species-specific feature such as GC distance versus MP methods. This occurs when the average content, which is usually unknown in nonmodel organ- evolutionary divergence is large (figs. 4b, 5b, and 6), or some isms. Second, the comparison among species differing in distantly related genotypes are present in the data set GC content will determine differential lineage-specific (Koopman and Gort 2004). This problem has a higher im- impacts of the lack of band homology, which will pact on true fragments than on experimental bands. This is certainly compromise the accuracy of phylogenetic because the comigration of fragments in experimental bands increases the similarity among taxa (Althoff et al.
Experimental approaches have been developed to re- 2007), making less likely the generation of undefined dis- duce the impact of the lack of band homology in AFLP- tances. Our results show that when this problem occurs, it based trees. Dasmahapatra et al. (2009) have recently is useful to use parsimony (fig. 4a). This will usually exclude proposed a protocol that minimizes the number of nonho- problematic characters that are very variable among taxa mologous AFLP bands in experimental studies by excluding by using only those characters that are parsimoniously those bands that are suspicious to be nonhomologous informative (Lockhart et al. 1994).
across species. This approach increases the phylogenetic Simmons et al. (2007) similarly noted that MP per- signal of experimental AFLP data sets and provides results formed comparably to ME in symmetric trees but that similar to those simulated here for true fragments, in which MP outperforms ME in asymmetric trees. Two reasons ex- nonhomology is nonexistent. This allowed them to con- plain this discrepancy with respect to our results in asym- struct an AFLP-based phylogeny of seals in which AFLP metric trees: 1) Their conclusions only apply to tree fragments maintained phylogenetic signal over a timescale topology and not to both topology and branch length of ;15 my (Dasmahapatra et al. 2009). The AFLP tree con- as is the case here and 2) the distribution of internode dis- structed for seals resembles the most favorable scenario for tances was not preserved by reducing the number of taxa in AFLP-based trees in terms of tree topology, that is, a tree asymmetric trees, thus making asymmetric trees compar- with recent radiation. As shown in our Figure 5, under re- atively much more difficult to reconstruct. Although we cent radiation scenarios, the topology of AFLP trees is well cannot overemphasize the lack of resolution provided preserved, even for large divergences. This agrees with the by undefined distances, these results do not permit making fact that the AFLP phylogeny of seals displays a topology any generalized statement about the relative accuracies of closely resembling those based on both nuclear (Fulton and character-based MP versus distance-based (ME and NJ) Strobeck 2006) and mitochondrial DNA (mtDNA) sequen- methods for building AFLP-based trees. However, the un- ces (Arnason et al. 2006). Even so, there are still some dis- certainty in deciding which phylogenetic method to use is crepancies with respect to the interspecific relationships of generally small relative to the uncertainty arising from some taxa. Thus, up to three taxa (Callorhinus ursinus, Pho- factors such as divergence, nonhomology of bands or ca hispida, and Ommatophoca rossi) display an incorrect the low information content of AFLP markers. Whatever interspecific relationship with respect to the nuclear phy- methods are used, robust interpretations of reconstructed logeny. However, the relative branch lengths of the recon- AFLP-based trees will require the confirmation by an structed AFLP phylogeny of seals are only rarely independent source of data.
concordant with those from the nuclear and mtDNA trees.
This agrees with our observation in Figure 4 that recon- Supplementary Material structions from AFLP trees are largely wrong in termsof branch lengths for recent radiation. The data from is available at Molecular Biology and Dasmahapatra et al. (2009), however, further emphasize Evolution online Phylogenetic Accuracy of AFLPs · doi:10.1093/molbev/msp315 Gort G, Koopman WJM, Stein A. 2006. Fragment length distributions and collision probabilities for AFLP markers.
We thank Jose´ Castresana and two anonymous referees for useful comments on the manuscript. This work was funded Hansen M, Kraft T, Christiansen M, Nilsson N-O. 1999. Evaluation of by the Ministerio de Ciencia e Innovacio ´n and Fondos Feder AFLP in Beta. Theor Appl Genet. 98:845–852.
(CGL2006-13445-CO2/BOS and CGL2009-13278-CO2) and Hasegawa M, Kishino H, Yano T. 1985. Dating the human-ape Xunta de Galicia.
splitting by a molecular clock of mitochondrial DNA. J Mol Evol.
22:160–174.
Holland B, Clarke AC, Meudth HM. 2008. Optimizing automated AFLP scoring parameters to improve phylogenetic resolution.
Syst Biol. 57:347–366.
Althoff DM, Gitzendanner MA, Segraves KA. 2007. The utility of Hollingsworth PM, Ennos RA. 2004. Neighbor joining trees, amplified fragment length polymorphisms in phylogenetics: dominant markers and population genetic structure. Heredity a comparison of homology within and between genomes. Syst Biol. 56:477–484.
Johns GC, Avise JC. 1998. A comparative summary of genetic Arnason U, Gullberg A, Janke A, Kullberg M, Lehman N, Petrov EA, distances in the vertebrates from mitochondrial cytochrome ¨la¨ R. 2006. Pinniped phylogeny and a new hypothesis for b gene. Mol Biol Evol. 15:1481–1490.
their origin and dispersal. Mol Phylogenet Evol. 41:345–354.
Jukes TH, Cantor CR. 1969. Evolution of protein molecules. Vol. 3.
˚ kesson M. 2005. Ten years of AFLP in ecology and Mammalian protein metabolism. New York: Academic Press. p.
evolution: why so few animals? Mol Ecol. 14:2899–2914.
Brouat C, McKey D, Douzery EJP. 2004. Differentiation in a geo- Karlin S, Burge C. 1995. Dinucleotide relative abundance extremes: graphical mosaic of plants coevolving with ants: phylogeny of a genomic signature. Trends Genet. 11:283–290.
the Leonardoxa africana complex (Fabaceae: Caesalpinioideae) Karp A, Seberg A, Buiatti M. 1996. Molecular techniques in the using amplified fragment length polymorphism markers. Mol assessment of botanical diversity. Ann Bot. 78:143–149.
¨ller S, Duftner N, Sefc KM, Aibara M, Stipacek M, Blanc M, Bussel JD, Waycott M, Chappill JA. 2005. Arbitrarily amplified DNA Egger B, Sturmbauer C. 2007. Reticulate phylogeny of gastropod- markers as characters for phylogenetic inference. Perspect Plant shell-breeding cichlids from Lake Tanganyika—the result of Ecol Evol Syst. 7:3–26.
repeated introgressive hybridization. BMC Evol Biol. 7:7.
Caballero A, Quesada H. 2010. Homoplasy and distribution of AFLP Koopman WJM. 2005. Phylogenetic signal in AFLP data sets. Syst fragments: an analysis in silico of the genome of different species.
Biol. 54:197–217.
Mol Biol Evol. Advance Access published January 7, 2010, Koopman WJM, Gort G. 2004. Significance tests and weighted values for AFLP similarities, based on Arabidopsis in Silico AFLP Caballero A, Quesada H, Rola´n-Alvarez E. 2008. Impact of AFLP fragment length distributions. Genetics 167:1915–1928.
fragment size homoplasy on the estimation of population Kosman E, Leonard KJ. 2005. Similarity coefficients for molecular genetic diversity and the detection of selective loci. Genetics markers in studies of genetic relationships between individuals for haploid, diploid, and polyploid species. Mol Ecol. 14:415–424.
Campbell D, Bernatchez L. 2004. Generic scan using AFLP markers as Kumar S, Gadagkar SR. 2000. Efficiency of the neighbor-joining a means to assess the role of directional selection in the divergence method in reconstructing deep and shallow evolutionary of sympatric whitefish ecotypes. Mol Biol Evol. 21:945–956.
relationships in large phylogenies. J Mol Evol. 51:544–553.
Cantarel BL, Morrison HG, Pearson W. 2006. Exploring the Lockhart PJ, Steel MJ, Hendy MD, Penny D. 1994. Recovering relationship between sequence similarity and accurate phyloge- evolutionary trees under a more realistic model of sequence netic trees. Mol Biol Evol. 23:2090–2100.
evolution. Mol Biol Evol. 11:605–612.
Chen F-C, Li W-H. 2001. Genomic divergences between humans and Luo R, Hipp AL, Larget B. 2007. A bayesian model of AFLP marker other homininds and the effective population size of the evolution and phylogenetic inference. Stat Appl Genet Mol Biol.
common ancestor of humans and chimpanzees. Am J Hum Marhold K, Lihova´ J, Perny´ M, Bleeker W. 2004. Comparative ITS and Dasmahapatra KK, Hoffman JI, Amos W. 2009. Pinniped phyloge- AFLP analysis of diploid Cardamine (Brassicaceae) taxa from netic relationships inferred using AFLP markers. Heredity closely related polyploid complexes. Ann Bot. 93:507–520.
Mechanda SM, Baum BR, Johnson DA, Arnason JT. 2004. Sequence Despres L, Gielly L, Redoutet B, Taberlet P. 2003. Using AFLP to assessment of comigrating AFLPä bands in Echinacea – resolve phylogenetic relationships in a morphologically di- implications for comparative biological studies. Genome 47: versified plant species complex when nuclear and chloroplast sequences fail to reveal variability. Mol Phylogenet Evol.
Meudt HM, Clarke AC. 2007. Almost forgotten or lattest practice? AFLP applications, analyses and advances. Trends Plant Sci.
El-Rabey HA, Badr A, Schaffer-Pregl R, Martin W, Salamini F. 2002.
Speciation and species separation in Hordeum L. (Poaceae) Mueller UG, Wolfenbarger LL. 1999. AFLP genotyping and resolved by discontinuous molecular markers. Plant Biol.
fingerprinting. Trends Ecol Evol. 14:389–394.
Nei M, Li W-H. 1979. Mathematical model for studying genetic Fulton TL, Strobeck C. 2006. Molecular phylogeny of the Arctoidea variation in terms of restriction endonucleases. Proc Natl Acad (Carnivora): effect of missing data on supertree and supermatrix Sci U S A. 76:5269–5273.
analyses of multiple gene data sets. Mol Phylogenet Evol.
O'Hanlon PC, Peakall R. 2000. A simple method for the detection of size homoplasy among amplified fragment length polymor- Giannasi N, Thorpe RS, Malhotra A. 2001. The use of amplified phism fragments. Mol Ecol. 9:815–816.
fragment length polymorphim in determining species trees at Pellmyr O, Segraves KA, Althoff DM, Balca´zar-Lara M, Leebens- fine taxonomic levels: analysis of a medically important snake, Mack J. 2007. The phylogeny of yuccas. Mol Phylogenet Evol.
Trimeresurus albolabris. Mol Ecol. 10:419–426.
Garcı´a-Pereira et al. · doi:10.1093/molbev/msp315 Peng J, Korol AB, Fahima T, Ro¨der MS, Ronin YI, Li YC, Nevo E. 2000.
Amplified-fragment length polymorphism analysis: the state of Molecular genetic maps in wild Emmer wheat, Triticum the art. J Clin Microbiol. 37:3083–3091.
dicoccoides: genome-wide coverage, massive negative interfer- Simmons MP, Zhang L-B, Webb CT, Mu ¨ller K. 2007. A penalty of ence, and putative quasi-linkage. Genome Res. 10:1509–1531.
using anonymous dominant markers (AFLPs, ISSRs, and RAPDs) Peters JL, Constandt H, Ney P, Cnops G, Zethof J, Zabeau M, for phylogenetic inference. Mol Mol Phylogenet Evol. 42:528–542.
Gerats T. 2001. A physical amplified fragment-length poly- Sites JW, Marshall JC. 2003. Delimiting species: a renaissance issue in morphism map of Arabidopsis. Plant Physiol. 127:1579–1589.
systematic biology. Trends Ecol Evol. 18:462–470.
Pompanon F, Bonin A, Bellemain E, Taberlet P. 2005. Genotyping Soria-Carrasco V, Talavera G, Igea J, Castresana J. 2007. The K tree errors: causes, consequences and solutions. Nat Rev Genet.
score: quantification of differences in the relative branch length and topology of phylogenetic trees. Bioinformatics 23:2954–2956.
Rambaut A, Grassly NC. 1997. Seq-Gen: an application for the Sullivan JP, Lavoue´ S, Arnegard ME, Hopkins CD. 2004. AFLPs resolve Monte Carlo simulation of DNA sequence evolution along phylogeny and reveal mitochondrial introgression within phylogenetic trees. Comput Appl Biosci. 13:235–238.
a species flock of African electric fish (Momyroidea: teleostei).
Robinson DF, Foulds LR. 1981. Comparison of phylogenetic trees.
Math Biosci. 53:131–147.
Swofford DL. 2002. PAUP*: phylogenetic analysis using parsimony Rombauts S, Van der Peer Y, Rouze´ P. 2003. AFLP in Silico, (*and other methods), version 4. Sunderland (MA): Sinauer simulating AFLP fingerprints. Bioinformatics 19:776–777.
Rouppe van der Voort JN, van Zandvoort P, van Eck HJ, Vekemans X, Beauwens T, Lemaire M, Rolda´n-Ruiz I. 2002. Data from amplified fragment length polymorphism (AFLP) markers Jacobsen E, Helder J, Bakker J. 1997. Use of allele specificity of show indication of size homoplasy and of a relationship between comigrating AFLP markers to align genetic maps from different degree of homoplasy and fragment size. Mol Ecol. 11:139–151.
potato genotypes. Mol Gen Genet. 255:438–447.
Vos P, Hogers R, Bleeker M, Reijans M, Van de Lee, Hornes M, Rzhetsky A, Nei M. 1992. Statistical properties of the ordinary least- Frijters A, Pot J, Peleman J, Kuiper M. 1995. AFLP: a new technique squares, generalized least-squares and minimum-evolution for DNA fingerprinting. Nucleic Acids Res. 23:4407–4414.
methods of phylogenetic inference. J Mol Evol. 35:367–375.
Young WP, Schupp JM, Keim P. 1999. DNA methylation and AFLP Savelkoul PHM, Aarts HJM, de Haas J, Dijkshoorn L, Duim B, marker distribution in the soybean genome. Theor Appl Genet.
Otsen M, Rademaker JLW, Schouls L, Lenstra JA. 1999.

Source: http://webc03.webs2.uvigo.es/XENETICA/XB2/pubs/par055.pdf