Trying to Understand the Complicated Taxonomy in Amaranthus (Amaranthaceae): Insights on Seeds Micromorphology

Amaranthus is a genus taxonomically complex because of its high morphological variability, which led to nomenclatural disorders, misapplication of names, and misidentifications. Floristic and taxonomic studies on this genus are still incomplete, and many questions remain open. Seed micromorphology has been shown to play an important role in the taxonomy of plants. Regarding Amaranthaceae and Amaranthus, investigations are rare, and they refer to one or a few species. With the primary aim to test if seed features are helpful in the taxonomy of Amaranthus, we here present a detailed SEM study on seed micromorphology in 25 Amaranthus taxa using morphometric methods. Seeds were collected from field surveys and herbarium specimens; 14 seed coat features (7 qualitative and 7 quantitative) were measured on 111 samples (up to 5 seeds per sample). The results obtained revealed that seeds micromorphology provides interesting new taxonomic data concerning some taxa (species and below ranks). In fact, we were able to distinguish a few seed types, including one or more taxa, i.e., blitum-type, crassipes-type, deflexus-type, tuberculatus-type, and viridis-type. On the other hand, seed features are not useful for other species, for example, those included in the deflexus-type (A. deflexus, A. vulgatissimus, A. cacciatoi, A. spinosus, A. dubius, and A. stadleyanus). A diagnostic key of the studied taxa is proposed. Subgenera cannot be distinguished using seed features, thus confirming the published molecular data. All these facts reveal, once again, the taxonomic complexity of the genus Amaranthus since, e.g., just a few seed types can be defined.

Mosyakin & Robertson [14] proposed a classification of Amaranthus recognizing three subgenera, i.e., subgenus Acnida (L.) Aellen ex K. R The identification of the species was made by using a stereomicroscope LEICA EZ4W and following literature [1,3,42]. The Nomenclature of the names follows [41], except for Amaranthus emarginatus Salzm. ex Uline & Bray (recognized as a subspecies of A. blitum L. by [41]) that is here accepted at species rank according to [43].

Scanning Electron Microscopic Analyses
Micro-morphological seed traits were examined by a scanning electron microscope SEM (JSM5910, 3 kv voltage, and secondary electron detector). The seeds for each Amaranthus species were mounted on metallic stubs using double adhesive tape and coated with gold for 6 min in a sputtering chamber followed by observation under SEM. The photographs were taken using different magnifications (from 50× to 10,000×) depending on the size of the seeds. For each seed, SEM micrographs were taken in lateral, frontal, and apical views, and on the hilum region.

Morphometric Analysis
In total, 25 taxa, up to 5 sites per taxon, and up to 5 seeds per sample (mature and not deformed/broken) were studied (Table 1). In total, 14 characters [7 qualitative, 7 quantitative (Table 2)] were measured on 111 samples (a total of 1554 measurements) using a scanning electron microscope SEM (JSM5910, 3 kv voltage, and secondary electron detector). Definitions of qualitative characters (excepting color) with associated typesimages are reported in Table 3.  Table 1. List of the Amaranthus taxa studied (alphabetical order). Native distribution areas from [41] and, concerning A. blitum, from [3]. S: South; W: West; N: North: E: East; C: Central.

Code Taxon
Native Distribution Area al Amaranthus albus L.
Ecuador to NW-Argentina cr Amaranthus centralis J. Palmer  S-America  Table 2. Characters measured for the morphometric analysis. The characters labelled with an asterisk (*) are qualitative, and the others are quantitative (see Table 3).

Present Not present
The data matrix (samples × variables) was processed using the software NCSS 2007. The variability of the characters has been examined by cluster analysis (UPGMA method), principal component analysis (PCA), discriminant analysis (DA), and box plots [in yellow boxes presented, illustrate interquartile ranges (= the range between the 25th and 75th percentile) and medians (horizontal line); vertical lines are the whiskers that represent the scores outside the middle 50% (i.e., the lower 25% of scores and the upper 25% of scores)]. PCA analysis was performed both by excluding the qualitative characters and by including them as binary variables according to [44]. DA was performed using the

Present Not present
The data matrix (samples × variables) was processed using the software NCSS 2007. The variability of the characters has been examined by cluster analysis (UPGMA method), principal component analysis (PCA), discriminant analysis (DA), and box plots [in yellow boxes presented, illustrate interquartile ranges (= the range between the 25th and 75th percentile) and medians (horizontal line); vertical lines are the whiskers that represent the scores outside the middle 50% (i.e., the lower 25% of scores and the upper 25% of scores)]. PCA analysis was performed both by excluding the qualitative characters and by including them as binary variables according to [44]. DA was performed using the boxes presented, illustrate interquartile ranges (=the range between the 25th and 75th percentile) and medians (horizontal line); vertical lines are the whiskers that represent the scores outside the middle 50% (i.e., the lower 25% of scores and the upper 25% of scores)]. PCA analysis was performed both by excluding the qualitative characters and by including them as binary variables according to [44]. DA was performed using the first six components derived from PCA, which explains about 69% of the total variability. The use of component scores (each other linearly independent by construction) allows obtaining an unbiased discriminant model both solving the indeterminacy due to the multicollinearity of the independent variables and receiving a more reliable prediction for the smaller number of involved variables [45][46][47]. We performed the DA on groups classified using both names of taxa (subgenera and species and below ranks) and major groups as a result of the molecular analyses by Waselkov et al. [15] (pag. 446, Figure 1A,B; hereafter reported as "molecular clades"). A k-means procedure (which is the most common unsupervised non-hierarchical clustering technique maximizing the between/within-cluster variance ratio (F-Ratio) for a given k number of clusters [48], was performed to identify the optimum number of groups without using "no prior knowledge".
Concerning qualitative variables (nominal), we also prepared a matrix including the percentage of each variable for each taxon.

Results and Discussion
As a whole, the results of the morphometric analyses revealed that seeds micromorphology provides some new interesting taxonomic data concerning some taxa; on the other hand, others seed features are not useful for other species.

Analyses of the Whole Dataset
Hierarchical clustering (UPGMA method) shows three main groups. The first one is very small (A), including only samples of Amaranthus crassipes, a second small group (B) comprising samples of A. viridis and A. muricatus, and a third large group (C), with all the other samples/taxa ( Figure 2). In group (C), several taxa (e.g., A. albus, A. hybridus, or A. hypochondriacus) are not grouped together, whereas some (A. centralis, A. graecizans subsp. sylvestris, A. induratus, A. retroflexus, and A. vulgatissimus) have a very low dissimilarity with each other and form separated subclades.
The three groups shown in Figure 2 do not correspond to the three subgenera of Amaranthus proposed by Mosyakin & Robertson [14] since clades (A) and (B) would be both parts of the subgen. Albersia, whereas in the large clade (C), all three subgenera (Acnida, Albersia, and Amaranthus) are represented and intermixed.
The hierarchical clustering (UPGMA method) performed on Waselkov's "molecular clades" [15] (note that all these Waselkov's clades are represented in our samples) showed that no well-defined and separated group can be identified ( Figure 3). Note that the dendrogram reported in Figure 4 refers to the molecular tree by Waselkov et al. [15] constructed based on nuclear genes. Hierarchical clustering performed on Waselkov's "molecular clades" referred to chloroplast regions (not showed) reveals the same results, i.e., the absence of well-distinct groups.
The PCA shows that the cumulative percentage of eigenvalues for the first six axes is 68.86%, with a higher contribution (more than 10%) given by the first four components (17.96%, 15.51%, 11.28%, and 10.33%, respectively). Examining the combined graphs among pairs of these six components shows three well-separated groups (Figure 4) along the first and second components for the muricatus/viridis-group and the sixth component for the crassipes-group. The highest contributions to axes were given by the following characteristics: seed coat ornamentation, the central cell's shape, and the flat border's length.  The three groups shown in Figure 2 do not correspond to the three subgenera of Amaranthus proposed by Mosyakin & Robertson [14] since clades (A) and (B) would be both parts of the subgen. Albersia, whereas in the large clade (C), all three subgenera (Acnida, Albersia, and Amaranthus) are represented and intermixed.
The hierarchical clustering (UPGMA method) performed on Waselkov's "molecular clades" [15] (note that all these Waselkov's clades are represented in our samples)  The DA shows different results depending on the use of the names of the taxa (species and below ranks or subgenera) or Waselkov's "molecular clades" [15]: (1) By classifying the samples using the taxa names [species and below ranks (25 groups); see Table 1], DA predicted two main groups ( Figure 5)  (2) When we classified the samples using the subgenera names (=three groups, i.e., Acnida, Albersia, and Amaranthus), DA do not predict any separate group, and the three groups completely overlapped each other ( Figure 6). The first two discriminant functions explain 100% of the total variation [eigenvalues: 86.0% (first function) and 14.0% (second function)]. The matrix of actual/predicted groups reveals that (a) all the actual samples included in the Acnida-group are predicted as included in the other two groups (80% in the Albersia-group, 20% in the Amaranthus-group) and (b) about the 25% of the actual observations of the Albersia-and Amaranthus-groups is predicted under, respectively, the Amaranthus-and Albersia-groups. As a whole, the value of correct classification is low (57.5%); (3) After running the DA procedure on samples classified using Waselkov's "molecular clades" [15] (pag. 446, Figure 1A, based on nuclear genes), two groups were predicted (Figure 7) based on the first two discriminant functions, which explain 88.5% of the total variation [eigenvalues: 58.8% (first function) and 29.7% (second function)]. These two groups correspond to (A) the Dioecious/Pumilus-group+ESA-group (=Eurasian/S-African/Australian group [15]) and (B) a larger group including the remaining "molecular clades" sensu Waselkov et al. [15]. Note that the ESA-group partially overlaps the residual-group. The matrix of actual/predicted groups displays high percentages along the diagonal (whose values reveal the matching of actual and predicted observations for each group) for the ESA-group (96%) and Furthermore, we performed the DA using the three groups generated from the PCA (muricatus/viridis-group, crassipes-group, residual-group). The result is that these three groups are statistically well supported, based on the two discriminant functions, which explain 100% of the total variation [eigenvalues: 97.4% (first function), and 2.6% (second function)] (Figure 8). The value of correct classification is high (94.3%).
K-means confirm the three clusters solution for the samples considered showing a high F-Ratio (197.13) of the first PCA component (which gives the higher contribution in PCA analysis, i.e., 17.96%) in the 3-clustered running procedure; in contrast, F-Ratios are 185.30, 132.16, and 99.9 in, respectively, 2-, 4-, and 5-clustered procedures. (Table 4).
Box plots, made on quantitative characters (see Table 2), show the following results: (1) Subgenus rank (sensu Mosyakin & Robertson [14]): no group can be distinguished using seed micromorphology ( Figure 9). (3) "Molecular clades" [sensu Waselkov et al. [15] (pag. 446, Figure 1A, based on nuclear genes)]: only the Dioecious/Pumilus-group can be distinguished based on micromorphology of seeds, i.e., by the length of the flat border of the seed and, partially, the length and width of the whole seed ( Figure 11). (4) Box plots (not shown) originated based on Waselkov's "molecular clades," which refer to chloroplast regions that do not reveal any separate group. The PCA shows that the cumulative percentage of eigenvalues for the first six axes is 68.86%, with a higher contribution (more than 10%) given by the first four components (  The DA shows different results depending on the use of the names of the taxa ( cies and below ranks or subgenera) or Waselkov's "molecular clades" [15]:

of 27
(1) By classifying the samples using the taxa names [species and below ranks groups); see Table 1], DA predicted two main groups ( Figure 5)         Furthermore, we performed the DA using the three groups generated from the PCA (muricatus/viridis-group, crassipes-group, residual-group). The result is that these three groups are statistically well supported, based on the two discriminant functions, which explain 100% of the total variation [eigenvalues: 97.4% (first function), and 2.6% (second function)] (Figure 8). The value of correct classification is high (94.3%).  Furthermore, we performed the DA using the three groups generated from the PCA (muricatus/viridis-group, crassipes-group, residual-group). The result is that these three groups are statistically well supported, based on the two discriminant functions, which explain 100% of the total variation [eigenvalues: 97.4% (first function), and 2.6% (second function)] (Figure 8). The value of correct classification is high (94.3%).        Concerning the qualitative characters (nominal variables; see Table 2), the synoptical matrix of taxa confirms the three main groups resulting from hierarchical clustering, PCA, and DA analyses. These groups can be distinguished based on seed coat ornamentation. The muricatus/viridis-group has a wrinkled coat, the crassipes-group shows a pebble-stoned coat, and the residual-group displays a puncticulate or colliculate coat (never wrinkled or pebble stoned). The shape of the central cells also allows for distinguishing the muricatus/viridis-group, which included taxa showing about 80% of seeds with an irregular shape. In contrast, the other two groups (crassipes and the residual group) have central cells with a regular shape (about 100%). Finally, no character is useful to distinguish groups using both subgenera and "molecular clades" classifications.

Analyses on Subdatasets
As stated in the previous paragraph, samples referred to some taxa, despite being included in the large so-called "residual-group" (see Figures 4 and 5), have a very low dissimilarity with each other in the hierarchical clustering procedure and form separated subclades. These taxa are A. centralis, A. graecizans subsp. sylvestris, A. induratus, A. retroflexus, and A. vulgatissimus. So, we try to understand if differential combinations of seed characters in comparison with the related taxa characterized them as follows: Amaranthus centralis: it is an endemic Australian species (Northern Territory, Queensland, South Australia, and Western Australia) macro-morphologically similar to A. induratus. According to Palmer [42], these two species differ from each other based on the shape of the leaves (linear to narrowly oblong or narrowly ovate in A. induratus vs. ovate or elliptic in A. centralis) and tepals (margins with a single or serrated tooth-like projection on each side vs. margins without tooth-like projections). Note that A. centralis and A. induratus are included in the same Waselkov's "molecular clade," i.e., the ESA-clade." (ESA = Eurasian/S-African/Australian group [15]). Moreover, our analyses reveal a micro-morphological similarity between these two taxa as follows:

>
In the hierarchical clustering procedure (UPGMA method), the centralis-group and induratus-group are part of the same subgroup of the large "residualgroup" (see Figure 3); > In K-means 10-clustered procedure (not shown), one of the clusters is composed of the samples of A. centralis and A. induratus; > In DA analysis performed on samples classified using taxa, actual/predicted observations match each other for these two species (100% percentages along the diagonal).  Table 1). Amaranthus graecizans subsp. sylvestris: it is a taxon native to central and southern Europe and north Africa, and it is (macro-) morphologically characterized by having leaves usually acute, flowers arranged in axillary glomerules, three tepals in pistillate flowers, and fruit as long as or longer than the perianth [3]. This morphological configuration is similar to the forms of the Asian A. tricolor L. without terminal synflorescences [3,7,49], which were originally published by Linnaeus as A. tristis L. [50], A. tricolor [50], and A. polygamus L. [51], but later synonymized with A. tricolor [52]. A. graecizans subsp. sylvestris is included in the ESA-clade by Waselkov et al. [15], where also A. tricolor occurs. A. graecizans subsp. sylvestris and A. tricolor can be distinguished using seed micromorphology by the seed length [(1.09-)1. 28 [41]. This species can be distinguished by its stem, usually tomentose-pubescent and erect, green synflorescence, and five tepals spathulate with obtuse to emarginate apex [1,3,7,53]. A. retroflexus is macromorphologically similar to A. wrightii S.Watson, which differs mainly by being glabrous or nearly so [1,7]. Other authors e.g., [53] highlighted the similarity between A. retroflexus and A. quitensis (Bolòs & Vigo [54] even proposed to treat A. quitensis as subspecies of A. retroflexus). All these three species (A. quitensis, A. retroflexus, and A. wrightii) belong to the Hybridus-clade sensu Waselkov et al. [15]. Our analyses show that A.  [7]. In Waselkov et al. [15], A. vulgatissimus belongs to the SA-clade, sister to A. muricatus. The latter species was part of the muricatus/viridis-group according to our analyses (see Figures 2,4,5 and 8), and it differs primarily by the seed coat ornamentation (wrinkled) and shape of the central cell (irregular), whereas A. vulgatissimus has colliculate seed coat and regular central cells. A further difference between these two species refers to the width of the seed [(0.84-)0.85-0.99(-1.07) vs. (1.11-)1.14-1.16 (-1.24)].
In addition to the above-discussed species, Amaranthus rajasekharii, A. spinosus, and A. tuberculatus can also be analyzed in detail by considering the results of DA using the names of the taxa. As highlighted in the previous paragraph, the actual/predicted observations for these three species match each other (100%). A. tuberculatus is discussed above (see paragraph "3.1. Analyses on the whole dataset"); some comments about the other two species and the related ones as follows: Amaranthus rajasekharii: a species recently described from India [55] morphologically related to A. dibius Mart. from which differ by the stem (reddish to purple vs. green in A. dubius), bracts (linear and up to 0.1 mm long vs. ovato-deltoid, 1.3-1.7 mm long), tepals shape (ovate to lanceolate vs. oblong-spatulate), gynoecium (whitish vs. green), and pollen grain [with 21-23 pores (vs. 27-30), 3-5 ektexinous bodies (vs. mostly 3), and margin of pores not depressed and without conspicuous ornamentation (vs. clearly depressed and with conspicuous ornamentation)]. A. dubius is included in the Hybridus-clade by Waselkov et al. [15], whereas A. rajasekharii did not appear in their work being published later. Our results reveal that these two species differ from each other by the ratio length/width of seeds [(0.98-)1.03-1.05(-1.11) in A. rajasekharii vs. (1.09-)1.11-1.14(-1.26) in A. dubius]; Amaranthus spinosus: a species native to tropical America (from Mexico to Argentina), which is easy to distinguish from the other ones by its spine-like structure (metamorphosed bracts of the first flower in the first cyme) [1,3,7,53]. Recently, a similar species (A. saradhiana Sindhu Arya, V.S.A.Kumar, W.K.Vishnu & Rajesh Kumar) was described from India. It has, along the stem, only two spines per node, whereas no spines occur in the synflorescence part (4 spines per node along the stem and spines in the synflorescence part in A. spinosus); further differences regard the stem and petiole color (purple vs. green in A. spinosus), the apex of tepals (acute vs. often spathulate), gynoecium (whitish to light green vs. dark green), and the number of pores in pollen grains (26-30 vs. 37-40) Table 1).  [15] (pag. 446, Figure 1A].

Seed Types and Taxonomic Key
The deep morphometric analyses on seeds micromorphology show that several seed characters are taxonomically useful to distinguish some taxa both by considering the whole dataset (25 taxa) and by comparing taxa related from morphological and/or molecular points of view. On the contrary, seed characters are not able to separate groups at a rank higher than species (subgenus), so confirming the molecular results by Waselkov et al. [15] (pag. 446, Figure 1A,B). Moreover, seed characters analysis does not allow to distinguish Waselkov's "molecular clades," except for the Dioecious/Pumilus-group (nuclear genes), which is distinguished by the length of the flat border of the seed and, partially, the length and width of the whole seed (see Figure 11). The morphological differences found (especially those statistically well supported) allow us to define some seed types. On the contrary, in other cases, we prefer to refrain from proposing defined types for the moment. A tentative taxonomic key of the studied taxa/seed types, based on micromorphological characters, is below presented. Concerning the seeds coat ornamentation dichotomous alternatives at the step no. 4 of the key (punticulate vs. colliculate), we introduced the term "mostly" since (1) type punticulate always occurs in three taxa (A. centralis, A. emarginatus subsp. pseudogracilis, and A. graecizans subsp. sylvestris), whereas A. albus, A. blitum, and A. sarahdiana have seeds mostly punticu-late (up to 80%); (2) type colliculate is typical (100% of the studied samples) of eight taxa (A. deflexus, A. palmeri, A. hybridus, A. hypochondriacus, A. induratus, A. quitensis, A. retroflexus, A. spinosus, and A. vulgatissimus), whereas A. blitoides, A. cacciatoi, A. caudatus, A. dubius, and A. rajasekarii have seed mostly punticulate (up to 80%); A. standleaynus, which shows 50% of seeds punticulate and the rest colliculate, was included under two parts of the key (steps nos. 5-9 and 10-19). Finally, note that A. deflexus, A. vulgatissimus, A. cacciatoi, A. spinosus, A. dubius, and A. stadleyanus (which we classified in the same seed deflexus-type) cannot be distinguished using seeds characters.

Conclusions
Seed micromorphology was shown to be a highly informative taxonomic criterion that helps to solve ambiguities in plant taxonomy. Concerning the genus Amaranthus, few articles have been published until now, and they refer to one to a few species. Our paper