Introduction

After being principally an academic subject for many decades, biodiversity declines are now also a major topic for decision makers (e.g. economists, CEOs, political leaders; World Economic Forum, 2020). This heightened interest has highlighted once more that one of the most important tasks in science is unfinished: the discovery and description of earth’s biodiversity (May, 2011). Most of the species-level diversity and a significant proportion of biomass are concentrated in taxonomically poorly known taxa that are nowadays often referred to as “dark taxa” (e.g. bacteria, fungi, invertebrates; Hausmann et al., 2020). Such groups are usually avoided by taxonomists because they combine high diversity with high abundance. Yet, dark taxa are a major cause for the “taxonomic impediment” that is the result of a range of issues. One is the large number of undescribed species (here called “dark taxon impediment”) whereas another is species descriptions that are too superficial to allow for the identification of the species without inspecting type specimens, collecting additional data and re-describing the species (here called “superficial description impediment”). Addressing the superficial description impediment is particularly time-consuming because it requires museum visits/loans, type digitization and possibly lengthy searches for misplaced specimens. Eventually, the species have to be re-described, making the total process of getting to a useful species description take twice the amount of time needed if there had not been a poor-quality legacy description. The reality is that new species descriptions could be accelerated manifold if taxonomists did not have to spend so much time on improving existing descriptions.

What the 21st Century undoubtedly needs is more species descriptions, but what should be avoided is creating new “superficial description impediments” in the form of large numbers of poorly supported and described species. Yet, this is what Sharkey et al. (2021) propose when they state: “… we view barcode-based descriptions as a first pass in an iterative approach to solve the taxonomic impediment of megadiverse and under-taxonomically resourced groups that standard technical and biopolitical approaches have not been able to tackle.” Sharkey et al. delegate the critical “iterative” work to future generations of taxonomists. They will have to start their own revisions by first revisiting the species descriptions, types and specimens in Sharkey et al. (2021) to resolve the species boundaries based on data that should have been collected and analyzed at the time of description.

Upon closer inspection, Sharkey et al.’s “minimalist revision” relies on three core assumptions/techniques which are here shown to be problematic. The first is that cytochrome oxidase I (COI) barcodes are sufficiently decisive and informative that they can be used as the only/main data source for delimiting and describing species. The second is that species can be diagnosed unambigously using only a consensus barcode. The third is that it is appropriate to use a particular type of barcode cluster (“BIN”) that is delimited by a proprietary algorithm only installed on BOLD Systems.

Are COI barcodes sufficiently informative for delimiting and describing species?

Sharkey et al.’s revision covers 390 BINs obtained from BOLD. Of these, >99% (387) are converted into species. This is surprising because virtually all studies that have looked into the stability of barcode clusters have found that the cluster composition varies depending on which barcode clustering algorithm is used (Virgilio et al., 2010; Kekkonen and Hebert, 2014; Srivathsan et al., 2019; Yeo et al., 2020; Hartop et al., 2021). This suggests that barcode data are not sufficiently decisive for delimiting species—rather, species delimitation must integrate multiple forms of data. We here first tested whether the barcode data used in Sharkey et al. (2021) are unusually decisive and could be used without validation from another source. Conventionally, such re-analyses of data from published studies are straightforward because the data are made available by the authors. However, the “molecular data” in the Supporting Information of Sharkey et al. (2021) consist only of neighbor-joining (NJ) trees. We thus proceeded to obtain the sequences directly from BOLD by consulting the BIN numbers in the revision and NJ trees in the supplement. Note, however, that BINs are calculated based on public and private data with the ratio between the two apparently being ≈3:1 (see below). In addition, the NJ trees in the supplement contained non-Costa Rican BINs from other Neotropical countries that were not covered by the revision. Consequently, we ended up with three datasets of increasing size: (I) public barcodes for only those BINs described by Sharkey et al. (2021); (II) public barcodes for all BINs found on the NJ trees in the Supporting Information; and (III) all public barcodes in the BOLD database for the 11 braconid subfamilies covered by the revision of the Costa Rican fauna (see Appendix S1).

We then proceeded to analyze all three datasets with several species delimitation methods as recommended by Carstens et al. (2013). The overall analyses followed Yeo et al. (2020): objective clustering with uncorrected p-distance thresholds at 1–5% using TaxonDNA (Meier et al., 2006), ASAP—a new implementation of ABGD—where we only kept the results obtained with the top five partitions (= lowest P-values; Puillandre et al., 2012a; Puillandre et al., 2021), and Single-rate PTP (Zhang et al., 2013) based on maximum-likelihood trees generated from the barcode data with RAxML v.8 (Stamatakis, 2014). These re-analyses were used to test whether the different methods unambiguously supported the BINs that were described as species by Sharkey et al. (2021). Arguably, this is a minimum requirement for using BIN = Species as default because it is hard to justify describing barcode clusters as species if the former are not consistently supported across methods.

We find that for Dataset I, 131 of the 401 BINs (32.7%) conflict with at least one of the species delimitation treatments. The incongruence of these BINs increases as more COI evidence is included (Fig. 1), with 283 of the 615 BINs (46.0%) from Dataset II, and 2 912 of the 3 896 BINs (74.7%) from Dataset III being in conflict with the results of at least one of the analyses. The corresponding numbers for only those 401 BINs described as species in Sharkey et al. (2021) are 138 (34.4%: dataset with all Neotropical barcodes from BOLD) and 276 (68.8% dataset with all braconid barcodes from BOLD). This means that as braconid barcodes are sampled more densely across the Neotropics and the world, an increasing number of the BINs described as species in Sharkey et al. (2021) contain specimens that are assigned to barcode clusters that disagree with the BIN assignment from BOLD. Overall, this reanalysis shows that COI barcodes contain only ambiguous signals when it comes to delimiting many species, and that this problem increases with sampling.

In order to obtain lower-bound estimates for incongruence, we identified the parameters for each species delimitation algorithm (objective clustering, ASAP) whose application yielded the largest number of barcode clusters that are identical to BINs; i.e., the parameters from each algorithm that maximized congruence with BINs. We find that 38 of the 401 BINs (9.48%) described in the paper are incongruent even after the methods were optimized to replicate BINs as well as possible (Fig. 2). Incongruence again increases as more of the available barcodes and BINs are included (Dataset II: 82 of 615 BINs incongruent (13.33%), Dataset III: 1 340 of 3 896 BINs incongruent (34.39%)). These analyses illustrate that the available COI data provide ambiguous support for 10–30% of the BINs described as new species in Sharkey et al. (2021). This finding is congruent with virtually all studies that have investigated the congruence between barcode clusters obtained with different clustering methods (Virgilio et al., 2010; Kekkonen and Hebert, 2014; Srivathsan et al., 2019; Yeo et al., 2020; Hartop et al., 2021). Furthermore, there are good reasons to assume that 10–30% is an underestimate of the true number of clusters that require additional data for resolving species boundaries because denser COI sampling will close barcoding gaps between species, thus leading to further uncertainty with regard to the boundaries of barcode clusters (Bergsten et al., 2012).

Sharkey et al. (2021) are aware that broader geographical sampling regularly leads to BINs that contain multiple species: “Braconid specimens from the following New World countries appear to be relatively well-sampled in BOLD: Canada, USA, Belize, Argentina, French Guiana, and Mexico. There is a small number of cases where specimens from these countries fall in the same BIN as one of our Costa Rican species, but they were not studied. More sampling between these disparate localities, and more genomic and/or morphological and behavioral data will help resolve these species-level cases, which are beyond the scope of this paper.” When they occasionally engage with the available evidence, additional problems with BINs are immediately found. For example, for BIN “BOLD:AAG8211” the authors state under “other material” that “BIOUG07751-A11 (Malaise-trapped) from Honduras was examined and considered conspecific, and it is included as a paratype. There are dozens of specimens in the same BIN that are from widespread locations across Canada and the USA. Several of these were examined and they differ significantly in colour from the Mesoamerican specimens and are not considered conspecific. See BOLD for details.” In particular, the North American specimens from Florida should have been consulted before delimiting Costa Rican species, but these specimens were nevertheless neglected.

We suspect that Sharkey et al. (2021) overlooked the problems with delimiting species with COI because they relied too much on BINs supplied by BOLD Systems. As in all scientific studies, it is important that authors analyze the underlying data thoroughly in order to present robust conclusions. For species descriptions involving molecular data, the first step is to evaluate the decisiveness of the data by applying different species delimitation algorithms (Virgilio et al., 2010; Carstens et al., 2013; Kekkonen and Hebert, 2014; Yeo et al., 2020). This analysis strategy also avoids publishing results obtained based on a mixture of public and private data.

All this highlights that relying heavily on BIN = Species for describing species is not advisable. Indeed, depending on the time of BIN description, two of Sharkey et al.’s (2021) wasp taxa would have been described as a fly or stonefly species, respectively. Likewise, biologists using BOLD as an identification tool should be aware that using BINs can lead to misleading conclusions. This makes it all the more important that BIN history can be traced easily. However, this is time-consuming in BOLD Systems. One can only reconstruct BIN evolution by tracing individual specimens—one at a time—and clicking through pages of changes over time; yet, this will still only reveal information on publicly available sequences. This is unfortunate because “taxon concepts” matter greatly in taxonomy (Franz, 2005; Meier, 2016; Packer et al., 2018) and it would be particularly straightforward to establish a “BIN-concept tracking” feature in BOLD Systems given that it is a relatively new database and deals only with one type of data.

The instability of species estimates obtained with barcodes is also evident when evaluating whether the original species diversity estimates in publications are consistent with species richness estimates based on the number of BINs in BOLD. We initially studied the data from publications that were highlighted in Sharkey et al. (2021). We then added the data for additional publications so long as they allowed for a direct comparison of species numbers (=authors provided a list of GenBank accession or BOLD sample IDs). We obtained the sequences in February–April 2021 using these numbers by querying GenBank, the Sequence-ID tool of GBIF and BOLD Systems. Table 1 compares the number of species and BINs for those species that had molecular data. We find that, for example, the barcodes for the ten species in Astraptes belong to five BINs (Hebert et al., 2004), the 32 species of Belvosia reported in Smith et al. (2006) belong to 20 BINs, and the barcodes in Smith et al. (2007) are in 61 BINs instead of the 73 species reported in the paper. These results illustrate how ambiguous COI can be when it comes to delimiting species. For the data from these studies, BINs are either not equal to species or the species reported in the publications were not species.

In this context it is important to remember that there are good theoretical and empirical reasons why one should not expect COI clusters to be equal to species (Meier et al., 2006; Puillandre et al., 2012a; Puillandre et al., 2012b; Zhang et al., 2013). It has long been known that introgression, lineage sorting and maternal inheritance of mt DNA can obfuscate the species-level signal in COI gene data (Will and Rubinoff, 2004; DeSalle et al., 2005). In addition, we know that for closely related taxa the COI protein is under strong stabilizing selection (Roe and Sperling, 2007; Kwong et al., 2012; Pentinsaari et al., 2016), which explains why most of the evolutionary change at the DNA level is synonymous and concentrated in 3rd positions. Nucleotide fixation in these positions is likely caused by genetic drift; thus, COI distances between BINs and closely related species mostly measure time of divergence which is only correlated and not identical to the probability of speciation.

This can be illustrated by the three datasets that we analyzed here. We find numerous BINs whose barcodes translate into identical amino acid sequences (Table S1: Dataset I: 11 cases involving 26 BINs; Dataset II: 21 cases and 49 BINs; Dataset III: 28 cases and 57 BINs—see Appendix S1 for details) and the proportion in other taxa can be as high as 80–90% (Kwong et al., 2012). These BINs have no known, functionally significant biological differences because the biology of these wasp “species” is not at all affected by which triplet codons are used to code for identical proteins. Given that most biologists associate speciation with the origin of biologically meaningful differences, describing such BINs as species rests on the assumption that the correlation between time of divergence and the origin of new species is sufficiently strong that biologically meaningful differences will later be found. However, most biologists are rightfully skeptical of results based on correlations. Correlations warrant more exploration; thus, barcode clusters should only be treated as unnamed first-pass groupings until they are validated with other data and described as species.

An even more serious concern is the nonrandom nature of the error that is caused by BIN = Species. Species originating quickly via an adaptive radiation will be systematically lumped, whereas old species will be systematically split. Evidence for these problems can be found throughout BOLD. Anyone who has worked with the database will have found many cases where one BIN contains sequences that were identified as coming from multiple species. Conversely, it is common to find species that have barcodes in multiple BINs. This can be a consequence of misidentification or genuine lumping/splitting. Fortunately, there are taxa for which there are sufficient morphological, genetic and behavioural data that species boundaries are well-understood. For these, one can test whether the sequences of recently diverged species end up in one BIN and the sequences for old species are found in several BINs. Such data are available, for example, for many species of Sepsidae (Diptera) (Puniamoorthy et al., 2009; Puniamoorthy et al., 2010; Tan et al., 2010; Ang et al., 2013a; Ang et al., 2013b; Araujo et al., 2014; Rohner et al., 2014; Ang et al., 2017). Known pairs of closely related species that are distinct with regard to morphology and behaviour are routinely found lumped into the same BIN (e.g. Sepsis neocynipsea and S. cynipsea: BOLD:AAC2855; Sepsis orthocnemis and S. fulgens: BOLD:AAJ7599; Themira lucida and T. flavicoxa: BOLD:AAD7140). Sepsis punctum is an example of the opposite. The populations in North America and Europe can interbreed but are split into two BINs (BOLD:AAG5639; BOLD:ACS2531). If taxonomists had applied BIN = Species as a default for delimiting these species, several putatively young species would have been overlooked and some old species would have been split. Some will argue that this degree of error is acceptable, but we would counter that the nonrandom nature of the error is a major concern because it systematically misinforms about speciation.

Can species diagnoses consist of only a consensus barcode?

There are additional oddities with regard to the use of consensus barcodes as diagnoses, which highlight that relying only on one type of data is ill-advised. BOLD:ABX6701 was described by Sharkey et al. (2021) as Plesiocoelus vanachterbergi. The consensus barcode includes two single indels and presumably was obtained from the seven “barcode-compliant” sequences in BOLD Systems (as of 18 April 2021). Only one of the seven barcodes is translatable to amino acids with the remaining six having deletions. Such sequences are typical for pseudogenes, and it is conceivable that P. vanachterbergi was delimited and is now diagnosed based on paralogues. In addition, the consensus barcode for Pseudorhysipolis mailyngonzalezae contains 104 indels, which may be related to the fact that the BOLD fasta download for this BIN consists of data for multiple genes (COI, 28S, 16S, Ef1a).

We discuss these issues here because diagnoses are critical elements of species descriptions. They render species hypotheses testable (Ahrens et al., 2021). A diagnosis consisting of only a consensus barcode cannot be tested with morphological data and vice versa. This is one reason why integrative taxonomy is preferred over other approaches. Sharkey et al.’s (2021) reliance on consensus barcodes is particularly problematic because they are static summaries of dynamically evolving barcode databases. Barcoding additional specimens will almost certainly reveal new synonymous nucleotide changes for the >200 species of braconids for which Sharkey et al. (2021) only provide consensus barcodes as diagnoses; thus, these new polymorphisms will falsify the diagnoses. By omitting morphological information for these species, Sharkey et al. also create a parallel taxonomy that does not allow for cross-referencing the new species to the described fauna (Ahrens et al., 2021; Zamani et al., 2021). It furthermore interferes with the work of those biologists, policymakers and citizen scientists in biodiverse countries who have no access to molecular data (Zamani et al., 2021).

Our discussion of consensus barcodes as diagnoses also highlights Sharkey et al.’s (2021) confusing relationship to morphological evidence. On the one hand, Sharkey et al. (2021) argue that morphological evidence should not be used for braconid taxonomy. On the other, a large number of diagnoses contain morphological information, and in many cases the morphological characters are essential for justifying the validity of the species. Morphology presumably also was used to identify those BINs that consist of multiple species. Despite this, no integrative process is described and the minimalist revision failed to provide proper morphological diagnoses/description for many species.

Should BINs obtained from BOLD Systems be used as species estimates?

There are three reasons why the BINs in the “minimalist revision” do not meet the minimum standards for reproducible science. All three are related to open science issues within BOLD Systems. First, the algorithm for calculating BINs (RESL: Refined Single Linkage) is not publicly available. Secondly, it is unclear how the underlying sequences were aligned. Thirdly, BINs are calculated based on public and private barcodes:

Final remarks and suggestions

Sharkey et al. (2021) are pessimistic about new 21st Century solutions to the dark-taxon impediment. We do not share this sentiment. Large throughput imaging and sequencing already are available––these are not vague promises for the future (Hebert et al., 2018; Ärje et al., 2020; Srivathsan et al., 2021; Wührl et al., 2021). Likewise, data now can be analyzed with increasingly sophisticated algorithms that will provide taxonomists with a solid foundation for species descriptions that can be based on multiple sources of data (Puillandre et al., 2012b; Hartop et al., 2021). These data will be particularly suitable for generating automatic species descriptions that are resilient and future-ready. Therefore, this is the worst time to propose “minimalist revisions” that trade scientific rigour for speed.

Conflict of interest

None declared.