Introduction

Species are the fundamental unit of biodiversity (Mayr 1982); therefore, proper species delimitation and identification are important prerequisites to population genetic, physiological, and ecological studies (Wiens 2007; Butlin et al. 2009). However, species–level taxonomy is rife with practical issues, especially for groups whose morphology is uninformative, plastic, and/or difficult to describe (Hebert et al. 2003). This contrasts with the relative ease with which conspicuously divergent taxa can be recognized, as is often the case when dealing with large terrestrial organisms such as birds, mammals, and butterflies (Gaston and Blackburn 2000). In contrary, marine invertebrate species are notoriously difficult to delineate; extreme cases of difficulties are represented, for instance, by corals that are morphologically plastic in response to variations in environmental conditions (Todd 2008), by assemblages of planktonic larval stages (Baretta-Bekker et al. 1998), or by microscopic meiofaunal organisms in the sediments (Curini-Galletti et al. 2012).

Even in well-studied and familiar organisms, questions arise regarding the diagnosis and interpretation of morphological and/or genetic divergence as intraspecific or interspecific (Hey 2009). The situation is complicated further by the variety of alternative theoretical approaches for defining species boundaries. Using different species delimitation criteria (de Queiroz 2007) and metrics (e.g., Tang et al. 2012) can produce contrasting assessments of diversity: for example, strict phylogenetic delimitation criteria may be more prone to overestimate the number of species compared to gene-flow–based biological criteria that put emphasis on reproductive isolation (or lack thereof) as the defining property of species (Flot et al. 2010, 2011).

In addition to the ambiguities inherent to the interpretation of biological phenomena such as species, another problem is the so-called taxonomic impediment, i.e., the gap between the small number of expert taxonomists and the large number of species to describe and specimens to identify (Rodman and Cody 2003). The taxonomic impediment is especially pervasive for the meiofauna (Giere 2009), which is composed of small animals with high levels of cryptic diversity and frequent morphological stasis (e.g., Fontaneto et al. 2009). The scarcity of available taxonomic expertise is one of the reasons for the abundant undescribed diversity that is characteristic of most meiofaunal taxa (Curini-Galletti et al. 2012; Fonseca et al. 2014). Furthermore, meiofaunal organisms may inhabit areas that are difficult to sample, such as the deep or open sea, or the sediment of remote caves. Getting a better understanding of meiofaunal diversity is especially important when studying marine benthic environments, where meiofaunal organisms play key ecological roles (Zeppilli et al. 2015). Thus, we will focus our review on the methods of DNA taxonomy applicable to meiofaunal species, particularly those inhabiting marine environments. Our hope is to help systematists and non-systematists alike use DNA information to obtain reliable data on the underexplored but highly varied group of organisms that compose the meiofauna. However, the methods we outline are not specific to the meiofauna and will be useful to a broader audience.

DNA taxonomy techniques offer taxonomists and ecologists fast and objective means to assess biodiversity. Here we define DNA taxonomy as the analysis of variation in genetic data (such as DNA sequences of selected loci or complete genomes, single-nucleotide polymorphisms, microsatellites, amplified fragment length polymorphisms, etc.) to delimit species (Tautz et al. 2003). These types of studies are now more accessible than ever, and numerous sequence-based approaches have been proposed to inform species diagnosis, mostly thanks to the increasing abundance of molecular data (McCormack et al. 2013), to the existence of large sequence data sets (e.g., BOLD; Ratnasingham and Hebert 2007), and to the rise of quantitative phylogenetic methods (Sites and Marshall 2004; Carstens et al. 2013).

For animals, most studies start by (and many rely solely on) sequencing one marker of suitable variability (e.g., the “barcode” portion of the mitochondrial cytochrome oxidase c subunit I gene, or the internal transcribed spacer 2 located in nuclear ribosomal DNA) to delineate species, yielding species hypotheses that can be compared to morphology, ecology, cross-mating experiments, physiology, distribution, and behaviour. However, it is increasingly recognised that accurate species delineation requires a multilocus approach taking into account two or more independent markers from the same individuals, which necessarily involves sequencing nuclear genes (since all mitochondrial markers are linked together and cannot be considered independent sources of information; Moore 1995). Sequencing nuclear markers is often considered difficult and costly because of their heterozygosis: hence, multilocus species delineation methods are most often used when single-locus approaches yield results that are ambiguous or inconsistent with morphology or with the other sources of information mentioned above. Multilocus approaches to species delimitation is an active field of research and will undoubtedly predominate in future studies, given the ever-decreasing per-base cost of sequencing and the ease with which such approaches should be scalable to entire genome sequences. Although most available multilocus species delineation methods are still experimentally and computationally expensive (and are therefore difficult to apply to studies of meiofaunal species), we include them in our review, albeit briefly, given their great potential.

Our review considers all the steps in DNA taxonomy, starting with data acquisition before detailing the various single-locus species delimitation methods available (Fig. 1). We then consider succinctly the more refined multilocus approaches, and conclude our review by emphasising some potential caveats and pitfalls. The various methods attempt to optimise different criteria for identifying and delimiting species: given that our review aims to provide guidelines for DNA taxonomy, we do not dwell on the differences in philosophy between these approaches, nor on their statistical properties; neither do we consider the confounding effect of interspecific gene flow, nor the possible differences between sexual and asexual organisms. Instead, we use a pragmatic approach, highlighting the previous usage of these methods in meiofaunal studies, their strengths, and their potential weaknesses.

Fig. 1
figure 1

Typical DNA taxonomy workflow. First row from top, laboratory procedures: (1) animals are isolated in the wild and identified to morphospecies level under light microscopy; (2) these animals are washed in double-distilled H20 and transferred to individual tubes in which the DNA is extracted. Second and third rows, data acquisition: (3) specific genetic loci are PCR-amplified, Sanger-sequenced, phased to isolate individual haplotypes (in the case of nuclear loci), and aligned. The two asterisks denote ambiguous, potentially heterozygous, base pairs (see Fig. 2). (4a) The alignment is used to construct a pairwise distance matrix and/or (4b) a phylogenetic tree and/or (4c) a haplotype network. Note that phylogenetic trees can be constructed either from sequence alignments or from distance matrices, whereas it is also possible to compute a patristic distance matrix from the branch lengths of a tree. Bottom rows, data processing: (5) the matrix and/or phylogeny and/or network serve as a basis for various species delimitation approaches, either based on distances (left, DNA barcoding and ABGD), on branching rates (centre, K/θ, GMYC and PTP), or on heterozygosis (right, haplowebs). In the latter case, the curves drawn on the right side of the tree connect sequences found co-occurring in heterozygous individuals

Full size image

Data acquisition

Sampling and DNA extraction

Data acquisition is the crucial first step in any species delimitation study; hence, careful consideration is required when designing a sampling scheme, as this choice can have a strong influence on the number of species delimited (Papadopoulou et al. 2008; Lohse 2009; Bergsten et al. 2012; Talavera et al. 2013). Most DNA taxonomy methods implicitly or explicitly assume that all populations and species are adequately sampled (Lim et al. 2012). In the traditional approach, when specimens are sorted and processed individually, it is advisable to plan a redundant sampling in which each species is represented by multiple specimens, if possible collected from different locations. Collecting large numbers of specimens is rarely a problem for meiofaunal species living in undisturbed marine littoral sediments, given their small size and abundance: non-selective methods such as aspiration of sediments followed by filtration usually yield numerous specimens, but specific methods targeting different meiofaunal groups are also available (Giere 2009; Curini-Galletti et al. 2012). Yet, given the uneven distribution of species abundances in the field, often characterised by a small number of common species and a very high number of rare species (Magurran and Henderson 2003), some clades are likely to be over-represented in any sampling strategy. This is even more problematic for rare species living in peculiar habitats such as marine caves and crevices or sediments in oligotrophic streams: collecting them may pose specific challenges, in which case a balanced sampling scheme is particularly difficult to achieve. Many different extraction methods are available to extract a reliable quantity of DNA from single individuals of the meiofauna, and the resulting DNA can be stored for years at −20 °C; however, when large numbers of minute individuals have to be processed, it is advisable to choose quick and inexpensive methods such as the HotSHOT protocol (Truett et al. 2000; Montero-Pau et al. 2008) or a combination of Chelex and proteinase K (Estoup et al. 1996), which work well even with an amount of material as small as a single rotifer egg (Montero-Pau et al. 2008).

In contrast with the traditional, specimen-per-specimen approach outlined above, an alternative that is increasingly being used is to collect samples of sediment and process them in bulk, without sorting the animals into single tubes (Creer et al. 2010; Fonseca et al. 2010, 2014). With this second approach, called “metagenetics” or “metabarcoding” (Taberlet et al. 2012), standardisation and planning of a balanced sampling design for a targeted group of species or species complexes is impossible, and no morphological information can be linked to each individual. Whereas the first method is most useful for detailed taxonomic analyses, the second one can provide quantitative genetic information suitable for species identification (provided that a reference sequence database is available) and ecological studies. Besides, analysing bulk environmental samples makes it possible to detect efficiently rare species and study their distribution (Zhan and MacIsaac 2015). Various kits are available to perform bulk DNA extraction from water or sediments, and the choice of a particular approach should be considered carefully, since different methods have been shown to yield different overall estimates of biological diversity (Knauth et al. 2013; Rees et al. 2014; Deiner et al. 2015). When the biomass of the meiofauna is very low compared to the sampled volume of soil, sediment, or water to be processed, animals are commonly isolated from their environment before pooling them and extracting bulk DNA (Creer et al. 2010; Fonseca et al. 2010, 2014).

Choice of marker(s)

Once an appropriate sampling and DNA extraction strategy has been chosen, the next important step is to decide which marker(s) will be analysed. The basic and most important feature of a marker suitable for DNA taxonomy is its variability, but other properties have to be considered as well, notably a marker’s propensity to convergent evolution. Although microsatellite markers are extremely variable, they are not the most suited for delimiting species because their number of repeats is always comprised between zero and an upper boundary of at most a few tens of repeats, meaning that two individuals may present the same number of repeats by convergence instead of by descent (Garza et al. 1995; Garza and Freimer 1996). The same convergence problem plagues single-nucleotide polymorphisms (SNPs), as they can only take one out of four possible states: A, C, G and T (or even only two states for diallelic SNPs). Convergence is also frequently observed with approaches that are based on the presence/absence of bands of specific lengths, such as randomly amplified polymorphic DNAs (RAPDs), single-strand conformation polymorphisms (SSCPs), and amplified fragment length polymorphisms (AFLPs). By contrast, the probability of obtaining two identical DNA sequences of several hundred base pairs by random convergence instead of by descent is so small that it is practically negligible (Tajima 1983), which is why most DNA taxonomy studies use sequence data to delineate species. Besides, DNA sequences from previous studies can easily be obtained from GenBank (Benson et al. 2013) and reused, which greatly improves the strength of approaches based on sequences compared with other types of genetic data.

Another important property of a marker suited for DNA taxonomy is its universality: markers that are only available for a subset of species are not particularly useful when dealing with the various taxa that make up the meiofauna. Therefore, although genomes comprise thousands of genes and lots of intergenic regions that can potentially be used as markers for DNA taxonomy, it is important that the variable region be flanked by two conserved ones that can be used as priming regions to ultimately amplify and sequence the marker by the polymerase chain reaction (PCR; Saiki et al. 1988). Introns of single-copy genes are ideal from this point of view, since their sequences experience little structural constraints and the exons on each of their sides are much more conserved, making it generally possible to design excellent PCR primers. Such markers are often called EPIC (exon-primed intron-crossing) in the literature (Palumbi and Baker 1994), and can be defined with increasing ease now that complete genome sequences are available for many groups. However, given that meiofaunal species are small and therefore contain little DNA per specimen, single-copy gene introns can be difficult to amplify consistently (notably if the DNA extracts contain inhibitors; Rameckers et al. 1997). This is why single-copy gene introns are rarely a first choice in DNA taxonomy studies dealing with the meiofauna; instead, most studies use markers that are present in many copies per cell, as part of either mitochondrial or ribosomal DNA.

The most commonly used marker for the DNA taxonomy of animals is a fragment of the cytochrome c oxidase subunit 1 gene (abbreviated variously as cox1 or COI), for which universal metazoan primers are available (Folmer et al. 1994) and for which other primers targeting specific groups are frequently redeveloped (e.g., Prosser et al. 2013 for freshwater microcrustaceans). This is a very variable marker in bilaterians, able to resolve not only species, but also populations within a given species (Avise et al. 1987). However, COI shows much less variation among non-bilaterian metazoans such as cnidarians (Shearer et al. 2002; Hellberg 2006; Huang et al. 2008). For these metazoans, the entire mitochondrial genome appears to be very stable: in a particularly striking example, populations of the deep-sea coral Lophelia pertusa collected 7,500 km apart in different oceanic basins (the Mediterranean Sea vs. the Barents Sea) were found to share near-identical complete mitochondrial genome sequences that differed by at most a single nucleotide position (Flot et al. 2013). As a result of this stability, COI sequences are not very useful for distinguishing cnidarians species (Shearer and Coffroth 2008), although other mitochondrial regions variable enough for this purpose have been found in some groups (Pont-Kingdon et al. 1995; Flot and Tillier 2007). Similar challenges are likely to occur in meiofaunal cnidarians such as the interstitial hydrozoan genus Halammohydra, which is prominently displayed on the logo of the International Association of Meiobenthologists, but may concern other meiofaunal groups as well.

Ribosomal DNA markers are also frequently used to delineate species. Like COI, ribosomal DNA is present in many copies per cell and can therefore be readily amplified using PCR. These copies generally evolve in a concerted fashion (Dover 1982; Hillis et al. 1991) that prevents them from diverging; as a result, although sequencing cloned PCR products of ribosomal DNA markers may yield a bewildering variety of sequences, it has been found that direct sequencing of the ITS (intergenic transcribed spacer), for example, tends to yield only one or two dominant types per specimen, rarely three (Flot et al. 2006). Some studies used a variable region of the 28S ribosomal DNA gene (e.g., Lorion et al. 2010) for which semi-universal primers are available (e.g., Verovnik et al. 2005 for isopods and amphipods), but this region is not always variable enough to distinguish closely related species; for instance, the two amphipod species Pontoniphargus racovitzai and Pontoniphargus ruffoi are morphologically distinct and reciprocally monophyletic for COI, but share the same 28S sequence (Flot et al. 2014). Hence, when sequencing a single marker, it may be preferable to target the internal transcribed spacers ITS1 and/or ITS2, located between the 18S and 28S genes in ribosomal DNA. These genes are much more variable than 28S and therefore more suitable to distinguish closely related species, although they can be difficult to amplify in some groups in which their sequences are very long; for example in niphargid amphipods, with a record length of 1159 bp for Niphargus plateaui, the longest ITS2 sequence ever reported for a metazoan (Kornobis and Pálsson 2013).

Sequencing

Sequencing strategies targeting a single marker or a handful of independent markers usually start with amplification of the target region(s) using PCR. Optimised PCR protocols have been developed for most taxa, the breadth of which is outside of the scope of this review; for specific protocols, it is best to consult the literature available for each group. We provide references to meiofaunal studies in the following paragraphs, and the same papers often report protocols on how the sequences were obtained. These protocols sometimes require some level of troubleshooting, with common approaches to consider. A common source of failure, notably when pooling individuals, is an excess of material during the DNA extraction step, which may result in the co-purification of various contaminants that hinder downstream PCR amplification. To avoid this problem, one should decrease the amount of material used for DNA extractions (it is perfectly fine if the amount of DNA recovered is so small that it is undetectable by spectrophotometry). For instance, when dealing with minute amphipods, it is possible to routinely use only one or two legs for DNA extractions (Flot 2010a; Flot et al. 2010); to the extreme, successful amplification of multiple markers can be obtained even from single microscopic individuals such as unicellular eukaryotes (Blin and Stafford 1976). A second important thing to check is whether the PCR amplification buffer contains dimethylsulfoxide (DMSO). Adding a final concentration of 5 % DMSO to PCR mixes helps amplifying DNA regions by preventing them from folding into secondary structures (Winship 1989), as ribosomal DNA and introns typically do. A third action to take if results are inconsistent across individuals is to increase the number of PCR cycles, to take into account the reduced PCR efficiency resulting from co-purified contaminants or DNA fragmentation (Rameckers et al. 1997): enduring 65 cycles of denaturation-annealing-elongation poses no problem to modern Taq polymerases. Even though increasing the number of cycles may result in high-molecular weight products visible as smears on gels (Bell and DeMarini 1991), this does negatively impact the Sanger-sequencing of these products. Finally, a failure to obtain PCR products in some or all individuals can be due to the Taq polymerase being too stringent, in which case trying a different, less stringent Taq polymerase (e.g., BioTaq or QbioTaq) may solve the problem, especially when using degenerate primers. Conversely, to get rid of multiple bands or smears, one should try a more stringent Taq (e.g., RedTaq). If all this fails, it probably means that the primers found in the literature are suboptimal, in which case it may be necessary to design new ones; an excellent tool to do so, yielding robust primers pairs that do not require further optimisation, is Primer3 (Rozen and Skaletsky 2000), which is freely available as an online web tool. Otherwise, the strategies listed in Roux (2009) can be used to get suboptimal primers to work. Another problem that can be encountered when sequencing mitochondrial markers in animals is the presence of nuclear mitochondrial paralogs, called pseudogenes or numts, an abbreviation for “nuclear mitochondrial DNA” (Richly and Leister 2004). These extra copies are often co-amplified with the mitochondrial marker, or may even be the only copies that are amplified in some cases, creating problems if they are not recognised as paralogous to the other sequences in the data set (Song et al. 2008). Discarding a posteriori sequences likely to be pseudogenes, for instance because of stop codons or an excess of non-synonymous mutations, can alleviate the problem, but some numts may remain undetected. Thus, the most reliable solution is to take measures such as pre-PCR dilution to avoid amplifying numts in the first place, as suggested by Calvignac et al. (2011).

Most DNA taxonomy projects adopt the Sanger technology (primer elongation using a mixture of deoxynucleotides and dideoxynucleotide terminators followed by electrophoresis; Sanger et al. 1977) to sequence PCR products. An important good practice is to systematically check the chromatograms obtained from Sanger sequencing and not to blindly use the FASTA files provided with the chromatograms. Automated base-calling programs often make mistakes, such as miscalling some bases, overlooking others and including artefactual ones. If such mistakes are not detected but instead are carried on into the downstream analyses, they will inevitably affect any further inference and may bias the results. For instance, Collins and Cruickshank (2014) report how DNA barcode sequences that contain stretches of incorrect bases because of common sequencing artefacts (“dye blobs”) are classified incorrectly. The best, standard way to avoid such problems is to systematically sequence each marker using several primers, so that each base is covered by at least two different chromatograms, ideally in different directions. Aligning the forward and reverse chromatograms together makes it possible to detect and correct errors automatically, in a very time-efficient way and without having to visually inspect all the bases. An open-source program for aligning forward and reverse chromatograms and checking for discrepancies between the two is SeqTrace (Stucky 2012), available online at http://seqtrace.googlecode.com/. Although sequencing PCR primers in both directions inevitably inflates the cost of DNA taxonomy, it is the best way to obtain sequences that are reliable and accurate. Spending extra time and money at the beginning of the process in order to obtain accurate sequences greatly diminishes the chances of carrying mistakes through the later stages of the analyses, thereby saving time and money overall.

As an alternative to traditional Sanger sequencing approaches, recent advances in sequencing technologies, commonly grouped under the terms next-generation sequencing (NGS) or high-throughput sequencing (HTS), have been revolutionizing DNA studies for the last decade (Kircher and Kelso 2010). These approaches can be used to analyse the complete genome of one individual at a time, but also to sequence PCR products for a fraction of the cost of traditional Sanger sequencing (Bik et al. 2012). In the most basic approach, PCR products are pooled together for sequencing; the PCR products of several individuals can be tagged using specific identification suites of nucleotides to tell them apart in the ensuing analyses (Meyer et al. 2008). It is also possible to perform hundreds of PCRs in parallel in a single test tube using microdroplets, and then simultaneously sequence all of the resulting amplicons (Tewhey et al. 2009). More refined reduced-representation approaches can be used to obtain the sequences of thousands of (supposedly) homologous loci across large numbers of individuals for a fraction of the cost of whole-genome sequencing (Van Tassell et al. 2008); notably, these approaches include RNAseq (only the messenger RNAs of genes that are expressed are sequenced; Wang et al. 2009), RADseq (only restriction-site–associated DNA regions are sequenced; Davey and Blaxter 2010), and exon capture (only exonic sequences complementary to specific oligonucleotidic probes are targeted; Hodges et al. 2007). One can already foresee that in the future, most of these approaches will be replaced by direct, full-genome sequencing of many individuals in the populations under study, generating a tsunami of data unprecedented in the history of biology. In preparation for this, Dowton et al. (2014) recently proposed a framework for next-generation DNA barcoding where multilocus data sets are coupled with coalescent-based species delimitation methods, which sparked an intriguing debate about the benefits and potential limitations of large next-generation data sets in DNA barcoding practises (Collins and Cruickshank 2014).

Data processing

Phasing heterozygous sequences

One thorny issue when sequencing nuclear markers is how to deal with heterozygosis. Indeed, markers that are variable enough to be useful for distinguishing species also present a good deal of intraspecific variations, which in the case of diploid organisms results in double peaks in the chromatograms obtained from Sanger-sequencing PCR products (Fig. 2). Although double peaks can be easily overlooked when they are few (particularly if one of the alleles gives a much stronger signal than the other one, resulting in peaks of unequal heights), a striking pattern occurs when the two alleles of an individual have different lengths: the chromatograms of such length-variant individuals display numerous double peaks in both their forward and reverse chromatograms (Flot et al. 2006). Such chromatograms were often discarded in the past as they looked hopelessly messy; however, the double peaks in the forward and reverse chromatograms are different, and it is therefore possible to combine them to reconstruct with certainty the sequences of the corresponding two alleles (Flot et al. 2006). This can be easily done by hand when the two alleles differ by only one insertion/deletion (indel), but when they differ by multiple indels it becomes quite complicated to perform the task manually, which led to the development of a web tool that automates the reconstruction process (Flot 2007; available online at http://jfflot.mnhn.fr/champuru/).

Fig. 2
figure 2

Example of forward (top) and reverse (bottom) chromatograms obtained from Sanger sequencing of a length-variant heterozygote. When the two alleles of the sequenced individual differ by one or several indels in addition to SNPs, a large number of double peaks are observed; however, the double peaks in the forward and reverse chromatograms contain different, complementary information, allowing reconstruction of the two haplotypes of the individual without cloning, either by hand (Flot et al. 2006) or using Champuru (Flot 2007, accessible online at http://jfflot.mnhn.fr/champuru/). In the example shown, the two haplotype sequences differed by one substitution and a 1-base indel: TAAATTCAAATCACACTCGCGAAAATCATGAA and TGAATTCAAATCACACTCGCGAAATCATGAA

Full size image

In a typical EPIC data set, about 30 % of the individuals are homozygous, 30 % are heterozygotes presenting a single double peak in their chromatogram (a case that is trivial to solve), and 30 % are length-variant heterozygotes that can be solved as outlined above. The remaining 10 % are heterozygotes with several double peaks, meaning that their two alleles have the same length but differ by more than one substitution. Phasing these individuals requires comparing their genotypes with those of other individuals sampled in order to infer the most likely haplotypes: this can be performed either by hand (Clark 1990) or computationally in a Bayesian framework using the programs SeqPHASE (Flot 2010b; available online at http://seqphase.mpg.de/seqphase) and PHASE (Stephens et al. 2001). Since the haplotypes of 90 % of the individuals in the data set are already known prior to running SeqPHASE and PHASE, the remaining 10 % are usually inferred with very high posterior probabilities. The rare individuals for whom uncertainties remain (as indicated by posterior probabilities lower than 0.9) may be solved by re-sequencing the PCR products using haplotype-specific primers (Hare and Palumbi 1999), or, as a last resort, by cloning them. Therefore, using this set of methods makes it possible to consider nuclear sequence markers as co-dominant, which was previously only the case for markers scored on gels (such as microsatellites and AFLPs).

For markers that present copy-number variations (CNVs; Freeman et al. 2006), individuals with more than two haplotypes are observed; the chromatograms of such individuals contain triple peaks if there are three haplotypes, and even quadruple peaks if more than three haplotypes are present. Paradoxically, this seems to happen more frequently with supposedly single-copy markers such as nuclear gene introns, rather than with multicopy markers such as ribosomal DNA (Flot et al. 2008). Resolving individuals with three or more haplotypes is difficult but not impossible, for instance using haplotype-specific primers or by cloning.

Aligning sequences

For protein-coding markers such as the widely used COI, alignment is straightforward since there are no single-nucleotide insertions or deletions (indels), which would result in loss-of-function frameshifts. Instead, indels always involve multiples of three bases and are typically rare among closely related species. Such sequences can easily be aligned manually using the alignment editor included in MEGA (Tamura et al. 2013) or in Mesquite (Maddison and Maddison 2014). The situation is different for markers that are not protein-coding, such as ribosomal genes and introns; non-coding markers often exhibit many indels of various sizes, even between closely related species. For small numbers of sequences with few indels, the implementation of MUSCLE (Edgar 2004) in MEGA comes in handy, but for large, complex data sets, we recommend using MAFFT (Katoh et al. 2009), which is conveniently available as a web server (http://mafft.cbrc.jp/alignment/server/). Alignments should be checked by eye for small errors in some sequences, particularly towards the ends of sequences where the quality of the chromatograms tails off; when such errors are suspected, the original chromatograms should be consulted so that bona fide sequence differences are not mistaken for sequencing errors. Whatever the type of marker and the alignment strategy, it is a good practice to curate the sequences by removing primers and (if cloning was performed) vector sequences, so that the first base in the alignment corresponds to the first base of the marker following the forward primer and the last base in the alignment corresponds to the last base of the marker preceding the reverse primer. Sequences alignments can then be processed in programs such as MEGA (Tamura et al. 2013) to generate phylogenetic trees and/or pairwise distance matrices under various evolutionary models. To determine the best-suited model, one may use the comparison tool included in MEGA or the standalone program jModelTest (Posada 2008). Patristic distances (i.e., distances between two tips measured along the tree) can be calculated atop a phylogenetic tree using the program Patristic (Fourment and Gibbs 2006). When numerous individuals of the same species are sequenced, it may be advantageous to present the data as a network instead of as a tree (Posada and Crandall 2001); among the various approaches available, we favour the median-joining algorithm (Bandelt et al. 1999) implemented in the program Network (Fluxus Technologies), which has been shown to perform well in a simulation study (Cassens et al. 2005). Whatever the methods used, a good practice in data sharing and data quality control is to make all chromatograms, alignments and trees/networks freely available online, either as supplementary material to the paper concerned or in dedicated online repositories such as GenBank (Benson et al. 2013), BOLD (Ratnasingham and Hebert 2007), TreeBase (http://treebase.org), or Dryad (http://www.datadryad.org).

In the case of NGS data, the millions of short reads obtained are commonly aligned against a reference sequence using fast, dedicated tools such as bowtie2 (Langmead and Salzberg 2012); then other programs such as SAMtools (Li et al. 2009) are used to generate read pileups and infer consensus sequences. We will not enter into details here as this would go beyond the scope of the present review, but interested readers may refer to the studies of Creer et al. (2010) and Fonseca et al. (2014) for more information on applying metagenetic approaches to meiofaunal studies.

Data analysis

Current approaches: single-locus species discovery

Up to now, the most commonly used methods in DNA taxonomy have been designed for data sets with one single marker sequenced across several individuals (single-locus datasets), although some may also be applied on concatenated alignments of several loci. The popular methods (or methods likely to become popular in the future) described below include those that require only a matrix of pairwise genetic distances, such as DNA barcoding (Hebert et al. 2003) and ABGD (automated barcode gap discovery; Puillandre et al. 2012a); those that require both a matrix of genetic distances and a phylogenetic tree, such as K/θ (formerly known as the 4X rule; Birky et al. 2010); those that require only a phylogenetic tree, such as GMYC (generalized mixed Yule–coalescent; Pons et al. 2006; Fujisawa and Barraclough 2013) and PTP (Poisson tree process; Zhang et al. 2013); and those that require phased heterozygous markers, such as haplowebs (Flot et al. 2010) (Fig. 1).

Using several of these approaches and looking for a consensus between the results obtained may increase our confidence regarding the outcome. However, these different methods use different criteria to delineate species; therefore, one can expect some degree of incongruence (especially when delineating recent species).

DNA barcoding

We define here DNA barcoding in its strictest and narrowest meaning as the use of a fixed, a priori defined threshold in genetic distances to identify units of diversity. DNA barcoding defined in this way groups two distinct, but often lumped disciplines (Hebert et al. 2003; Vogler and Monaghan 2007; Collins and Cruickshank 2012): (1) DNA barcoding sensu stricto, which is the identification of individuals of already known species, and (2) the discovery of new species, which is a branch of the large field of DNA taxonomy. The former consists in comparing standardised stretches of DNA (barcodes) to reference databases to identify unknown specimens, and has been particularly useful for the identification of juvenile stages (e.g., Webb et al. 2006) or of processed animals in the food industry (e.g., Rasmussen et al. 2009). This method is very widely used in the applied fields and in forensic science, and has a lot of added infrastructure around it. Additional methods such as ad hoc distance thresholds (Sonet et al. 2013) have been developed to account for false positives (e.g., erroneous attribution of a specimen of a new species to an already known species), a whole new system called Barcode Index Number (BIN; Ratnasingham and Hebert 2013) has been put in place to organise and register the barcoding data for all animals, and consortia have been assembled across the world to barcode different taxa (e.g., Consortium for the Barcode of Life; CBOL). A detailed description of the rationale and the caveats of DNA barcoding can be found in Casiraghi et al. (2010).

The second aspect, DNA taxonomy through DNA barcoding, is more controversial but easy to implement, and is likely to be the first step in molecular studies of meiofaunal diversity. This approach posits an a priori nucleotide distance threshold, below which specimens are considered conspecific and above which they are considered to belong to different species. The major assumption behind it is that intraspecific and interspecific variations do not overlap; that is to say, individuals of a given species are more similar molecularly than individuals belonging to different species. The existence of a “barcode gap” is a prerequisite for such an approach to work, but many studies employ a predefined threshold without checking whether it exists or not. Thus, a better approach is to start by plotting the distribution of pairwise distances between sequences in a data set; when this distribution reveals a clear gap, a threshold placed in this gap can be used to delineate species (Lefébure et al. 2006). However, the barcode gap is often difficult to detect or even non-existent, in which case species delimitation using this approach becomes quite arbitrary, and changing the stringency of the threshold changes the estimated diversity (e.g., Creer et al. 2010). Originally, a 3 % nucleotide divergence threshold was proposed for Lepidoptera COI sequences (Hebert et al. 2003); as an alternative, a relative threshold of ten times the mean intraspecific variation for the group under study was subsequently proposed (Hebert et al. 2004). Whether such a threshold is applicable to all groups and whether the initial design of the Lepidoptera study was representative of a natural sample has been debated and has led to suggestions that barcode gaps do not exist (Meyer and Paulay 2005; Wiemers and Fiedler 2007). The application of the DNA barcoding approach with a fixed pre-determined threshold to assess diversity in understudied organisms is certainly appealing, but its use is not as straightforward as it may seem, and the a priori decision of a cut-off threshold is ambiguous and subjective. A valid resource to analyse the results of the application of different barcoding thresholds and other parameters is jMOTU (Jones et al. 2011). Nevertheless, even this approach does not explicitly test if a barcoding gap actually exists in the data set, nor if a more likely threshold exists.

According to simulations, clear barcode gaps are only observed when species have small effective populations sizes and new species originate infrequently (Dellicour and Flot 2015). Indeed, successful studies using DNA barcoding for species discoveries are mostly reported for well-known, large organisms with relatively small effective population sizes, for which a large amount of information on putative species boundaries is already present, and when analyses are performed at a small spatial scale (Bergsten et al. 2012). This is often not the case for meiofaunal organisms since they are small, may have very large population sizes, and are understudied (Curini-Galletti et al. 2012; Fonseca et al. 2014). These features recently led to a discussion on whether delineating species of microscopic organisms using large DNA barcoding data sets is warranted (Rossberg et al. 2013, 2014; Morgan et al. 2014). Consequently, any attempt to apply DNA barcoding to the meiofauna should be crosschecked against the result of other methods (e.g., Tang et al. 2012).

Automatic barcode gap discovery

A less subjective means of defining a barcoding threshold for a given data set is the automatic barcode gap discovery tool ABGD (Puillandre et al. 2012a). Instead of using one or several predefined distance thresholds to delimit species, ABGD attempts to determine directly the threshold that is optimal for a given data set. If no satisfying threshold is detected, it concludes that all the specimens sequenced are conspecific. This method accepts an alignment as input to generate a distance matrix, either raw or corrected following the JC69 (Jukes and Cantor 1969) or K2P (Kimura 1980) models of sequence evolution; alternatively, a user-made distance matrix can be uploaded. ABGD requires users to specify one or a range of upper bounds on intraspecific genetic distances. From each of these priors and the distance matrix, it estimates a 95 % confidence interval for the population mutation rate θ (equal to 4μNe for nuclear markers and to 2μNe for mitochondrial ones, where μ is the mutation rate and Ne is the effective population size) using coalescent theory, and then looks for gaps in the distribution of pairwise distances that fall outside of the confidence interval for θ (Puillandre et al. 2012a). When one such gap is detected, ABGD uses it as a threshold to delimit primary species hypotheses (PSHs). Like fixed-threshold DNA barcoding methods, and in contrast to most other single-locus methods except haplowebs, ABGD does not require monophyly to delineate species (Fig. 3).

Fig. 3
figure 3

Hypothetical phylogenetic tree illustrating how distance-based approaches (DNA barcoding, ABGD) can potentially delineate species that are not monophyletic; here, such approaches will group all “A” individuals (A1 to A7) in one species and all the “B” individuals (B1 to B5) in another one, even though the hypothetical species A is not monophyletic

Full size image

The ABGD method has been used predominantly to define metazoan PSHs. In some cases, it has been found to delimit groups identical or similar to GMYC-based approaches and K/θ (Kekkonen and Hebert 2014); however, in other cases, its results have been more divergent (e.g., Tang et al. 2012). Meiofaunal studies using ABGD have included rotifers (Leasi et al. 2013), nemerteans (Leasi and Norenburg 2014) and molluscs (Jörger et al. 2012), but also nematodes, tardigrades, gastrotrichs, acoels, and flatworms (Tang et al. 2012). A rotifer COI data set with a detailed explanation on how to apply the ABGD approach to it is available in Fontaneto (2014).

K/θ

This method, described by Birky et al. (2005; 2010), uses population genetic theory to propose that, for sister clades of a given marker, interclade divergence (K) at least four times greater than intraclade variation (θ) means that these clades have more than 95 % chances to correspond to different species. In other words, clades meeting this “4X rule” are unlikely to have arisen solely by neutral genetic drift within a single population, but probably experienced barriers to gene flow, such as physical separation (allopatry), divergent selection for adaptation to distinct niches, or both. To date, the method has been used predominantly for asexual taxa, including some marine meiofaunal species, as their population genetics are arguably simpler (Birky et al. 2010; Tang et al. 2012), but the principles behind the method are applicable to sexual taxa as well (Birky 2013). Although the application of this method has been limited to mitochondrial or chloroplast sequences so far, it can be used with nuclear markers as well (C. William Birky Jr., personal communication).

The K/θ approach requires a gene tree to identify putative sister clades, and distance matrices to estimate genetic variation within and between these clades. Currently, the method requires the user to manually identify which clades are to be tested. The procedure comprises seven steps: (1) generate a gene tree, typically with neighbour joining (but maximum likelihood or Bayesian inference are also acceptable); (2) identify pairs of sister clades that have high support values (> 70 % using bootstrap); then for each pair, (3) calculate the nucleotide diversity π of each clade (equal to its mean pairwise distance corrected for sample size by n/(n-1)); (4) calculate θ = 2Neμ by π/(1-4π/3); (5) calculate the mean pairwise difference between the two clades (K); (6) calculate K/θ; and (7) find the probability of the two clades being compared to be distinct species. As previously stated, ratios higher than 4 mean that sister clades have more than 95 % chances of being distinct species; the probabilities for other ratios are given in a table available from C. William Birky Jr.

This method has already been applied to several meiofaunal groups, including gastrotrichs (Kånneby et al. 2012; Kieneke et al. 2012), rotifers (Birky et al. 2011; Iakovenko et al. 2013; Leasi et al. 2013), copepods (Marrone et al. 2010, 2013), ostracods (Martens et al. 2012, 2013; Shearn et al. 2012; Schön et al. 2012), but also nematodes, tardigrades, nemerteans, acoels, and flatworms (Tang et al. 2012). Birky et al. (2010) used rotifer data sets to develop the method, and Birky (2013) provides a detailed explanation on how to perform the analyses.

GMYC-based approaches

The generalized mixed Yule–coalescent model (GMYC; Pons et al. 2006; Fujisawa and Barraclough 2013) is a coalescent-based phylogenetic method that sets a threshold between coalescent and species-level processes in order to delineate evolutionary significant units (ESUs) akin to species (Simpson 1951). Approaches based on the GMYC model rely on the expectation that intraspecific coalescent branching proceeds discernibly quicker than speciation, which is modelled as a Yule process; therefore, species can be identified in gene tree as clusters of terminals separated by longer internal branches. On an ultrametric tree (i.e., a tree whose branch lengths are proportional to time), changes in branching rates indicative of a shift from species-level processes (i.e., coalescent) to population-level processes (i.e., Yule) can be used to delimit ESUs. To do so, GMYC-based approaches cycle through each node and separately model coalescent and Yule processes, and given the observed branching processes, calculate the most likely threshold(s) between species-level and population-level branching rates.

The single-threshold model (ST-GMYC) was the first one to be proposed (Pons et al. 2006). With this approach, the most likely solution identifying Yule and coalescent processes is compared to the null hypothesis (a single branching rate within a single species) using a χ 2 test. If significant, the threshold is used to delimit ESUs. Given that likelihood values are available for all possible solutions, one may also, as a second step, assess whether other solutions are significantly less likely than the one favoured by the method. Such an approach provides very useful confidence intervals around the most likely solution, which makes it possible to determine whether the species delimitations inferred from the data are reliable. ST-GMYC can be applied using the splits package (Ezard et al. 2009) in R (R Core Team 2014) or using a webserver (http://species.h-its.org/gmyc/), with step-by-step guides available in Fontaneto (2014) or on Tomochika Fujisawa’s webpage (http://tmfujis.wordpress.com/2013/04/23/how-to-run-gmyc/).

Several further GMYC approaches have been developed: the multiple-threshold GMYC model (MT-GMYC; Monaghan et al. 2009) allows for rate heterogeneity among species and does not assume that the same threshold applies to all the parts of the gene tree; the multimodel-averaging approach (MM-GMYC; Powell 2012) accounts for uncertainty in GMYC model selection; and a Bayesian implementation of GMYC (bGMYC; Reid and Carstens 2012) considers uncertainty in phylogenetic reconstruction. The MT-GMYC and MM-GMYC approaches are included in the updated splits R package (Fujisawa and Barraclough 2013), whereas bGMYC is available for download from Noah Reid’s website (https://sites.google.com/site/noahmreid/home/software). A study comparing ST-GMYC, MT-GMYC and bGMYC on simulated data sets found bGMYC to outperform its forerunners in most cases (Dellicour and Flot 2015).

The ultrametric trees needed for the GMYC methods are typically reconstructed using either maximum likelihood with post hoc branch smoothing or using BEAST (Drummond and Rambaut 2007; Bouckaert et al. 2014; Tang et al. 2014a). A recent meta-analysis shows that the ST-GMYC applied on BEAST trees provides the most robust diversity estimates in terms of both richness and identity (Tang et al. 2014a). The GMYC model has been applied to several meiofaunal groups, including gastrotrichs (Kånneby et al. 2012; Kieneke et al. 2012), rotifers (Fontaneto et al. 2007, 2011; Birky et al. 2011; Obertegger et al. 2012, 2014; Leasi et al. 2013; Tang et al. 2014b; Malekzadeh-Viayeh et al. 2014), copepods (Gollner et al. 2011; Cornils and Held 2014), ostracods (Adolfsson et al. 2010; Brandão et al. 2010; Bode et al. 2010; Martens et al. 2012; Schön et al. 2012), nemerteans (Leasi and Norenburg 2014), flatworms (Sluys et al. 2013), and molluscs (Jörger et al. 2012), as well as nematodes, tardigrades, and acoels (Tang et al. 2012). A rotifer COI data set with an explanation on how to analyse it using the single-threshold GMYC approach can be found in Fontaneto (2014).

Poisson tree process model

The Poisson tree process model (PTP; Zhang et al. 2013) is another tree-based species delimitation method that uses coalescence theory to distinguish between population-level and species-level processes. It assumes that intraspecific and interspecific substitutions follow two distinct Poisson processes, and that intraspecific substitutions are discernibly fewer than interspecific substitutions because they have less time to accumulate; this method uses substitutions directly to represent time rather than via a method that corrects for rate variation, such as GMYC. This coalescent-based method is very fast as it does not require ultrametric trees as input (as opposed to GMYC), just a regular rooted gene tree; it has been shown to produce species delimitations matching traditional taxonomic groupings (Tang et al. 2014a). This method is implemented as a standalone program and as a web server (http://species.h-its.org/). The most recent version of the method includes both maximum-likelihood and Bayesian searches for species boundaries, and returns Bayesian support values for those delimited species. Albeit very recent, this method has already found several applications in meiofaunal taxonomic studies, notably on rotifers (Tang et al. 2014a; Velasco-Castrillón et al. 2014), on nemerteans (Leasi and Norenburg 2014), and on copepods (Blanco-Bercial et al. 2014).

Haplowebs

Haplowebs rely on a different species delineation criterion, mutual allelic exclusivity (Doyle 1995; Flot et al. 2010), to delineate species of diploid organisms that have been reproductively isolated long enough to not share any identical sequence for the marker under investigation. Newly diverged species always reach mutual allelic exclusivity prior to, or at the same time as, reaching reciprocal monophyly (Flot et al. 2010); besides, the time needed for newly diverged species to reach mutual allelic exclusivity only depends on the length of the marker and its mutation rate, not on the effective size of the populations. In contrast, the time needed for newly diverged species to reach monophyly is strongly dependent on genetic drift, and therefore on the effective population size of the species. Using shared alleles to delineate species is only applicable to nuclear markers that do not exhibit homoplasy (convergence), which is why mutual allelic exclusivity was shown to perform poorly on microsatellites, RFLPs, RAPDs and AFLPs (Miller and Spooner 1999; Hausdorf and Hennig 2010)—all types of data that, in contrary to DNA sequences, exhibit frequent convergence between species.

The criterion of mutual allelic exclusivity was implemented in a graphical approach called haplowebs (short for “haplotype webs”; Flot et al. 2010): starting from a network or a tree of nuclear haplotypes (the method used for obtaining the tree or network does not really matter), connections are added between haplotypes found to co-occur in heterozygous individuals. Once all connections have been added, inspection of the graph reveals discrete pools of interconnected alleles, each of which corresponds to a group of individuals that appears to be reproductively isolated from the others; each such group is called a “field for recombination” (Carson 1957; Doyle 1995), i.e., a putative species. Since this method is based on the co-occurrence of haplotypes in heterozygous individuals, it cannot be applied to metagenetic data (in which haplotypes cannot be traced to individuals) and it requires that a sufficient number of heterozygotes be sequenced. Hence, a large data set comprising several individuals per species should be collected. Moreover, the haplowebs approach is rooted in the biological species criterion that delineates species based on the presence/absence of gene flow, and as with all approaches that follow this line of thought, it may incorrectly lump species that occasionally hybridise. This method was originally developed for cnidarian DNA taxonomy (Flot et al. 2008, 2010, 2011), and its applications have thus far dealt with cnidarians (Flot et al. 2013; Schmidt et al. 2013; Adjeroud et al. 2014; Schmidt-Roach et al. 2014), crustaceans (Flot et al. 2014) and rotifers (Li 2012).

Comparison between the different single-locus approaches

The methods mentioned above can be classified as tree-based and non–tree-based (Sites and Marshall 2003). GMYC, K/θ, and PTP are tree-based, and as a result, only delineate species that are monophyletic in the gene trees used to run the method (Table 1); whereas DNA barcoding, ABGD and haplowebs are non–tree-based and do not require monophyly. Surveys found that 15–40 % of species of various groups of animals are not monophyletic in mitochondrial gene trees (Funk and Omland 2003; Ross 2014), and the situation is probably much worse with nuclear markers, since acquisition of monophyly is expected to be four times slower for nuclear genes than for mitochondrial ones (Moore 1995). Besides, the highest percentage of non-monophyletic species reported by Funk and Omland (2003) was for non-insect invertebrates, a category encompassing most meiofaunal taxa. Hence, DNA taxonomic results obtained using a tree-based approach should be crosschecked against a non–tree-based approach so that non-monophyletic species are not overlooked.

Table 1 Comparison of popular and/or promising single-locus approaches in DNA taxonomy
Full size table

These methods also differ in the criterion they use to delineate species: a criterion based on genetic distances for distance-based DNA barcoding; a quantitative approach based on coalescent for ABGD; a phylogenetic criterion based on branching rates for GMYC and PTP; and a population genetic criterion based on genetic isolation for haplowebs and K/θ. When the divergence between species is large, sampling within species is comprehensive, and the effective population sizes are small, then these methods are generally congruent (Tang et al. 2012, 2014a; Carstens et al. 2013; Dellicour and Flot 2015). However, given that each of these methods uses either different criteria or inputs, incongruence between the methods is expected under certain conditions. For example, the stringency of the K/θ method in terms of the separation between lineages and the single thresholds of the ST-GMYC and PTP approaches are expected to lump potentially distinct species when lineages have recently diverged. Undersampling within and between species will likely introduce biases in the coalescent approaches; nevertheless, likelihood-based methods such as GMYC, which provide confidence intervals for the most likely solutions, are potentially able to suggest when undersampling may affect the results. Undersampling is even more problematic for haplowebs, potentially leading to oversplitting (Dellicour and Flot 2015). Furthermore, different rates of substitution among lineages make predefined thresholds inappropriate, as these are likely to either lump species in rapidly evolving lineages or split species in slowly evolving lineages.

Several studies have evaluated factors that could decrease the accuracy of some of these methods. For GMYC, simulation studies have addressed the effects of various aspects of sampling (Papadopoulou et al. 2008; Bergsten et al. 2012; Reid and Carstens 2012; Talavera et al. 2013), population size and speciation rates (Esselstyn et al. 2012; Fujisawa and Barraclough 2013; Dellicour and Flot 2015). For PTP, simulations have been used to evaluate the effect of birth rates (i.e., evolutionary distances between species) and sampling unevenness (Zhang et al. 2013). In general, it seems that GMYC based on BEAST trees provides results highly congruent with PTP (Tang et al. 2014a). A recent simulation study (Dellicour and Flot 2015) compared barcode gap detection, GMYC and haplowebs, and found a “sweet spot” (characterized by small effective population sizes and low speciation rates, resulting in large interspecific divergence and low intraspecific diversity) where all the methods tested performed well. However, none of these single-locus methods was able to delineate species properly when effective population sizes and speciation rates were both large (in which case divergence between species was small and intraspecific diversity was high), emphasizing the need for multilocus approaches to tackle such difficult cases.

Besides, single-locus approaches fail to account for possible discrepancies between markers: if the examined marker exhibits an idiosyncratic evolutionary history (for instance, because it is subjected to interspecific introgressions or gene captures, or because the data set includes paralogues and pseudogenes that obscure the signal), this will directly impact the inferred species boundaries. Most of the single-locus methods mentioned above could be used on concatenated data sets originating from independent markers, as was done for ST-GMYC on rotifers with COI+28S (Fontaneto et al. 2007) and on non-meiofaunal organisms (e.g., Williams et al. 2011; Bellati et al. 2015). Nevertheless, this cannot be considered a bona fide multilocus approach, since it does not take into account the potential discordance between the signals given by these markers. If the markers disagree with each other, it may be expected that one of them will swamp the signal from the other ones and imprint its own history on the resulting delimitation, or the contradictions to be so strong that no significant delineation will be proposed as an outcome (Bull et al. 1993). In what follows, we present some multilocus methods that make it possible to overcome this problem.

One step further: multilocus species delimitation

Some multilocus approaches looking for congruence between gene genealogies are fairly old (e.g., Koufopanou et al. 1997), but there has been a recent surge of interest in these methods, leading to the publication of several key articles in the last few months. Multilocus species discovery methods undoubtedly represent the future of DNA taxonomy, as the use of a large number of independent markers made possible by technological advances in sequencing will allow researchers to tackle several of the commonly encountered problems in species delimitation (but see Collins and Cruickshank 2014). Multilocus species delimitation methods can account for non-monophyletic species, gene tree discordance, incomplete lineage sorting, gene flow after divergence, and other confounding factors that may create problems in single-locus DNA taxonomy (Camargo et al. 2012; Fujita et al. 2012). Yet, these methods have been rarely used in meiofaunal studies so far; hence, we will just briefly mention them without assessing their strengths and weaknesses, and without providing suggestions on their use.

Structure and Structurama

The programs Structure and Structurama use Bayesian clustering algorithms (Pritchard et al. 2000; Falush et al. 2003; Huelsenbeck et al. 2011) to detect population structure in co-dominant genetic data such as nuclear sequences or microsatellites. They were originally developed to detect intraspecific population structure caused for example by geographic distance, but they are also frequently used to detect species boundaries (even though their suitability for this purpose is somewhat questionable). One minor drawback of Structure is that it requires the number of populations to be specified beforehand: some methods have been proposed to find the best value of this parameter (Evanno et al. 2005), or one may use the number of species suggested by DNA barcoding for instance. A different approach has been implemented in Structurama, an extension of Structure that uses a Dirichlet-process prior in order to estimate the number of populations as part of the algorithm (Huelsenbeck and Andolfatto 2007; Huelsenbeck et al. 2011). Structure is available for download from http://pritchardlab.stanford.edu/structure.html, and Structurama from http://cteg.berkeley.edu/~structurama/. An example of the application of Structure to meiofaunal studies can be found in Tulchinsky et al. (2012), which used inter-simple-sequence-repeat (ISSR) markers to delineate marine nemertean species.

Bayesian phylogenetics & phylogeography

The Bayesian phylogenetics & phylogeography (BP&P) method uses Bayesian modelling of the multispecies coalescent to generate the posterior probabilities of species assignments (Yang and Rannala 2010; Rannala and Yang 2013). It accounts for uncertainties in gene tree reconstruction and, unlike tree-based single-locus methods, is designed to deal with non-monophyletic species arising from incomplete lineage sorting. The input of the method consists of multiple gene trees, but in the classical usage of the method it also needs a user-specified guide tree to avoid integrating over all possible species delimitations; however, a new version of the program BP&P was recently published (Yang and Rannala 2014) that is able to delineate species in an unguided way. A program implementing this method is available from Ziheng Yang’s webpage (BP&P: http://abacus.gene.ucl.ac.uk/software.html). BP&P has already been used in meiofaunal studies: the original description of the method used a rotifer data set to estimate its performance (Yang and Rannala 2010), and more recently, it was applied to molluscs (Jörger et al. 2012) and to several species complexes of nemertodermatids (Meyer-Wachsmuth et al. 2014).

SpedeSTEM

SpedeSTEM (Ence and Carstens 2011) uses a maximum-likelihood approach to perform species delimitation using STEM (species tree estimation; Kubatko et al. 2009). This approach calculates the probability of different models containing various numbers of evolutionary lineages, and then ranks these models according to information theory criteria. The inputs of SpedeSTEM are single gene trees (no global guide topology is needed). According to simulations performed by the authors of the method, SpedeSTEM can work using as little as five loci, but this method has yet to be applied to meiofaunal species. A program implementing this approach can be downloaded from https://spedestem.osu.edu/.

O‘Meara’s heuristic search

O’Meara (2010) introduced two methods (one parametric and the other non-parametric) with heuristic search strategies to delimit species using multiple trees from individual genetic markers as input. The parametric method, also called “KC delimitation”, seeks to find the delimited species tree that maximises the probability of the gene trees. The non-parametric method quantifies two metrics, called “gene tree conflict” and “excess structure”, in order to minimise their costs. According to the original paper, the non-parametric method performs better, albeit inconsistencies can be found. These analyses are implemented in Brownie 2.0 (http://www.brianomeara.info/brownie), but have yet to be applied to meiofaunal studies.

New Bayesian methods using the multispecies coalescent

The multispecies coalescent was already used to obtain species trees from gene trees on multilocus data sets by approaches such as BEST (Liu et al. 2008) and *BEAST (Heled and Drummond 2010), and was successfully applied to species delimitation in BP&P (Yang and Rannala 2010). Other recent methods are continuously appearing: three articles using Bayes Factor Delimitation (BFD; Grummer et al. 2014), Bayes Factor Delimitation with genomic data (BFD*; Leaché et al. 2014) and an assignment-free Bayesian method for species discovery called DISSECT (Jones et al. 2014) were published in the last few months. Approximate Bayesian computation (ABC) can also be used to answer several topics in eco-evolutionary studies, and it has already been applied to species delimitation in animals (Camargo et al. 2012). As these methods are all very new, it is too early to write about their respective pros and cons, but this profusion of new Bayesian approaches using the multispecies coalescent suggests that other very significant advances in this field are likely to come out in the near future as well.

Caveats and perspectives

Herein we have described popular metrics and techniques that taxonomists and non-taxonomists alike can use to obtain DNA-based working hypotheses regarding species boundaries. These methods offer effective species proxies that are quick, easy to implement, and relatively robust when the assumptions of the methods used to obtain them are met by the data sets on which they are applied. Hence, DNA taxonomy is a useful springboard to gauge the diversity of groups where morphological studies are painstakingly difficult and/or where the number of species far outweighs the availability of taxonomists to investigate them, as is generally the case for the meiofauna. Still, all the methods described here should be used with caution. It is generally accepted that DNA taxonomy is not a substitute for taxonomic descriptions (Wiens and Servedio 2000; Sites and Marshall 2004), and indeed, most DNA taxonomy methods yield only primary species hypotheses (PSHs) that require further testing and validation (Puillandre et al. 2012b; Pante et al. 2015). Because different methods can provide different conclusions, it is advisable to use several approaches, look at the congruence between the results obtained from each of them (Carstens et al. 2013), and try to understand the reasons for the observed incongruences. In an optimal scenario, one may adopt an integrative, iterative approach including genetics, morphology, ecology, behaviour, geography, as well as other sources of data to support species identities (Padial et al. 2010; Schlick-Steiner et al. 2010). The drawback of including so many approaches, however, is that some level of incongruence between them will show up in most cases, and there is then no obvious, objective way to decide which results to trust and which ones to discard. This is particularly likely when dealing with meiofaunal species, since their potentially large population sizes and dispersal abilities makes them prone to incomplete lineage sorting (Rossberg et al. 2013). On the bright side, meiofaunal species that are easy to collect have often been used as test data sets when proposing new methods for species delimitation, and it is likely that this trend will continue in the future, keeping meiofaunal studies at the cutting edge of DNA taxonomy.