To evaluate the effect of intraspecific sampling size on DNA barcoding, two to three individuals per population, and three to eight populations for each species were collected representing its entire distribution range. In total, 103 individuals from 39 populations of nine species of Eurasian yews identified in Liu et al. [29] (with the exception of T. sumatrana and the Emei type, for which no population samples were available.), were used in the present study (Supplementary Information, Table S1). Among the 103 individuals, 156 sequences from 39 individuals were based on Liu et al. [29], and 256 sequences from 64 individuals were newly sequenced in the present study (Supplementary Information, Table S1). Voucher specimens were deposited in the herbaria of Kunming Institute of Botany (KUN) and/or Royal Botanic Garden Edinburgh (E).

Total genomic DNA was isolated using a modified 4 × CTAB method [31] from silica-gel dried leaf materials. PCR amplification and sequencing of the matK, trnL-trnF, trnH-psbA and ITS regions were performed as described by Liu et al. [29].

Sequences were assembled and edited using SeqMan (DNA STAR package, DNAStar Inc., Madison, WI, USA). Each DNA region was aligned using CLUSTAL × 2.0 with default parameters [32], and the alignment modified manually where required under EditPlus Text Editor 3.20 [33]. Sequences were checked against those in GenBank using the BLAST algorithm. All sequences have been submitted to GenBank (Supplementary Information, Table S1).

For the sampling strategy, we used three criteria to identify the optimal intraspecific sampling. The first was the genetic distance at the population and species level; the second was the haplotype diversity of given species, and the last was clustering of the accessions using a tree building approach. Inter- and intra-specific P-distances for each locus were calculated using MEGA 4.0 [34] at both the population and species levels. Populations WX and KV were excluded from the intra-population analyses due to only a single individual being sequenced successfully. The number of haplotypes per locus per species was determined using MEGA 4.0. The neighbor-joining (NJ) trees were constructed with the P-distance model in MEGA 4.0. Bootstrap value for all clades was assessed with 5000 bootstrap replicates. The optimal sampling strategy for DNA barcoding was identified as the point where increasing the numbers of individuals at the population or species level did not change further the genetic distance, and clustering relationships of the given species. Of course, we also considered the number of haplotypes found in each species when we defined the idea sampling strategy.

Alignment of the two non-coding chloroplast regions (trnL-trnF and trnH-psbA) revealed the occurrence of many indels. Thus, we used these two data sets to test the utility of the indels for barcoding closely related species by using four different indel treatments, (1) complete deletion (abbreviated hereafter as CD), sites containing alignment were removed prior to the analysis; (2) pairwise deletion (PWD), in which indels were removed during the analysis as the need arises (e.g., pairwise distance computation); (3) simple indel coding (SIC), which was implemented by coding all indels that have different 5′ and/or 3′ termini as separate presence/absence characters; in this method, whenever a gap from one sequence contains a smaller gap in another sequence, the longer, completely overlapping gap is coded as inapplicable (Figure 1 in [22]); (4) modified complex indel coding (MCIC), which not only corrected costs downwards compared to the complex indel coding method by Simmons and Ochoterena [22], but also maintained symmetry for all step matrices, and treated overlapping indels as multistate characters [35]. To compare these four indel treating schemes, we used a tree-building method to identify the species identification success by assessing bootstrap values as described by Liu et al. [29]. If all of the individuals of the given species were clustered in a clade with a nodal support value of greater than 50%, we considered this as successful sequence identification. NJ trees were constructed using the P-distance model in MEGA 4.0 as described above. Numbers of variable sites and parsimony-informative sites for each region were also estimated with MEGA 4.0.

The matK alignment was 1533 bp in length, with no length variation in any of the sampled individuals (Table 1). Length variation was observed in the trnL-trnF, trnH-psbA and ITS sequences (Table 1), which ranged from 797 bp to 852 bp, 532 bp to 981 bp and 1135 bp to 1141 bp, respectively. The aligned matrix of trnL-trnF was 869 characters in length with 11 indels which ranged from 1 to 41 bp. The length of trnH-psbA matrix had 1321 characters including 13 indels, ranging from 1 to 474 bp in length. The length of ITS aligned matrix was 1143 with 5 mononucleotide indels.

Mean interspecific and intraspecific P-distances differed among the four regions (Table 1). Seven of the nine species exhibited no intra-specific variation for matK, whilst four and five species lacked intra-specific variation for both the trnL-trnF and trnH-psbA regions (Table 2). For ITS, three species showed no intra-specific variation (Table 2). One to three haplotypes per species were found in the matK region, and one to eight, one to ten and one to six were found for trnL-trnF, trnH-psbA and ITS, respectively (Table 2). Among the 37 multiple-sampled populations of the nine species, only one population of T. fuana (GL) exhibited polymorphism for matK, eight populations were polymorphic for trnH-psbA, and eleven for both trnL-trnF and ITS (Supplementary Information, Table S2). For the polymorphic populations within each species for the four regions, intra-population P-distances were generally less than inter-population and intra-specific P-distances (Table 2).

The NJ tree of the nine species based on combination of the four DNA regions is shown in Figure 1. All the 103 individuals fell into distinct clades with high bootstrap support values corresponding to the nine species. The clustering relationships of the species were similar to those found by Liu et al. [29], which utilized smaller sample sizes for each species.

The numbers of variable and parsimony-informative characters varied between the two indel coding approaches (SIC and MCIC) for trnL-trnF and trnH-psbA (Table 1). The variable sites and parsimony-informative characters ranged from 2.88% to 4.09% with MCIC, and 2.53% to 3.30% with SIC for trnL-trnF, and ranged from 1.29% to 2.75% and 0.98% to 1.97% with MCIC and SIC for trnH-psbA, respectively (Table 1). When comparing numbers of the variable sites and parsimony-informative characters, CD = PWD < MCIC < SIC. Indel-coding increased the numbers of variable and parsimony-informative characters, and SIC generally resulted in more variable and parsimony-informative sites than MCIC (Table 1).

NJ trees for each region were constructed with different indel treatments by MEGA 4.0. Clades were recovered for each species and nodal support values are shown in Table 3. Nodal support values base on SIC and MCIC were always higher than those using CD and PWD in the trnL-trnF NJ tree, except for T. cuspidata which had the highest bootstrap support with PWD. All the nodal support values for each clade were over 70% except for a value of 58% in the Qinling type. Seven out of the nine species were identified with CD and PWD, eight with SIC, and all nine species with MCIC. For trnH-psbA, PWD, SIC, MCIC all performed better than CD on species resolution and nodal support. Only one of the nine species was discriminated with CD, and three with PWD. SIC and MCIC provided matching species discriminatory power, identifying four of the nine species (44%) (Table 3). It is of note that two and one of the nine species could be distinguished using the trnL-trnF and trnH-psbA regions respectively when considering only indels variation (Table 3).

The number of individuals required to create a reliable reference for valid species identification has been one of the basic issues considered since the beginning of the DNA barcoding initiative [18,21]. Generally, the depth of individuals per species sampled for barcoding is usually sacrificed in favor of greater taxonomic coverage [36]. In the present study, levels of sequence divergence within and between species differed among the four DNA regions in this study, most likely as a result of different evolutionary processes/levels of functional constraint. Little or no intraspecific sequence divergence was observed in matK, indicating that 2–3 individuals from different populations would be representative of the genetic diversity of each species (Table 2). For more quickly evolving regions, such as trnH-psbA, trnL-trnF and ITS, high levels of intraspecific sequence divergence mean that more individuals (a maximum of 10, 8 and 6) from different populations are required to represent the majority of the genetic variation at these loci (Table 2). Although a sample size of 12 individuals per species was proposed for barcoding animals by Matz & Nielsen [18], a study based on simulated data and real data from the mtDNA COI region for the skipper butterfly [21] suggested that much larger sample sizes were required to be representative of the total genetic diversity. Based on the results in this study and those from a separate phylogeographic study of Taxus (Liu et al., unpublished data), 8–10 individuals per species from the entire geographic distribution of the species analysed appear to be sufficient for plant DNA barcoding. Nevertheless, more detailed studies on this issue are required using simulated data and real data on a large sample scale in future.

In our study, intraspecific sampling was conducted in a hierarchical fashion to elucidate whether variation between populations could influence sampling strategies for barcoding. Most of the populations studied for multiple individuals were monomorphic for all four DNA regions (Supplementary Information, Table S2). Intra-population sequence divergence was usually less than that observed between populations within species (Table 2), and mean interspecific distances (Table 1) and clustering relationships of the species in the NJ tree (Figure 1) were similar to a previous study which only considered a single individual per population [29]. Thus, increasing numbers of individuals within populations does not influence barcoding accuracy, and one individual per population is likely to be adequate for the majority of plant DNA barcoding projects, especially for the closely related taxa. Most of the variation within species is generally due to differences between geographically distant populations. Thus it is more important to randomly sample multiple individuals across the whole geographical distribution of a species, a similar finding to that from a previous study [21].

As methods for coding indels have become more sophisticated, the inclusion of indels as characters in phylogenetic analyses has gained increasing popularity [37]. Many studies concluded that gaps should be included, in some way, in order to provide additional phylogenetic information (e.g., [23,24,28]). In plant DNA barcoding studies, indels have generally not been taken into account due to the lack of a standard approach to the utilization of such regions. Studies have generally treated indels as missing data [6,11,12], in a few cases, as a fifth character [28] or as some indel coding method [26].

Length variation as a result of mononucleotide repeat expansion and contraction in the chloroplast genome has frequently been utilized as a marker system in population genetic studies [38]. The bidirectional nature of the mutational processes operating at these regions, however, leads to homoplasy, particularly above the species level, and thus they are not generally considered as informative in phylogenetic studies, or even in phylogeographic studies across large geographic scales, e.g., [39]. In the present study, as in previous other studies, e.g., [40], similarly, polymorphic mononucleotide regions have been identified in the nuclear ITS. Given that these result from the same mutational mechanisms as their chloroplast counterparts, they are unlikely to provide stable information for inter-specific barcoding studies.

Rapidly evolving non-coding plant DNA regions, such as trnH-psbA and trnL-trnF, usually exhibit length variation due to the occurrence of indels, e.g., [7,41]. Where such length variation makes it difficult to unambiguously align sequences, this represents a disadvantage in DNA barcoding [11], although this is primarily a problem in distantly related taxa. Conversely, the potentially diagnostic nature of indels for closely related species is highlighted in the present study, as well as in previous barcoding researches [6,12,42]. Our study shows that some species have private intraspecific indels that distinguish them from other closely related species. For instance, T. chinensis has a specific insertion at 514 position in trnH-psbA matrix which differentiated it from the other species studied. Likewise, T. mairei can only be distinguished from the ‘Tonkin type’ by a private indel in the trnL-trnF region.

Given that indels are known to provide some level of information for phylogenetic studies, e.g., [24,37], we utilised four different methods of indel treatment to test their performance in DNA barcoding analysis. Based on our results, the two gap coding methods (SIC or MCIC) performed better in species discrimination and nodal support than CD and PWD, and both approaches usually obtained similar species resolution and nodal support values. A comparison of the CD and PWD approaches revealed that they performed similarly in trnL-trnF, but that PWD performed better in species identification than CD in trnH-psbA. Overall, the CD method showed the lowest species discrimination and nodal support. Using this approach, where indels were effectively treated as missing data, the resulting phylogentic trees were less accurate than those obtained using gap coding treatments [24]. In general, gap coding methods of SIC and MCIC increased the number of variable and parsimony-informative sites, which gave more accurate clustering relationships and a stronger nodal support value. Where the indel was a species-specific diagnostic character, this always led to the accessions of the species forming a clade and subsequent species identification success. Thus, for the barcoding of the closely related species of Taxus examined in this study, it was preferable to use a gap-coding method (SIC or MCIC) rather than the CD approach.

The findings of the study confirm the potential utility of indels in plant barcoding studies, as previously suggested in other species [42]. Considering that the rate of base pair substitutions in the chloroplast genome is fairly low [43,44], longer stretches of DNA may be required to generate sufficient barcoding information, which will in turn increase the cost of such ventures. Given the high frequency of indels, particularly in the two regions studied here, and their phylogenetic utility in species discrimination, we suggest that with the appropriate treatment, they may provide a valuable addition to many plant barcoding studies.