Phylogeny, Floral Evolution, and Inter-Island Dispersal in Hawaiian Clermontia (Campanulaceae) Based on ISSR Variation and Plastid Spacer Sequences

Previous studies based on DNA restriction-site and sequence variation have shown that the Hawaiian lobeliads are monophyletic and that the two largest genera, Cyanea and Clermontia, diverged from each other ca. 9.7 Mya. Sequence divergence among species of Clermontia is quite limited, however, and extensive hybridization is suspected, which has interfered with production of a well-resolved molecular phylogeny for the genus. Clermontia is of considerable interest because several species posses petal-like sepals, raising the question of whether such a homeotic mutation has arisen once or several times. In addition, morphological and molecular studies have implied different patterns of inter-island dispersal within the genus. Here we use nuclear ISSRs (inter-simple sequence repeat polymorphisms) and five plastid non-coding sequences to derive biparental and maternal phylogenies for Clermontia. Our findings imply that (1) Clermontia is not monophyletic, with Cl. pyrularia nested within Cyanea and apparently an intergeneric hybrid; (2) the earliest divergent clades within Clermontia are native to Kauài, then Òahu, then Maui, supporting the progression rule of dispersal down the chain toward progressively younger islands, although that rule is violated in later-evolving taxa in the ISSR tree; (3) almost no sequence divergence among several Clermontia species in 4.5 kb of rapidly evolving plastid DNA; (4) several apparent cases of hybridization/introgression or incomplete lineage sorting (i.e., Cl. oblongifolia, peleana, persicifolia, pyrularia, samuelii, tuberculata), based on extensive conflict between the ISSR and plastid phylogenies; and (5) two origins and two losses of petaloid sepals, or—perhaps more plausibly—a single origin and two losses of this homeotic mutation, with its introgression into Cl. persicifolia. Our phylogenies are better resolved and geographically more informative than others based on ITS and 5S-NTS sequences and nuclear SNPs, but agree with them in supporting Clermontia's origin on Kauài or some older island and dispersal down the chain subsequently.


Introduction
The Hawaiian lobeliads (Campanulales: Campanulaceae) are the largest family of flowering plants (6 genera, ca. 130 species) native to the Hawaiian Islands, and have undergone specta cular adaptive radiations in growth habit, floral morphology, habitat, leaf form, and photosynthetic physiology [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16]. Recent analyses based on DNA sequence variation indicate that this group is the product of a single colonization event about 13 million years ago (Mya), and that the two largest genera -Cyanea (ca. 80 spp.) and Clermontia (22 spp.) -diverged from each other roughly 9.7 Mya near the now eroded seamounts of the French Frigate Shoals and Gardner Pinnacles, long before the oldest tall island today (Kauài) emerged above the Pacific 4.7 Mya [14]. Previous research on Cyanea, the largest genus of flowering plants native to Hawaii, produced a partial phylogeny based on plastid DNA restrictionsite variation which showed a strong relationship between geographic distribution and ancestry [8,9]. Most of the dispersal events implied by that tree were from one island to the next, younger island in the Hawaiian chain, supporting the so-called ''progression rule'' [17], with most of the inferred inter-island dispersal events having been from one island to the next younger, ecologically unsaturated one in the chain, as expected from ecological theory [8,9].
Extant Cyanea species began diversifying soon after the genus split from Clermontia, whereas extant Clermontia began diverging from each other only in the last 5 My [14]. As a consequence, genetic divergence among species of Clermontia is quite limited, and it has proven difficult to derive a phylogeny for Clermontia based on either morphological or molecular variation. Lammers [6] derived a moderately resolved phylogeny of Clermontia based on a maximum-parsimony analysis of morphological characters, and used that phylogeny to infer the origin of the genus (or at least the extant lineage) on the youngest island of Hawai`i, which emerged from the Pacific no more than 0.5 Mya, and to imply several dispersal events back up the Hawaiian chain to older islands. This contrasts with the conclusion by Givnish et al. [9,14] -based on plastid restriction and sequence data, though with few species of Clermontia included -that Clermontia and its sister genus Cyanea arose ca. 9.7 Mya, on islands now reduced to pinnacles and reefs, with most dispersal events down the chain to younger islands. However, six of Lammers' 32 characters -including sepal size, shape, color, texture, fusion, and persistence -hinge simply on whether the sepals are like or unlike the petals. Hofer et al. [18] recently proposed that petaloid sepals in Clermontia represent a homeotic mutation, involving the ectopic expression of PISTIL-LATA B-function MADS box gene homologs in the outermost floral whorl, and present gene-expression data supporting their view. Given this, Lammers' six sepal characters should be coded as only one -petaloid sepals present or absent. When we re-analyze Lammers' data making this change, almost all resolution disappears in the unweighted strict consensus tree, including the distinction between the two traditional sections based on sepaloid petals being present (sect. Clermontia) vs. absent (sect. Clermontioides) (see Results). Hofer et al. [18] presented a molecular phylogeny for Clermontia based on cloned sequences of the non-transcribed spacer of 5S rDNA. Although poorly resolved, their tree does place the Kauài endemic Clermontia fauriei sister to all other Clermontia species, consistent with an origin for the genus on Kauài or some older island. Pillon et al. [19] also present a poorly resolved phylogeny for Clermontia, based on 31 single-nucleotide polymorphisms (SNPs) for .200 accessions; their analysis places Cl. fauriei sister to all other members of the genus except Cl. pyrularia, which is instead resolved as part of Cyanea. Several Clermontia hybrids of recent origin have been suspected based on their having a morphology intermediate to that of the putative parental taxa, including apparent crosses involving Cl. calophylla, Cl. drepanomorpha, Cl. hawaiiensis, Cl. kohalae, Cl. montis-loa, and Cl. parviflora on Hawai`i, as well as between Cl. kakeana and Cl. micrantha on Maui [1,2,5,6,7].
Studies of evolution within Clermontia have been stymied by a lack of a reliable, well-resolved phylogeny based on markers that evolve with sufficient rapidity. The lack of comparisons between trees for Clermontia based on maternally inherited plastid markers and biparentally inherited nuclear markers has similarly hampered progress toward identifying and understanding instances of potentially widespread reticulate evolution within the genus. Here we present data from rapidly evolving, predominantly nuclear ISSRs (inter-simple sequence repeat polymorphisms) and five plastid spacer or intron sequences, and employ them to create biparental and maternal phylogenies for Clermontia, respectively. We then use these trees to test the monophyly of Clermontia, evaluate the extent of genetic divergence within the genus, infer the pattern of inter-island dispersal during its evolution, reconstruct potential cases of reticulate evolution, and identify the number and timing of origins and losses of the apparent homeotic mutation for petaloid sepals.

Phylogenetic analysis of morphological characters
We re-scored Lammers' morphological dataset (32 characters 6 25 taxa, including one generalized Cyanea and two geographic populations each for Cl. oblongifolia and Cl. peleana) for Clermontia [6] by excluding five of the six characters representing the presence/ absence of petaloid sepals, retaining sepal texture. The remaining 27 characters were then re-analyzed using maximum parsimony (MP) in PAUP* 4.0b8 [20]. We initiated 10,000 replicate MP searches with random starting trees and used TBR branchswapping. This approach increased the chance of detecting multiple islands of equally parsimonious trees. We formed the strict consensus of all shortest trees recovered. Bootstrap support for each node was estimated using 1000 random resamplings of the data, with each replicate search seeded by a new random-addition tree. Given the few morphological characters, we also used successive reweighting of characters based on the consistency index (CI) of each character [21,22], and recalculated the strict consensus tree and bootstrap values for each node using the reweighted data.

Phylogenetic analysis of ISSR and cpDNA characters
Genomic DNAs used in this study were collected and extracted for previous analyses [8,9]. Thirty-four accessions representing all 21 extant species of Clermontia were surveyed, along with two outgroups consisting species of the sister genus Cyanea (Cy. leptostegia and Cy. pilosa ssp. longipedunculata, representing the purple-and orange-fruited clades [8,9]). For each species of Clermontia which occurs on more than one island, or in more than one biogeographic region within an island (e.g., East vs. West Ò ahu), samples from each island and biogeographic region were included whenever possible. Table 1 lists the populations represented; nomenclature follows Lammers [7]. Thirty-two accessions were included in the ISSR survey; 31, in the plastid sequence study (Table 1).
ISSR variation. The ISSR technique is based on surveying variation in the length of the spacers between short simple repeats (SSRs) of a particular kind [23,24]. In practice, this was achieved by using PCR primers that anneal to a specific SSR to amplify the intervening spacers. After amplification, PCR products were sizeseparated by electrophoresis; bands of ISSR DNA of different lengths were visualized by staining the gel with ethidium bromide and photographing it on a UV light table. Different taxa were scored by running ISSRs generated by a given primer pair side-byside on the same gel, and then tallying the presence/absence of different bands across taxa.
We screened 58 different 2-and 3-base-pair repeat primers for their effectiveness in producing useful, repeatable, interpretable levels of variation ISSR length in four taxa ( Table 2) were chosen to survey ISSR variation across populations because they produced repeatable band polymorphisms, involving a large number of co-migrating bands that were easily scorable by eye. Given the relative sizes of the nuclear genome of Hawaiian lobeliads (4C = 4106 Mb for Brighamia insignis, Kew C-value database, http://data.kew.org/cvalues/) vs. that of the plastid genome (ca. 0.15 Mb), essentially all ISSR bands should reflect variation in nuclear DNA.
We used a slight modification of the Tsumara et al. [25] protocol for ISSR amplification. PCR reactions consisted of 28.5 ml double-distilled H 2 O, 4.0 ml 25 mM MgCl 2 , 5.0 ml 10X Mg-free reaction buffer, 8.0 ml 1.25 mM dNTPs, 2.5 ml 4 mg/ml BSA, 1.0 ml formamide, 0.5 ml 20 M primer, 0.25 ml Taq polymerase, and 0.5 ml template DNA. PCR was conducted using a Perkin-Elmer Gene Amp 2400 thermal cycler, using the following program: initial denaturation of 7 min at 94uC; then 45 cycles of 30 s at 94uC, 45 s at 52uC, and 120 s at 72uC; and a final extension period of 7 min at 72uC. Amplified products were size-separated on 2% agarose gels using 0.5X TBE buffer, stained with ethidium bromide, photographed on a UV light table, and visually scored for the presence/absence of individual bands.
Plastid sequence data. DNA samples were sequenced for four rapidly evolving plastid intergenic spacers -atpB-rbcL, trnL-trnF, trnT-trnL, trnV-trnK -and the rpl16 intron following Givnish et al. [14]. Three taxa surveyed for ISSR variation (Cl. grandiflora from East Maui, Cl. micrantha, and Cl. oblongifolia from East Ò ahu) were not sequenced for lack of DNA, leaving a total of 31 accessions. All new sequences were uploaded to GenBank and accession numbers obtained (Table 1). Phylogenetic analyses. ISSR and plastid sequence data were analyzed independently using MP in PAUP*, using the same approach as for morphological data (see above), employing Cyanea leptostegia and Cy. pilosa ssp. longipedunculata as outgroups, and using unweighted and (for ISSR data) successively weighted analyses to calculate strict consensus trees and bootstrap values. Plastid sequence data were also analyzed using Bayesian inference (BI) and maximum likelihood (ML). BI was implemented in MrBayes 3.2 [49], using a model of sequence evolution (GTR + G + I model) that best fit the entire concatenated plastid data identified with jModelTest [50]. Posterior probabilities were approximated in a search over 4,000,000 generations via four simultaneous MCMC chains with every 1,000 th tree saved. Default values were used for MCMC parameters. Of the resulting trees the first 20% (burn in) were discarded. The remaining trees were summarized in a majority rule consensus tree yielding the probabilities for each clade to be monophyletic. To ensure that the MCMC chains explored total tree space, we replicated these runs with four different random starting trees. ML used Garli 2.0 [51] implemented in the Cyberinfrastructure for Phylogenetic Research (CIPRES) portal 2 teragrid [52]. Maximum-likelihood bootstrapping was executed using RAxML 7.0.4 [53,54].
To assess conflict in phylogenetic structure between the unweighted ISSR and plastid data, we conducted an incongruence length difference (ILD) test [55] in PAUP*. It is widely recognized that the ILD test has several limitations [56,57], but is can be useful to identify broad-scale incongruence between datasets. Incongruence was also evaluated based on the presence of moderately supported ($ 70% BS value) differences between nodes of the plastid and ISSR phylogenies. Incongruence between maternally inherited vs. biparentally inherited markers can reflect differences in their evolutionary history, reflecting hybridization, introgression, or chloroplast capture. To the extent that incongruence reflects such processes, plastid and nuclear data should not be combined. If there is no significant conflict, concatenating the data sets can improve their power to resolve relationships.
We utilized the following steps to assess the degree of incongruence between plastid sequence and ISSR data. First, taxa not scored for both data sets were removed; second, missing plastid sequences for one population each of Cl. oblongifolia and Cl. grandifolia were replaced with those from another population of each of the species from the same island. If these data sets then showed significant incongruence, we pruned taxa whose position on the ISSR and plastid sequence trees conflicted with bootstrap support $ 70%. We also pruned the single taxon (Cl. pyrularia) placed by morphology in Clermontia but by molecular data in Cyanea. We then re-analyzed the pruned data sets separately and pruned any additional taxa whose position on the ISSR and plastid trees conflicted with BS $ 70%. We combined the pruned ISSR and plastid data sets if they did not then exhibit significant incongruence under the ILD test, and analyzed the joint data using MP and BI. For BI analyses in MrBayes, the binary ISSR data were included as a separate partition of sequence characters with only that partition evaluated under the simplistic Jukes-Cantor model. The combined data set was also analyzed using BI in BEAST [58,59] as described below.

Historical biogeography
Ancestral area reconstruction (AAR) and between-island dispersal events were analyzed using the contrasting methods and accompanying assumptions implemented in BEAST [58,59], DEC [60], MacClade [61], and BayesTraits [62]. We categorized the eight current tall islands of the Hawaiian chain into four groups, reflecting terrestrial connections during sea-level lows during the Pleistocene: (i) Kauài and Ni`ihau; (ii) Ò ahu; (iii) Molokài, Lanài, Maui, and Kahòolawe; and (ii) Hawai`i. Islands within group (i) and (iii) were interconnected in the recent past, with connections having been lost partly due to island subsidence and cyclical sea-level change. Pleistocene glacial cycles lowered ocean levels by as much as 120 m, resulting in the islands of Maui Nui (group iii) being connected as recently as last glacial period, and having been interconnected with each other for about 75% of the last million years [63].
In order to obtain a time-calibrated phylogenetic tree for Clermontia, we used a relaxed clock approach as implemented in BEAST v1.7.4 with the uncorrelated lognormal model on the combined, taxon-reduced data set. The two data sets were brought in as separate partitions. The ''simple substitution model'' was used for the binary ISSR data partition, whereas the full GTR substition model was used for the plastid data partition. We explored both the Yule process and the birth-death speciation tree priors. Trial runs were used to determine the number of generations necessary to achieve an effective sample size (ESS) of at least 200 in Tracer v1.5, and to optimize the operator settings for our final analyses. For each of the two speciation tree priors, we ran 15 million generations on two computers, each starting with a randomly generated tree. Samples were taken every 1,000 generations, and the first 1.5 million generations of each run were discarded as burn-in. The resulting 13,500 trees from each run were combined with LogCombiner v1.6.1. The trees were then interpreted by TreeAnnotator v1.6.1 prior to visualization in FigTree v1.3.1. We constrained three nodes of Clermontia and Cyanea with normal priors using stem and crown ages of Clermontia and Cyanea derived from the Hawaiian lobeliad phylogeny based on sequences of eight plastid spacers and calibrated with asterid fossils [14]. The root of Clermontia + Cyanea, the crown of Cyanea, and the crown of Clermontia were set at 9.74, 9.58, and 4.33 Mya, respectively, and with standard deviations to permit age ranges matching the 95% confidence intervals in the Hawaiian lobeliad chronogram [14]. We also constrained the nodal ages of two clades that radiated exclusively on Hawai`i and Maui Nui, respectively, based of post-emergence ages of the two islands of 0.6 6 0.1 Mya and 2.0 6 0.1 Mya, respectively [64]. Additionally, the MP tree (see above) was converted to a chronogram in BEAST by inputting the tree as a start tree and not allowing the topology to change by removing all the operators that act on the treemodel (narrowExchange, wideExchange, wilsonBalding, and subtree-Slide). MP reconstructions of geographic distribution used MacClade applied to the MP and BI chronograms based on the combined molecular data. We assumed a basal condition for Cyanea of Kauài or some older, former tall island, following Givnish et al. [9] and the estimated stem age of Clermontia and Cyanea being 9.74 Mya, long before any of the current tall islands rose above the Pacific [14]. We used the resolving option of ''all most parsimoniuous states at each node'' for the initial mapping, and then used delayed transformation (DELTRAN) to produce a wholly resolved set of transitions and a map summarizing them, based on the conservative assumption that inter-island dispersal is an unlikely phenomenon and thus should be assumed only when absolutely necessary. In practice, but not through constraint, DELTRAN created the same resolution of dispersal events as would have a resolution of any uncertainties favoring the progression rule. Colonization and dispersal across the Hawaiian Islands were also reconstructed under BI criteria using the BayesMultiState option in BayesTraits applied to the combined molecular tree. To reduce some of the uncertainty and arbitrariness of choosing priors under MCMC, we used the recommended hyperprior approach (rjhp command) [62]. We used a random subset of 100 post burn-in Bayesian trees (Perl script from [65]).
We implemented a DEC (dispersal-extinction-cladogenesis) analysis in Lagrange following the method used with Hawaiian Psychotria [60]. We modeled dispersal across the Hawaiian chain by employing five time intervals (. 5.1 Mya for ''pre-Kauài'', 5.1-3.9 Mya for Kauài, 3.9-2.2 Mya for O'ahu [66], 2.2-0.7 for Maui Nui, and , 0.7 Mya for Hawai`i). In each time interval the possibility of forward dispersal to islands presently above water was 1.0, and 0.0 to all other islands. Backward dispersal to older islands was permitted, but at a reduced probability of 0.1. The adjacency matrix was restricted to neighboring islands, although this configuration had no impact on the results. The BEAST tree derived from both molecular data sets was used as the basis for the DEC analysis.

Petaloid sepal character-state reconstruction
The presence/absence of petaloid sepals was overlaid on the MP and BI chronograms based on the combined molecular data using MP implemented in MacClade [37], with the resolving option of ''all most parsimoniuous states at each node''. BI overlays were assembled using MultiState and a random set of 100 Bayesian PP trees in BayesTraits [38]. Ancestral reconstruction of character evolution under BI with the 100 random PP trees was represented by pie charts indicating state probabilities at each node in the combined DNA chronogram.

Morphology-based phylogeny
The morphological data matrix yielded one island of 4928 shortest trees, each 49 steps long (CI = 0.69; CI' = 0.61 excluding autapomorphies), and a strict consensus phylogeny that resolved only 5 of 22 potential nodes within Clermontia (Fig. 1A). Seventeen of 27 characters were parsimony-informative. Unweighted morphological variation grouped together three pairs of species (arborescens-tuberculata, fauriei-peleana, and pallida-persicifolia) and one unresolved quartet (calophylla-micrantha-multiflora-parviflora). Successive reweighting produced, after three iterations, a single stable island of 100 trees, each 34.3 steps long, resolving 11 of 22 potential nodes within Clermontia (Fig. 1B). Seven characters were down-weighted in the latter analysis: petiole length, number of flowers per inflorescence, pedicel length, hypanthium shape, sepal texture, wine-colored corollas, and anther tube length. The strict consensus of the weighted analysis produced a tree that is slightly more resolved than that for the unweighted analysis, but which is otherwise consistent with the latter.

DNA-based phylogenies
The nine ISSR primers yielded 58 bands, of which 3 were constant and 55 were phylogenetically informative. The initial heuristic search produced 19 islands with a total of 2296 most parsimonious trees, each of length 219 steps (CI = 0.25; CI' = 0.13). The strict consensus of these trees resolved only peleanamontis-loa, oblongifolia-persicifolia, and pyrularia sister to Cy. leptostegia within Cyanea (Fig. 2A). Two cycles of successive reweighting produced one stable island of six most parsimonious trees, each of length 55.0 steps (CI = 0.41, CI' = 0.28 excluding autapomorphies). Few polytomies remained in the strict consensus of these six trees, which placed Cl. fauriei sister to all other members of Clermontia proper, then Cl. persicifolia from West Ò ahu, then a grade consisting of Cl. arborescens, Cl. tuberculata, and Cl. calophylla (Fig. 2B). While the strict consensus of the weighted ISSR analysis is nearly completely resolved, most nodes are poorly supported, with a bootstrap support value , 50%.
The five non-coding regions of plastid DNA produced 4532 aligned nucleotides, of which 4355 were constant, 98 were parsimony-uninformative, and 79 were parsimony-informative. A heuristic search produced one island of 76 shortest trees, each 194 steps long, with CI = 0.94 and CI' = 0.87 and a strict consensus tree with only 9 of 29 potential nodes within Clermontia resolved (tree not shown). In the plastid strict consensus, Cl. fauriei was sister to all Clermontia except Cl. pyrularia (which itself is sister to Cyanea leptostegia within Cyanea), and Cl. persicifolia-oblongifolia was sister to a largely unresolved clade including all other species sequenced. The plastid majority-rule tree, by comparison, had substantially greater structure, with peleana-tuberculata sister to other elements of the large, mostly unresolved clade in the strict consensus, and those other elements being divided into three weakly supported clades (Fig. 3A). The ML tree placed peleana-tuberculata sister to the same large clade, and divided the latter into two weakly supported subclades, with clade A including Cl. pallida and all populations of Cl. arborescens and Cl grandiflora, and clade B containing all remaining species, with all species except Cl. clermontioides and Cl. samuelii showing little or no molecular divergence from each other (Fig. 3B). This topology is essentially identical to the MP majorityrule consensus, except that clermontioides-samuelii moved from clade B in the ML tree to an unresolved trichotomy involving it, clade A, and the little-divergent remainder of clade B (Cl. calophylla, drepanomorpha, hawaiiensis, kakeana, kohalae, lindseyana, montis-loa, parviflora, waimeae). Note that clermontioides-samuelii sits on an incongruously long branch in the ML tree. The BI tree (not shown) had a topology identical to the ML tree.
The ILD test revealed highly significant (P,0.001) incongruence in phylogenetic structure between the largely nuclear ISSR data and the plastid sequence data. Moderately ($ 70% BS) to strongly ($ 90% BS) supported conflicts in position between the ISSR and plastid trees occur in several taxa, including (1) Cl. persicifolia-Cl. oblongifolia from East Ò ahu and Cl. oblongifolia from West Ò ahu; (2) Cl. tuberculata from East Maui; and (3) Cl. peleana from Hawai`i ( Figs. 2 and 3). We also removed (4) Cl. pyrularia from Hawai`i, based on the conflict in its position in Clermontia based on morphology and in Cyanea based on molecular data. Removing these taxa left a substantial conflict between the ISSR and plastid trees in the position of Cl. samuelii, with that species sister to Cl. clermontioides with 93% BS in the plastid tree, and sister to populations of Cl. grandiflora in the ISSR tree (trees not shown). We therefore also pruned Cl. samuelii from the data set; a final ILD test show no significant incongruence between the ISSR and plastid data for the remaining 24 taxa of Clermontia and two of Cyanea (P.0.21).
The combined ISSR and plastid sequence data included 86 parsimony-informative characters and produced one shortest MP tree, 339 steps long with CI = 0.61 and CI' = 0.40 (Fig. 4A). In this tree, Cl. fauriei was sister to all other Clermontia, with Cl. persicifolia, a monophyletic Cl. arborescens, and a paraphyletic Cl. grandiflora sister to progressively more inclusive, nested clades. The innermost clade included a grade of nine species from Hawai`i, in which was embedded a complex involving Cl. kakeana, Cl. lindseyana, and Cl. pallida, almost exclusively from Maui Nui; branches within this innermost clade are all short and weakly supported. The MP tree based on the combined molecular data strongly supported the monophyly of the non-reticulate taxa of Clermontia and the clade sister to Cl. persicifolia, as well as the positions of Cl. fauriei and Cl. persicifolia. The ML tree based on the combined data (Fig. 4B) was quite similar to the MP tree, but (with weak support) identified Cl. grandiflora as being monophyletic and placed it sister to Cl. arborescens. In addition, ML placed Cl. lindseyana from Hawai`i into the kakeana-lindseyana-pallida complex from Maui Nui. Finally, BI produced a tree (not shown) quite similar to the ML tree. The most notable differences between the BI and ML trees are (1) the shift of pallida sister to the arborescensgrandiflora clade, and (2) the relatively high posterior support values for the pallida-arborescens-grandiflora clade, its sister clade, and the kakeana-lindseyana clade.

Historical biogeography
In the MP combined molecular phylogeny, inter-island dispersal events among non-reticulate taxa appeared to be initially consistent with the progression rule under parsimony mapping, but with later reversal on a few short, weakly supported branches (Fig. 5A). Clermontia fauriei from Kauài diverged from the deepest node; Cl. persicifolia from Ò ahu, from the next shallowest node; and Cl. arborescens and then Cl. grandiflora from Maui Nui, from the next several shallowest nodes. Dispersal to Hawai`i was inferred for the clade sister to Cl. grandiflora from East and West Maui, with backdispersal to Maui Nui by the large kakeana-lindseyana-pallida clade, and back-dispersal to O'ahu in one population of Cl. kakeana (Fig. 5A).
Essentially the same pattern was reconstructed using BEAST, with an Ò ahu-Maui Nui uncertainty at the node joining Cl. persicifolia to its sister clade (Fig. 5B) (Fig. 5B). The four nodes marked with an asterisk in Fig. 5B actually were not resolved as being Maui Nui vs. a 50:50 chance of O'ahu vs. Maui Nui in the most likely analysis, but the simpler form shown (where all four nodes were resolved as Maui Nui) did not differ significantly from the more complex resolution.

Petaloid sepal character-state reconstruction
Parsimony mapping of sepal characteristics onto the MP combined molecular phylogeny implies that petaloid sepals arose twice, in Cl. persicifolia on Ò ahu by 3.3 Mya, and in Cl. grandifloraand the taxa it subtends -on Maui Nui by 2.0 Mya (Fig. 6A). Petaloid sepals appear to have been lost independently on Hawai`i twice, in Cl. clermontioides by 0.9 Mya and in Cl. waimeae by 0.5 Mya. Bayesian mapping implies essentially the same pattern (Fig. 6A). Alternatively, and perhaps more plausibly, the conflict between the ISSR and plastid trees and the crossing it implies between Cl. grandiflora and Cl. oblongifolia, and between Cl. oblongifolia and Cl. persicifolia, suggests that petaloid sepals may have arisen only once, in the ancestor of Cl. grandiflora and the taxa the latter subtends, with petaloid sepals occurring in Cl. persicifolia via introgression from Cl. grandiflora and Cl. oblongifolia.
Parsimony mapping onto the BI combined tree implied two independent origins of petaloid sepals, on Ò ahu no later than 3.9 Mya in the clade subtended by Cl. persicifolia, and on Lanài no later than 1.6 Mya in Cl. grandiflora (Fig. 6B). This mapping implies three secondary losses of petaloid sepals, in Cl. arborescens on Maui Nui no later than 2.4 Mya, in Cl. clermontioides on Hawai`i no later than 0.8 Mya, and in Cl. waimeae on Hawai`i no later than 0.5 Mya. Bayesian mapping yields essentially the same pattern. Again, a more plausible scenario might involve an origin of petaloid sepals in Cl. persicifolia via introgression. Overall, the scenario for the evolutionary gain and loss of petaloid sepals appears to be simplest if we assume the MP combined molecular phylogeny and the introgressive origin of such sepals in Cl. persicifolia.

Species relationships and apparent patterns of hybridization/introgression
Neither morphology, nor ISSR variation, nor plastid sequences, nor the combined molecular data supported the monophyly of Clermontia sect. Clermontioides, defined by a lack of petaloid sepals, or of Clermontia sect. Clermontia, marked by their presence. Unweighted morphological data resolved only four of 22 possible nodes within Clermontia under maximum parsimony; sequential reweighting resolved 11 nodes in the MP strict consensus tree (Fig. 1A,B). Both analyses supported the monophyly of only three of the six series described by Lammers [5] based on his cladistic analysis of morphology, including Unilabiatae (Cl. fauriei, Cl. peleana), Sarcanthae (Cl. arborescens, Cl. tuberculata), and Parviflorae (Cl. calophylla, Cl. micrantha, Cl. multiflora, Cl. parviflora) (Fig. 1).
None of the morphological series proposed by Lammers [5] was supported as monophyletic by the ISSR, plastid, or combined molecular phylogenies (Figs. 2, 3). The closest approach to support of a series was the identification of a grade including Cl. arborescens, Cl. tuberculata, and Cl. calophylla in the ISSR tree, approaching series Sarcanthae; and Clade A in the ML and BI plastid trees, approaching the large series Kakeanae but including Cl. calophylla (excluded by morphology by its small flowers), Cl. hawaiiensis (excluded based by its tubular corolla), and Cl. clermontioides and Cl. waimeae (excluded by lack of petaloid sepals), and excluding Cl. persicifolia (initially included based on petaloid sepals). Interestingly, the ISSR data are not consistent with the monophyly of any individual species sampled more than once (Cl. grandiflora, Cl. kakeana, Cl. oblongifolia, Cl. persicifolia), while the plastid sequence data are consistent with monophyly in each of these cases but fail to demonstrate it positively in any one (Figs. 2, 3). The MP combined molecular tree identified Cl. arborescens as monophyletic, Cl. grandiflora and Cl. kakeana as paraphyletic, and Cl. lindseyana as polyphyletic (Fig. 4A). The ML and BI combined molecular trees fail even to resolve Cl. arborescens as monophyletic (Fig 4B).
The ISSR and plastid sequence data showed a highly significant degree of incongruence, with conflict between the strict consensus trees suggesting several instances of hybridization, introgression, or incomplete lineage sorting, involving the five species Cl. persicifolia, Cl. oblongifolia, Cl. tuberculata, Cl. peleana, and Cl. samuelii. Conflict between morphology and molecular data also point to intergeneric hybridization in Cl. pyrularia and Cl. tuberculata. Given that several previous authors have made compelling cases for crossing between species of Clermontia [1,2,5,6,7], we believe that these taxa are likely to be products of recent hybridization and/or introgression, as detailed here: 1. We propose that Clermontia pyrularia is an intergeneric hybrid, involving a cross between a species of the purple-fruited clade of Cyanea and Clermontia clermontioides subsp. clermontioides. Clermontia pyrularia was strongly supported as closely related to Cy. leptostegia, a representative of the purple-fruited clade, in both the ISSR and plastid sequence trees. Morphologically, however, Cl. pyrularia it is most closely similar to Cl. clermontioides subsp. clermontioides, with which it is known to co-occur on leeward Hualalai and Mauna Loa [5].  2. We propose that Cl. oblongifolia is a hybrid involving Cl. persicifolia and Cl. grandiflora, based on the disparate positions of oblongifolia in the ISSR and plastid trees. Lammers [5] noted that Cl. oblongifolia is similar vegetatively to Cl. persicifolia, and similar in many floral traits to Cl. grandifolia. Cl. grandifolia and Cl. oblongifolia share the unusual characteristic of pendent flowers. In addition, Cl. oblongifolia is the most widespread species of Clermontia, occurring on four islands, and presumably is highly mobile. A cross with Cl. grandifolia may have introduced the gene(s) controlling sepaloid petals into Cl. oblongifolia, and introgression from Cl. oblongifolia into Cl. persicifolia may have transferred the gene(s) into the latter as well, which would account for what would otherwise be an extraordinary additional origin of petaloid sepals in Cl. persicifolia. A careful analysis of PISTILLATA sequences [18] might be used to test this hypothesis. 3. Clermontia tuberculata may represent the product of intergeneric hybridization involving Cl. arborescens and Cyanea aculeatiflora.
The ISSR data clearly identify Cl. tuberculata as part of the Cl. arborescens complex, while the plastid data (Fig. 3) place Cl. tuberculata closest to Cl. peleana. Morphologically, Cl. tuberculata is most like Cl. arborescens, and occurs in and near populations of Cl. arborescens subsp. waihiae on East Maui [5]. However, unlike any other member of Clermontia -or, indeed, almost any other angiosperm species -Cl. tuberculata has prickly (muricate) inflorescence axes, sepals, and petals [5]. Clermontia tuberculata is wholly sympatric with Cyanea aculeatiflora, with which it shares the highly unusual trait of muricate flowers. Cyanea aculeatiflora, in turn, is partly sympatric with Cy. macrostegia ssp. macrostegia, which lacks muricate flowers but is otherwise similar to Cy. aculeatiflora. Lammers [5] suggested that hybridization/introgression may have transferred muricate flowers into Cl. tuberculata from populations of Cy. aculeatiflora that, in turn, were recently derived from Cy. macrostegia. We do not yet have molecular data from Cyanea to test this hypothesis, however. Convergence seems far less likely an explanation for the muricate inflorescence axes, sepals, and petals of Cl. tuberculata. 4. We propose that Clermontia peleana is the result of hybridization involving a member of the arborescens-tuberculata complex (with which it is associated in the plastid tree) and Cl. montis-loa (to which it is sister in the ISSR tree) (Figs. 2, 3). In fact, Cl. peleana is sister to Cl. tuberculata itself in the plastid tree, but it lacks the distinctive muricate flowers of Cl. tuberculata, suggesting that a close relative within the arborescens-tuberculata complex may have been the ancestor. Prior to 1930, Cl. peleana subsp. singuliflora grew on East Maui, home to Cl. tuberculata and one of the latter's inferred ancestors, Cl. arborescens subsp. waihiae (see above). Clermontia peleana subsp. singuliflora also was native to windward Mauna Kea on Hawai`i, and in 2010 was rediscovered in the Kohala Mountains; its former haunts on This scenario may help explain the origin of wine-colored corollas in Cl. peleana subsp. peleana (which shares this trait with Cl. montis-loa), and the association of Cl. fauriei and Cl. peleana and of Cl. arborescens and Cl. tuberculata in the morphology-based trees (Fig. 1A,B) and of Cl. arborescens and Cl. tuberculata in the weighted ISSR tree. All four species share fleshy corollas, consistent with Cl. Faurń ei and Cl. arborescens emerging from two of the three deepest nodes in Clermontia, and with Cl. peleana and Cl. tuberculata emerg ning from Cl. arborescens via hybridization/introgression. Unilabiate corollas are unique to Cl. fauriei and Cl. peleana within the genus, separate from the bilabiate Cl. arborescens and Cl. tuberculata. Clermontia must have undergone at least two transitions in floral zygomorphy, given that it includes taxa with unilabiate, bilabiate, and rotate (nearly actinomorphic) corollas. Its sister genus Cyanea, like most members of Lobeliaceae, generally have bilabiate corollas, but one group of palmiform species also evolved unilabiate corollas, as did distantly related Lobelia sect. Tupa from South America, sect. Tylomium from the Greater Antilles, sect. Heteroclita from South America, and sect. Jasionopsis from South Africa, while closely related Hawaiian genus Brighamia has corollas that are nearly actinomorphic (see phylogenies and character-state descriptions in [5,14,67,68]). The pattern of zygomorphy, in other words, shows substantial evolutionary lability across lobeliads as a whole, although a broader and more complete phylogeny will be needed to analyze the precise number of independent origins of different zygomorphic patterns.
1. We propose that Cl. samuelii is the result of hybridization between Cl. grandiflora (with which it is associated in the ISSR tree) and Cl. clermontioides (with which it is associated in the plastid tree, and in the initially pruned MP tree based on the combined molecular data). Clermontia samuelii has flowers quite similar in shape to those of Cl. grandiflora, but smaller; its range immediately abuts that of Cl. grandiflora subsp.

Historical biogeography
Both phylogenies based on the combined molecular data, pruned of apparently reticulate taxa, provide substantial support for the progression rule. Contrary to Lammers [6], our MP analysis places the origin of Clermontia on Kauài or an older island, with Cl. fauriei (the only species native to Kauài, but once also found on the older, western side of the next younger island, Ò ahu) emerging from the oldest node, Cl. persicifolia (the only species restricted to Ò ahu) emerging from the next oldest node, then Cl. arborescens and Cl. grandiflora from Maui Nui at successively younger nodes, then a large grade endemic to Hawai`i, and finally backdispersal to Maui Nui and then Ò ahu implied by inclusion of the Cl. kakeana-lindseyana-pallida complex (Fig. 5A). The current distribution of species and the MP analysis based on the combined molecular data are not adequate, however, to conclude with certainty that Cl. persicifolia resulted via dispersal from Kauài to Ò ahu.
The estimated dates of initial dispersal from one island to the next youngest in the chain based on the MP chronogram are largely consistent with the known ages of each target group of tall islands: 3.3 Mya for dispersal to Ò ahu, vs. its maximum age of 3.9 Mya; and 2.0 Mya for dispersal to Maui Nui, vs. its maximum age of 2.2 Mya (Fig. 5A). However, the estimated date of initial dispersal to Hawai`i of 1.4 Mya substantially antedates the emergence of the youngest island not earlier than 0.6 Mya. Our DEC analysis eliminates this conflict, placing the initial arrival of Clermontia on Hawai`i from Maui Nui at 0.6 Mya, and a subsequent independent arrival of Cl. lindseyana on Hawai`i from Maui Nui sometime in the last 0.6 My (Fig. 5B) Fig. 5B). Thus, five of six inter-island dispersal events implied by DEC analysis for nonreticulate taxa are consistent with the progression rule, compared with three of five such events implied by MP analysis. By contrast, Lammers [5] used his morphology-based analysis to infer 12 interisland dispersal events, only five of which were consistent with the progression rule.
Proposed reticulation events suggest, in addition, one dispersal event from Ò ahu to Maui Nui and a return flight from Maui Nui to Ò ahu to account for dispersal of Cl. oblongifolia down the chain, and subsequent introgression of gene(s) derived from Cl. grandiflora into Cl. persicifolia on Ò ahu (Fig. 7). An additional dispersal from East Maui to Hawai`i is required to account for the origin of Cl. peleana via introgression between Cl. arborescens subsp. waihiae and Cl. montis-loa. Finally, back dispersal (or, at last, gene flow) of Cl. clermontioides from Hawai`i to East Maui is required for the origin of Cl. samuelii there via hybridization with Cl. grandiflora (Fig. 7). Overall, based on the DEC analysis, seven of the proposed interisland dispersal events were down the chain, consistent with the progression rule, and three involved back dispersal up the chain. Most of the latter appear to have occurred relatively late in the evolution of Clermontia, with the possible exception of the (undated, Figure 5. Overlay of geographic distribution of Clermontia on chronograms based on the combined molecular data. Inset maps shows the inferred number of inter-island dispersal events needed to account for the distribution of present-day non-reticulate taxa based on the MP reconstruction on the MP and BI trees, respectively, using delayed transformation. (A) Shifts in insular distribution plotted on the MP chronogram; branch colors represent inferred ancestral states based on parsimony mapping. Grey represents uncertainty. Pie diagrams represent inferred ancestral states under Bayesian inference. Horizontal razor lines represent the 95% confidence intervals about each date.. (B) Shifts in distribution plotted on the BI chronogram. doi:10.1371/journal.pone.0062566.g005 5.

6.
unstudied) back-dispersal of introgressed Cl. oblongifolia from Maui Nui to Ò ahu. Based on MP, five of nine dispersal events were consistent with the progression rule. In both cases, early dispersal events appear to have been almost entirely consistent with the progression rule.

Petaloid sepal character-state reconstruction
Based on the MP combined-data tree, the simplest scenario for the evolution of petaloid sepals is that they were initially absent in Clermontia, arose once in the clade subtended by Cl. grandiflora, spread into Cl. oblongifolia via introgression from Cl. grandiflora, and into Cl. persicifolia from Cl. oblongifolia, and with subsequent losses of such sepals in Cl. waimeae and Cl. clermontioides. Hybridization/ introgression would account for the absence of petaloid sepals in Cl. peleana and Cl. tuberculata derived from Cl. arborescens bearing sepaloid sepals, crossed with Cl. montis-loa (which does have petaloid sepals) and Cy. aculeatiflora (which does not). This scenario is also largely consistent with those based on the ISSR or plastid data alone, but appears to be more plausible than either of them.
The ISSR tree implies that absence of petaloid sepals was the ancestral state, that there was one gain of petaloid sepals in the clade sister to Cl. arborescens from Molokài and East Maui, followed by three losses of such traits in Cl. calophylla, Cl. peleana, and Cl. waimeae, with introgression carrying the trait into Cl. oblongifolia and C. persicifolia. The latter, of course, is not based on the ISSR tree itself, but on the conflict between the ISSR and plastid trees, and the morphological similarities and distributions of Cl. oblongifolia, Cl. persicifolia, and Cl. grandiflora.
The plastid tree alone implies a much less plausible scenario, involving at least three independent gains of petaloid sepals and three losses. Given the obvious presence of reticulation in Clermontia, and the high likelihood that petaloid sepals involve a homeotic mutation in nuclear-encoded genes [18], it seems wiser to trace their evolution on the largely nuclear ISSR tree or -bestto trace it on the combined-data tree once the apparently reticulate taxa have been been pruned from the analysis.
Our results for the origin and spread of petaloid sepals, are consistent with but far more resolved than those presented by Hofer et al. [18] based on sequences of the non-transcribed spacer of 5S rDNA, and by Pillon et al. [19] based on SNPs. All four studies place Cl. fauriei sister to all other core Clermontia; Pillon et al. and our current paper identify Cl. pyrularia as falling within Cyanea. Our study differs from all previously published studies, however, in proposing that Cy. pyrularia is a hybrid including Cl. clermontioides and Cy. hamatiflora subsp. carlsonii of the purple-fruited clade of Cyanea, and in the details of relationships and at least six reticulation events within the rest of Clermontia, with important implications for the reconstruction of geographic spread and morphological evolution within the genus. Given the far greater number of informative characters in the combined molecular data Figure 6. Overlay of petaloid vs. sepaloid sepals on chronograms based on the combined molecular data. (A) Shifts in sepal nature plotted on the MP chronogram; branch colors represent inferred ancestral states based on parsimony mapping. Grey represents uncertainty. Pie diagrams represent inferred ancestral states under Bayesian inference. Horizontal razor lines represent the 95% confidence intervals about each date. (B) Shifts in sepal nature plotted on the BI chronogram. doi:10.1371/journal.pone.0062566.g006 Figure 7. Summary of patterns of geographic dispersal, sepal evolution, and reticulate speciation in Clermontia inferred from ISSR, plastid, and combined molecular data and morphological variation. Cladogram illustrates phylogenetic relationships among non-reticulate taxa based on MP. Branch color represents inferred ancestral distribution under parsimony; presence/absence of dashed line represents presence/ absence of petaloid sepals. Curved arrows, colors, and dashed lines leading to the right represent apparent hybrid/introgressive origins of five reticulate taxa and the distribution of parental taxa. Superscript +'s on the latter indicate the presence of petaloid sepals. Curved arrow leading to the left represents apparent introgression of petaloid sepals into Cl. persicifolia from Cl. oblongifolia (and, ultimately, from Cl. grandiflora). The basal split of Cyanea into orange-and purple-fruited clades is shown, together with the particular species of Cyanea thought to be involved in the origins of Cl. pyrularia and Cl. tuberculata based on morphology and/or geographic distribution. doi:10.1371/journal.pone.0062566.g007 set vs. morphology, the inability of the unweighted morphological data to recover almost any phylogenetic structure in Clermontia, and the power of the molecular data to pinpoint apparent hybridization/introgression events, we believe that the molecular data provide a much stronger basis for inferring phylogeny than do the morphological data.
Our current molecular data, however, are not adequate to fully resolve the phylogeny of Clermontia and clarify all questions regarding the diversification of this genus. Because we lack fully resolved, strongly supported trees based on plastid sequence data and multiple, individually powerful nuclear data sets, as well as coverage of several critical taxa in the sister genus Cyanea, we are not yet in a position to identify and unambiguously reconstruct many cladogenetic events and reticulations, or assemble the detailed history of geographic diversification and character evolution of Clermontia. As a result, we are uncertain whether we have been able to remove all reticulate taxa from our trees, or to identify all repeated layers of reticulation, and thus to derive a single, natural phylogeny based on combined nuclear and plastid data from which incongruence has been purged. Given the low levels of support and/or resolution in both our nuclear and plastid trees to date, it is clear that future progress on Clermontia evolution will require us to increase greatly the amount of data per taxon, to include multiple, individually highly informative nuclear loci, and to sample populations of each taxon more densely on each island on which it occurs. Given that the extant species of Clermontia have diverged from each other in less than 4.7 My, are known to cross with each other (and even with Cyanea) in several instances, and appear to be highly mobile within the Hawaiian archipelago, the requirement for greater density and extent of data to reconstruct evolution within Clermontia should come as no surprise. Such increases in data density per taxon and in taxon sampling will almost surely require using the power of next-generation sequencing, to which our labs are now turning.