Quantifying morphological variation in the Castilleja pilosa species complex (Orobanchaceae)

Introduction

Because they provide the basis for the recognition of one of the primary units of biodiversity, the species, classifications are the cornerstone of the biodiversity sciences. As such, classifications are vital to our understanding of biodiversity and the process of speciation. Therefore, the careful and robust delimitation of species is imperative. Species delimitation relies on our ability to define boundaries between one population and the next. Historically, this has been done using morphological evidence (Sneath & Sokal, 1973), ecological evidence (Van Valen, 1976), and more recently, in light of technical and analytical advances, molecular evidence (Baum & Shaw, 1995). Each criterion has limitations for being widely applied across the tree of life (de Queiroz, 2007), and no one criterion has been universally applied to defining species boundaries (de Queiroz, 2009). Instead, there has been a movement to include multiple lines of evidence in the delimitation of species (Padial et al., 2010; Schlick-Steiner et al., 2010; Carstens et al., 2013; Dejaco et al., 2016; Freudenstein et al., 2016). In addition to describing newly discovered species, species delimitation methods are often applied to existing classifications where species boundaries are poorly defined and/or sample assignment to species is difficult (e.g., Barley et al., 2013; Giarla, Voss & Jansa, 2014). In these cases, species delimitation is used in a validation context (where taxonomic boundaries are validated—i.e., individuals are assigned to a group a priori (Ence & Carstens, 2010)) and attempts to clarify species boundaries and which lines of evidence (morphological, ecological, molecular) do and do not describe species.

Traditionally, species have been described and established, in part, by designating a type specimen. From the type collection and other specimens examined during species discovery, morphological, ecological, and distributional traits are used to create a species description, providing a central reference point or general conceptualization of the new species, around which some amount of variation occurs (Fig. 1). However, the characteristics of this variation (the amount, the direction, etc.) are not static, and additional collections assigned to a species can shift the conceptual boundaries of the species, in particular how this is applied on-the-ground. For example, regional and floristic studies can result in treatments and species descriptions that incorporate variation observed in the field on a local scale. Revisionary work, typically happening at a broader scale (e.g., the Flora of North America), often recognizes overlapping variation between similar species and synonymizes names where appropriate. As a result, there can be a shift of species boundaries and known variation. In essence, these shifts can inflate or deflate the taxonomic concept of an entity, sometimes greatly outside the realm of its original description. Taxonomic drift such as this can result in a species description that no longer represents the range of described variation for a species, and instead only represents a portion of that variation (Fig. 1).

Often, the species involved in these taxonomic fluctuations are characterized as species complexes (i.e., groups of species that are difficult to distinguish from one another), and are already known to have overlapping variation that is difficult to classify. The shifting of recognized and ascribed variation through time and across treatments can further increase the fuzziness of species boundaries, making the identification of unknown individuals (and therefore the usefulness of the classification) even more difficult. This is further complicated when an unknown comes from a geographic boundary (or, conversely, one that is widespread but has varieties that occur in geographically restricted areas), served by two or more regional or localized treatments that have varying interpretations of variation within a taxon. This requires choices to be made by the identifier in preferring one treatment to another when treatments are in conflict (e.g., one treatment recognizes varieties, while another does not). Cases such as these—species complexes with a great deal of taxonomic confusion—are good targets for robust species delimitation. By clarifying and determining which lines of evidence distinguish species, classifications can be updated to reflect more accurate estimates of species boundaries.

Recognizing species in the genus Castilleja Mutis ex L.f. (Orobanchaceae Vent.) is notoriously difficult, particularly in the field. These difficulties largely stem from a nearly continuous range of variation both within and across taxonomic boundaries (Cronquist et al., 1984). The source of this morphological continuity is likely a combination of the young age of the lineage, the widespread and highly variable instances of polyploidy, and interspecific gene flow when species co-occur (Heckard & Chuang, 1977; Tank & Olmstead, 2008; Tank, Egger & Olmstead, 2009). This means that most often the characters that diagnose species are slight and often overlapping.

A good example of these difficulties can be found in the Pilosae alliance. Composed of approximately eight taxa and several varieties (Castilleja arachnoidea Greenm., C. cinerea A. Gray, C. nana Eastw., C. pilosa (S. Watson) Rydb. (with three named varieties: var. pilosa, var. longispica (S. Watson) N.H. Holmgren, and var. steenensis (Pennell) N.H. Holmgren), C. praeterita Heckard & Bacig., C. rubida Piper, C. salsuginosa N.H. Holmgren, and C. schizotricha Greenm.), members of this group have tubular flowers with less-showy corollas, short beaks, and a pouchy, lower corolla lip that has somewhat petaloid teeth (Fig. 2B) (Cronquist et al., 1984; Hitchcock et al., 1984; Wetherwax, Chuang & Heckard, 2012). This morphological alliance is unique among most Castilleja alliances in that it is composed of perennial diploids, with no documented polyploidy (Heckard & Chuang, 1977).

Within the alliance, two widespread species are especially difficult to distinguish, particularly in the field—C. pilosa and C. nana. These taxa compose the core members of the pilosa species complex. C. pilosa is composed of three taxonomically recognized varieties, distinguished primarily by geography, in addition to slight variations in a suite of morphological characters. Castilleja pilosa var. pilosa is found in the Sierra Nevada, north and east into Oregon; Castilleja pilosa (S. Watson) Rydb. var. steenensis (Pennell) N.H. Holmgren is endemic to the high ridges of Steens Mountain in southeastern Oregon; Castilleja pilosa (S. Watson) Rydb. var. longispica (A. Nelson) N.H. Holmgren occurs in the southern half of Idaho, east into western Wyoming and Montana, and has a disjunct population in northern Idaho. These varieties are distinguished by calyx length, herbage pubescence, elevation, and geographic position (Fig. 2A) (Cronquist et al., 1984). Castilleja nana occurs throughout the central and southern Sierra Nevada range of eastern California and extends eastward on the high ridges of Nevada’s basin and range topography. Castilleja nana is primarily distinguished from C. pilosa by elevation, occurring between 2,400 and 4,200 meters, while C. pilosa is found primarily at lower elevations between 1,200 and 3,400 meters. Additionally, C. nana is often a smaller plant with decumbent branches and smaller features (Fig. 2A).

When C. pilosa and C. nana occur in sympatry and at the same elevation, it is often quite difficult to distinguish the two species. Additionally, many of the members of this complex occur across geographic and political boundaries and are represented in multiple, overlapping regional and floristic treatments (Cronquist et al., 1984; Hitchcock et al., 1984; Wetherwax, Chuang & Heckard, 2012). Subsequently, there has been a great deal of taxonomic confusion, demonstrated by the number of synonyms associated with C. pilosa and C. nana. Several of these incorporations are based on collections made in the Sierra Nevada where both species occur in sympatry, as well as northern California at the border with Oregon and Nevada. These regions also lie at the boundary between the Great Basin, the Pacific Northwest, and the Sierra Nevada and California floristic province, where a great deal of taxonomic work has been done. The taxonomic confusion in this group could be the result of any of the following factors: the young age of the lineage, the propensity for gene flow when species are sympatric, little to no morphological distinction between species, and/or the absence of species in the complex (i.e., the entire complex is actually a single lineage). As such, this complex is in great need for robust species delimitation.

Here we begin this process by quantifying morphological variation in the complex and assessing its correlation (or not) with the current taxonomy. By sampling many populations across the known ranges of these entities, identifying them using regional treatments, and measuring and analyzing a suite of morphological traits, we test the assumption that there are morphological clusters that correspond to taxonomic entities. We perform principal coordinate analyses to understand the position of individuals in morphospace, and then apply a non-hierarchical clustering method to assess the signal of morphological similarity that exists among these entities. In this way, we aim to quantify and begin to characterize the morphological variation in this species complex, information that will ultimately become part of a robust delimitation of species boundaries in this group.

Materials and Methods

Results

Quantifying morphological variation

Violin dot plots of individual quantitative trait values, grouped by taxonomic identity revealed a great deal of overlap in raw trait values for each taxon across many traits. In some cases, this overlap occurs across all focal taxa, as in bract width and leaf width (Fig. 3), where all taxa have widely overlapping trait values. In other cases, the distribution of trait values distinguishes one of the focal taxa from the remaining three. For example, C. pilosa var. steenensis has a larger beak to tube ratio than the remaining taxa (meaning that the difference in length between the tube and the beak is greater), C. nana has a longer bract than all varieties of C. pilosa and most C. pilosa var. longispica have shorter calyces than other varieties of C. pilosa and C. nana. There are also cases of interspecific overlapping trait distributions, as in plant height where C. nana and C. pilosa var. steenensis are generally shorter in height than C. pilosa var. longispica and C. pilosa var pilosa. We see a similar pattern of overlap in traits across taxa in our qualitative data. With the exception of the decumbent habit, no one qualitative trait is found primarily in one taxon, let alone exclusively (Fig. 4). In general, pubescence traits were equally variable across taxa, C. nana was the only taxon that occasionally lacked lobes on the leaves, and C. pilosa var. pilosa and C. pilosa var. steenensis were the only focal taxa that were never scored as having broader, deltoid shaped calyx lobes. Summary statistics for raw values of continuous traits and raw counts of categorical traits can be found in the supplement (Tables S2 and S3, respectively).

Principal coordinate analysis

A Cailliez correction, equal to D′ = −0.5 * (D + 0.57237)², was applied to all negative eigenvalues. The position of each individual in the first two and three principal coordinates are shown in Fig. 5, with 95% confidence ellipses around the mean position of each focal taxon in morphospace. An examination of axes 1 through 106 does not change interpretation of the results presented here; the first two principal coordinate axes represent the maximum morphological distance among individuals sampled, and the third axis reveals no further distinction (Fig. 5).

In general, and considering all three principal coordinate axes, individuals identified as C. nana (yellow) occupy a different part of the scatterplot than those identified as C. pilosa, including its named varieties (blue (var. pilosa), orange (var. longispica), and red (var. steenensis)). Considering only those individuals identified as C. pilosa, there is a large amount of overlap with no discernible position in morphospace unique to any variety (Fig. 5). Confidence ellipses lend support to this conclusion and further suggests a greater distinction of C. pilosa var. steenensis (in red) from any other focal taxon. The variation in distances of these individuals lies along a different axis than the rest of the focal taxa; however, the effect of sample size (n = 4) cannot be discounted.

Fuzzy clustering

We performed seven fuzzy-clustering analyses (corresponding to seven different values of the membership exponent variable; values between 1.1 and 1.7, in increments of 0.1) for each of three possible numbers of clusters (k = 4, 3, and 2). Different values of the membership exponent produced consistent results within each “k = X number” of clusters. For simplicity, we present the results from all clustering scenarios with a membership exponent of 1.3.

Fuzzy-clustering analyses, regardless of the number of clusters considered, resulted in clusters with small silhouette coefficients (both within and across clusters), and low values for the normalized Dunn coefficient (Fig. 6; Table 2). As cluster number was reduced, there appeared to be some small improvement in these measures (average silhouette coefficient increased from 0.2 (k = 4) to 0.22 (k = 3), and to 0.25 (k = 2) and normalized Dunn coefficient increased from 0.37 (k = 4) to 0.38 (k = 3), and 0.44 (k = 2)); however, overall these values are extremely low. Generally speaking, regions of overlap in morphospace coincide with lower probabilities of membership of each individual to each cluster (i.e., the probability of membership to each cluster was higher, as opposed to having an overwhelming probability of membership to any one single cluster), consistent across all clustering scenarios (Fig. S2).

Table 2:

Results of fuzzy clustering analyses.

	k = 4			k = 3			k = 2
	n	Avg s(i)	stdev s(i)	n	Avg s(i)	stdev s(i)	n	Avg s(i)	stdev s(i)
Cluster 1	26	0.24	0.11	32	0.2	0.11	57	0.23	0.13
Cluster 2	24	0.28	0.11	47	0.21	0.08	51	0.27	0.1
Cluster 3	31	0.21	0.11	29	0.24	0.11
Cluster 4	27	0.1	0.09
Avg s(i) across analysis		0.2			0.22			0.25
Normalized Dunn Coefficient		0.3768			0.3879			0.4424

DOI: 10.7717/peerj.7090/table-2

Note:

Results of fuzzy clustering analyses with k = 4, 3, and 2 clusters. Here we report average silhouette coefficients within and across clusters in analyses, as well as normalized Dunn coefficients for each analysis. Silhouette coefficients close to 0 represent less similarity, those close to 1 represent high similarity, and negative silhouette coefficients indicate likely misassignment to a cluster. The normalized Dunn coefficient is a measure of the overall fuzziness of an analysis. Values close to 0 indicate high levels of fuzziness (near equal membership of individuals to all clusters) and values close to 1 indicate very low levels of fuzziness (i.e., hard partitions).

A somewhat subjective approach to quantifying the structure in a dataset is to calculate the silhouette coefficient (SC) of the dataset (Kaufman & Rousseeuw, 2005). This value is the maximum, average silhouette coefficient of all possible numbers of clusters, from k = 2 as a minimum, to k = n as a maximum (n = 108, in this study). At k = 53, our standard 100,000 iterations of clustering were not enough to satisfy fuzzy-clustering objectives, and we ran into convergence issues. However, considering k = 2 through k = 53 clusters, the average silhouette coefficients were highest at k = 2 (average s(i) = 0.25), and steadily dropped as values of k increased.

To visualize the taxonomic composition of clusters, we painted the silhouettes with colors corresponding to the taxonomic identity of each individual. Across all three clustering schemes, one cluster is consistently composed of mostly C. nana individuals, with the remaining clusters being variously composed of all three varieties of C. pilosa. When we restrict the cluster number to two, the C. nana cluster begins to be more heavily composed of C. pilosa individuals (Fig. 6).

Discussion

Classifications are useful when they organize objects based on relationships, when they reflect similarities and differences among the constituent parts, and when they aid in the identification and placement of unknowns within the classification (Sokal, 1974; de Queiroz & Donoghue, 2011; de Queiroz & Donoghue, 2013). The species description, based in part on the type specimen, plays an important role in the creation and implementation of classifications, but with a reliance on it comes the challenge of tracing and managing type collections and species descriptions through time—a problem that we are still dealing with (Hitchcock, 1905; Dayrat, 2005). In addition, when objects are discrete and discontinuous, classifications are easy to build and use; however, when there is continuous variation in characters used in the classification, this becomes more difficult.

In this study we have closely examined morphology—a commonly used character for describing taxonomic boundaries—for four named taxa, from across their ranges, in a species complex known to be taxonomically difficult to diagnose. Here we have quantified a great deal of overlap in raw character traits that are typically used to diagnose species (when using taxonomic keys) in Castilleja (Figs. 3 and 4). In some cases, these traits are continuous across taxonomic boundaries (Fig. 3), emphasizing the extreme morphological similarity among these named entities. We continue to see little distinction among current taxonomic groups when we examine the morphospace described by all morphological characters that we measure here (Fig. 5). For example, C. pilosa, where we are essentially incapable of distinguishing taxonomic varieties using morphology alone (Fig. 5), even in C. pilosa var. steenensis, considered the most distinctive of the three varieties due to its isolation on Steens Mountain in SE Oregon (Hitchcock et al., 1984). Finally, when we interrogate morphospace for evidence of structure, we find little support (low silhouette widths for each cluster and low average silhouette widths for each clustering scenario, Fig. 6) and equal assignment probabilities of individuals that occur in areas of overlap to each cluster (Fig. S2).

And yet, despite the overall high levels of similarity we observe some consistent distinction between individuals of C. nana and C. pilosa, indicating some morphological distinction between taxa (Fig. 5). This is also supported by the results of fuzzy-clustering analyses that, regardless of the number of clusters considered, recover a cluster composed primarily of C. nana, with C. pilosa individuals variously scattered among the remaining clusters (Fig. 6). Several continuous traits distinguish C. nana from C. pilosa (Fig. 3; see also Fig. S3), however, the overlapping tails of these distributions, and the nature of these distinguishing traits (i.e., size and length traits that could be environmentally plastic), goes a long way towards explaining the morphological confusion that has plagued this complex historically.

It is clear that geographic and ecological characters must have played a dominant role in shaping the species descriptions in this complex. This is apparent from the species descriptions included both in regional and genus-wide treatments (Cronquist et al., 1984; Hitchcock et al., 1984; Wetherwax, Chuang & Heckard, 2012), as well as the inferred species ranges (Fig. 2). For example, C. nana does not occur in the northern limits of the C. pilosa range. So, if you encounter a relatively small individual in Idaho, there is no way to confuse it with C. nana (a California and Nevada species), as the ranges do not overlap, and the regional treatment does not consider C. nana (Hitchcock et al., 1984). Similarly, C. pilosa var. steenensis only occurs on Steens Mountain in Eastern Oregon. If you found a relatively small individual in central Oregon, you could only classify it as C. pilosa var. pilosa, using these regional treatments.

When species occur sympatrically, however, the distinction between named entities becomes much more difficult to parse. In the Sierra Nevada, C. pilosa var. pilosa (a moderate elevation taxon) and C. nana (a high elevation taxon) can co-occur at the limits of their elevational ranges (high and low, respectively) where environments are heterogeneous. Similarly, C. pilosa var. pilosa and C. pilosa var. steenensis can co-occur on the western slopes of Steens Mountain in the transition area between the high, exposed ridge and the surrounding lower elevation steppe. In heterogeneous habitats and at ecological boundaries, phenotypes can be accentuated and variable (van Kleunen & Fischer, 2005), potentially in response to local microhabitat conditions such as light availability and precipitation (Schlichting, 1986; Dorn, Pyle & Schmitt, 2000; van Kleunen, Fischer & Schmid, 2000; Nicotra et al., 2010). As a result, it is possible that in these areas of sympatry that correspond with environmental transitions, individuals could experience extreme conditions that may affect the morphological traits that we examine when we try to identify unknowns. We see this in several individuals from the Sierra Nevada that have extreme values in the traits that distinguish C. nana and C. pilosa (Fig. 7). Furthermore, these are the individuals that occur in the region of overlap in morphospace between these two taxa (Fig. 7).

In some cases, these regions of sympatry also correspond with hotspots of taxonomic synonymy historically—i.e., these sympatric areas are places where synonyms of currently accepted taxa were described (Fig. 8). For example, the area surrounding Lake Tahoe has seen the description of four distinct taxa (C. jusselii Eastw. (Eastwood, 1940), Orthocarpus pilosus S. Wats. (Watson, 1871), C. inconspicua A.Nelson & P.B.Kenn (Nelson & Kennedy, 1906), C. nana Eastw. (Eastwood, 1902a)), two of which (O. pilosus and C. nana) are the type specimens for C. nana and C. pilosa (Fig. 8). The remaining two taxa were later incorporated into C. nana (C. inconspicua) and C. pilosa (C. jusselii), effectively meaning that these entities are no different from C. nana and C. pilosa. However, when we place our best approximation of C. inconspicua in morphospace (i.e., a specimen of the same taxon (C. inconspicua is a synonym of C. nana) that was measured by us that is as geographically close to the type collection of C. inconspicua as possible), we find that this collection occupies a region of morphospace very different from that of the type collection of C. nana (Fig. 8). By including this species into the concept of C. nana through synonymization in the Intermountain Flora (Cronquist et al., 1984), the amount of variation attributed to C. nana likely expanded.

Areas of sympatry are not the only source of potential confusion in the taxonomic history of either taxon. For example, the synonymization of C. lapidicola A.Heller (Heller, 1912) in eastern Nevada with C. nana also expanded the region of morphospace attributed to C. nana ((Cronquist et al., 1984), Fig. 8). Similarly, in northern California the inclusion of C. ochracea Eastw. (Eastwood, 1941) and C. pisttacinus (Orthocarpus psittacinus Eastw. (Eastwood, 1902b); C. psittacina (Eastw.) Pennell (Abrams, 1951)) increased the area of morphospace occupied by C. pilosa ((Cronquist et al., 1984); Fig. 8). Ultimately, the qualitative decisions made about species boundaries based on regional treatments have extended and inflated the morphological concepts of both taxa. By going through this procedure of quantifying morphological variation, we can visualize what morphological variation the taxonomy currently embodies. It is apparent that the morphological concept of both C. nana and C. pilosa have expanded through the incorporation of additional taxa as synonyms, and it is possible that the species description of both taxa may no longer represent the features of either taxon.

Conclusion

The inflation of morphological variation attributed to C. nana and C. pilosa during species level revisions, much of them regionally based, in addition to an apparent reliance on potentially plastic morphological characters to distinguish species in sympatry, has resulted in a great deal of morphological confusion in this complex. This likely contributes to the tumultuous taxonomic history of these taxa, and suggests that relying on morphology alone to define species boundaries in this complex is problematic. This is where molecular and ecological lines of evidence will be incredibly important to delimit species. In a robust and integrated delimitation of species, we may find that taxa that have been synonymized are not truly part of their corresponding taxa, or vice versa. Subsequent classifications should reflect these boundaries and highlight the similarities and differences between them.

Here we have begun that process by quantifying morphological variation in this species complex and we have estimated the position of type specimens in that space. The next steps in this group will be to gather molecular and ecological evidence to contribute to a robust species delimitation that is based on multiple lines of evidence. With all data in hand, we can more confidently apply names, whether that is applying an old name, a new name, or combining them all in one.

Supplemental Information