Background

In phylogenomic studies, gene tree-species tree incongruence can occur due to incomplete lineage sorting (ILS) and introgression, which operate through separate mechanisms. Incomplete lineage sorting happens when alleles within populations fail to coalesce due to the retention of ancestral polymorphisms. This can lead to hemiplasy (Avise, Robinson 2008), resulting in gene trees that do not align with the species tree. Incomplete lineage sorting is more common in large populations with short speciation intervals (Maddison, Knowles 2006) and is prevalent in sexually reproducing organisms, especially recent divergences (Steenwyk et al. 2023).

On the other hand, introgression occurs during hybridization when genetic material is exchanged between species, introducing alleles with evolutionary histories that differ from the species tree (Hibbins and Hahn 2022). Detecting introgression can be challenging, especially for large genomic regions, and methods like the D-statistic (Durand et al. 2011; Pease and Hahn 2015) and branch-length tests (Hahn, Hibbins 2019) are used. Both ILS and introgression contribute to gene tree-species tree incongruence, but they are distinct evolutionary processes that have unique implications for understanding the evolutionary history of organisms.

Natural selection is another factor that may influence phylogenetic conflicts (Steenwyk et al. 2023). The influence of natural selection on molecular evolution can introduce complexities into phylogenetic analyses. When positive selection leads to convergent evolution of sequences, it can mislead phylogenetic reconstructions by grouping sequences with similar traits due to selection rather than common ancestry (Sackton and Clark 2019). Accurately unraveling the evolutionary relationships among species and genes requires understanding the potential effects of natural selection on phylogenetic inference.

Although using genome-scale data for molecular phylogenies has potential benefits, recent studies have highlighted conflicts in large-scale phylogenomics (Steenwyk et al. 2023). The main challenge is that current methods may not be suitable for processing vast data, leading to systematic biases and inconsistent clade support (Philippe et al. 2011; Reddy et al. 2017; Salichos and Rokas 2013). Additionally, errors may arise from poor taxonomic sampling, varied gene orthology criteria, and complex interactions among these factors. These issues can be reduced by filtering datasets and focusing on informative loci, but incongruences may still persist (Reddy et al. 2017). As a result, hypothesis-testing procedures tailored for phylogenomics, such as gene genealogy interrogation (GGI), have become increasingly important (Betancur-R et al. 2019).

Aspidistra Ker Gawl. (Asparagaceae) is a subtropical to tropical genus, mainly distributed in eastern and southeastern Asia, comprising more than 200 species (Kalyuzhny et al. 2022). Species belonging to this genus are usually found in lowlands below an altitude of 1500 m (Tillich 2014). Most species have limited ranges; some are exclusively found in a single location. They grow in the understory of forests and usually flower on the ground, often concealed by litter material. As a result, new species are continuously being discovered and described.

The plant body of Aspidistra species is composed of a vertical rhizome with alternate leaves. The flowers are solitary or multiple and are located near the terminal nodes of the rhizome. The most crucial characteristic of Aspidistra is the flower structure, particularly the stigma, which varies greatly (Li 2004; Tillich 2014). An infragenus classification system for Aspidistra was proposed by Li (2004) based on stigma morphology, but this system was not supported by phylogenetic studies (Huang et al. 2013). Although most species have been described based on morphology, molecular biological methods have rarely been used to discuss Aspidistra species.

Taiwan, a continental island situated southeast off mainland Asia, harbours five taxa (three species and two varieties), including A. attenuata Hayata, A. daibuensis Hayata var. daibuensis, A. daibuensis Hayata var. longkiauensis C.T.Lu, Ming Jen Yang & J.C.Wang, A. mushaensis Hayata var. mushaensis, and A. mushaensis Hayata var. longiconnectiva (C.T.Lu, K.C.Chuang & J.C.Wang) C.T.Lu & J.C.Wang (Lu et al. 2020, 2023). These five taxa are all endemic to Taiwan. Due to morphological and genetic similarity, A. longiconnectiva (Lu et al. 2020) was recently reduced to A. mushaensis var. longiconnectiva (Lu et al. 2023). Another variety, A. daibuensis var. longkiauensis, is genetically distinguishable, revealed by simple-sequence repeats (SSR), despite similar morphology (Lu et al. 2023). The consolidation or separation of taxa, although only changes in rank, can affect evolutionary inferences and conservation applications. Therefore, it is necessary to provide more evidence to reexamine their taxonomic ranks.

When examining the evolutionary history of a species, it is expected that the different varieties of a species will be grouped in a monophyletic manner. The genetic separation of two varieties of A. daibuensis (Lu et al. 2023) may lead to two hypotheses: (1) var. daibuensis and var. longkiauensis are reproductively isolated and monophyletic, or (2) they have convergent phenotypes and are non-monophyletic. If the latter is true, these two taxa cannot be classified within a single species. However, the change in the taxonomic rank of Taiwanese Aspidistra has yet to be verified through phylogenetic analysis. Furthermore, A. mushaensis var. mushaensis and var. longiconnectiva have mixed genetic components from A. daibuensis var. longkiauensis and A. attenuata (Lu et al. 2023), implying a possibility of a hybrid origin of A. mushaensis.

Lu et al. (2023) conducted morphological analyses on Taiwanese Aspidistra and found minor genetic inconsistencies. The environment can influence many morphological characteristics, particularly vegetative organs (Shimizu-Inatsugi et al. 2023). While some traits may help identify differences between certain species, they may not apply to other taxa due to their lack of analytical power in reflecting phylogenetic relationships and evolutionary rates. On the other hand, phylogenetically conservative characteristics can provide more accurate information about evolutionary relationships and species divergence rates, making them better suited for systematic classification (Zhang et al. 2016). With this in mind, we will revisit Lu et al.‘s (2023) research to identify the most reliable diagnostic characteristics of Aspidistra in Taiwan.

In this study, we specifically conducted phylogenetic analysis on five taxa of Aspidistra from Taiwan. This study examined the taxonomic challenges and evolutionary relationships of the Aspidistra species in Taiwan. We employed both genomic analysis and morphological studies to clarify the taxonomic relationships among these species, which have previously been obscured by unclear classifications in Taiwan. Our research, which builds upon the findings of Lu et al. (2023), specifically examines the relationship between A. daibuensis var. daibuensis and A. daibuensis var. longkiauensis. Our analysis of thousands of genes obtained from transcriptomes highlights the impact of ILS, introgression, and positive selection on the conflict between gene trees and the species tree of our study subjects.

Materials and methods

Sampling and data collection

We collected nine samples from the field, encompassing five taxa in Taiwan. The plants collected were planted in a common garden, and fresh tissues were collected from the young shoots or root apical meristems for RNA extraction. Six species, A. zongbayi, A. elatior, A. fenghuangensis, Tupistra fungilliformis, Reineckea carnea, and Rohdea japonica were used as outgroups. The detailed information is listed in Table S1.

RNA extraction

Total RNA was extracted from plant powder using the modified CTAB method with NaCl and PVPP to remove polysaccharides and polyphenols. The extraction buffer contained 2% CTAB, 2% PVPP, 2 M NaCl, 100 mM Tris-base, 20 mM EDTA, pH 7.5, and 2% of β-mercaptoethanol. After heating and centrifuging, the supernatant was mixed with acid phenol-chloroform and centrifuged again. The aqueous phase was transferred to another tube with isopropanol and LiCl, and the mixture was centrifuged again before washing the pellet with 70% ethanol and resuspending in DEPC water.

RNA sequencing, filtering, and de novo assembly

The RNA samples were sequenced by Illumina NovaSeq 6000 platform with 150 bp pair-end sequencing at Genomics, Taiwan. Transcriptome files of A. fenghuangensis were downloaded from NCBI and converted to fastq format using the SRA toolkit. Data quality was assessed using FastQC v0.11.9 (Andrews 2010) and MultiQC v1.10 (Ewels et al. 2016). We removed adapters and low-quality reads by Trimmomatic v0.39 (Bolger et al. 2014). We performed de novo assembly using Trinity v2.12 (Grabherr et al. 2011) and removed redundant transcripts (> 99% similarity) using CD-HIT v4.8.1 (Fu et al. 2012; Li and Godzik 2006). TransDecoder v5.5.0 (Haas 2013) and OrthoFinder v2.5.4 (Emms and Kelly 2019) were used to predict genes and infer orthologs, respectively.

RNAseq dataset analyses

Species tree reconstruction

From orthologs inference, we obtained 1,029 single-copy genes. ASTRAL-II v5.7.7 (Mirarab and Warnow 2015) provides a method to infer the species tree from gene trees using a coalescent-based approach. To prepare for ASTRAL-II input, we constructed maximum likelihood gene trees by IQ-TREE v2.1.2 (Minh et al. 2020; Nguyen et al. 2015). The species tree was then examined with the “-q” option to determine the normalized quartet score and the “-t 3” option, providing the local posterior probability for each branch in ASTRAL-II.

We employed Bayesian multispecies coalescent (MSC) analysis to estimate species tree and molecular dating with a strick clock model in StarBEAST3 v1.0.5 (Bouckaert et al. 2019; Douglas et al. 2022). The prior clock rate was set at 0.005/site/million years, referring to the average rate of silent-site divergence (µ) in angiosperms (De La Torre et al. 2017). We retained a strict clock with a fixed substitution rate in StarBEAST3 because our focus was on relative divergence times among closely related Taiwanese taxa, for which a well-established angiosperm-wide average rate is appropriate. Due to the lack of suitable fossil evidence for calibration points and the limited analysis of 189 genes, the relaxed clock model’s effectiveness is restricted. Thus, we used a strict clock for preliminary computations. To address rate variation more fully, we complemented this approach with MCMCTree analyses using a relaxed-clock prior (see below). Due to the Bayesian Markov chain Monte Carlo being a time-consuming task, we only used the 189 genes whose topology aligns with the result of ASTRAL estimating by topological test (see below). For each gene, we used Jmodeltest v2.1.10 (Darriba et al. 2012; Guindon and Gascuel 2003) to select the best-fit substitution model according to BIC. Analyses were run for 100,000,000 Markov Chain Monte Carlo (MCMC) generations, sampling every 25,000th generation. We used Tracer v1.7 (Rambaut et al. 2018) to check the effective sample size (ESS) of parameters and discard 10% of the trees as burn-in. DensiTree (Bouckaert 2010) was used to visualize the posterior probability distributions of trees.

Given that Aspidistra represents a rapidly radiating lineage, substitution rate heterogeneity may bias molecular dating. To improve temporal calibration, we included outgroup species from closely related genera. Four secondary calibration points, based on Ji et al. (2022), were applied. For divergence time estimation, we identified 341 single-copy genes from 15 samples representing 11 operational taxonomic units (OTUs). Species tree topology was inferred using ASTRAL-II, followed by divergence time estimation using MCMCTree (Rannala and Yang 2007; Reis, Yang 2011). For OTUs represented by multiple samples, we generated consensus sequences using the most frequent base at each position, implemented with the R package Biostrings (Pagès et al. 2025). In MCMCTree, we implemented a relaxed-clock model with a lognormal prior on substitution rates, which allows for among-lineage rate variation. For the overall substitution rate, we estimated from the sequences by Baseml in PAML. We specified a diffuse gamma prior (G(1, 12.5), corresponding to Gamma (α = 1, β = 12.5) with a mean of 0.08 substitution per site per year, in order to avoid overly restrictive assumptions. This setting follows recommendations for datasets with potential rate heterogeneity. Two independent MCMC runs were conducted, each for two million generations with samples taken every 1,000 iterations, resulting in 20,000 samples per run. Convergence was assessed by comparing the posterior mean node ages between the two runs (Figure S1, and prior versus posterior distributions of calibrated node ages were evaluated (Figure S2).

Topological test

In order to comprehend phylogenetic conflicts, we utilized IQ-TREE v2.1.2 (Minh et al. 2020; Nguyen et al. 2015) to conduct a topological test. The topological test is the GGI strategy (Arcila et al. 2017). Our topological test focused on two key inquiries: (1) how many gene trees fit the dai-lgk morphological resemblance? Moreover, (2) how many gene trees fit the species trees reconstructed by ASTRAL-II? We combined the variety lcn into mus and tested the remaining four taxa for 15 hypothetical tree topologies for each gene (Fig. 3). Samples of the same species were constrained within a clade. The tests were evaluated using an approximately unbiased (AU) test (Shimodaira 2002) with 10,000 resamplings of estimated log-likelihoods (RELL) replicates (Kishino et al. 1990) in IQ-TREE. It is imperative to note that we selected the best-fit tree topology for each gene with the lowest ΔL value. Gene annotation was conducted for those genes that fit the dai-lgk morphological resemblance.

Positive selection

The positive selection was tested for 214 genes (see results) that did not reject the hypothesis of dai-lgk morphological resemblance. The branch model of PAML was conducted with a priori of the ASTRAL species tree and the dai and lgk lineages as foregrounds. The constant model was set as the null model. The likelihood ratio test assessed the significance of positive selection (Ka/Ks > 1).

Phylogenetic signals

We calculated the phylogenetic signal of 21 morphological traits measured by Lu et al. (2023) with Blomberg’s K (Blomberg et al. 2003) within R package geiger (Harmon et al. 2009). The species tree, reconstructed by StarBEAST, served as the input tree. We employed the Brownian model as the null model and conducted 10,000 simulations for a randomization test.

Gene annotation

We used MACSE v2 (Ranwez et al. 2011) to align the nucleotide sequences and translate them to amino acid sequences. Subsequently, BLASTP v2.11.0+ (Camacho et al. 2009) was used to compare amino acid sequences to the Arabidopsis Information Resource (TAIR) database (Berardini et al. 2015) with a threshold E-value of 1e-6.

SSR dataset analyses

STRUCTURE analysis

Lu et al. (2023) conducted a STRUCTURE analysis but only presented K = 2 and K = 4 without showing the hierarchical genetic structure. We replicated their analysis with their SSR data and presented the hierarchical STRUCTURE result (K = 2 ~ 6).

Network analysis

The SSR data was transformed to a distance tri-matrix and output nexus format using GenAlex v6.503 (Peakall and Smouse 2006) to reconstruct the network employing the NeighborNet algorithms within SplitsTree v4.14.4 (Huson and Bryant 2006). The output files generated by SplitsTree were then visualized using the R package ggnetworx (Paradis and Schliep 2019; Schliep et al. 2017; Schliep 2011; Yu et al. 2017).

Principal component analysis

The principal component analysis (PCA) was conducted by R package adegenet (Jombart 2008) and visualized by R package ggplot2 (Wickham et al. 2016).

Approximate bayesian computation (ABC)

According to the STRUCTURE analysis, two distinct groups, dai-mus and att-lgk, emerge when K = 2. However, at K = 4, the genetic resemblance of att-lgk disappears, while the dai-mus genetic cluster continues to persist. The dai-mus genetic cluster was also evident in the NeighborNet analysis. This observation leads to the formulation of a hypothesis suggesting a genetic resemblance between dai and mus. Additionally, when K = 4, it becomes evident that mus originates as a hybrid of dai and certain components of att, implying a hybridization scenario. Considering morphological characteristics, lgk resembles dai, classifying these two taxa as varieties of A. daibuensis. The species tree reconstruction, utilizing 1,029 genes with the MSC method, proposes a hypothesis wherein att and lgk share genetic similarities with mus (including its variety lcn), while dai shows a more distant relationship (see Results). Based on these findings, four plausible hypotheses are postulated:

Hypothesis 1

mus originated through hybridization between att and dai (hybrid scenarios: a and b).

Hypothesis 2

The species tree hypothesis suggests (((att, lgk), mus), dai) (1029-genes species tree scenario: c).

Hypothesis 3

dai exhibits greater similarity to mus than to lgk (dai-mus genetic resemblance scenarios: d, e, and f).

Hypothesis 4

dai and lgk form a monophyletic group (dai-lgk morphological resemblance scenarios: g, h, and i).

In total, nine evolutionary hypotheses were compared using Approximate Bayesian Computation (ABC). The summary statistics of observed data were computed using arlsumstat v3.5.2 (Excoffier and Lischer 2010). For each model, one million simulations were performed by fastsimcoal26 (Excoffier, Foll 2011; Excoffier et al. 2013). After simulations, we removed the highly correlated summary statistics with a Pearson correlation coefficient |r| > 0.8. The following model selection and parameter estimation were performed using the R package abc (Csilléry et al. 2012). Out of one million simulations, subsets of five thousand were retained for model selection, and the model with the highest posterior probability was considered the best model. The parameters were log-transformed and estimated from a subset of two thousand simulations.

Results

Species tree reconstruction and topological tests

The species tree reconstructed by ASTRAL with 1,029 genes and by the MSC method with 189 genes showed att-lgk and mus-lcn as sister groups, while dai appeared more distantly related (Fig. S3 and 1a). This tree topology supports the taxonomic treatment of mus (var. mushaensis) and lcn (var. longiconnectiva) within a species (A. mushaensis) but suggests lgk (var. longkiauensis) should not be a variety of A. daibuensis. The molecular dating indicates that the Taiwanese Aspidistra diverged from its sister A. elatior at 0.42 Ma (95% confidence interval (CI): 0.36–0.48 Ma). The divergence between mus and lcn has relatively low supporting value (0.509), suggesting their incomplete differentiation (Fig. 1a). The relationships were also recovered in the tree calibrated using secondary points; however, the estimated divergence times were older with the crown age of Aspidistra was at 3.27 Ma (95%CI: 2.34–3.97 Ma) and the origin of the Taiwanese Aspidistra lineage at 1.71 Ma (95%CI: 1.22–2.09 Ma) (Fig. 2).

Fig. 1
figure 1

(a) Species tree topology and molecular dating reconstructed by StarBEAST3 with 189 genes and (b) stigma width value of five taxa in Taiwan. The number on the node represents the supporting value/median value of molecular dating, and the number in the parentheses is the 95% confidence interval of molecular dating. The confusion between lcn and mus on DensiTree and their low supporting value shows they may not be completely differentiated

Full size image
Fig. 2
figure 2

Species tree inferred using ASTRAL, with divergence times estimated by MCMCTree. Red dots indicate the four secondary calibration points. Blue bars at the nodes represent the 95% credibility intervals of divergence time estimatess

Full size image

We performed a topological test to determine how many genes support the MSC species tree and merged lcn with mus to simplify the process. After analyzing 15 different topologies (Fig. 3), we found that the species tree topology was the most supported hypothesis, receiving the fewest rejections from the genes. It was found that 18.37% of genes (189 out of 1029) aligned with the species tree hypothesis (Fig. 3, H1). However, more than 80% of gene trees did not match the species tree topology, suggesting complex reticulate evolutionary connections as revealed by NeighborNet (Fig. 4a). This also suggests the presence of incomplete lineage sorting (ILS), introgression, or selection on these genes.

Fig. 3
figure 3

The summary of the AU test results includes 15 hypothetical tree topologies, along with the number of genes in each tree that were unable to reject the topology. It’s worth noting that a single gene may not reject just one tree topology, leading to the sum of genes (1466) exceeding 1029. Abbreviations att, dai, lgk, mus, and ela denote A. attenuate, A. daibuensis var. daibuensis, A. daibuensis var. longkiauensis, A. mushaensis (including var. mushaensis and var. longiconnectiva), and A. elatior, respectively

Full size image
Fig. 4
figure 4

Genetic clustering of Aspidistra in Taiwan. (a) NeighborNet network; (b) Principal component analysis (PCA). Abbreviations att, dai, lgk, mus, and lcn denote A. attenuata, A. daibuensis var. daibuensis, A. daibuensis var. longkiauensis, A. mushaensis var. mushaensis, and A. mushaensis var. longiconnectiva, respectively

Full size image

dai-lgk resemblance

Since the current taxonomic treatment that dai and lgk were within a species, we performed the topological test to examine how many genes were consistent with the sister relationship of dai and lgk. The result shows that 20.80% of genes (214 out of 1029) did not reject the monophyletic relationship between dai and lgk (69, 84, and 95 gene trees for H6, H7, and H14, respectively). These genes revealed diverse functions (Table S2). Among these genes, 19 were found to be chloroplastic, suggesting that photosynthesis-related functions highly contribute to dai and lgk grouping. Additionally, at least seven genes are ubiquitin-related genes, including E3 ubiquitin-protein ligase, Light-mediated development protein DET1, Ubiquitin system component Cue protein, Ubiquitin family protein AT2G35360, E3 ubiquitin-protein ligase RNF170-like protein, E2 ubiquitin-conjugating enzyme UBC19, and E3 ubiquitin-protein ligase UPL5 (Table S2). Ubiquitin mediates plant cell signaling, stress responses, immunity, and morphogenesis (Doroodian, Hua 2021; Li and Li 2014; Pokhilko et al. 2011; Wang and Deng 2011). We are uncertain about how these genes resemble the morphology of dai and lgk, but their responses to the environments may play important roles.

Positive selection for the dai-lgk resemblance

The PAML branch model observed that the dai and lgk lineages showed significant positive selection signals (Ka/Ks > 1) in four genes. These genes are EARLY LIGHT-INDUCIBLE PROTEIN (ELIP1), Iojap (Ij), PROTEASOME BIOGENESIS-ASSOCIATED CHAPERONE 5 (PBAC5), and Protein arginine N-methyltransferase 1.1 (PRMT11). ELIP1 and Ij determine plant color, PBAC5 encodes chaperone function, and PRMT11 involves arginine methyltransferase activity.

ILS and introgression

Among the high proportion of gene trees conflicting with the species tree, most were considered as a consequence of the ILS. The QuIBL analysis suggests that, among 12 triplet tree topologies, only four are detected to have significant ILS + introgression events (ΔBIC: BICILS + intro – BICILS only < −10 (Edelman et al. 2019), while the others cannot reject the ILS-only hypothesis. In these four ILS + introgression triplet tree topologies, only a few proportions of genes can be detected to involve introgression (1, 4, 6, and 12 genes, Table 1). Compared with hundreds of ILS genes, the proportion of these introgression genes is only < 2% (0.10%~1.17%). This shows that the recent rapid speciation of these Taiwanese Aspidistra has left numerous genes unable to be completely sorted.

Table 1 Summary statistics of the proportion of ILS and introgression (non-ILS) genes inferred by QuIBL. The results showed that the ILS-only model causes most phylogenetic discordances of gene trees, and only 4, 6, 12, and 1 genes (BICILS + intro – BICILS only < −10 (Edelman et al. 2019) resulted from introgression under the ILS + introgression model
Full size table

Phylogenetic signal

According to the analysis of the phylogenetic signal, only the stigma width (SW) showed a significant Blomberg et al. (2003) K value greater than 1, indicating that this trait evolved with a closer adherence to the phylogeny, that is, the trait is primarily driven by evolutionary constraints or shared ancestry, and tends to follow the phylogenetic tree closely (i.e., phylogenetic conservatism, Table 2; Fig. 2b). The other traits did not show significant estimates deviating from 1 (P > 0.05), suggesting that they evolved following the Brownian motion model and were under genetic drift or neutral evolution.

Table 2 Phylogenetic signals of quantitative traits estimated by Blomberg et al. (2003) K with 10,000 randomization. All trait measurements were adopted from Lu et al. (2023)
Full size table

Network analysis

The NeighborNet analysis clearly demonstrates species differentiation, with only a few misidentified samples, such as between lgk and att (Fig. 4a), despite confirmed morphology. Moreover, lcn was appropriately grouped with mus, supporting its taxonomic status as an A. mushaensis variety. The intricate reticulate structures in the network strongly imply interbreeding or introgression events between these species.

Principal component analysis

According to the PCA, att, lgk, dai, and mus were separated, with lcn being clustered with mus (Fig. 4b), confirming that mus and lcn are of the same species. The first axis of PCA, which accounts for 32.5% of genetic variations, failed to distinguish dai and mus, while lgk is a subset of att. This finding aligns with NeighborNet, which clustered lcn with mus and placed dai closer to mus. These results demonstrate a genetic similarity between dai and mus.

Bayesian clustering analysis revealed by STRUCTURE

We analyzed the SSR data of Lu et al. (2023) in different cluster numbers, ranging from K = 2 to K = 6 (Fig. 5). Our findings show that att and lgk belong to the same cluster, while dai, mus, and lcn belong to a different cluster when K = 2. However, when K = 3, lgk and dai showed mixed genetic components with a part of att and mus + lcn components, respectively. When K = 4, as Lu et al. (2023) interpreted, mus and lcn comprised genetic components from dai and att, which led to the hypothesis of hybridization (hypothesis 1 in ABC). However, when K = 5, the genetic components in att that contributed to mus became relatively small, and when K = 6, the genetic contribution of dai into mus was small. Our analysis of the genetic contributions at different hierarchical clustering levels suggests that the genetic mixing may not solely be attributed to hybridization but could be more complicated by ancestral polymorphism or ILS.

Fig. 5
figure 5

Hierarchical clustering of the STRUCTURE from K = 2 to K = 6. Data were collected from Lu et al. (2023). Abbreviations att, dai, lgk, mus, and lcn denote A. attenuata, A. daibuensis var. daibuensis, A. daibuensis var. longkiauensis, A. mushaensis var. mushaensis, and A. mushaensis var. longiconnectiva, respectively

Full size image

Approximate bayesian computation

The ABC suggests scenario c as the most probable (posterior probability 0.6711) (Fig. 6). Parameter estimation of the best model shows att diverged from lgk 8.8 kya, mus 18.6 kya, and dai 93.1 kya, given a generation time of one year (Fig. 6c). Scenario d of the dai-mus genetic resemblance scenarios scored the second-highest probability of 0.1856 (Fig. 6d). The top two scenarios cluster lgk with att instead of with dai that reject the morphological resemblance scenarios between dai and lgk, which classify them as varieties of A. daibuensis.

Fig. 6
figure 6

Nine scenarios of the evolutionary relationship of Aspidistra in Taiwan. For simplify the model, A. mushaensis var. longiconnectiva (lcn) was incoporated into A. mushaensis (mus). These nine scenarios represent four evolutionary hypotheses: H1: hybrid scenarios (a and b); H2: species tree scenario (c); H3: dai-mus genetic resemblance scenarios (d, e, and f); H4: dai-lgk morphological resemblance scenarios (g, h, and i). The divergence time estimations were conducted for the best scenario of each hypothesis. The best scenario of the nine scenarios was coloured black, while the others were grey. Abbreviations att, dai, lgk, and mus denote A. attenuata, A. daibuensis var. daibuensis, A. daibuensis var. longkiauensis, and A. mushaensis, respectively

Full size image

Taxonomic treatment

Aspidistra longkiauensis (C.T. Lu, Ming Jen Yang & J.C. Wang) C.T. Lu & J.C. Wang, stat. nov.

Basionym Aspidistra daibuensis Hayata var. longkiauensis C.T.Lu, Ming Jen Yang & J.C.Wang, PhytoKeys 222: 129–151. 2023.

Type material examined Taiwan. Pintung County, Shuangliou National Forest Recreation Area, Banyan Trail, elev. 200–300 m, 12 Jun 2020, M.J.Yang s.n. (holotype: TAIF; isotype: TNU).

Distribution and habitat

Aspidistra longkiauensis is native to the Hengchun Peninsula and southeastern Taiwan. It grows on forested slopes at elevations of 200 to 500 m.

Note

This species resembles A. daibuensis, but it has a shorter leaf length (72.24 ± 11.68 cm vs. 95.46 ± 21.04 cm), a smaller leaf blade length-to-width ratio (3.73 ± 1.43 vs. 8.05 ± 1.73), thinner perianth lobes (1.28 ± 0.38 mm vs. 2.79 ± 0.64 mm), and a longer stigma curvature (10.98 ± 1.38 mm vs. 16.23 ± 1.45 mm) and narrower stigma width (10.33 ± 1.79 mm vs. 16.43 ± 2.21 mm). It also has a larger pistil-to-stamen height ratio (2.49 ± 0.59 vs. 2.03 ± 0.78). Additionally, it is also similar to A. mushaensis, but this species differs in its smaller leaf blade ratio (3.73 ± 1.43 vs. 7.66 ± 1.35), narrower perianth lobe base (5.12 ± 1.12 mm vs. 7.27 ± 1.19 mm), and larger pistil-to-stamen height ratio (2.49 ± 0.59 vs. 1.59 ± 0.26).

Discussion

Phylogenetic inference by gene genealogy interrogation (GGI)

While the gene tree topology, which comprises 189 genes, is consistent with the species tree derived from the MSC method inferred by ASTRAL, more than 80% of the genes support alternative tree topologies in the AU test. As the GGI strategy, the AU test ranks alternative resolutions supported by individual genes, revealing support distribution for different topologies and mitigating gene tree estimation errors in coalescent analyses (Betancur-R et al. 2019). Thus, only 189 genes were utilized to determine the divergence times among species through dating with the MSC tree. This revealed very recent divergent times among species (Bagley et al. 2020). On the other hand, the genes that did not conform to the MSC species tree topology indicate a notable occurrence of ILS in these closely related species (Degnan and Salter 2005). However, the proportion of ILS in this case (> 90%, Table 1) is much higher than the estimation of approximately 10% by Rosenfeld et al. (2012).

The AU test inferred that the top three supported topologies clustered mus, att, and lgk, with a distant phylogenetic taxon, dai, backed by 189, 150, and 119 genes. It is noteworthy that these three species (mus, att, and lgk) are predominantly located on the western side of Taiwan’s Central Mountain Range, compared to the east-distributed dai, indicating that these 458 genes could potentially distinguish the diverse environments between the eastern and western Taiwan, although the common characteristics of these genes are yet to be determined. In addition, the ABC result also suggests that mus, att, and lgk had a very recent coalescence time (18.6 kya) during the LGM period, in contrast to the relatively long divergence of dai (93.1 kya) (Fig. 6c). Even in the BEAST analysis, the divergence among mus, att, and lgk was 154.3 kya, much shorter than the divergence time of 281.3 kya from dai (Fig. 2a).

It is believed that the eastern-distributed dai was the first to diverge from the other species. As the diversity center of Aspidistra was located in southern regions like Guangxi Province of China and Vietnam, the ancestral area of mus, lcn, att, and lgk was likely in the south of Taiwan. Moreover, since the Center Mountain Range of Taiwan acts as a barrier, it is less feasible for the species to disperse westward across the mountains. Therefore, based on the tree topology, after mus and lcn branched off from att and lgk, they may have expanded northward in western Taiwan. The current distribution trends are a direct result of the postglacial dispersion of the common ancestor, with the environment potentially playing a significant role.

It is important to note that only stigma width demonstrates a strong correlation with species relationships among the morphological traits considered. While other features may differentiate species, they are all subject to random variation under the Brownian motion model. Speciation is the natural ecological process that leads to biodiversity synthesis (Swenson 2011), and traits that have evolved alongside this process, such as the stigma width of Taiwanese Aspidistra, may be more suitable as diagnostic characters for taxonomy (e.g., Zhang et al. 2016). In traditional taxonomical studies, the shape of the stigma has been used as a critical character for the species-level classification of Aspidistra. This study supports this view.

The width of the stigma plays a crucial role in filtering the body size of pollinators since the anthers of Aspidistra are below the stigma. The evolution of stigma width could be critical for the reproductive barriers between Aspidistra species. Importantly, the conservative nature of stigma width suggests that this trait may provide a practical and reliable morphological marker in future taxonomic revisions of Aspidistra. When combined with broader geographic and genomic sampling, stigma width has the potential to facilitate integrative approaches to resolving species boundaries across the genus. On the other hand, the other morphological traits can exhibit plasticity during evolution in response to environmental changes. Such changes are not necessarily evolutionarily constrained and can lead to homoplasious characters (Brandley et al. 2009). Although these traits may help distinguish a single species, they are unsuitable for identifying an entire species group.

Raising the taxonomic rank of A. daibuensis var. longkiauensis to A. longkiauensis

Genetic evidence suggests that dai and lgk may not be considered a single but two potentially independent species. However, because of the morphological resemblance, dai (A. daibuensis var. daibuensis) and lgk (var. longkiauensis) were considered as varieties of the same species previously (i.e., dai-lgk morphological resemblance hypothesis). This distinction could be attributed to similarities in certain genes, leading to homoplasious characteristics within a phylogenetic context (Wake et al. 2011). In our study system, the SSR dataset partially supports the hypothesis that dai and lgk share morphological similarities, with an approximate 10% posterior probability in ABC analysis. On the other hand, the RNAseq dataset indicates that 20.8% of genes do not reject the idea of clustering dai and lgk together in their phylogeny. Among these genes, we have identified multiple genes that play a role in chloroplastic functions, suggesting a convergence in photosynthetic processes, such as electron transfer (AOX4 gene) and redox reactions (ACHT2 gene), as well as chlorophyll-related processes, including chlorophyll degradation (PAO gene), in both taxa. Additionally, the nuclear gene DET1, responsible for encoding a light-mediated development protein involved in fruit pigmentation and nutritional quality, also appears to be linked to morphogenesis. Specifically, four other genes, ELIP1, Ij, PBAC5, and PRMT11, were detected to be positively selected. ELIP1 plays a crucial role in the cellular response to light and regulates chlorophyll biosynthesis (Casazza et al. 2005) and seed germination (Rizza et al. 2011). Iojap is a chloroplastic gene known for its unique green and white striped leaf pattern (Han et al. 1992). PBAC5 is a period circadian protein highly expressed in the floral, leaves, and seed parts (Klepikova et al. 2016) and is involved in the function of the chaperone (Gemperline et al. 2019). These three genes share a light response trait, potentially impacting the plant’s coloration. PRMT11 functions in epigenetics (DNA methylation) and arginine methylation (Scebba et al. 2007). The Aspidistra plant thrives in shaded areas of forest floors and is not well-suited to direct sunlight. Changes in amino acids on certain genes indicate adaptive signals for the dai and lgk branches. Although the exact link between these genes and the physical similarities of dai and lgk is still not fully comprehended, their resemblance may be related to their physiological and photomorphogenic co-adaptation.

Despite morphological similarity, the ABC analysis supports a non-monophyletic relationship between A. daibuensis var. daibuensis and var. longkiauensis. The result matches the MSC species tree from 1,029 coding sequences, although only 18.4% of genes confirmed the species tree topology in the AU test. It is possible that the ILS and ancient interbreeding could be the reason for the high percentage of genetic inconsistency with the species tree topology (Edwards et al. 2007). A relatively short divergence time (< 1 Ma) implied that multiple genes may not have had enough time to differentiate completely, leading to ILS (Bagley et al. 2020; e.g., Murillo-A et al. 2022). This, in turn, means no substantial genetic barriers preventing gene flow between species, which can cause genetic leaks across reproductive barriers (Wu and Ting 2004).

However, the hybrid scenario inferred by STRUCTURE was rejected in ABC with a relatively low posterior probability. Lu et al. (2023) had shown that mus was comprised of genetic components from att and dai. Thus, a hybrid hypothesis could have arisen. However, only 4% posterior probability supports this hypothesis in ABC. QuIBL also suggests a high frequency of ILS rather than introgression, indicating that the genetic leaks across the reproductive barrier only occurred between gene levels instead of species levels. Thus, the frequent hybridization may not explain the reticulate evolution revealed in the NeighborNet. It may result from the ILS due to the short divergence time. In fact, the interbreeding between species was also less observed in the field. Evident geographical isolation confined small-area distribution may accelerate species differentiation, which may be explained by genetic drift of small populations and local adaptation (e.g., Liu et al. 2022), but further research is needed.

Conclusion

In conclusion, our study provides valuable insights into the complex evolutionary dynamics and taxonomic relationships within the Taiwanese Aspidistra species group. Phylogenetic inference through GGI revealed that while the gene tree topology generally aligns with the species tree derived from the MSC method, a substantial proportion of genes supported alternative tree topologies, suggesting the presence of ILS in these closely related species. The genetic evidence suggests that A. daibuensis var. daibuensis and var. longkiauensis should not be considered a single species, as previously believed, due to morphological resemblance. Instead, they exhibit non-monophyletic relationships.

Interestingly, certain genes related to chloroplastic functions and photomorphogenic co-adaptation were identified, shedding light on potential physiological and adaptive similarities between these taxa. Additionally, our findings emphasize the importance of stigma shape, particularly stigma width, as a diagnostic character in traditional species-level taxonomy. This trait is also recognized as phylogenetically conservative. It may play a crucial role in future integrative taxonomic revisions, especially when combined with more extensive sampling across Asia. This characteristic might offer stronger indications of speciation events compared to other morphological features subject to random variation.

Notably, our phylogenomic analyses indicate that A. longkiauensis is distinct from A. daibuensis within the Taiwanese clade. However, our findings are based solely on endemic species from Taiwan. Given that Aspidistra includes over 200 species across Asia, our conclusions are provisional and require broader pan-Asian sampling. Expanding taxonomic and geographic coverage will be crucial to assess whether the evolutionary patterns observed are consistent across the genus and to improve the systematic framework of Aspidistra. Overall, this study highlights the intricate interplay of genetics, morphology, and ecology in the evolution of closely related plant species and calls for further research to elucidate the underlying mechanisms driving their diversification and adaptation.