Hedysarum L. (Fabaceae: Hedysareae) Is Not Monophyletic – Evidence from Phylogenetic Analyses Based on Five Nuclear and Five Plastid Sequences

The legume family (Fabaceae) exhibits a high level of species diversity and evolutionary success worldwide. Previous phylogenetic studies of the genus Hedysarum L. (Fabaceae: Hedysareae) showed that the nuclear and the plastid topologies might be incongruent, and the systematic position of the Hedysarum sect. Stracheya clade was uncertain. In this study, phylogenetic relationships of Hedysarum were investigated based on the nuclear ITS, ETS, PGDH, SQD1, TRPT and the plastid psbA-trnH, trnC-petN, trnL-trnF, trnS-trnG, petN-psbM sequences. Both nuclear and plastid data support two major lineages in Hedysarum: the Hedysarum s.s. clade and the Sartoria clade. In the nuclear tree, Hedysarum is biphyletic with the Hedysarum s.s. clade sister to the Corethrodendron + Eversmannia + Greuteria + Onobrychis clade (the CEGO clade), whereas the Sartoria clade is sister to the genus Taverniera DC. In the plastid tree, Hedysarum is monophyletic and sister to Taverniera. The incongruent position of the Hedysarum s.s. clade between the nuclear and plastid trees may be best explained by a chloroplast capture hypothesis via introgression. The Hedysarum sect. Stracheya clade is resolved as sister to the H. sect. Hedysarum clade in both nuclear and plastid trees, and our analyses support merging Stracheya into Hedysarum. Based on our new evidence from multiple sequences, Hedysarum is not monophyletic, and its generic delimitation needs to be reconsidered.


Introduction
The legume family (Fabaceae) is the third largest flowering plant family with about 19,500 species in about 751 genera. The family exhibits a high level of species diversity and evolutionary success in various ecosystems worldwide [1]. The genus Hedysarum L. (Fabaceae: Hedysareae) a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 evolutionary rate at least as fast as the ITS sequence [16]. Furthermore, several studies have developed other nuclear sequences for phylogenetic reconstruction in various taxonomic groups of Fabaceae [12,17,18]. The SQD1 (UDP sulfoquinovose synthase gene) belongs to the low copy conserved ortholog set (COS) genes [19]. It has been used in a phylogenetic study of the caesalpinioid legumes [18]. The nuclear coding PGDH (putative phosphogluconate dehydrogenase gene) and TRPT (putative triosephosphate translocator gene) sequences are exonderived, putative orthologs, single copy genes [12], and were used in a phylogenetic study of the Hologalegina legumes [17].

Ethics statement
The plant materials used in this study did not involve protected or endangered species. No specific permits were required for the collection of samples. Voucher information of all the samples was given in S1 Appendix.

Taxon sampling
A total of 58 accessions were included in this study (S1 Appendix), representing all genera in the tribe Hedysareae and all sections and major infra-sectional clades in Hedysarum as recognized by Duan et al. [11]. Hedysarum boveanum Bunge ex Basiner, H. denticulatum Regel, H. minjanense Rech.f., H. poncinsii Franch., and H. syriacum Boiss. were placed in sect. Multicaulia s.l. [2] based on morphological characters [20][21][22], and they were sampled for the first time to test their phylogenetic positions. Alhagi sparsifolia Shap. ex Keller & Shap. was selected as the outgroup based on previous results [10,11]. Some of the ITS, psbA-trnH, and trnL-trnF sequences used in this study were published by Duan et al. [11] and Amirahmadi et al. [10], while all other sequences were generated by the present study (S1 Appendix). Voucher information of DNA sequences was listed in S1 Appendix.

DNA extraction, PCR and sequencing
Total genomic DNAs were extracted from silica-gel dried leaf material or herbarium specimen using either the Plant DNA Extraction Kit AGP965/960 (AutoGen, Holliston, Massachusetts, USA) or the DNeasy Plant Mini Kit (Qiagen, Valencia, California, USA).
Polymerase chain reaction (PCR) was performed in a 25μl volume with the following components: 10× reaction buffer, 200μmolÁL -1 of each dNTP, 10μg BSA, 0.4μmolÁL -1 of each primer, 2.5mmolÁL -1 MgCl 2 , 1U of BIOLASE DNA Polymerase (Bioline USA Inc., Taunton, Massachusetts), 1-7.5μL template DNA. The amplification conditions were 3 min at 95˚C, followed by 36-40 cycles of 1 min at 94˚C, 1 min at 50-60˚C, and 1-1.5 min at 72˚C, then a final extension at 72˚C for 7-10 min. The PCR products were purified using ExoSAP-IT (USB Corporation, Cleveland, Ohio, USA) or the polyethylene glycol (PEG) precipitation procedure [31]. Cycle sequencing reactions were conducted in both directions using the amplification primers and the BigDye 3.1 reagents. After being cleaned up by the Sephadex columns, the sequencing products were run on an ABI 3730 automated sequencer (Applied Biosystems, Foster City, California, USA). Sequences were assembled using the program Geneious v.8.1.2 [32] (http://www.geneious.com/). All sequences have been deposited in GenBank and the accession numbers were listed in S1 Appendix.

Phylogenetic analysis
Multiple sequence alignments were conducted in Geneious v.8.1.2 using MUSCLE [33] with default settings, followed by manual adjustments. For the non-coding sequences, ambiguously aligned regions were removed from the matrix prior to phylogenetic analysis. Insertions and deletions (indels) in the data matrices were coded as binary characters using the program SeqState [34] according to the "simple coding" method [35]. The binary characters were combined with the DNA data as the additional partition of the matrix. SequenceMatrix [36] was employed to assemble combined datasets.

Single locus analysis
Phylogenetic reconstructions were first conducted by using single locus datasets. The best-fit nucleotide substitution model for each of the 10 individual sequences was determined using the Bayesian Information Criterion (BIC) in jModelTest v.2.1.7 [37]. For the ITS dataset, boundaries of the 5.8S, ITS1 and ITS2 regions were determined by comparing with the published 5.8S sequence of Vicia faba L. [38], and models for the ITS1, 5.8S and ITS2 regions were determined separately (see Table 1).
Bayesian inferences (BI) were conducted in MrBayes v.3.2.5 [39,40]. When models (such as TIM3ef, TrNef, TPM1, TVM) determined by BIC were not directly available in MrBayes, we transformed a GTR model in MrBayes by fixing the six nucleotide substitution rates and four nucleotide state frequencies to the values as calculated by jModelTest for each dataset. The model applied to all the coded binary partitions was a default Standard Discrete Model in MrBayes [41]. For the ITS dataset, the partitions were done for ITS1, 5.8S and ITS2 separately. In the Bayesian inference, two independent analyses starting from different random trees with three heated and one cold chain were run for 10,000,000 generations, and trees were sampled every 1,000 generations (10,000 trees sampled in total). The first 2,500 trees (25%) were discarded as burn-in, and the remaining trees were used to construct a 50% majority-rule consensus tree and posterior probabilities (PP). Tree visualization was achieved in FigTree v1.4.3 (http://tree.bio.ed.ac.uk/software/figtree/). Phylogenetic reconstructions were also carried out with maximum parsimony (MP) and maximum likelihood (ML) analyses using PAUP Ã 4.0b10 [42] and RAxML v.8.2 [43], respectively. The MP bootstrap analysis was performed with the following settings: heuristic search, TBR branch-swapping, 1,000 bootstrap replicates, random addition sequence with 10 replicates, a maximum of 1,000 trees saved per round. The ML rapid bootstrap analysis was conducted with a random seed, 1,000 alternative runs, and the same partition scheme as in the Bayesian analysis. Model parameters were estimated and optimized for each partition of the dataset by RAxML with the GTRCAT commands. Bootstrap support values from the MP (PBS) and ML (LBS) analyses were labeled on the corresponding branches of the BI trees.

Concatenated data analysis
We conducted incongruence length difference (ILD) test [44] to assess whether different datasets could be concatenated for phylogenetic reconstruction. The ILD tests were performed in PAUP Ã 4.0b10 [42] by using only the informative characters [45] with heuristic search, 1,000 replicates, simple addition sequence (with a maximum of 1,000 trees saved per replicate), and tree-bisection-reconnection (TBR) branch-swapping algorithm. When the ILD test found a p value greater than 0.01, datasets were concatenated for phylogenetic analysis [46]. BI, MP and ML analyses were conducted for the concatenated datasets by using the same methods as in the single locus analyses. For BI and ML, the best partitioning scheme and substitution model were determined by using PartitionFinder v.1.1.1 [47] with the following settings: linked branch length, BIC metric for model and partitioning selection, and the greedy algorithm. Data blocks in PartitionFinder were set as the followings: the ETS, ITS1, 5.8S and ITS2 datasets, the first, second and third codon position of each of the PGDH, SQD1 and TRPT datasets, and the psbA-trnH, trnC-petN, trnL-trnF, trnS-trnG and petN-psbM datasets.
We performed the approximately unbiased (AU) tests [48] using CONSEL v.0.2 [49] to test statistical support of incongruence between nuclear and plastid topologies. Topological constraints were generated from the ML trees using MEGA v.6.0 [50]. Per site log likelihoods for each topology were estimated in RAxML v.8.2 [43] under the GTR + G model. The dataset partition scheme was the same as in the ML analysis.

Coalescent analysis
The Bayesian coalescent-based analyses for multi-locus data were performed using Ã BEAST [51] as implemented in the BEAST v.2.4.3 package [52]. Two sets of analyses were conducted: one with all five nuclear sequences, another with all five plastid sequences. The substitution model for each sequence was that used in the single locus analysis, and model parameters were unlinked among partitions. The lognormal relaxed clock model, the linear with constant root population function and the Yule speciation model were used. Four independent runs of 100,000,000 and 500,000,000 generations were performed for the nuclear data and the plastid data, respectively. Samples were stored every 5,000 generations. Tracer v.1.6 (http://beast.bio. ed.ac.uk/Tracer) was used to check the effective sample sizes (ESSs) of the parameters and the convergence of different runs. After removing the 10% burn-in of each run, the log and tree files from different runs were combined by using LogCombiner [52]. In the combined results, the ESSs of the sampled parameters all exceeded 200. The maximum clade credibility tree was annotated using TreeAnnotator [52].

Bayesian concordance analysis
We carried out the Bayesian Concordance Analysis (BCA) [53] for the five nuclear loci as well as the five plastid loci using the BUCKy v.1.1.4 package [54]. Gene trees of each locus were obtained from the single locus BI (.t files generated by MrBayes), and were summarized by using the program mbsum [54] with a 25% burn-in. The output files of mbsum were then used to construct the primary concordance tree and concordance factor (CF) in the program bucky [54]. The a priori level of discordance among loci (α) was set to 0.01, 1 and 100 in alternative runs to assess the effect of α on the BCA results. Each run was conducted with 10,000,000 updates, four replicates, one cold and three hot chains of the Metropolis coupled Markov chain Monte Carlo (MCMCMC). In MCMCMC, a swap was proposed once every 100 updates, and the α multiplier was set to 10. Different runs with α set to 0.01, 1 and 100 all recovered the same primary concordance tree topology and CFs, hence the BCA results obtained with the default α = 1 were reported. The primary concordance tree was visualized in FigTree, and branch labels were annotated as the sample-wide posterior mean CFs and their 95% credibility intervals.

Nuclear data
Characteristics of the individual nuclear datasets were summarized in Table 1. Treating indels as missing data or binary characters did not change the topologies, but the analyses that included indels as characters increased support values in general. The results presented here were based on the datasets with indels treated as binary characters. Trees generated from each individual datasets supported the monophyly of each of the genera Ebenus L., Greuteria, Sulla, and Taverniera DC., and several small lineages within Hedysarum and Onobrychis. Hedysarum denticulatum and H. minjanense were nested with H. sect. Stracheya as redefined by Duan et al. [11] in all five nuclear trees. Deep relationships within Hedysareae were largely unsolved by single sequences. However, Hedysarum was not supported to be monophyletic in any of the five single-sequence trees.
In order to search trees with better resolution, we concatenated individual datasets in our analyses. To explore whether two or more nuclear datasets could be concatenated, we conducted 10 pairwise ILD tests for the five nuclear datasets. The p values were shown in Table 2. Congruence was detected between any pair in the ETS, ITS, PGDH and TRPT datasets (all with p > 0.01). We first concatenated the non-coding ribosomal ETS and ITS datasets (pairwise ILD: p = 0.13), as well as the coding PGDH and TRPT datasets (pairwise ILD: p = 0.061) for two separate analyses.
In the concatenated ETS + ITS tree (S1 Fig), Corethrodendron, Ebenus, Greuteria, Onobrychis, Sulla and Taverniera were each supported to be monophyletic. Corethrodendron,  Although support values varied in these two trees, Hedysarum was shown to be biphyletic. Pairwise ILD test between the ETS + ITS dataset and the PGDH + TRPT dataset found p = 0.703, thus we concatenated these four sequences in our further analyses.
In the concatenated ETS + ITS + PGDH + TRPT tree (  Table 1). Hedysarum was, nevertheless, not monophyletic in the SQD1 tree. Members of the Sartoria clade grouped with Taverniera In the concatenated ETS + ITS + PGDH + TRPT + SQD1 tree (Fig 1), relationships of the CEGO, Hedysarum s.s., Sartoria clades and Taverniera remained the same as in the concatenated ETS + ITS + PGDH + TRPT tree (S3 Fig). With the SQD1 dataset added, the support values were even higher (Fig 1)

Plastid data
Characteristics of plastid datasets were presented in Table 1. Trees based on individual plastid datasets had limited resolution. The 10 pairwise ILD tests for the five plastid datasets all found p > 0.01 (Table 3), therefore we concatenated these five plastid datasets in our analyses. Because there was incongruence between the plastid and nuclear trees, the plastid and nuclear datasets were not concatenated in our analysis.
The AU tests of the concatenated nuclear ETS + ITS + PGDH + TRPT dataset against the concatenated five plastid sequence dataset were performed. The nuclear dataset rejected the plastid topology (p < 0.01), and the plastid dataset rejected the nuclear topology (p < 0.01) as well, suggesting incongruence between the nuclear and plastid datasets. The concatenated five nuclear sequence dataset (ETS + ITS + PGDH + TRPT + SQD1) also rejected the plastid topology (p < 0.01), and vice versa.

Non-monophyly of Hedysarum
Our results revealed incongruence between the nuclear and the plastid trees in Hedysarum (Fig 3). In the nuclear trees (Figs 1 and 3), Hedysarum is biphyletic; the Sartoria clade is sister to Taverniera; and the Hedysarum s.s. clade is sister to the CEGO clade. On the other hand, Hedysarum is monophyletic and sister to Taverniera in the plastid tree (Figs 2 and 3). The Table 3. psbA-trnH, trnC-petN, trnL-trnF, trnS-trnG and  petN- topological incongruence concerning Hedysarum is robustly supported (Figs 1 and 2). The difference between the nuclear and the plastid topologies lies on the position of the Hedysarum s. s. clade, with the Hedysarum s.s. clade sister to the CEGO clade in the nuclear tree, in contrast with its sister relationship with the Sartoria clade in the plastid tree (Fig 3).

P values of pairwise ILD tests for the plastid
Our trees based on concatenated nuclear and plastid data are well supported concerning the position of the Hedysarum s.s. clade (Figs 1 and 2). However, there is concern that data  psbA-trnH, trnC-petN, trnL-trnF, trnS-trnG, and petN-psbM sequences. The Bayesian posterior probabilities are above the branches, and the maximum parsimony (left) and maximum likelihood (right) bootstrap support values are below the branches. A dash indicates a branch that is not found in the maximum parsimony tree. Non-Monophyly of Hedysarum concatenation may lead to overconfidence in incorrect species tree [55]. Thus we also used the coalescent method and BCA to infer the phylogenetic relationships using multi-locus data. Our coalescent trees (S5 and S7 Figs) agree with the concatenated trees (Figs 1 and 2), and they are also well supported. In BCA, the concordance factor (CF) of a certain clade is the proportion of the genes or the genome for which that given clade is true, and should not be confused with a clade support value [53,56]. In our BCA results (S6 and S8 Figs), CFs of many clades are below 1, which is likely due to the limited number of phylogenetic informative sites in each individual locus especially in the PGDH, SQD1, TRPT and psbA-trnH sequences ( Table 1). The single locus trees show limited resolution with many polytomies, such as the SQD1 tree (S4 Fig). The nuclear sequences have fewer informative sites than the plastid ones (Table 1), and the CFs from the nuclear BCA are lower than those from the plastid BCA (S6 and S8 Figs). However, in the nuclear as well as in the plastid BCA results, the primary concordance tree is built from the clades with highest CFs and shows the relationships shared by a large proportion of the sampled genes [56]. It agrees with the concatenated data tree and the multi-locus coalescent tree. Closer inspection of the nuclear BCA result shows that in the list of clades which are not in the nuclear primary concordance tree but with estimated CF > 0.05, there is no such Non-Monophyly of Hedysarum clade of a monophyletic Hedysarum as in the plastid phylogeny. Also, in the list of clades which are not in the plastid primary concordance tree but with estimated CF > 0.05, there is no contradictory clade of a biphyletic Hedysarum as in the nuclear phylogeny. In our analyses, various methods of phylogenetic reconstructions, i.e., BI, MP and ML based on concatenated data, multi-locus coalescent, and BCA, all uncover the same incongruent pattern between the nuclear and the plastid trees (Fig 3).
Insufficient data is probably responsible for the low resolution of the SQD1 tree, however it can be ruled out from the candidate mechanisms that explain the incongruent position of the Hedysarum s.s. clade because both the nuclear and plastid topologies are robustly supported (Figs 1 and 2, S3 Fig). Gene choice and rate of molecular evolution [58] can also be ruled out for the same reason. Sequencing error probably affects phylogenetic accuracy where there are a small number of informative sites [58]. In our trees, sequencing error at most only influences the closely related terminals where the support is low, but should not affect well supported deep relationships. Our sampling covers all genera of the tribe Hedysareae and all sections and major infra-sectional clades in Hedysarum as recognized by Duan et al. [11]. Furthermore the different sequences were generated from the same DNA extraction (except for the PGDH sequence of Sulla coronaria, see S1 Appendix) for each accession. Insufficient taxon sampling and sample misidentification are thus not likely to be the causes of the deep incongruence.
Incomplete homogenization of tandem repeats and the possibly uniparental inheritance of the nuclear ribosomal gene region, and orthology/paralogy conflation of the nuclear low or single copy genes, may affect the accuracy of phylogenetic reconstruction [13,58,60,61]. These issues may lead to incongruence between different nuclear gene trees. We cannot fully rule out these issues in our nuclear data. However, the results based on ribosomal gene regions (S1 Fig) and low or single copy genes (S2 and S4 Figs) all support a biphyletic Hedysarum. Furthermore, the monophyly of Hedysarum is not detected in any of the nuclear trees. Therefore, the incongruence between the nuclear and the plastid trees cannot be simply explained by such mechanisms.
Allopolyploidy can be excluded from being considered as a mechanism to explain the incongruent position of the Hedysarum s.s. clade. We checked the chromosome number counts for Hedysareae in the Index to Plant Chromosome Numbers (IPCN, http://www. tropicos.org/Project/IPCN). Most species in Hedysarum as well as those of other genera in Hedysareae have been reported to have chromosome numbers 2n = 16 or 2n = 14, suggesting that they are diploid because their basic chromosome numbers are indicated as x = 8 or x = 7 [2,62]. To summarize the available data from IPCN, deep lineages of Hedysareae are predominantly diploid, e.g., 2n = 2x = 16 for Alhagi, Sulla (species name still under Hedysarum in IPCN; with one report of 2n = 18), Taverniera [63].
Occasionally horizontal gene transfer between species through nonsexual means may explain gene tree incongruence [58]. But the variational pattern in our study would require multiple horizontal transfers in the nuclear genes (ETS, ITS, PGDH, TRPT, SQD1) in a particular ancestor of Hedysarum, which simultaneously leads to discordant gene trees with the plastid tree. This hypothesis is viewed difficult, as many unlinked genes are not likely to be horizontally transferred in parallel [64]. Stegemann et al. [65] showed that an entire chloroplast genome could be horizontally transferred between sexually incompatible species through grafting. Their results were based on artificial grafting and selection process. It is not known whether chloroplast genome could be horizontally transferred through natural grafting. Additionally, unlike in woody plants, species of Hedysareae are not likely to form natural grafting because most species are herbaceous perennials.
Lineage sorting of ancestral polymorphisms can cause incongruent gene trees. The phylogenetic consequence of lineage sorting is usually difficult to distinguish from that of the other processes, such as hybridization/introgression [58,66]. In general, lineage sorting is expected to be a possible cause of incongruence at lower taxonomic ranks often in population and/or species-level studies [58]. But the incongruence in our results involves deep relationships among different genera. Lineage sorting may contribute to incongruence at higher taxonomic ranks when polymorphisms are maintained by balancing selection through many speciation events [58]. Concerning the position of the Hedysarum s.s. clade, however, our tree based on coding genes (PGDH, TRPT) agrees with the tree based on non-coding sequences (ETS, ITS), which are thought to be neutral. Our tree based on the plastid non-coding data (Fig 2) also agrees with the tree based on the plastid coding matK gene in a previous study [10]. Hence lineage sorting may not be the best explanation for the incongruence in our results.
Our phylogenetic pattern may be best explained by a chloroplast capture hypothesis via introgression. In most angiosperm taxa, the plastid genome is uniparentally inherited, whereas the nuclear genome is biparentally inherited [13,67]. After many generations of backcrossing through introgression, the introgressant inherited the plastid genome from one parent, and nearly all nuclear genes from the other parent. A significant feature of chloroplast capture is that plastid genome introgression is always not accompanied by nuclear introgression [14,58,65]. Our results are consistent with this scenario, i.e., multiple nuclear datasets are not supporting the monophyletic Hedysarum as suggested by the plastid tree. Chloroplast capture is probably one of the most common causes of phylogenetic incongruence [58], and has been reported in a wide range of taxa at various taxonomic levels [14,15,[68][69][70][71][72][73][74][75].
Our phylogenetic analyses of the nuclear sequences show that Hedysarum is not monophyletic. Its generic delimitation needs to be reconsidered. The Sartoria clade and the Hedysarum s.s. clade can be distinguished from each other by several morphological characters. Most species in the Sartoria clade have grayish-green leaves, obscure lateral veins in the leaflets, biconvex and non-winged loments which always possess prickles, bristles, and ribs, whereas species in the Hedysarum s.s. clade always have bright-green leaves, visible lateral veins in the leaflets, compressed and more or less winged loments which are always unarmed [2,3,5]. Loment morphology of the Sartoria clade is very similar to that of Taverniera by its prickles, bristles, and ribs [2,76], corresponding to the sister relationship between the Sartoria clade and Taverniera in the nuclear tree (Fig 3). The compressed, unarmed, more or less winged loment is also found in Greuteria and Eversmannia in the CEGO clade (Fig 3) [10]. Mironov [77] found that species of H. sect. Gamotion (members of the Hedysarum s.s. clade) differed from species of H. sect. Multicaulia (members of the Sartoria clade) in several pericarp anatomical characters that were thought to be of systematic significance.
Additionally, the Hedysarum s.s. clade and the Sartoria clade also show eco-geographical differentiation. In general, members of the Hedysarum s.s. clade adapt to mesic and/or psychric habitats in temperate montane forests, alpine and arctic regions of Eurasia and North America, whereas species of the Sartoria clade are distributed in xeric habitats in arid/semi-arid areas and steppes of central and western Asia, the Mediterranean region and western North America [11,78].
Systematic position of the Hedysarum sect. Stracheya clade Stracheya was originally established as a monotypic genus [9], which was then treated as a section of Hedysarum based on comprehensive morphological analyses [2,3]. Hedysarum tibeticum was the sole member of H. sect. Stracheya when Choi and Ohashi first constructed this section in Hedysarum [2]. However, previous phylogenetic analyses showed different systematic positions of the H. sect. Stracheya clade in Hedysareae with low support values [10,11]. With more sequence data, our nuclear and plastid trees resolved congruent sister relationship between the H. sect. Hedysarum clade and a clade of five species including H. tibeticum, H. kumaonense, H. lehmannianum, H. denticulatum, and H. minjanense (Figs 1 and 2, S3 Fig). Our analysis supports merging Stracheya with Hedysarum and expanding H. sect. Stracheya to include four additional species. These four species were previously treated as members of H. sect. Subacaulia because of their strongly reduced stems [5,20]. However, they did not form a clade with other members of H. sect. Subacaulia such as H. petrovii, H. ferganense Korsh., and H. poncinsii (Figs 1 and 2, S3 Fig). Some species of H. sect. Hedysarum, for example, H. tanguticum B.Fedtsch. and H. pseudoastragalus Ulbr., also have reduced stems [5]. Thus, the strongly reduced stem is most likely a homoplasious convergence in Hedysarum. The dwarf habit of the H. sect. Stracheya lineage might be an adaptation to the cold and arid habitats in the pan-Himalayan region and the adjacent eastern part of central Asia [79,80].
Supporting Information S1 Appendix. Taxon names, geographical locality, voucher information and herbarium codes, and GenBank accession numbers for the sequences used in this study. Herbarium codes follow the Index Herbariorum (http://sweetgum.nybg.org/ih/). For each taxon, Gen-Bank accession numbers of the ten sequences are given in the sequence of ETS, ITS, PGDH, SQD1, TRPT, trnL-trnF, psbA-trnH, trnS-trnG, trnC-petN, and petN-psbM. New sequences generated in this study are indicated by an asterisk ( Ã ). Missing sequences are indicated by a dash (-).