1. Top of page
  2. Abstract
  3. Introduction
  4. Conclusions
  5. References

The recently introduced term ‘integrative taxonomy’ refers to taxonomy that integrates all available data sources to frame species limits. We survey current taxonomic methods available to delimit species that integrate a variety of data, including molecular and morphological characters. A literature review of empirical studies using the term ‘integrative taxonomy’ assessed the kinds of data being used to frame species limits, and methods of integration. Almost all studies are qualitative and comparative – we are a long way from a repeatable, quantitative method of truly ‘integrative taxonomy’. The usual methods for integrating data in phylogenetic and population genetic paradigms are not appropriate for integrative taxonomy, either because of the diverse range of data used or because of the special challenges that arise when working at the species/population boundary. We identify two challenges that, if met, will facilitate the development of a more complete toolkit and a more robust research programme in integrative taxonomy using species tree approaches. We propose the term ‘iterative taxonomy’ for current practice that treats species boundaries as hypotheses to be tested with new evidence. A search for biological or evolutionary explanations for discordant evidence can be used to distinguish between competing species boundary hypotheses. We identify two recent empirical examples that use the process of iterative taxonomy.


  1. Top of page
  2. Abstract
  3. Introduction
  4. Conclusions
  5. References

Modern taxonomy and the species concept

Although 2010 has been designated International Year of Biodiversity by the Convention on Biological Diversity and the United Nations (Johns, 2010), taxonomy, one of the major scientific disciplines that underpins our knowledge of biodiversity and its management, has been in decline for many decades (Wilson, 1985, 2004). A core priority of taxonomic research is the discovery and delimitation of species; therefore, it is essential to biodiversity research. Recent developments in thinking about species concepts, and the analysis and interpretation of data used to delimit species, have profound implications for taxonomic research that have not been realized fully, much less resolved. Such approaches to species delimitation have major implications for any attempts to understand and manage the biodiversity crisis. Most eukaryote species awaiting description are insects (Raven & Yeates, 2007; Yeates, 2009), thus it is appropriate that systematic entomologists critically appraise current proposals for species delimitation.

The last decade has seen real progress in our thinking about what species are (Wiens, 2007). The distinction between species concepts and the criteria or evidence used to delimit them (Hey, 2006; De Queiroz, 2007; Knowles & Carstens, 2007) has revealed the underlying unity of modern species concepts, that is, species are lineages (Mayden, 1997; De Queiroz, 1998, 2007). In this sense, many of the species concepts in operation today represent the use of different evidence and criteria to distinguish species lineages. It is instructive that these species concepts can lead to the recognition of different species boundaries (Wiens & Penkrot, 2002; Laamanen et al., 2003; Agapow et al., 2004), and that very few taxonomists report the kind of criteria or evidence that are critical to their species concept (Schlick-Steiner et al., 2010). The two questions of discovering species boundaries and the relationships between species are biologically linked, but are rarely methodologically coupled; however, O’Meara (2010) recently introduced new heuristic methods that jointly estimate species limits and relationships from individual gene tree input. These methods may provide insights into these questions, but amply demonstrate the complexity of the issue in theory and practice.

An extreme example of the ‘species as lineages' concept has been pursued by proponents of DNA barcoding (Hebert et al., 2003), in which the discovery of species is as simple as the recognition of deeply diverged, monophyletic lineages within a single genetic marker (cytochrome c oxidase I, COI, for animals; several chloroplast genes for plants). Among philosophical (Lipscomb et al., 2003; Will & Rubinoff, 2004) and practical criticisms of barcoding (Cameron et al., 2006; Hickerson et al., 2006; Meier et al., 2006; Rubinoff et al., 2006), perhaps the most potent criticism is its reliance on a single data source (Rubinoff & Holland, 2005). More generally, circumstances can be envisioned in which almost any single taxonomic data source would fail to resolve the true set of species as lineages (Schlick-Steiner et al., 2010). As the ways in which any data source may fail are almost limitless, it is increasingly common for taxonomists to use multiple independent lines of evidence in framing species hypotheses.

Evidence-based taxonomy

One expression of this desire for multiple lines of evidence is the call for ‘integrative taxonomy’, a term that was introduced independently into the literature by Dayrat (2005) and Will et al. (2005) to mean slightly different things. In what was primarily a critique of DNA barcoding, Will et al. (2005) used the term integrative taxonomy to mean a taxonomic process that was inclusive of all available data sources, not just mitochondrial DNA (mtDNA) COI barcode data. For Will et al. (2005), integrative taxonomy gives primacy to morphological characters because of their greater complexity and presumed multigenic origin, which are thought to constitute a more secure basis for separating species than small fragments of DNA sequence. Subsequent proposals require character ‘discrimination’ (i.e. a measure of the quality of the character evidence) before the integration of all information (Valdecasas et al., 2008). In a contrasting paper published almost simultaneously to Will et al. (2005), Dayrat (2005) used the term integrative taxonomy to postulate a set of guidelines that taxonomists should follow when proposing names that aimed to facilitate the integration of data from different sources. Although his six guidelines may have merit in some circumstances (e.g. no. 5 ‘species should only be named when their limits are supported by multiple lines of evidence’), and some are common practice for many taxonomists (e.g. no. 4 ‘differentiated species need not be formally described’ and no. 6 ‘types preserved so that molecular data can be extracted’), in general, his set of six guidelines were overly operational and proscriptive (e.g. no. 1 ‘no new species should be described unless the group has recently been monographed’, no. 3 ‘no new species should be described with only one or a few specimens') (Valdecasas et al., 2008). Although differing in many respects, these two papers coined the term integrative taxonomy independently, and both focused on a taxonomy that integrated all available sources of data to develop hypotheses of species-level taxa. Neither paper dealt explicitly with the operational integration of taxonomic data.

The quest for a truly integrative taxonomy

To achieve the goal of an integrative taxonomy, it is necessary to elucidate the desirable characteristics. Obviously, such a taxonomy should be applicable to the broadest range of taxa and character types as possible. In terms of taxa, it should apply to plants and animals, and asexual and sexually reproducing groups. It should be amenable at least to the common types of available characters that are applied routinely to species delimitation, in particular morphological attributes and molecular sequence data. It should also be able to incorporate other features that can be coded as categorical character states in a matrix, such as attributes of the extended phenotype, behaviour (Wenzel, 1992) and ecology (Miller & Wenzel, 1995). It should also be able to cope with qualitative, polymorphic, meristic and continuous traits (Thiele, 1993; Wiens, 2001). Ideally, integrative taxonomy should also be able to provide some level of statistical confidence regarding the species limits that it proposes. The method should give greater confidence to species limits that are supported by more and more diverse evidence (Dayrat 2005: guideline no. 5; Schlick-Steiner et al., 2010). Because species limits should be treated as testable hypotheses (Wheeler, 2004; Fitzhugh, 2005), the method must produce species boundaries that can be tested by application to datasets other than those used to develop the hypothesis.

Any study linking different kinds of data by, for example, mapping morphological diversity on to a molecular phylogeny is therefore, arguably, integrative. Equally, however, one could argue that all such studies are correlative or corroborative rather than truly integrative. In practice, a spectrum of integration exists along which any study employing different kinds of data actually falls. At one end are studies that verbally and qualitatively compare data classes. Several studies cited as examples of integrative taxonomy by Schlick-Steiner et al. (2010) used correlative approaches to compare molecular data and morphology (Malhotra & Thorpe, 2005; Yoder et al. 2005; Rissler & Apodaca, 2007; Roe & Sperling, 2007). In our sense of the word integrate, integrative taxonomy should aim for quantitative methods that allow different data partitions to contribute to the delimitation of species boundaries. Although concatenation is one such method, a striking example of an integrative method that has emerged recently from within molecular systematics is the species-tree approach, for example Belfiore et al. (2008) and Linnen & Farrell (2008).

To date, Schlick-Steiner et al. (2010) have offered the most detailed treatment of the process of integrative taxonomy. They stressed that integrative taxonomy does not replace traditional taxonomy, but that it uses complementarity among disciplines to improve rigour. Their integrative procedure relies on an agreement among three ‘conclusive’ disciplines that is both proscriptive and restrictive, and disciplines and datasets are defined rather arbitrarily. For example, mtDNA is one discipline, and all nuclear DNA is another; morphological data from males is one dataset, and from females is another. A discipline is conclusive or at least consistent when analyses of two or more datasets from it delimit the same species taxa, but inconclusive if they significantly support different species. Resolution of the disagreement among datasets (in inconclusive disciplines) involves deciding which species hypotheses are plausible and which are not. They do not consider concatenation or species-tree approaches, and we consider their approach more correlative than integrative, seeking correlation in species boundaries between different data sources.

The method of Schlick-Steiner et al. (2010) proposes a ‘stopping rule’ when three disciplines agree about a species boundary, but we view species as hypotheses in a hypothetico-deductive framework, which are always subject to testing and rebuttal in the face of new evidence. Prior to formal publication, individual taxonomists decide on the weight of evidence for their species hypotheses.

One aspect of Schlick-Steiner's (2010) approach that we do endorse is their search for a biological or evolutionary explanation for datasets that disagree (their p. 425). For example, a morphologically-diagnosed putative species may be paraphyletic with respect to mtDNA, but coalescent theory explains this difference and allows us to choose between competing hypotheses.

Integrative taxonomy: a survey of current practice

We conducted a survey of the literature using the Thompson ISI Web of Science, searching for the term ‘integrative taxonomy’ to establish the methods used by researchers employing this term. The search, conducted in May 2010, resulted in 28 empirical studies that used multiple data sources to frame species limits (Table 1). The majority of publications (15) gathered data from two or more molecular markers. Mitochondrial gene sequences were the most commonly collected, with relatively short (600–800 bp) sequences from COI (16) and 16S (7) being the most frequently used. Although at most 43 non-molecular characters were used, they came from a diversity of sources, ranging from standard morphological attributes, bioacoustic variables, ecophysiological traits and karyotype differences. Two studies used the results of interbreeding experiments and molecular markers. We divided the methods of integration into four types as follows: (i) qualitative verbal descriptions of data from different partitions; (ii) non-molecular data ‘mapped’ onto a molecular tree; (iii) molecular partitions concatenated (iv); and morphological and molecular data compared using a quantitative, statistical measure of correlation.

Table 1. Results of literature survey for empirical studies in integrative taxonomy.
AuthorData typeIntegration level (1–4)b
Molecular (# characters)aNon-molecular
  1. a+, partitions were concatenated.

  2. bData integration categories: 1, verbal descriptions of data from different partitions, qualitative comparison; 2, non-molecular data ‘mapped’ onto molecular tree; 3, concatenation of at least some molecular datasets; 4, some statistical testing of association between data types. Some studies fall into two categories: for example, concatenated sequence data, but qualitatively comparing results of molecular and morphological data.

  3. cSeparated by the authors using traits including body size, coloration, surface structure and male genital morphology.

Ahrens et al. (2007) COI (826)+ rrnL (853) 28S (834)43 morphological typesc2, 3
Cardoso et al. (2009) ITS1 (300)+ COI (717)41 morphological characters3, 4
Castroviejo-Fisher et al. (2009a) 16S (850)9 morphological characters; 5 bioacoustic variables1
Castroviejo-Fisher et al. (2009b) 16S (850)11 morphological characters; 8 bioacoustic variables2
Casu et al. (2009) SSU (828) Allozymes (11) ISSRs (5)11 morphological characters; crossbreeding experiments1
Fonseca et al. (2008) COI (396) ITS (911) D2D3 (597)Morphology (10 characters for one species, 5 for the other); interbreeding experiments4
Francuski et al. (2009) Allozymes (10) COI (770)Morphological: wing size and shape variation4
Gibbs (2009) COI (657)Species' morphology described fully, some distribution data1
Haase et al. (2007) COI (638)Seven shell characters, canonical variates analysis1
Lanzone et al. (2007) Cyt b (687)Karyotype analysis, morphology (5 external and 21 cranial characters)2
Leite et al. (2008) Cyt b (1140)23 cranial measurements, karyotype analysis1
Lewis & Karageorgopoulos (2008) RAPD-PCR (542)Morphological and ecophysiological (12 characters)1
Mengual et al. (2006) COI (784)+ 28S (396) ITS2 (496)Morphological identification of specimens (based on previous literature)1, 3
Mikkola & Ståhls (2008) COI (729) ITS2 (484)Species' morphology described fully1
Milankov et al. (2008) Allozymes (15) COI (720)6 morphological characters1
Milankov et al. (2009) COI (768)Morphometric analysis (wing size and shape plus 4 other characters).1
Nitta (2008) rbcL (1238)+ trnSGG (2334)+ trnH-psbA (782)Morphology, distribution and ecology1, 3
Padial & de la Riva (2009) 16S (591)13 morphological characters, morphometrics and bioacoustics (advertisement calls)1
Padial et al. (2007) 16S (591)15 morphological characters, morphometrics and bioacoustics (advertisement calls)1
Padial et al. (2008) 16S (591)+ Cyt b (350)13 morphological characters, morphometrics and bioacoustics (advertisement calls)1, 3
Padial et al. (2009) 16S (528)10 morphological characters, bioacoustics (advertisement calls)2
Puillandre et al. (2009) COI (658)+ 28S (900)Shell morphology2, 3
Roe & Sperling (2007) COI (475) ITS2 (499) EF1α (365)Morphology (4 forewing characters) and larval host plant association1
Schmidt (2009) COI (658)Full morphological descriptions1
Schmitz & Rubinoff (2008) COI (773) EF1α (762)Full morphological descriptions1
Silva-Brandão et al. (2008) COI (1520)+ COII (640)+ EF1α (1700)Some morphology based on previous papers (nothing quantitative)1, 3
Vieites et al. (2009) 16S (514)Morphological (qualitative), bioacoustic (call variation) and geographical distribution2
Walker et al. (2009) COI (539) 18S (1686)Morphology (12 characters described)1

Our results show that, in almost all cases (25), researchers ‘integrate’ different data sources using qualitative techniques, inspecting the groupings produced from different data and crafting species limits in a post hoc fashion, either comparing data partitions verbally (integration type 1; e.g. Lanzone et al., 2007), or mapping morphological data onto a tree generated from a molecular partition (integration type 2; e.g. Ahrens et al., 2007). Seven studies concatenated their molecular data (integration type 3), but none concatenated molecular and morphological partitions. In only three cases (e.g. Cardoso et al., 2009) were statistical comparisons carried out to assess the correlation between partitions (integration type 4).

In order for data integration or comparison to be most accurate and appropriate, data from different partitions should be sampled from the same individuals. Few authors (e.g. Cardoso et al., 2009; Puillandre et al., 2009) specified that the morphological examination and DNA extraction were undertaken on the same individuals. Furthermore, few studies presented the non-molecular data in a tabulated or matrix format that would facilitate future use.

Diverse data is desirable, but how can it be integrated?

It is now commonplace to assemble multiple lines of evidence when attempting to draw species boundaries. Typically, data from one or more molecular loci plus morphology are collected (for example, the studies in Table 1). As the genomic revolution continues, it will be commonplace to assemble increasingly large datasets containing separate lines of evidence from independent loci (e.g. multiple nuclear genes). It will be increasingly unlikely that all lines of evidence will be perfectly congruent with one another (Padial et al., 2010), and circumscribe exactly the same species boundaries. How should incongruence be addressed in this context? Concatenation (an implementation of total evidence; Kluge, 1989; De Queiroz et al., 1995; Baker & DeSalle, 1997; Wiens, 1998; De Queiroz & Gatesy, 2007) is one possible way of integrating information from different data sources. One of the great advantages of concatenation is that, when combining all data, a common signal may emerge even if a weakly contradictory signal is contained in the individual data partitions. That is, the species tree emerges because data insufficiencies in the subsets are overcome (Gatesy et al., 1999). This presumes that all data have the same phylogenetic history and the reconstruction of the tree improves when datasets are added.

However, this positive effect of using a range of data sources is undermined when working near the species boundary, because incomplete lineage sorting and gene flow can cause true differences in the history of various genetic loci. In other words, different molecular markers may have evolved on different trees (Maddison, 1997; Slowinski & Page, 1999). As an illustration of this point, the non-monophyly of species on mitochondrial haplotype trees is very common in congeneric animal species (Funk & Omland, 2003; Will et al., 2005; Joseph & Omland, 2009). In this context, concatenation is not warranted, because a single underlying tree common to all available genes may not be recoverable. When very shallow species trees (short internodes compared with the effective population size) are present, gene trees that do not match the species tree are more probable than matching gene trees, and concatenation would be expected to perform poorly in these situations (Carstens & Knowles, 2007; Edwards et al., 2007; Kubatko & Degnan, 2007; reviewed in Degnan & Rosenberg, 2009). Simulation studies suggest that concatenation will tend to give the correct species tree only in situations where the divergences are deep, perhaps as much as eight times the effective population size in generations (8N; Carstens & Knowles, 2007).

‘Species tree’ approaches

First proposed over a decade ago by Maddison (1997) and Slowinski & Page (1999), ‘species tree’ approaches aim to estimate the species tree based on data from multiple gene trees without concatenation of the data (Liu et al., 2008; Edwards, 2009). Although these methods have been framed explicitly in terms of integrating data from different genetic markers, these methods could be extended to incorporate other classes of phylogenetic information, and thus form a workable solution for integrative taxonomy. To date, however, the software implementations of these methods address only molecular data. Most of these methods use gene trees as input data and the congruence between individual input trees calculated via various optimality criteria, e.g. likelihood in stem (Kubatko et al., 2009), or maximising the coalescence probabilities of gene trees for a given species tree (esp-coal;Carstens & Knowles, 2007). BUCKy (Ane et al., 2007) and AUGIST (Oliver, 2008) use consensus methods, and both simultaneously accommodate the uncertainty in gene genealogies and providing measures of confidence in species tree estimates. Computationally there is little to stop one from using a morphology tree or tree derived from any non-molecular data source as input trees for these methods, although such an approach may violate some assumptions. For example, stem makes use of branch lengths in its likelihood calculations, and although these can be calculated for morphological trees, few workers would consider these reliable branch-length estimates (but see Omland, 1997). Concordance factors for morphological characters used in BUCKy may not be interpreted easily. Other approaches such as those implemented in best (Edwards et al., 2007; Liu & Pearl, 2007) and *beast (Heled & Drummond, 2008), model the coalescent for sequence data across different input genes to estimate the species tree. Integrative taxonomy is not possible for these other methods, as they cannot handle non-molecular data types in their current implementations. Indeed, we question whether standard character state-based morphological data could be modelled in a coalescent framework.

Although acceptance of species tree methods is growing among phylogeneticists (Edwards, 2009), these methods assume that each terminal is a single metapopulation lineage. In other words, they rely on input species boundaries and are not designed to find them. Carstens & Dewey (2010) are the first to empirically explore statistical tests of species limits by calculating the probability of competing models of lineage composition using AUGIST, stem and best.

Iterative taxonomy

We advocate a process of iterative taxonomy to refine and define species boundaries using multiple lines of evidence. In many cases, different data partitions will circumscribe the same species boundaries (e.g. Tan et al., 2010). For example, variation in molecular data may suggest the presence of multiple sibling species, and congruent species boundaries can be established by re-examining morphological data (e.g. Carew et al., 2005). In cases where different markers, or groups of markers, disagree as to the position of species boundaries, species limits can be tested iteratively. There may be evolutionary interpretations for discordance among data partitions that are consistent with the species boundaries suggested by some evidence (Padial et al., 2010; Schlick-Steiner et al., 2010). In an integrative taxonomy, these explanations have logical consequences that can be tested.

The process of iterative taxonomy can be formulated as follows:

  • 1
    Establish a prima facie (H0) estimate of species boundaries based on one data source using a repeatable protocol, for example morphology using population aggregation analysis (PAA).
  • 2
    Test these boundaries with results from a different dataset (e.g. a fragment of mtDNA), with taxon sampling based on H0, to produce H1.
  • 3
    If H0 and H1 are the same, the current species boundaries have survived testing and iteration can end. When new data become available, return to step 2.
  • 4
    If H0 and H1 propose different species boundaries, search for a biological or evolutionary explanation for the source of discordance.
  • 5
    Based on the results of step 4, refine species boundaries to produce a new hypothesis of species boundaries H2. The H2 species boundaries have survived testing and iteration can end. When new data become available, return to step 2.

The basis of iterative taxonomy is continually to test and retest species boundary hypotheses with new data sources. If data sources are discordant (step 4), then a biological or evolutionary explanation may exist for the disagreement. As a trivial example, PAA may produce two species, but mtDNA may suggest one shared haplotype, with the disagreement explained if the PAA results have overlooked sexual dimorphism within one of the species. In another example, PAA may produce two species, but mtDNA analyses produces haplotypes shared between these species, explicable by incomplete lineage sorting or introgression. The process of iterative taxonomy is independent of the criteria used to establish species boundaries, as the preferred criteria will inform how to address step 4.

Two recent examples illustrate the iterative process of species taxonomy that we advocate. Lumley & Sperling (2010) used larval host association and adult pheromone attraction to establish null hypotheses of species boundaries in spruce budworm moths belonging to the genus Choristoneura. These hypotheses were tested with a combination of mtDNA (a 470-bp region of COI) and linear discriminant analysis of 47 forewing morphometric measurements. Although two of the species hypotheses were corroborated on the Bayesian mitochondrial haplotype trees and networks, three species shared haplotypes. However, these three species could be distinguished using linear discriminant analysis of the forewing morphometrics, thus the combined evidence diagnosed five species in the complex. Introgression or lineage sorting were considered the most likely evolutionary reason for the discordance between the species hypotheses defined by mtDNA and morphological characters. Null hypotheses of species boundaries in the mealybug genus Ferrisia were established using disjunctions in morphological character data (Gullan et al., 2010). These hypotheses were tested by searching for the least inclusive clades among one mtDNA and two nuclear DNA (nDNA) markers. Clades in the combined maximum likelihood (ML) molecular data tree were reconciled with morphological disjunctions to formulate final species-boundary hypotheses, and morphological character evidence was established to diagnose each species. There was no discordance between datasets, and new morphological characters were used for distinguishing species from the positions and characteristics of minute pores occurring ventrally and dorsally on the adult female scale.


  1. Top of page
  2. Abstract
  3. Introduction
  4. Conclusions
  5. References

Our survey indicates that most authors of ‘integrative taxonomy’ have adopted a qualitative inspection of the results of separate partition analyses, and only the most sophisticated analyses conduct these comparisons in a statistical framework. Although species delimitation is emerging as a major issue (Wiens, 2007), we conclude that a repeatable, quantifiable method of truly ‘integrative taxonomy’ is not available yet. The development of an integrative taxonomy is hindered when species delimitation lies at the interface of population genetics and phylogenetics. Concatenation may be the best possible method of integrative taxonomy available currently because it can handle different data types and will work when the underlying species tree is the same as most sampled gene trees. Current surveys suggest that this will be the case most of the time, for mtDNA at least (Funk & Omland, 2003; Joseph & Omland, 2009). Although promising, coalescent species tree methods assume species boundaries, can only be used to infer them with difficulty and are designed for use on molecular sequence data. For the development of these methods as tools in integrative taxonomy, we advocate the further refinement of species tree methods in two ways: (i) the inclusion of non-molecular data types where possible; and (ii) their expansion to include tests of species boundaries by calculating the probability of competing hypotheses of species composition (cf. Carstens & Dewey, 2010).

If morphological data alone are available, these should probably be analysed best using a method such as PAA (Davis & Nixon, 1992; Brower, 1999), which searches for unique attributes of individuals that are fixed in putative species. This is in essence a codification of a standard traditional method for delimiting species based on one or more diagnostic character differences (Sites & Marshall, 2003).

For the present we advocate a process of iterative taxonomy to refine and define species boundaries using multiple lines of evidence. In cases in which different markers, or groups of markers, disagree as to the position of species boundaries, species limits can be tested iteratively. Evolutionary interpretations may exist for discordant data partitions that are consistent with the species boundaries suggested by other evidence. In an interative taxonomy, these explanations have logical consequences that can be tested, with results that allow choice between competing species boundary hypotheses. Iterative taxonomy has the advantage of placing species boundaries in a hypothetico-deductive framework, and this resonates with current practice. Until advances in analytical methods allow the application of a truly ‘integrative taxonomy’, iterative taxonomy as proposed here will represent both the most practical approach and the most defensible, on the basis of classical scientific hypothesis testing.


  1. Top of page
  2. Abstract
  3. Introduction
  4. Conclusions
  5. References