Parsimony (MP) and maximum likelihood (ML) analyses of combined sequences from the chloroplast genes rbcL and matK, the chloroplast intron trnL, and the spacers rpl20-rps12 and trnL-F, under the GTR + G + I model yielded a topology (Fig. 1) that was congruent with that obtained from the nuclear internal transcribed spacer region (Fig. 2). Chloroplast and nuclear data were therefore combined, and a parsimony tree from the combined data with MP and ML bootstrap support is shown as Fig. 3 (seven of the species lack ITS sequences, Table 1). In the family-wide analysis (with 123 of 130 genera of Cucurbitaceae sequenced), Cucumis is sister to Muellerargia (Fig. 4). The genera Cucumella, Dicaelospermum, Mukia, Myrmecosicyos, and Oreosyce are embedded among species of Cucumis (Fig. 3). The remaining genera of Cucumerinae sensu C. Jeffrey 11, Cucumeropsis Naudin, Melancium Naudin, Melothria L., Posadaea Cogn., and Zehneria Endl. (plus Neoachmandra and Scopellaria 24) group far from Cucumis (Fig. 4). This fits with their geographic concentration in the New World (where Cucumis is absent): Melancium is a monotypic genus from Brazil, Posadaea a monotypic genus from tropical America, and Melothria has ten species in Central America and South America. However, Cucumeropsis, with a single species from tropical Africa, and Zehneria, Neoachmandra, and Scopellaria, with 66 species in tropical and subtropical Africa, Madagascar, Asia, New Guinea, and Australia 24 overlap with the natural range of Cucumis.

The sister genus to Cucumis, Muellerargia, consists of one species in Madagascar and one in Indonesia and Queensland. Both are herbaceous trailers or climbers with straight or apically reflexed anthers and softly spinose fruits. Muellerargia has never been recognized as closely related to Cucumis 1225, perhaps because it is extremely poorly collected, with but a few specimens even in major herbaria: The Madagascan species, Muellerargia jeffreyana Keraudren, is known from three collections (in the Paris herbarium), and permission was not granted to sacrifice material for this study. It is morphologically similar to the Indonesian-Australian species M. timorensis Cogn. 26. The poor documentation of the genus in herbaria also led to the Australian species being described at least three times; first as Muellerargia timorensis Cogn., then as Melothria subpellucida Cogn., and then as Zehneria ejecta Bailey (syn. Melothria ejecta (Bailey) Cogn.).

The Cucumis species relationships found here differ from those found in earlier studies 1202122. An unrooted nuclear isozyme tree 21 showed C. sativus as the genetically most distant species, while C. melo was sister to an African clade. The neighbor-joining tree from nuclear ITS sequences of Garcia-Mas et al. 22 was rooted on Citrullus lanatus and Cucurbita pepo, and showed C. sativus as the first-branching species in the genus, while C. melo was sister to a large African clade. Finally, the chloroplast tree of Chung et al. 1 also was rooted on Citrullus and showed C. sativus and C. hystrix as sister to C. melo (as did studies focusing on C. sativus; e.g., 27). By contrast, the data presented here (Figs. 1, 2, 3, 4) indicate that (i) the deepest divergence in Cucumis is between C. hirsutus and C. humifructus on the one hand and all other species on the other, (ii) C. sativus (cucumber) and C. hystrix are closer to Dicaelospermum and Mukia than they are to any species of Cucumis, and (iii) C. melo (melon) is sister to a clade comprising Dicaelospermum, Mukia, C. sativus, C. hystrix, and a new species from Australia (HS414).

There are several possible explanations for the contrasting findings of the earlier phylogenetic studies. First, the use of distant outgroups might have "attracted" the long-branched (i.e., mutation-rich) C. sativus, pulling it to the base of the tree. Garcia-Mas et al. 22 and Chung et al. 1 used Citrullus lanatus and/or Cucurbita pepo as sole outgroups. Both taxa are many clades, and millions of years of evolution, removed from the Cucumis clade (Fig. 4) and therefore add long branches to neighbor-joining and parsimony analyses 122. The inclusion of these long branches could have caused long-branch attraction between them and C. sativus.

A second reason why previous molecular phylogenetic studies were unable to test the monophyly of Cucumis and to infer the sister clades of cucumber and melon is that they did not include a sufficiently broad sample of taxa. For example, rigidly testing the monophyly of Cucumis section Melo required sequencing all of its species, C. melo, C. hirsutus, C. humifructus, and C. sagittatus. Results (Figs. 1, 2, 3) show that C. hirsutus and C. humifructus, rather than being close to C. melo, are sister to all other species of Cucumis sensu lato, that is, including all five genera nested in Cucumis.

Another possible reason for apparent differences between earlier topologies and the phylogeny found here is insufficient signal in the data and misidentified material. Comparison of the ITS sequences of Garcia-Mas et al. 22 to our ITS sequences showed that the sequence labeled Cucumis membranifolius in GenBank (AJ488223) and Oreosyce africana in the published paper (these names refer to the same species fide 12), does not represent Oreosyce africana. The sequence came from a seed provided by the North Central Regional Plant Introduction Station in Ames, Iowa, and since there is no voucher, its identification cannot be verified. We also could not reproduce the topology and bootstrap support obtained in the original paper 22, partly probably because the phylogenetic signal in the data is weak, resulting in many equally likely trees. Garcia-Mas et al. 22 included sequences resulting from direct sequencing as well as sequences obtained by pGEM-T Easy Vector cloning and found sequences from multiple accessions generally grouping by species. Our Cucumis ITS sequencing confirmed these authors' assessment that ITS lineage sorting is not a problem in Cucumis. The two C. ficifolius sequences obtained by Garcia-Mas et al. 22 that do not group (Fig. 2) come from different plants and may simply represent different species; however, since the material is unvouchered, the identifications cannot be checked.

The phylogeny from the combined nuclear and chloroplast data (Fig. 3) implies that the deepest divergence lies between the common ancestor of C. hirsutus and C. humifructus and the stem lineage of the remainder of the genus. From the geographic ranges of the species of Cucumis sensu lato (i.e., the natural clade identified here) and its sister genus Muellerargia (with one species in tropical Australia and Indonesia, the other in Madagascar), the area where Cucumis may have originated cannot reliably be inferred. Strict and semi-parametric molecular clocks indicated that the deepest divergence in Cucumis may date back to 48–45 my and that the split between the C. melo lineage and its Australian/Asian sister clade is only slightly younger. The divergence of C. sativus from C. hystrix may be about 8 my old and that of their common ancestor from the ancestor of Dicaelospermum and Mukia maderaspatana about 19 my. The bulk of the African species appears to have evolved more recently. The absence of a fossil constraint within Cucumis, however, cautions against over-confidence in the molecular clock estimates.

Based on the tree (Fig. 3), the earliest divergence events in Cucumis likely took place in Africa. However, contrary to the traditional classification 12, which groups C. melo with the African C. hirsutus, C. humifructus, and C. sagittatus, melon is closest to an Australian/Asian clade (marked in grey-green in Fig. 3) that comprises an undescribed Australian species 28, species currently placed in Mukia (M. javanica, M. maderaspatana), Dicaelospermum ritchiei from Western India (recently transferred to Mukia 29), and Cucumis sativus and C. hystrix from India, China, Burma, and Thailand. In addition to the two species we sequenced, Mukia comprises three others 29, its overall geographic range extending from Indo-China southeast to Java, Borneo, and the Philippines, and west through India, Pakistan, and the Yemen into sub-Saharan Africa. Given the geographic distribution of its extant closest relatives (Fig. 3), C. melo itself could have originated somewhere in Asia and then reached Africa from there, rather than originating in Africa as traditionally assumed 1415. Notably, Indian melon landraces exhibit the largest isozyme variation among Asian melons 16 and Australia is a center of complex morphological variation of C. melo 28.

The evolution of morphological traits relevant for Cucumis breeders, for example fruit type, habit, and sexual system, will need to be reinterpreted based on the phylogenetic relationships presented here. Most of the 52 described species in the Cucumis clade are monoecious perennials, and the monoecious sexual system and perennial habit may be the ancestral condition from which an annual habit and dioecy appear to have evolved several times. However, the sexual system and habit of key taxa, such as Dicaelospermum, Muellerargia, and the as yet undescribed species from Australia 28 (Fig. 3) are not reliably known because species are under-collected and have not been studied in the field. Of the 17 species of Cucumis not yet sequenced, most are monoecious perennials; only C. kalahariensis A. Meeuse and C. rigidus Sond. are dioecious and perennial. Species currently placed in Mukia 29 and Cucumella 13 are mostly monoecious and perennial. The evolution of smooth fruits from spiny fruits, a traditional key character in Cucumis, and the mode of fruit opening are much more plastic than formerly thought. For example, in Oreosyce africana and Muellerargia timorensis the fruits open explosively [30; I. Telford, Beadle Herbarium, Armidale, personal communication, Feb. 2007); in C. humifructus, fruits mature below ground and are then dug up and the seeds dispersed by antbears, Orycteropus afer 31; in the new species from Australia (HS414 in Figs. 1 and 2), the developing fruit is pushed into rock crevices by the elongating pedicel and also matures below ground; and in Myrmecosicyos messorius the fruits are tiny and apparently dispersed by harvesting ants around whose nest entrances the species grows.

Table 1 lists all species sampled with authors, status as generic types where applicable, plant sources, and GenBank accession numbers [TreeBASE: study accession S1604, matrix accession M2887, M3250 and M3251]; 79 chloroplast and 20 ITS sequences were newly generated for this study. Species concepts and generic assignments throughout this study follow recent classifications 11121329, although as a result of this study, species in several genera have been transferred into Cucumis (34; this also provides a morphological key to the 52 described species). To resolve species relationships within Cucumis, we added sequences from the nuclear internal transcribed spacer region (220 nt of ITS 1, 163 nt of the 5.8S gene, and 240 nt of ITS 2) for the same species for which chloroplast data were generated. DNA extraction, purification, and sequencing of the selected loci followed standard procedures 10. All PCR products were sequenced in both directions. Direct PCR amplification of ITS yielded single bands and unambiguous base calls, except in C. ficifolia, the sequences of which were therefore not used. Sequences were edited and assembled with the Sequencher software (Gene Codes) and aligned by eye, using MacClade 35. The aligned chloroplast matrix comprised 4,375 positions after exclusion of a poly-T run in the matK gene, a poly-A run in the trnL intron, a TATATA microsatellite region in the trnL-F intergenic spacer and a poly-A run in the rpl20-rps12 intergenic spacer. The aligned ITS matrix comprised 677 aligned positions, and we excluded a poly-G stretch of 25 nt and a poly-C stretch of 17 nt from the ITS1 and a poly-C stretch of 11 nt from ITS2.

Equally weighted parsimony analyses were conducted using PAUP 4.0b10 36. The search strategy involved 100 random taxon addition replicates with tree-bisection-reconnection branch swapping, MulTrees and Steepest Descent in effect, no limit on trees in memory, and saving all optimal trees. For MP analyses, gaps were treated as missing data, while for ML searches (below) they were mostly removed. To assess node support, parsimony bootstrap analyses were performed using 1000 replicate heuristic searches, each with 10 random addition replicates and otherwise the same settings as used for tree searches. More computationally intensive heuristic approaches have been found not to increase the reliability of bootstrapping 37. Maximum likelihood analyses and bootstrapping were performed using GARLI 0.951 38. GARLI searches relied on the GTR + G + P-invar model, which ModelTest 3.06 39 selected as the best fitting model for the combined data. Parameters were estimated over the duration of specified runs.

Molecular clock dating in Cucurbitaceae is problematic because of the family's scarce fossil record. Without multiple calibrations, such as could come from several securely assigned fossils, relaxed molecular clock methods have been shown to perform poorly 404142. We therefore relied on a strict clock approach and compared it with results obtained with the semi-parametric penalized likelihood approach 40 implemented in r8s vs. 1.7). For strict clock dating, we employed the maximum likelihood topology obtained (under GTR + G) with the family data set of Kocyan et al. 10 augmented by the Cucumis sequences generated for this study for a total of 193 taxa and 5,028 aligned nucleotide positions. The tree was imported into PAUP 36, rooted on Corynocarpaceae, and rbcL branch lengths were then calculated under a GTR + G + I + strict clock model. Branch lengths were saved and a mutation rate obtained by dividing the distance from the most recent common ancestor (mrca) of Trichosanthes to the present (0.01416) by 65 my, based on the oldest seeds assigned to this genus 43. Using the resulting rate of 0.000218 substitutions/site/my, we obtained an age of 47.6 my for the mrca of Cucumis by dividing the distance from the basal divergence in Cucumis to the present (0.01037) by 0.000218. The time of the divergence of C. melo from its sister clade was calculated accordingly (0.00975 : 0.000218 = 44.7 my). To check this strict clock estimate based on rbcL, we imported the 193-taxon-ML tree with branch lengths from the combined chloroplast data (5,028 nt) into r8s and ran a cross validation analysis, using the following upper and lower temporal constraints. The mrca of the family Cucurbitaceae was constrained to maximally 100 my and minimally 65 my old based on Cucurbitales family relationships and fossil records 943; the mrca of Trichosanthes was constrained to minimally 65 my 42; and the mrca of an endemic clade of two species occurring on Hispaniola was constrained to maximally 30 my old based on the oldest ages of Dominican amber 44. Penalized likelihood yielded an age of 44.9 my for the mrca of Cucumis.