The Catalog of the Mosquitoes of the World, published in 1977 by Knight and Stone [1] and its three supplements [2–5], which were based on an earlier 1959 catalog [6] and supplements [7–11], are the primary reference point for all modern mosquito systematics. These catalogs summarize the nomenclatorial organization of the entire Culicidae family and list all references that established those names. The family Culicidae includes 3,601 described species and subspecies (www.mosquitocatalog.org accessed 22 Jan., 2014), a number that steadily increases due to the intense medical and veterinary importance of mosquitoes. The family is organized into two subfamilies, the Anophelinae (482 species) and the Culicinae (3,119 species). Aedini, with 1,261 species, is the largest of the 11 tribes within Culicinae. The genus Aedes in the tribe Aedini is one of three best-known genera of mosquitoes (along with Culex and Anopheles) since many of its species are important vectors of arboviruses and pathogens (Table 1). Because of its medical importance, Aedes is recognized by thousands of researchers and operators concerned with mosquito control, public health, and veterinary health.

This publication restores the generic structure of the tribe Aedini (Diptera: Culicidae) to its status prior to the year 2000. Specifically, all species in the tribe are placed in the following 10 genera: Aedes, Armigeres, Eretmapodites, Haemagogus, Heizmannia, Opifex, Psorophora, Udaya, Verrallina, and Zeugnomyia. Aedine genera created in and after 2000 are placed as subgenera of genus Aedes; subgenera created after 2000 are placed in informal species groupings (S1 Appendix, S2 Appendix). The classification proposed here, and further updates to culicid classification, will be reflected in the online taxonomic catalog at www.mosquitocatalog.org.

The 1977 catalog [1] lists nine genera in the tribe Aedini (Aedes Meigen, 1818; Armigeres Theobald, 1901; Eretmapodites Theobald, 1901; Haemagogus Williston, 1896; Heizmannia Ludlow, 1905; Opifex Hutton, 1902; Psorophora Robineau-Desvoidy, 1827; Udaya Thurman, 1954; and Zeugnomyia Leicester, 1908). Of the 51 subgenera in Aedini, 38 subgenera with about 800 species were placed in the genus Aedes. An “Aedes Group” was mentioned by Edwards in 1932 [58] in a compendium of mosquito taxonomy that preceded the 1959 and 1977 catalogs. His Aedes Group included all of the above except Udaya, which had not yet been described, and Zeugnomyia, which was then considered to be in the Uranotaenia group. Tribe Aedini with all nine genera was first mentioned by Belkin in 1962 [27] in a two-volume work on mosquitoes of the South Pacific that also discussed mosquito systematics and biogeography in a world-wide context.

Since Knight and Stone’s catalog and supplements, five new aedine subgenera were added (Isoaedes [59], Belkinius [60], Albuginosus [61], Kenknightia [62] and Zavortinkius [63]). Also during this period three subgenera of Aedes were elevated to genus, one in 1999 (Verrallina [64]) and two in 2000 (Ayurakitia [65] and Ochlerotatus [66]). Elevation of Verrallina and Ayurakitia to genus went largely unnoticed except by taxonomists. In contrast, because Ochlerotatus contained some well-known species, e.g. Ae. (Och.) sollicitans (Walker), a salt-marsh mosquito and important nuisance and disease vector (Table 1), there was a good deal of debate as to the merit of elevating it to a genus [67]. The classification that included Ochlerotatus as a genus was followed by some workers but ignored by others [68,69]. The recognition of these additional genera and subgenera in Aedini was based on traditional taxonomic judgments of similarities and relationships rather than on quantitative cladistic analysis.

In a series of papers published from 2004 through 2009 [70–73], Reinert, Harbach, and Kitching (henceforth RH&K) made many changes to the nomenclature of the tribe Aedini. The RH&K reclassifications [70–73] were based on comprehensive morphological studies of all life stages of the tribe. Their papers report many entirely new morphological observations, especially on the female genitalia. The authors were not able to examine every species in the tribe, but instead chose those species that were available to them and that they considered representative of previously established groups, usually choosing the type species from a genus or subgenus. Based on their results, RH&K identified monophyletic groups and declared new classifications that were further refined in each publication. The series of papers established 74 genera (S1 Fig, Fig 1) from what was formerly genus Aedes, essentially tripling the number of genera in the family Culicidae. Most of the decisions on classification, particularly in the 2009 paper [73], were based on quantitative cladistic analyses. These decisions were primarily anchored to an objectively defined measure of "Groups present/Contradicted" (GC) [74]. Those groups were identified by pseudo-replication, or jackknifing, using symmetric resampling in an implied-weights parsimony analysis in the computer program TNT [75]. When a grouping was supported by a GC threshold ≥ 40, it was considered cladistically significant [73]. However, the authors departed from this quantitative scheme by recognizing clades with GC<40 when the “number and nature of the characters that support the branches” were subjectively considered significant (Fig 2). RH&K did not propose diagnostic characters for the new genera, though they did assign many species in Aedini to genus or subgenus based on accepted groupings and their own analyses of characters. Exceptions were: genus Ochlerotatus, 138 species unassigned to subgenus; ‘Ochlerotatus’ (‘Finlaya’) sensu auctorum (= a name used by subsequent authors in a sense different from that established by the original author [76]), 28 species; ‘Ochlerotatus’ (‘Protomacleaya’) sensu auctorum, 39 species; and genus Stegomyia, 104 species unassigned to subgenus.

Although phylogenetic studies of DNA sequences should ultimately be helpful in determining taxonomic relationships between groups within Aedini, the few studies to date have not been able to test the validity of the reclassification proposed by RH&K based on morphological characters. DNA barcoding studies have examined 217 aedine species in China [77], India [78], and Afghanistan [79]. Of note, barcoding involves comparison of short sequences of the COI mitochondrial gene, producing distance measures that are only useful for delineating and differentiating species-level taxa, not for estimating their relationships. Combining COI and COII mitochondrial genes from 17 aedine species in western Africa [80] failed to distinguish them as a group from three non-aedine species and also failed to show significant groupings of species. Restriction mapping of rDNA [81] was similarly unsuccessful for a study of 12 culicine species representing six genera. Analysis of the white gene [82] successfully distinguished three aedine species from nine others, but the sample size was too small to examine relationships among the aedine species.

The Linnaean system of nomenclature that has served biology since 1758 was formalized in the International Code of Zoological Nomenclature (ICZN), a basic set of rules designed to promote stability and communication [76]. Classifications governed by the ICZN are meant to be “natural,” a term that since the translation of Hennig’s “Phylogenetic Systematics” [83] has been defined as strictly monophyletic, meaning “including all descendants from a common ancestor.” Wiley [84,85], one of the early theorists of phylogenetic systematics, asserted that an annotated Linnaean classification can take more than one form as long as each is logically consistent with a phylogeny, and that “… minimum taxonomic decisions will be made whenever possible to construct a classification or to modify an existing classification.” Wiley and Liebermann [86] also stated that “the ranks of well-known clades will be retained whenever possible.” Although explicitly "tree-based" classification systems have been suggested as alternatives to Linnaean classification (e.g., numerical prefix schemes [87] and the PhyloCode [88]), these alternative schemes require a well-supported phylogeny based on multiple lines of evidence, a condition not yet reached in Aedini as discussed above.

Regardless of the system, some accommodations are necessary for serving the sometimes non-overlapping requirements of taxonomy, information retrieval, and communication [86]. Recently, Vences and colleagues [89] discussed the relationship between nomenclatorial utility and phylogenetic accuracy in an expansion on previous treatments of the subject [84–86]. They acknowledged that taxonomic instability “…can become a serious problem consuming public and private resources…”. To address this and other issues, they proposed several levels of taxon naming criteria (TNCs), the priority criteria being monophyly, clade stability, and phenotypic diagnosability. Accessory criteria included, among others, the medical or economic significance of a taxon. These authors emphasized that “…except for monophyly, the priority TNCs are not proposed as mandatory requirements of a Linnaean taxon but as yardsticks to allow for an informed choice among various clades in a tree…” and warn that “…taxa of unstable monophyly or poor diagnosability reduce the information content and hence the utility of the Linnaean system.”

After reviewing the history of the classification of tribe Aedini and previous morphological and molecular phylogenetic analyses, we reanalyzed RH&K data and performed a critical assessment of their methods and results in the context of Wiley [84,85], Wiley and Lieberman [86]’s, “conventions for annotated Linnaean classifications”, and also Vences et al. [89]’s, “taxon naming criteria”. To preserve information about morphological evolution in mosquitoes, we suggest alternative methods for reflecting RH&K’s findings in a classification that retains the stability necessary for effective information retrieval and communication.

We reanalyzed RH&K’s data in the program TNT [75,90], not for the purpose of producing a classification, but rather to test the robustness of RH&K’s [73] generic groups (Fig 3; S1 Tree, S2 Tree, S3 Tree, S4 Tree) under alternative, but similar, cladistic analyses. We assumed that if proposed relationships among genera are not robust when subjected to closely related analytical methods, they are unlikely to be robust to the introduction of new data. To gauge the stability of RH&K’s [73] generic groupings to standard parsimony-based analytical methods, we reanalyzed their data with two alternative methods: (i) unweighted parsimony jackknife analyses, identifying clades and their GC (Groups present/Contradicted) support values [74] on the resulting jackknife tree, and (ii) unweighted maximum-parsimony analyses, identifying clades occurring on the strict consensus of the resulting most parsimonious trees. For each of these analytical methods, we conducted two sets of multiple analyses, one in which all characters were unordered and one in which 14 characters were ordered, as was the case in a subset of the original analyses [73]. We performed all analyses in TNT v.1.1 [75,90], although in some cases we carried out tree-bisection and reconnection (TBR) branch-swapping on accumulated trees in PAUP* [91].

Jackknife analyses. We conducted unweighted parsimony jackknife analyses consisting of 1,000 pseudoreplicates, each pseudoreplicate consisting of 10 separate tree searches utilizing TBR branch-swapping on a random-taxon-addition Wagner tree and limited to finding 10 most-parsimonious trees. Support for clades was measured with the GC (Groups present/Contradicted) criterion [74] based on the TNT default tree output, which displays a tree consisting of all groups with positive GC values (S1 Tree, S2 Tree). Following RH&K’s criterion, we report nodes with GC≥ 40.

Maximum parsimony analyses. We conducted multiple, progressive, unweighted parsimony analyses in which resulting trees were saved, accumulated trees were subjected to TBR branch-swapping, and duplicate trees were discarded. These analyses employed varying combinations of TNT "new technology" algorithms, including the parsimony ratchet (5 to 100,000 iterations), sectorial searching (RSS and CSS), tree drifting (10–1000 cycles), and tree fusing (5–100 rounds), in which accumulating trees from prior searches were carried over into subsequent searches. Multiple driven searches were also conducted, including (i) 1,000 initial random-taxon-addition sequences with five rounds of ratcheting and three rounds of tree-fusing and requiring that minimum-length trees be found 10 times; and (ii) combined sectorial, ratchet, tree-drifting, and tree-fusion searches requiring the consensus tree to stabilize four times (S3 Tree, S4 Tree).

GC support criterion. GC is calculated as the difference between the frequency with which a given group is retrieved (e.g., in the jackknife replicates) and the frequency with which its most frequent contradictory group is retrieved [74]. Absolute frequencies (the usual method of counting frequencies in jackknife or bootstrap analyses) do not distinguish between a group with a frequency of 60% that is uncontradicted by an alternative grouping (GC = 60) and a group with a frequency of 60% that is contradicted by an alternative grouping that may occur, e.g., 20% of the time, which would produce GC = 40 [74].

Ordered characters. Following RH&K, the 14 ordered characters (out of a total of 336 characters) are all measurement or ratio characters, ordered as suggested by Thiele [92] (characters 13, 14, 17, 20, 62, 67, 72, 84, 107, 138, 161, 273, 280, 318; for TNT characters 12, 13, 16, 19, 61, 66, 71, 83, 106, 137, 160, 272, 279, 317).

A series of 20 maximum-parsimony searches in which all characters were unordered converged multiple times on the same set of 832 most-parsimonious trees of unweighted length = 6,609 (S4 Tree). A series of 51 maximum-parsimony searches in which 14 characters were ordered produced 4,000 most-parsimonious trees of unweighted length = 6,654 (S3 Tree), but searches continued to find additional trees, so it is likely that not all of the most-parsimonious trees were found. Finding all most-parsimonious trees, however, is not necessary for calculating the strict consensus [93]. The majority of tree islands were likely visited and the strict consensus is likely stable because all 51 searches, including the driven consensus-stabilization searches, employed the parsimony ratchet [94]. At best, any additional most-parsimonious trees would have no effect on the consensus and, at worst, they would further erode its resolution.

Regardless of the specific analysis used (unweighted parsimony or strict consensus, ordered or unordered: see S1 Tree, S2 Tree, S3 Tree, S4 Tree), 48 supported clades, including all 10 pre-2000 Aedine genera, were supported by GC values ≥ 40 (Fig 3). These included 45 of the 46 clades proposed by RH&K, with the exception of the proposed genus Catageiomyia that split into two unsupported clades in our analysis (Fig 3). In addition, our analysis supports three clades corresponding to the proposed genera Danielsia, Hopkinsius and Petermattinglyius (Fig 3), which were not supported in the analysis of RH&K (Fig 1 in [73]). Significantly, our analyses split Aedimorphus into four distinct clades, none of which were supported by GC values ≥ 40 (Fig 3). Numerous instances of species groupings unsupported according to RH&K’s criteria but nonetheless ultimately elevated to generic status [73] were also not supported in our analyses (S1 Tree, S2 Tree, S3 Tree, S4 Tree, Fig 3). Most importantly, we found minimal support for the phylogenetic relationships among taxa (i.e., very low GC or jackknife support at branches), underscoring the low phylogenetic resolution of the character set used to assign ranks among the taxa. Detailed examples are discussed fully in the next section.

As we would expect, the implied-weight parsimony method used by RH&K [73] resulted in a more resolved phylogeny (Fig 2) than the standard equally weighted parsimony strategy we employed (Fig 3). In particular, and as admitted by the authors [73], the high level of precision of the floating-point arithmetic used to calculate fit in the TNT implementation of the implied-weights method is unlikely to find more than a single optimal cladogram. A single-solution, highly resolved cladogram may serve a valuable purpose in posing bold hypotheses maximally vulnerable to refutation by the addition of new data (i.e., taxa or characters) [95]. However, the potential instability that makes monophyletic groups on optimal trees ideal for the generation of bold hypotheses can make those same groups poor candidates upon which to base stable classifications. Because they recognized this problem, RH&K [73] utilized the support criterion of "Groups present/Contradicted" (GC) ≥ 40, based on jackknifing with symmetric resampling, to judge clades on their implied-weights parsimony tree (Fig 2). Although we applied this same criterion (i.e., GC ≥ 40) in our alternative unweighted parsimony analyses, we found negligible support for the basal branches uniting the terminal taxa, i.e., for higher-level relationships between supraspecific groups (Fig 3). Without a better understanding of the relationships among terminal taxa, application of the principle of equivalent ranking advocated by RH&K is unfeasible.

In light of these results and using Vences et al. [89] as a guide, we make a point-by-point evaluation to assess whether the classification of Aedini as proposed by RH&K [73] meets the taxonomic naming criteria (TNCs) described. Of 11 TNCs discussed by Vences et al. [89] the following seven are applicable to our argument (quotes are from Vences et al. [89] unless otherwise noted).