We excluded the poly A or poly T, and the ambiguous alignment from the data sets (rps16, 25 bp between 195 and 219, 505 bp between 245 and 749, 29 bp between 791 and 819, and 6 bp between 1224 and 1229; trnL-F, 6 bp between 73 and 78, 4 bp between 126 and 129, 14 bp between 523 and 536, 211 bp between 696 and 906, 3 bp between 1096 and 1098, and 6 bp between 1102 and 1107). The aligned length, the numbers of variable and parsimony informative sites, and the best-fit model determined by Modeltest for each plastid region were given (Table 1). The variation among three regions is different and the rps16 intron is the most variable region. The single-gene analysis using maximum parsimony and Bayesian methods demonstrated no significant incongruences for the phylogeny of Remusatia among the three regions (Figure S1–S3). Moreover, a quantitative approach using the incongruence length difference (ILD) test also indicated no conflict among the three data sets (p = 0.04). We thus concatenated the three plastid data set in our analysis.

The aligned length, the numbers of variable sites, and the numbers of parsimony informative sites for the concatenated plastid regions were given (Table 1). Treating gaps as missing data, the maximum parsimony analysis produced 3 most parsimonious trees (MPTs) of 412 steps, with a consistency index (CI) of 0.83, a CI excluding uninformative characters of 0.71, a retention index (RI) of 0.84, and a rescaled consistency index (RC) of 0.70. The Bayesian tree was nearly identical to the strict consensus tree of MPTs (Figure 1). Treating the gaps as new characters, the parsimony strict consensus tree also did not support the monophyly of Remusatia (Figure S4). Moreover, the resulting tree when gaps were treated as missing data is better resolved than the tree obtained with gaps as coded characters. For this reason, we discuss all results in this paper based on the analysis treating gaps as missing data (Figure 1).

The concatenated plastid data supported the monophyly of the Colocasia clade (PB = 77%, PP = 1.0). Remusatia was part of the Colocasia clade (Figure 1). The monophyly of Remusatia was not supported because R. hookeriana did not form a clade with other three species R. vivipara, R. yunnanensis, and R. pumila (see Figure 1). Excluding Remusatia hookeriana, the remaining Remusatia species formed a supported clade (PB = 60%, PP = 0.97) (Figure 1). Within this clade, Remusatia pumila is sister to the clade containing R. vivipara and R. yunnanensis (Figure 1).

Analyses constraining all Remusatia species into a clade generated MPTs that were the same tree length with the unconstrained MPTs. The SH test suggests topological congruence is well supported (p > 0.01) between constrained analysis and unconstrained analysis (Figure S5).

Thirty accessions representing four species of Remusatia and 20 related taxa were included in this study (Table 2). Based on recent phylogenetic analyses of Araceae [9], the following genera were used as closely related taxa that represent major lineage within the large Pistia clade: Ariopsis, Steudnera, Colocasia, Alocasia (Schott) G. Don, Arisaema Martius, and Pinellia Tenore. Species of Pistia L. and Protarum Engl. were selected as outgroups because they have been shown to be outside the Colocasia clade within the large Pistia clade. The wide range of multiple taxa within the large Pistia clade was selected to further test the monophyly of Remusatia within a broader phylogenetic framework.

Total DNA was extracted from about 15 mg silica-gel dried leaf material using the DNeasy plant mini kits (QIAGEN, Mississauga, Ontario) following the manufacturer’s protocol or the modified CTAB extraction method [24].

Three coding or non-coding plastid regions (the rbcL gene, the trnL-trnF intergenic spacer, and the rps16 intron) markers were employed in this study. The following primers were used for both amplification and sequencing: “1F” and “1460R” for the rbcL gene [25], “c” and “f” for the single trnL-F region or as two fragments with “c + d” and “e + f” [26]; when this region could not be amplified successful using primer “c”, we used instead primer “c2” [27], “F” and “R2” for the rps16 intron [28]. Polymerase chain reaction (PCR) amplifications were performed in a 25 μL volume containing 1.5 mM MgCl₂, 0.2 mM of each dNTP, 0.4 mM of each primer, 1 U of Taq polymerase (Bioline), and about 10–50 ng of DNA template under the following conditions: 3 min at 95 °C, followed by 37 cycles of 20 s at 94 °C, 30 s at 50 °C, and 40 s at 72 °C, and then a final 5 min extension at 72 °C.

The PCR products were purified using the polyethylene glycol (PEG) precipitation procedure following the protocol of Sambrook et al. [29]. Cycle sequencing was conducted using BigDye 3.1 reagents and carried out using the following profile: 35 cycles of 97 °C for 15 s, 50 °C for 5 s, and 60 °C for 4 min. The products of cycle-sequencing reactions were cleaned using the Sephadex columns (Amersham Pharmacia Biotech, Piscataway, New Jersey). The sequences were generated on an ABI prism 3730XL capillary sequencer (Applied Biosystems, Foster City, California). All sequences were newly generated in this study and have been deposited in GenBank (Table 2).

The program Sequencher 4.5 (Gene Codes Corporation, Ann Arbor, Michigan) was used to evaluate chromatograms for base confirmation and to edit contiguous sequences. Sequences were initially aligned with ClustalX version 1.83 [30], followed by manual adjustments on Se-Al v2.0a11 [31].

To evaluate congruence of the three plastid (rbcL, trnL-F, rps16) data sets, we employed the partition homogeneity test or the incongruence length difference (ILD) test [32]. The partition homogeneity test was conducted with PAUP* version 4.0b10 [33] with 100 replicates, each with 100 random addition sequence replicates, tree bisection-reconnection (TBR) branch swapping, and keeping no more than 100 trees per random addition replicate. Following Cunningham [34], a significance level of p = 0.01 was adopted for this test.

Phylogenetic trees for each plastid region and the combined data set (concatenating the rbcL, trnL-F, and rps16) were constructed using maximum parsimony (MP) and Bayesian methods. The MP analyses was conducted using PAUP* version 4.0b10 [33]. All characters were weighted equally and gaps were treated as missing data and coded as binary characters for the concatenated plastid data set using the “simple gap coding” method [35]. The program GapCoder [36] was employed to score the insertions and deletions (indels). The most parsimonious trees were obtained with heuristic searches of 1000 replicates with random stepwise sequence addition, tree bisection-reconnection (TBR) branch swapping, collapse of zero-length branches, multiple tree option in effect, saving 100 trees from each random sequence addition. Parsimony bootstrap values (PB) for the clades [37] revealed in the most parsimonious trees (MPTs) were calculated with 500 bootstrap replicates. In each replicate, we performed 1000 random sequence addition replicates followed by tree bisection-reconnection (TBR) swapping, keeping no more than 10 trees per replicate. Tree statistics, including consistency index and the retention index, were calculated using PAUP*.

Modeltest 3.7 [38,39] was used to determine the optimal model of molecular evolution and gamma rate heterogeneity using the Akaike Information Criterion (AIC). Bayesian inference was implemented with MrBayes 3.1.2 [40] using a mixed model Bayesian analysis strategy. We assigned model parameters for each gene partition identified by AIC in Modeltest (Table 1). The Markov chain Monte Carlo (MCMC) algorithm was run for 10,000,000 generations with one cold and three heated chains, starting from random trees and sampling one out of every 1000 generations. Runs were repeated twice. The average standard deviation of split frequencies below 0.01 was examined to evaluate the convergence between the runs. The program Tracer 1.5 [41] was used to ensure that plots of the two analyses were converging on the same area and the log likelihoods had stabilized. The value of the effective sample size (ESS) for each statistic was above 200 after excluding the 25% burn-in. After discarding the trees saved prior to this point as burn-in, the remaining 7500 trees were used to construct majority-rule consensus trees using PAUP*. Nodes with posterior probabilities (PP) ≥ 0.95 in the consensus trees were considered statistically significant.

With Remusatia shown to be non-monophyletic in our initial analysis, we performed a constraint analysis using the concatenated plastid data set. With Remusatia constrained to be monophyletic, a parsimony analysis was performed with the heuristic search option using 1000 random sequence additions, TBR, and saving 100 trees from each random sequence addition. Shimodaira-Hasegawa (SH) test [42] was used to evaluate the topological congruence between gene trees produced by the likelihood method. The SH test was implemented in PAUP* with the best-fit model estimated using Modeltest 3.7 [38,39], RELL optimization, and 1000 bootstrap replicates to compare the difference between the RELL optimization and the computationally much more intensive full optimization. We compared the optimal trees (unconstrained) with constraint trees from the maximum likelihood analysis separately.