Continuous Characters in Phylogenetic Analyses: Patterns of Corolla Tube Length Evolution in Lithospermum L

bs_bs_banner

Biological Journal of the Linnean Society, 2012, ••, ••–••. With 3 ﬁgures

Continuous characters in phylogenetic analyses: patterns of corolla tube length evolution in Lithospermum L. (Boraginaceae)

JAMES I. COHEN*

Texas A&M International University, 5201 University Boulevard, 379D LBVSC, Laredo, TX 78041, USA

Received 20 February 2012; revised 15 April 2012; accepted for publication 15 April 2012

The present study comprises an analysis of six different scoring schemes and eight different types of analytic methods aiming to investigate the evolution of a continuous character (i.e. corolla tube length) in Lithospermum L. (Boraginaceae). Corolla tube length in the genus is quite variable, ranging from 1 mm to 75 mm, and the length of the corolla tube has implications for pollination biology, such as longer corolla tubes (> 25 mm in length) being pollinated by hummingbirds or moths. In general, the various methods resolve similar ancestral character states; however, different states are reconstructed at nodes in which the descendants greatly differ in corolla tube length. Additionally, it is suggested that all of the variation of a continuous character should be included in analyses, and this may necessitate multiple analyses with different partitions of the data. The various analyses provide evidence that two maximum parsimony methods, linear parsimony and the TNT method, minimize the number of different rates of evolution. In Lithospermum, six origins of corolla tubes > 20 mm in length are resolved, and these origins occurred at two different times periods: (1) in the shadow of hummingbird diversi- ﬁcation in North America (approximately 6–8 Mya) and (2) more recently (approximately 1–1.5 MyA). Four substantial decreases in corolla tube length also are reconstructed, and these may be associated with the origin of self-pollination. © 2012 The Linnean Society of London, Biological Journal of the Linnean Society, 2012, ••, ••–••.

ADDITIONAL KEYWORDS: ancestral character state reconstruction – evolution – molecular dating – pollination.

INTRODUCTION and states often overlap, making effective state assig- nation problematic (Felsenstein, 1988; Stevens, 1991, Continuous or quantitative characters, such as corolla 2000; Gift & Stevens, 1997; Swiderski et al., 1998; length or leaf width, are commonly included in phy- Garcia-Cruz & Sosa, 2006). Presently, numerous logenetic analyses, particularly at lower taxonomic methods exist that allow for the investigation of evo- levels (e.g. among species or genera; Stevens, 1991). lutionary patterns of continuous characters without Utilizing these types of characters has often involved the need to create discrete states. These currently forcing quantitative variation into discrete (noncon- available phylogenetic methods differ as to whether or tinuous or binned) character states (Mickevich & not branch lengths are used, if and how estimation Johnson, 1976; Archie, 1985; Farris, 1990; Stevens, accuracy is calculated, which evolutionary model is 1991, 2000; Gift & Stevens, 1997; Rae, 1998; Swider- employed, and whether the data are analyzed in a ski, Zelditch & Fink, 1998; Garcia-Cruz & Sosa, 2006; maximum parsimony (MP), maximum likelihood Goloboff, Mattoni & Quinteros, 2006). Determining (ML) or Bayesian inference (BI) framework (Table 1). the appropriate manner in which to convert continu- Because of the particular parameters included in the ous characters into discrete states can be difficult, methodologies, each may reconstruct, for the same character, different ancestral values at the same node (Butler & Losos, 1997; Martins, 1999; Webster & *E-mail: [email protected] Purvis, 2002a, b).

MP methods attempt to minimize change, either locally (among adjacent branches) or globally (across the entire tree), throughout the tree (Table 1), and they do so without invoking an explicit model of

BAYESTRAITS model B evolution (Webster & Purvis, 2002a). By contrast, ML cies; MAX/MIN, and BI methods involve global reconstructions and utilize a speciﬁc model of evolution [i.e. Brownian motion (Felsenstein, 1985) or some form of selection (Martins, 2004)] to resolve ancestral character states (Webster & Purvis, 2002a; Pagel & Meade, 2011). Additionally, some MP methods as well as all ML and

BAYESTRAITS model A BI methods utilize branch lengths during the inference of ancestral states (Table 1). The objectives of the present study are two-fold: (1) to compare and evalu- ate different methods for ancestral character state reconstruction of continuous characters as well as different schemes for scoring continuous character Compare exponential model 20 mm in length; RANGE, the range of values for each

> data, and (2) to investigate the patterns of evolution of a particular continuous character (i.e. corolla tube length) as it relates to pollination biology in Lithos- permum L. (Boraginaceae).

Compare linear model Butler & Losos (1997), Martins (1999), and Webster & Purvis (2002a, b) have undertaken comparative studies on the utility of various methods employed to examine the evolution of continuous characters, and each reached a different conclusion. Butler & Losos

Weighted squared- change parsimony (1997) used three different MP methods (linear parsimony, squared-change parsimony, and weighted squared-change parsimony) and found that none accurately reﬂected the evolution of body size in Anolis lizards. Martins (1999) and Webster & Purvis

Squared- change parsimony (2002a, b) analyzed data in both MP and ML frameworks. Martins (1999) favoured a ML approach because ML methods, compared to MP methods, provide greater ﬂexibility in the inputs for evolutionary models. In addition, ML methods can provide SDs

Linear parsimony or SEs for reconstructed ancestral values, which MP methods do not (although some may provide a range of values; Butler & Losos, 1997; Goloboff et al., 2006) (Table 1). In more recent studies, Webster & Purvis

TNT method (2002a, b) found conﬂicting results. The ﬁrst (Webster & Purvis, 2002a) suggests that, according to residual sum of squares calculations, MP methods reconstruct the most predictive ancestral states. However, the second (Webster & Purvis, 2002b) concluded that, although a two-parameter ML model performed best, no model consistently provided good estimates of ancestral values. Given these varying results, each method for estimating ancestral character states appears to have its strengths and weaknesses, with no method consistently found to be best at reconstructing evolutionary history.

Analytical methods, scoring schemes, and type of phylogenetic tree used in presented analyses, and properties of various analytical methods In the aforementioned studies, the authors only used single point values at the terminals (although, for the use of two values, see Webster & Purvis, Error estimates includedType of reconstructionExplicit model of evolutionBranch lengths used NA NA Local NA NA Local NA NA NA Global NA Global NA NA NA Global X X Global X X X Global X X X Global X X X X X MINMEANMAXMAX/MINRANGE±1/2 RANGEType of phylogeny usedProperties for of reconstruction analytical methods MP MP X X X X NA MP X X X X NA X MP NA X X NA X X BI NA X X NA X BI NA X X X X NA X BI X X X X NA X X BI NA NA X X X X NA NA X X X X maximum value for species with corolla tubes less than 20 mm in length, and minimum value for species with corolla tubes Table 1. Scoring Scheme MP, maximum parsimony; BI, Bayesian inference. MIN, minimum value for each species; MEAN, mean value for each species; MAX, maximum value for each spe species; ± 1/2 RANGE, range/2) ± half of2002b). the range; NA, not applicable. However, the use of only a single point does

B. C. D.

E. F. G.

Figure 1. Flowers of species of Lithospermum.A,Lithospermum leonotis, corolla tubes 49–75 mm long. B, Lithospermum macromeria, corolla tubes 47–70 mm long. C, Lithospermum exsertum, corolla tubes 47–60 mm long. D, Lithospermum matamorense, corolla tubes 1–2 mm long. E, Lithospermum nelsonii, corolla tubes 11–14 mm long. F, Lithosper- mum multiflorum, corolla tubes 6–13 mm long. G, Lithospermum trinervium, corolla tubes 11–16 mm long. Lithospermum leonotis (A) and L. nelsonii (E) are sisters, and so are L. macromeria (B) and L. multiflorum (F). not capture the variation of a quantitative character 1985; Kittelson & Handler, 2006). Therefore, a change (Stevens, 1991; Gift & Stevens, 1997; Swiderski in the length of the corolla may be associated with a et al., 1998). To include all variation, analyses utiliz- shift in the type of pollinator that visits the flower. By ing either a range of values or multiple single points placing, in a phylogenetic framework, corolla tube should be undertaken (Swiderski et al., 1998; Goloboff length, it is possible to hypothesize patterns of evo- et al., 2006). In the present study, six different scoring lution and rates of change for corolla tube length, as schemes and eight different types of analytic methods well as identify locations on the phylogenetic tree (as implemented in four software packages) are where pollinator shifts may have occurred. utilized. These different coding schemes are needed because most of the available software for ancestral character state reconstruction only accepts single MATERIAL AND METHODS point values, and not a range; therefore, multiple analyses of single points are necessary to encompass TAXON SAMPLING the entire range of values for a character. Forty-two species were included in the analyses In the present study, corolla tube length is investi- (Table 2). Thirty-seven belong to Lithospermum, and gated in Lithospermum, a cosmopolitan genus that this sampling represents both the morphological and comprises approximately 60 species, with a centre geographic range of variation within Lithospermum of diversity in Mexico and the south-western USA s.l. (Cohen & Davis, 2009). The outgroup comprises (Johnston, 1952, 1954; Cohen & Davis, 2009). The three species of Buglossoides Moench – Buglossoides genus exhibits diverse floral morphology, including arvensis (L.) I. M. Johnst., Buglossoides purpureo- variation in the extent of herkogamy, as well as caerulea (L.) I. M. Johnst., and Buglossoides tenu- corolla shape, colour, symmetry, and tube length iflora (L.f.) I. M. Johnst. – and two species of (Johnston, 1952, 1954; Cohen & Davis, 2009; Weigend Glandora D. C. Thomas, Weigend, & Hilger – Glan- et al., 2009; Cohen, 2011) (Fig. 1). The genus includes dora diffusa (Lag.) D. C. Thomas and Glandora corolla tubes that are as short as 1 mm and as long as oleifolia (Lapeyr.) D. C. Thomas – two genera that 75 mm (Fig. 1, Table 2), and corolla tubes of dissimi- are close relatives of Lithospermum s.l. (Thomas, lar lengths often are associated with different polli- Weigend & Hilger, 2008; Ferrero et al., 2009; Cohen & nators (Boyd, 2002; Cohen, 2011). For example, Davis, 2009; Weigend et al., 2009); Cohen & Davis, flowers with long corolla tubes (> 25 mm in length) 2012). The majority of the species were collected from tend to be pollinated by hummingbirds or moths wild populations. For these taxa, herbarium speci- (Grant & Grant, 1970; Boyd, 2002; Cohen, 2011), mens were collected and deposited at the L. H. Bailey whereas flowers with shorter corolla tubes (< 20 mm Hortorium Herbarium at Cornell University, and in length) are visited by bees and butterflies (Weller, leaf tissue was dried and preserved in silica gel for

Table 2. Species included in study, range of corolla tube lengths of each species, and origin of measurements

Range of corolla tube Number of Origin of Species length (mm) specimens specimens

Outgroup Buglossoides arvensis (L.) I. M. Johnst. 4–9 Literature Buglossoides purpureo-caerulea (L.) I. M. Johnst. 15–19 Literature Buglossoides tenuiflora (L.f.) I. M. Johnst. 4–9 Literature Glandora diffusa (Lag.) D. C. Thomas 16–17 Literature Glandora oleifolia (Lapeyr.) D. C. Thomas 17–18 Literature Ingroup Lithospermum calcicola B. L. Rob 4–5 19 Herbarium Lithospermum californicum A. Gray 10–15 10 Herbarium, Literature Lithospermum calycosum (J. F. Macbr.) I. M. Johnst. 6–13 11 Herbarium Lithospermum canescens Lehm. 7–8 Literature Lithospermum caroliniense MacMill. 7–12 Literature Lithospermum cinereum DC. 2 Literature Lithospermum cobrense Greene 5–10 13 Herbarium, Fresh Lithospermum discolor M. Martens & Galeotti 7–13 24 Herbarium Lithospermum distichum Ortega 3–5 25 Herbarium, Fresh Lithospermum erythrorhizon Siebold & Zucc. 3–4 Literature Lithospermum exsertum (D. Don) J. I. Cohen 47–60 17 Herbarium Lithospermum flavum Sessé & Moc. 28–45 21 Herbarium Lithospermum gayanum I. M. Johnst. 3 Literature Lithospermum helleri (Small) J. I. Cohen 6–9 Literature Lithospermum johnstonii J. I. Cohen 35–53 17 Herbarium Lithospermum latifolium Michx. 3–4 Literature Lithospermum leonotis (I. M. Johnst.) J. I. Cohen 49–75 10 Herbarium Lithospermum macromeria J. I. Cohen 47–70 13 Herbarium, Fresh Lithospermum matamorense DC. 1–2 8 Herbarium Lithospermum mirabile Small 10–15 11 Herbarium Lithospermum mirabile ¥ incisum 10–15 11 Herbarium Lithospermum molle (Michx.) Muhl. 7–9 Literature Lithospermum multiflorum Torr.ex A. Gray 6–13 22 Herbarium, Fresh Lithospermum nelsonii Greenm, 11–14 9 Fresh Lithospermum notatum (I. M. Johnst.) J. I. Cohen 35–48 5 Herbarium Lithospermum oblongifolium Greenm. 20–37 20 Herbarium Lithospermum obovatum J. F. Macbr. 7–13 9 Herbarium Lithospermum officinale L. 3–4 Literature Lithospermum revolutum B. L. Rob. 7–9 8 Herbarium Lithospermum rosei (I. M. Johnst.) J. I. Cohen 12–16 5 Herbarium Lithospermum ruderale Douglas ex Lehm. 9–12 Literature Lithospermum scabrum Thunb. 5–6 Literature Lithospermum strictum Lehm. 7–12 23 Herbarium Lithospermum trinervium (Lehm.) J. I. Cohen 11–16 23 Herbarium, Fresh Lithospermum tuberosum Rugel ex DC. 5–6 Literature Lithospermum tubuliflorum Greene 8–13 8 Herbarium Lithospermum viride Greene 23–32 14 Herbarium

© 2012 The Linnean Society of London, Biological Journal of the Linnean Society, 2012, ••, ••–•• EVOLUTION OF COROLLA TUBE LENGTH 5 subsequent DNA isolation. Species not collected from than approximately 10% from those reported in the natural populations were obtained from plants at literature (Johnston, 1952; Zhu et al., 1995). There- the Cornell Plantations or as DNA isolations from fore, the different origins of the measurements did not the DNA bank of Royal Botanic Gardens, Kew or the greatly influence the results of the ancestral charac- South African National Biodiversity Institute. For ter state reconstructions. The range of corolla tube three species, Lithospermum cinereum DC., Lithos- length for each species is provided in Table 2. permum gayanum I. M. Johnst., and Buglossoides In the present study, corolla tube length, rather than tenuiflora, sequence data were obtained from corolla length (which includes the corolla tube and GenBank (see Supporting information, Table S1). the corolla lobes), was used for multiple reasons. The primary reason involves the fact that the lengths of the corolla tube and corolla lobes may be under different MOLECULAR DATA genetic controls (Leins & Erbar, 2003); therefore, the For 39 species of Lithospermum and related genera, inclusion of both in the same measurement may not DNA isolations, polymerase chain reaction (PCR) allow for a comparison of homologous features of the amplifications, and sequencing reactions were under- flower. Additionally, the vast majority of variation in taken for ten plastid DNA regions (matK, ndhF-rpl32, corolla length among species of Lithospermum is a psbA-trnH, psbJ-petA, the rpl16 intron, trnK-rps16, result of differences in corolla tube length, not corolla trnL-rpl32, trnQ-rps16, ycf6-psbM, and trnL-trnF), in lobe length (J. I. Cohen, unpubl. data). Finally, most of accordance with the methods described by Cohen & the species included in the present study bear lobes Davis (2009, 2012). Sequence trace files were com- that, at anthesis, are flared or reflexed (Fig. 1) (Cohen, piled, examined, and edited with SEQUENCHER, 2011); consequently, the corolla lobes would not con- version 4.6–4.8 (Gene Codes Corporation). Sequences tribute to the overall length of the tubular portion of were deposited in GenBank (see Supporting informa- the corolla, and their inclusion could result in an tion, Table S1), and the matrix is available at Tree- overestimate of this length. base (S12700).

MATRIX CONSTRUCTION COROLLA TUBE LENGTH MEASUREMENTS Initial alignments were performed with MUSCLE AND DEFINITION (Edgar, 2004) as implemented by the European The range in corolla tube length for 42 species of Bioinformatics Institute’s MUSCLE server (http:// Lithospermum and related genera was determined www.ebi.ac.uk/Tools/msa/muscle/) using the default either through direct observation or by consultation of settings. Subsequent adjustments were made in the literature. The corolla tube is defined as the fused, BIOEDIT, version 7.0.5.3 (Hall, 1999) and WIN- tubular portion of the corolla. The corolla tube CLADA, version 1.7 (Nixon, 2002). Gaps were coded extends from the base of the corolla to the base of the using simple indel coding (Simmons & Ochoterena, corolla lobes. For 25 species, the lengths of the corolla 2000). Inversions were coded as present/absent. Each tubes were measured on herbarium specimens or inverted sequence was recoded with the reverse com- from pickled flowers. Between 13 and 25 specimens plement and included in the analyses in this manner were examined for each species, and for each speci- (Graham et al., 2000; Ochoterena, 2009; Davis & men, the lengths of four to six corolla tubes were Soreng, 2010). All characters were weighted equally measured on mature flowers (Table 2). For some and treated as unordered (i.e. non-additive). Regions species, such as Lithospermum rosei (I. M. Johnst.) that aligned ambiguously were excluded from analy- J. I. Cohen and Lithospermum revolutum B. L. Rob., ses. The ten DNA regions were included into the same fewer specimens are available and, in these cases, matrix, and prior to concatenation, an inconguence approximately five to ten specimens were observed. length difference (ILD) test (Farris et al., 1995) was For the other 17 species, the ranges of corolla tube conducted, in WINCLADA (Nixon, 2002), for each pair lengths were taken directly from measurements of of DNA regions. The same molecular matrix was used corolla tube lengths listed in species descriptions of in all phylogenetic analyses. floras (Valentine & Chater, 1972; Cronquist et al., 1984; Al-Shehbaz, 1991; Gleason & Cronquist, 1991; Kelley, 1993; Zhu, Riedl & Kamelin, 1995) and revi- PHYLOGENETIC ANALYSIS sions (Johnston, 1952, 1954). Although the present MP phylogenetic analyses were conducted in TNT study includes measurements of corolla tube length (Goloboff, Farris & Nixon, 2008) with the same search from different sources (i.e. herbarium specimens, strategy as described by Cohen (2011) and Cohen pickled flowers, and published descriptions), measure- & Davis (2012). BI phylogenetic analyses were ments from the various sources do not differ by more conducted in BEAST, version 1.6.1 (Drummond &

Rambaut, 2007) using the BioHPC web computing different coding schemes analyzed in MP, ML, and BI resources of the Computational Biology Service Unit frameworks, as implemented in MESQUITE (MP) at Cornell University (http://cbsuapps.tc.cornell.edu/ (Maddison & Maddison, 2010), TNT (MP) (Goloboff index.aspx). Prior to these analyses, the matrix was et al., 2008), COMPARE, version 4.6b (ML) (Martins, examined with MODELTEST (Posada & Crandall, 2004), and BAYESTRAITS (BI) (Pagel, 1997, 1999; 1998) to identify the appropriate model for phyloge- Pagel & Meade, 2011). For ancestral character state netic analysis. Because the results of the ILD test reconstructions, the MP tree with the fewest steps (Farris et al., 1995) suggest that the ten regions was used for MP analyses, and the BI consensus tree are not significantly incongruent, all ten regions was used for ML and BI analyses (Fig. 2; Table 1). were concatenated into a single matrix that was Only these two trees were included in the present analyzed in MODELTEST. After the appropriate study to minimize potential topological incongruence, model (TVM + I +G) was determined, a search strat- and therefore allow for effective and equivalent com- egy was conducted in BEAST: one chain run for parisons of the various types of analytical methods 20 000 000 generations and sampled every 1000 gen- and scoring schemes. erations. This analysis was undertaken ten times, To include in the analyses the entire range of mor- and each time employed the following search param- phological variation in corolla tube length, six coding eters: a GTR + I +Gmodel (the closest approximation, schemes were necessary (Table 1). This is because in BEAST, of TVM + I +G) of with six gamma catego- only two software packages, COMPARE and TNT, can ries, an uncorrelated exponential clock, and a yule accept a range of values for each terminal. The other process model. two only accept single values for each analysis. The During these analyses, Lithospermum dakotense six coding schemes are: M. L. Gabel, a fossil species of Lithospermum (Gabel, 1. minimum value for each species (MIN); 1987), was placed at the stem of the New World 2. mean value for each species (MEAN); members of Lithospermum. Placing the fossil in this 3. maximum value for each species (MAX); position is most appropriate for three reasons: (1) 4. maximum value for species with corolla tubes less L. dakotense was found in the New World (Gabel, than 20 mm in length, and minimum value for 1987); (2) the New World members of Lithospermum species with corolla tubes > 20 mm in length form a monophyletic group (Cohen & Davis, 2009, (MAX/MIN); 2012); and (3) currently it is not possible to place the 5. the range of values for each species (RANGE); and fossil among the species of a particular clade of New 6. (range/2) ± half of the range (±1/2 RANGE). World members of Lithospermum. This fossil, which was found in the middle of the Ash Hollow Formation This scoring scheme was only used for analyses in in South Dakota (Gabel, 1987), is approximately COMPARE (Martins, 2004) (Table 1), and it is not the 8 Mya) (M. Gabel, pers. comm.), although the Ash standard manner in which to input intraspecific data Hollow Formation has a minimum age of 6.6 Mya into the software. The standard manner in which to (Skinner & Johnson, 1984). Consequently, to account do so is as 1 SE). The use of range encompasses all for the presumed age of the fossil, as well as the intraspecific variation, whereas 1 SE only accounts possibility of it being slightly younger or older, the for a portion of the observed intraspecific variation age of this fossil was estimated as 8 ± 1.5 Mya, and (Samuels & Witmer, 1999). Therefore, the inclusion of using a lognormal distribution, the fossil was incor- range data, rather than the mean ± SE, results in an porated into the BI analyses. Multiple preliminary even more critical test of change in corolla tube length analyses were undertaken in BEAST to optimize the between ancestors and descendants (Swiderski et al., operator values for the final analyses. To determine 1998). whether the analyses sufficiently searched tree space, the log files were examined with TRACER, version Maximum parsimony 1.5 (Rambaut & Drummond, 2007). Using TREEAN- Two different software packages, MESQUITE (Mad- NOTATOR, version 1.6.1 (Drummond & Rambaut, dison & Maddison, 2010) and TNT (Goloboff et al., 2007), a maximum clade credibility tree was recon- 2008), were used to investigate ancestral character structed using a posterior probability limit of 0.5 and state reconstructions in an MP framework. MIN, a 30% burn-in. This percentage was determined via MEAN, MAX, and MAX/MIN values were investi- observation of the log-likelihood plots. gated in MESQUITE using three different methods: linear parsimony (Swofford & Maddison, 1987), squared-change parsimony, and weighted squared- ANCESTRAL CHARACTER STATE RECONSTRUCTION change parsimony (Maddison, 1991). Ancestral values The ancestral values of corolla tube length in Lithos- were recorded at each node. Although these three permum and related taxa were examined with six methods utilize MP criteria, the methods are not

© 2012 The Linnean Society of London, Biological Journal of the Linnean Society, 2012, ••, ••–•• 02TeLnenSceyo London, of Society Linnean The 2012 © ilgclJunlo h ina Society Linnean the of Journal Biological VLTO FCRLATB LENGTH TUBE COROLLA OF EVOLUTION 2012, , •• ••–•• ,

Figure 2. 50% majority rule consensus tree (posterior probability =-17 596.602, likelihood =-17 449.003) of Lithospermum and related taxa. Phylogenetic tree is similar to one of 20 most-parsimonious trees (442 steps, consistency index = 0.55, retention index = 0.61). Numbers above branches are jackknife values Ն 50%/posterior probabilities Ն 0.95. Triangles denote origin of corollas > 20 mm. Numbers below branches are identiﬁcation numbers for nodes (see text). Hashmarks denote nodes from TNT reconstructions in which range of ancestor and that of at least one descendant do not overlap, and asterisks and diamonds represent nodes from COMPARE and BAYESTRAITS reconstructions, respectively, in which ranges of ancestor and that of at least one descendant differ byat least 1.96 SEs or SDs, respectively (for additional information, see Table 3). Node 2 is not included because of polytomy at base of tree. Dotted horizontal lines are branches that collapse in maximum parsimony strict consensus tree. Dashed vertical line denotes two different placements of Lithospermum ruderale, depending on method used. The scale at the bottom corresponds to millions of years before the present. Graph on right provides range of corolla tube lengths (in mm) for each species. L. = Lithospermum, G. = Glandora, and B. = Buglossoides. 7 8 J. I. COHEN equivalent to parsimony optimization (Swofford & were run, with the ﬁrst 50 000 iterations treated as Maddison, 1987; Maddison, 1991; Webster & Purvis, burnin. Ancestral values at each node were sampled 2002a), unlike the TNT method which is based on every 100 iterations, for a total of 10 000 estimations. Farris optimization (Farris, 1970; Goloboff et al., From these estimations, EXCEL (Microsoft) was used 2006). to calculate the mean and standard deviation (SD) for In TNT, patterns of character evolution for RANGE, each node. For all analyses, the values for RateDev MIN, MEAN, MAX, and MAX/MIN were examined and DataDev were optimized to accept the proposed using parsimony optimization (Farris, 1970; Goloboff character state between 20% and 40% of the time, as et al., 2006). Ancestral character states were recorded recommend (http://www.evolution.rdg.ac.uk). at each node, and adjacent nodes or terminals with For each analysis, values for the three scaling ranges that do not overlap were noted. parameters in BAYESTRAITS (i.e. delta, kappa, and For analyses in TNT, the maximum value of the lambda) were estimated. EXCEL was used to calcu- corolla tube lengths of Lithospermum leonotis (I. M. late the mean ± SD for delta, kappa, and lambda, and Johnst.) J. I. Cohen and Lithospermum macromeria these values were recorded. J. I. Cohen was scored as 65 mm, rather than 75 mm and 70 mm, respectively. This is because only 66 Rates of evolution states are available for each character in TNT, and it For each analysis, the rate of evolution of corolla tube is more prudent to change the maximum values of length was calculated for each branch. The rate was two species than to scale all values. Decreasing the determined by subtracting the corolla tube length of maximum values of these two species did not affect the descendant from that of the ancestor and dividing the character state reconstructions. In addition, only this change by the length of time (Myr) between the whole numbers were analyzed with TNT; conse- ancestor and descendant. If the corolla tube length of quently, when corolla tube lengths were Ն 0.5, values the ancestor was a range, then the mean value was were rounded up to one, and when these lengths were used for these calculations. < 0.5, they were rounded down to zero.

Maximum likelihood RESULTS The evolutionary patterns of corolla tube length values for MIN, MEAN, MAX, MAX/MIN, and ±1/2 MATRIX AND PHYLOGENETIC ANALYSIS RANGE were investigated with COMPARE, version The sequences of the ten plastid-DNA regions from 42 4.6b (Martins, 2004), using the BI consensus tree species of Lithospermum and related genera provided (Table 1). Each data partition was analyzed with the 7978 aligned nucleotides. From this aligned matrix, phylogenetic generalized least squares (PGLS) ances- 225 characters are parsimony-informative, with 196 tor method (Martins & Hansen, 1997), with both the (87%) being nucleotide substitutions and 29 (13%) linear and exponential models. For each analysis, being structural characters (e.g. gaps and inversions). 10 000 iterations were attempted to estimate ␴2.In MP analyses of the DNA sequence data recovered all analyses, except those with range information, 20 MP trees of 442 steps (consistency index = 0.55, within-species variation was assumed to be negligi- retention index = 0.61), and the strict consensus is ble. Additionally, PGLS – Rates of Evolution analyses well-resolved (Fig. 2). Both Lithospermum and the were conducted to determine whether the data better New World members of Lithospermum are resolved as ﬁt a model of Brownian motion (Felsenstein, 1985) or monophyletic, although these two clades, along with one of stabilizing selection (Martins, 1999). Ancestral most others of the ingroup, receive weak jackknife character states were recorded at each node, and support (jk). Despite this generally weak support, adjacent nodes that differ by 1.96 ¥ (range/2) or 1.96 each of eight species pairs is supported by > 70% jk SE were noted. (Fig. 2). BI analyses resulted in a consensus tree with a posterior probability of -17 596.602 and a likelihood Bayesian inference of -17 449.003. The consensus tree is identical to the With the consensus tree from the BI analyses, MIN, MP tree in Figure 2, with one exception: the place- MEAN, MAX, and MAX/MIN values were investi- ment of Lithospermum ruderale Douglas ex Lehm. gated in BAYESTRAITS (Pagel & Meade, 2011). (Fig. 2, dashed vertical line). In the MP tree, L. rud- The ancestral character states for each node were erale is resolved as sister to a clade composed of estimated using the protocol recommended in the Lithospermum californicum A. Gray and Lithosper- BAYESTRAITS manual (http://www.evolution.rdg.ac. mum matamorense DC. but, in the BI analyses, uk). The data were examined using model A (Brown- L. ruderale is sister to a clade that includes Lithos- ian motion; Felsenstein, 1985) and model B (direc- permum canescens Lehm., Lithospermum molle tional change). For each analysis, 1 050 000 iterations Muhl., and Lithospermum helleri (Small) J. I. Cohen.

Compared to the MP consensus tree, the BI consensus varied to a greater extent among nodes than did those tree has greater resolution and a greater number of from the exponential model, which were very similar well-supported clades, with seventeen clades having among all nodes (Table 3). The ancestral values > 0.95 posterior probability (pp). This includes all of derived from the linear model are similar to those from the species pairs that receive > 70% jk as well as nine squared-change parsimony. For both the linear and additional clades (Fig. 2). the exponential models, the addition of range data From the inclusion of fossil data in the BI analyses, resulted in smaller and less variable ancestral charac- the root of the phylogenetic tree is dated to approxi- ter state values than occurred with the exclusion of mately 16.9 Mya. The most recent common ancestor range data. Depending on the data partition and which (MRCA) of Lithospermum is approximately 12.1 Mya, of the two models was employed, at six to 12 nodes and the MRCA of the New World members of the the value of the ancestor differs by 1.96 SE or genus is approximately 8.4 Mya. Most extant species 1.96 ¥ (range/2) from that of one of the descendants, of Lithospermum included in the present analyses whereas, at four to ten nodes, the value of the ancestor originated < 5 Mya. differs by 1.96 SE or 1.96 ¥ (range/2) from that of both descendants (Fig. 2, asterisks; Table 3). The results of the PGLS – Rates of Evolution analyses provide evi- ANCESTRAL CHARACTER STATE RECONSTRUCTION dence that, for all permutations of the data, a model Corolla tube lengths in Lithospermum and related of Brownian motion, rather than one of stabilizing taxa range from 1 mm in L. matamorense to 75 mm in selection, better fits the data. L. leonotis. The reconstructed values for ancestors are provided in Table 3 (see also Supporting information, Bayesian inference Table S2). In BAYESTRAITS, just as with COMPARE, the pattern of corolla tube length evolution in Lithosper- Maximum parsimony mum better fits a model of Brownian motion than one In all analyses in Mesquite (MIN, MEAN, MAX, and of directional change. In general, the values from MAX/MIN), squared-change parsimony and weighted BAYESTRAITS are similar to those from squared- squared-change parsimony resolve a greater number change parsimony, weighted squared-change parsi- of shifts in corolla tube lengths compared to linear mony, and the linear model of COMPARE. However, at parsimony (Table 3). At nodes in which the two nodes where the descendants greatly differ in corolla descendants greatly differ in corolla tube length tube length (e.g. nodes 36 and 41) (Fig. 2, triangles), (Fig. 2, triangles), squared-change parsimony and the values reconstructed by BAYESTRAITS are fre- weighted squared-change parsimony often recon- quently smaller than those of these other methods. The struct values intermediate between those of the two ancestral values reconstructed by BAYESTRAITS tend descendants. By contrast, the values derived from to have large standard deviations, sometimes resulting linear parsimony frequently are more similar to the in negative corolla tube lengths. Additionally, for all values of the species with shorter corolla tubes than permutations of the data and for both models (A and to those of the species with longer ones (Table 3). B), the SD values for the ancestral character states are Optimizing MIN, MEAN, MAX, MAX/MIN, and greatest at the tips of the phylogenetic tree and RANGE values for corolla tube length onto the MP smallest at the base (Table 3). tree with the fewest number of steps for corolla tube Depending on the data and type of model employed, length, six origins of corollas > 20 mm are recon- at one to five nodes the SD of the corolla tubes of the structed (Fig. 2, triangles). At 11 nodes, the lengths of ancestor and the SD of that of one of the descendants the corolla tubes of the ancestor and one descendant do not overlap (Fig. 2, diamonds). It was never the do not overlap, but at only one node (16), the lengths case that the SD of the corolla tubes of the ancestor of the corolla tubes of the ancestor and both descend- did not overlap with the SD of the corolla tubes of ants do not overlap (Fig. 2, hashmarks; Table 3). The both descendants (Table 3). The values for the scaling values reconstructed from the TNT method tend to be parameters kappa, delta, and lambda estimated from more similar to those of linear parsimony than to the analyses that employed a model of Brownian those from squared-change parsimony or weighted motion (model A) are presented in Table 4. squared-change parsimony. Rates of evolution Maximum likelihood The rates of evolution ranged from -63.02 mm Myr–1 In COMPARE, two models (i.e. linear and exponential) to 43.41 mm Myr–1, depending on the type of method were used to investigate character evolution for MIN, used and data partition employed. The TNT method, MEAN, MAX, MAX/MIN, and ±1/2 RANGE (Table 1). linear parsimony, and the exponential model of Of the two, the linear model resulted in values that COMPARE reconstructed the greatest number of

RANGE MEAN COHEN I. J.

MP MP ML ML (±1/2 RANGE) BI

Weighted Squared- squared- Model A Model B Linear change change (Brownian (Directional Node TNT TNT parsimony parsimony parsimony Linear Exponential Linear Exponential motion) change) 02TeLnenSceyo London, of Society Linnean The 2012 © 1 15–16 17 16.5–17 14.81 15.81 12.38 ± 3.95 15.59 ± 3.34 11.95 ± 3.53 9.66 ± 3.81* 12.32 ± 1.42 10.42 ± 10.05 3 9‡ 7 6.5 8.54 7.48 7.62 ± 5.53 15.47 ± 3.34† 8.1 ± 4.81 9.66 ± 3.81 10.21 ± 4.49 10.84 ± 10.54 4 15–16 17 16.5 14.52 15.17 14.74 ± 2.60 15.59 ± 3.34 13.38 ± 2.40* 9.66 ± 3.81* 12.82 ± 2.61 12.54 ± 9.86 5 15–16 17 16.5 14.44 15.23 14.85 ± 2.44 15.59 ± 3.34 13.28 ± 2.26* 9.66 ± 3.81* 13.23 ± 3.74 13.18 ± 10.65 6 5–12 6–10 5.5–10 11.31 14.27 14.60 ± 1.96* 15.59 ± 3.34 12.42 ± 1.81* 9.66 ± 3.81 12.17 ± 4.87 11.70 ± 10.48 7 5–6‡ 6 5.5 6.27 3.75 4.48 ± 5.86 13.76 ± 4.06† 4.16 ± 5.47 8.57 ± 4.89* 8.93 ± 6.95 11.61 ± 11.29 8 5–12‡ 6–10 5.5–10 13.21 14.95 15.44 ± 1.83* 15.59 ± 3.34 12.84 ± 1.66* 9.66 ± 3.81 13.17 ± 5.88 13.50 ± 11.63 9 4 4 3.5 6.74 3.50 3.65 ± 6.14 4.68 ± 13.85 3.86 ± 5.58 5.44 ± 10.68 8.44 ± 7.87 10.58 ± 12.57 10 9–12‡ 10 9.5–10 21.59 16.90 17.55 ± 1.76† 15.59 ± 3.34 14.21 ± 1.52* 9.66 ± 3.81 16.83 ± 6.89 15.92 ± 12.26 11 9–12 10 9.5–10 18.34 16.37 16.66 ± 1.64 15.59 ± 3.34 13.49 ± 1.48 9.66 ± 3.81 16.47 ± 8.00 15.51 ± 12.40 12 28–35 37 36.5 33.20 31.08 29.57 ± 4.00* 15.59 ± 3.34† 22.76 ± 2.60† 9.66 ± 3.81† 23.5 ± 9.35 18.58 ± 13.22 13 9–12 10 9.5–10 19.56 16.92 16.66 ± 1.58 15.59 ± 3.34 13.52 ± 1.46 9.66 ± 3.81 16.99 ± 9.02 15.82 ± 13.06 14 9–12 10 9.5–10 13.89 14.20 14.47 ± 2.57 15.59 ± 3.34 11.62 ± 2.19 9.66 ± 3.81 14.87 ± 9.40 14.83 ± 13.10 15 9–12 10 9.5–10 13.32 13.28 13.81 ± 3.04 15.59 ± 3.34* 10.57 ± 0.2.67 9.66 ± 3.81 13.78 ± 10.46 14.52 ± 13.81 16 9–12§ 10 9.5–10 16.56 3.00 14.57 ± 4.53† 15.33 ± 3.46* 7.86 ± 4.47† 9.37 ± 4.02* 14.37 ± 11.87 14.67 ± 14.50 17 3–5 4 4 7.85 3.00 5.93 ± 5.54 9.45 ± 8.29 4.6 ± 5.74 6.82 ± 8.18 10.92 ± 13.29 13.38 ± 15.27 18 9–12‡ 10–14 9.5–13.5 25.07 17.91 19.16 ± 2.08* 15.59 ± 3.34 15.47 ± 1.84* 9.66 ± 3.81 19.55 ± 10.13 16.85 ± 14.00 19 35–47 44 44 40.86 41.20 40.90 ± 4.91* 16.00 ± 3.35† 29.45 ± 3.00† 9.77 ± 3.81† 29.61 ± 12.98 21.90 ± 15.68 20 9–12 10–14 9.5–13.5 14.76 17.05 16.13 ± 2.71* 15.59 ± 3.34 13.84 ± 2.43* 9.66 ± 3.81 16.78 ± 11.14 17.27 ± 14.22 ilgclJunlo h ina Society Linnean the of Journal Biological 21 11–12 10–14 10–13.5 13.72 16.29 14.93 ± 3.27 15.59 ± 3.34* 13.54 ± 2.77 9.66 ± 3.81 15.81 ± 12.65 15.64 ± 14.86 22 11–12 10–14 10–13.5 12.41 14.21 13.36 ± 4.35 15.57 ± 3.34 12.95 ± 3.34 9.67 ± 3.81 14.38 ± 13.98 15.55 ± 15.98 23 9–12 10 9.5–10 15.34 15.25 15.00 ± 1.61 15.59 ± 3.34 12.26 ± 1.49 9.66 ± 3.81 15.25 ± 9.99 15.16 ± 13.51 24 9–10 10 9.5–10 11.82 13.97 12.93 ± 2.24 15.59 ± 3.34 11.29 ± 1.94 9.66 ± 3.81 13.57 ± 11.23 14.88 ± 14.36 25 9–10 10 9.5–10 9.94 13.24 11.14 ± 3.74 15.59 ± 3.34* 10.52 ± 2.92 9.66 ± 3.81 12.07 ± 12.33 13.96 ± 14.98 26 9–10 10 9.5 10.18 11.18 11.80 ± 2.85 15.59 ± 3.34* 10.73 ± 2.41 9.66 ± 3.81 12.68 ± 12.47 13.36 ± 14.83 27 9–10 10 9.5 9.23 9.46 9.36 ± 5.31 15.01 ± 3.57* 9.29 ± 4.53 9.61 ± 3.84 11.50 ± 14.04 12.83 ± 15.36 28 9–12 10 9.5–10 14.62 15.25 14.76 ± 1.72 15.59 ± 3.34 11.53 ± 1.63 9.66 ± 3.81 15.69 ± 11.14 15.68 ± 14.39 29 9–10 10 9.5–10 11.13 12.01 11.83 ± 2.37 15.59 ± 3.34 9.63 ± 2.43 9.66 ± 3.81 13.26 ± 12.01 14.45 ± 14.60 30 8–9 8 7.5–8 8.93 10.50 9.94 ± 3.20 15.57 ± 3.34* 9.07 ± 2.88 9.66 ± 3.81 10.96 ± 14.28 12.92 ± 16.65 31 8–9 8 7.5–8 8.14 8.79 8.44 ± 4.78 15.10 ± 3.47† 8.35 ± 4.09 9.58 ± 3.84 10.12 ± 15.82 13.12 ± 17.04 32 9–10 10–11 9.5–10.5 9.86 11.49 11.02 ± 2.37 15.59 ± 3.34* 9.95 ± 2.15 9.66 ± 3.81 12.63 ± 12.21 13.76 ± 15.25 33 9–10‡ 10–11 9.5–10.5 7.95 11.11 9.50 ± 3.87* 15.57 ± 3.34* 8.31 ± 3.30* 9.65 ± 3.81* 11.51 ± 13.77 13.77 ± 15.85 34 9–13 10 9.5–10 17.37 16.11 17.48 ± 2.08 15.60 ± 3.34 12.29 ± 1.95 9.66 ± 3.81 17.44 ± 12.20 16.13 ± 15.05 35 9–13‡ 10 9.5–10 16.70 16.15 19.08 ± 3.33† 15.62 ± 3.34* 12.06 ± 3.02* 9.66 ± 3.81* 17.52 ± 13.49 16.08 ± 15.67 36 9–13‡ 10 9.5–10 28.23 21.29 29.41 ± 5.03† 17.13 ± 3.57† 15.70 ± 3.49† 9.90 ± 3.78* 22.58 ± 15.21* 19.05 ± 16.24* 37 10–14‡ 13 12.5 20.79 16.80 19.47 ± 2.74* 15.62 ± 3.34* 13.42 ± 2.37† 9.67 ± 3.81* 18.10 ± 13.35 16.64 ± 15.33 38 10–14 13 12.5 17.51 16.30 19.08 ± 2.96 15.67 ± 3.34 12.88 ± 2.79 9.68 ± 3.81 18.82 ± 14.27 18.20 ± 15.84 39 10–14 13 12.5 14.17 13.57 14.22 ± 5.09 15.04 ± 4.30 12.70 ± 3.73 9.86 ± 3.81 16.56 ± 16.01 17.33 ± 16.58

2012, , 40 10–14‡ 13 12.5 17.57 16.25 20.98 ± 3.71† 16.10 ± 3.39* 12.32 ± 3.58* 9.70 ± 3.84* 19.47 ± 15.92 17.72 ± 16.68 41 11–14‡ 13 12.5 30.69 28.38 32.29 ± 5.11† 21.74 ± 5.40† 14.93 ± 4.19† 10.53 ± 4.02† 24.98 ± 17.49* 20.78 ± 17.97*

•• Node 2 not included because of polytomy at base of tree (Fig. 2). MP, maximum parsimony; ML, maximum likelihood; BI, Bayesian inference. For discussion of methods and coding schemes, see text. Reconstructed values of all coding schemes in the Supporting information, Table S2. ••–•• , *At least 1.96 SE or SD between ancestor and one descendant. †At least 1.96 SE or SD between ancestor and both descendants. ‡Ranges of ancestor and one descendant do not overlap. §Ranges of ancestor and both descendants do not overlap. EVOLUTION OF COROLLA TUBE LENGTH 11 branches with a rate of 0 mm Myr–1 (no change the entire phylogenetic tree. By contrast, the other between nodes). This was the case for all permuta- methods resulted in either a decrease or no change in tions of the data (Fig. 3; see also Supporting informa- corolla tube length across the entire tree (see Sup- tion, Table S2). The use of the TNT method, linear porting information, Table S2). parsimony, and weighted squared-change parsimony resulted in an increase in corolla tube length across DISCUSSION METHODS OF ANCESTRAL CHARACTER Table 4. Values for scaling parameters estimated from STATE RECONSTRUCTION analyses of model A (Brownian motion) in BAYESTRAITS For the same data, the various methods of ancestral character state reconstruction resolved different Data type Kappa Delta Lambda values at the same node. In most instances, these hypothesized values differed only minimally among MIN 0.24 ± 0.12 1.71 ± 0.45 0.57 ± 0.26 MEAN 0.37 ± 0.62 1.07 ± 0.69 0.36 ± 0.25 methods but, in other cases, the resulting values MAX 0.22 ± 0.11 2.36 ± 0.35 0.71 ± 0.21 differed by two- or three-fold (Table 3). For example, MAX/MIN 0.26 ± 0.22 1.83 ± 0.61 0.59 ± 0.27 with the use of MEAN analyses with TNT and linear parsimony resulted in an ancestral corolla tube MIN, minimum value for each species; MEAN, mean value length of 9.5–10 mm at node 36, whereas analyses for each species; MAX, maximum value for each species; with weighted squared-change parsimony, model A in MAX/MIN, maximum value for species with corolla tubes BAYESTRAITS, and two other methods reconstructed less than 20 mm in length, and minimum value for species corolla tube lengths > 20 mm. Although this example with corolla tubes > 20 mm in length. is at the extreme end of the spectrum, large, although

Figure 3. Rates of evolution (in mm Myr–1) of mean corolla tube length for species of Lithospermum. Numbers after species are ranges in corolla tube length (in mm). A, rates derived from analyses of TNT/linear parsimony/weighted squared-change parsimony. B, rates derived from analyses of linear model of COMPARE, no range data/linear model of COMPARE, with range data/model A (Brownian motion) of BAYESTRAITS. Negative values are increase in corolla tube length between ancestor and descendant, and positive values are decrease in corolla tube length. A value of zero (0) is no change in corolla tube length.

© 2012 The Linnean Society of London, Biological Journal of the Linnean Society, 2012, ••, ••–•• 12 J. I. COHEN less drastic, differences are observed at other nodes Given that the TNT method and linear parsimony (Table 3). Even when different types of analyses do include the fewest assumptions and minimize the not produce such large differences, the various number of different evolutionary rates, these two methods still can result in ancestral values that differ methods should be preferred for the inference of by approximately 50% (e.g. nodes 14 and 34). For ancestral character states of continuous characters. example, the use of MEAN analyses of linear parsi- These advantages aside, the TNT method and linear mony and weighted squared-change parsimony parsimony are disadvantaged because they do not resulted in ancestral corolla tube lengths of 9.5– determine SD or SE, although they can provide 10 mm and 14.20 mm, respectively, at node 14. This ranges of ancestral values. Should a measure of esti- example demonstrates that these different ancestral mation accuracy (rather than a range of values) be reconstructions can occur even between methods with important, the results of the linear model of the same underlying evolutionary framework (i.e. MP, COMPARE appear to be most useful for ancestral ML or BI). character state reconstruction of corolla tube length Although results can vary among the MP methods, in Lithospermum. This is because the consistent SE MP methods differ in two apparent manners from the values from analyses in COMPARE provide more ML and BI methods. The ML and BI methods, in useful information than the more variable SD values comparison to the MP ones, incorporate models that from analyses in BAYESTRAITS (Table 3). have a greater number of assumptions concerning patterns of evolution (Pagel, 1997, 1999; Martins, 1999), and these two methods can resolve both mean DATA FOR ANCESTRAL CHARACTER and SE or SD values at nodes, which the MP methods STATE RECONSTRUCTION do not (Table 1). The SE reconstructed by COMPARE Corolla tube length data was partitioned into six remain fairly constant (approximately 3 mm) at each different coding schemes. Two of these schemes (i.e. node on the phylogenetic tree, although the SD of RANGE and ±1/2 RANGE) are each restricted to only BAYESTRAITS do not. At the base of the tree, the SD one type of analysis, whereas the other four (i.e. MIN, reconstructed by BAYESTRAITS are small, whereas, MEAN, MAX, MAX/MIN) were examined in an MP, towards the tips of the tree, the SD are much larger ML, and BI framework (Table 1). Of these four latter (Table 3). At some nodes, these values are greater schemes, MEAN and MAX/MIN appear most useful than the mean, which can result in negative corolla for studies on the patterns of evolution of continuous tube lengths. Consequently, the SD derived from characters. MEAN, as it is the average of the values BAYESTRAITS appear less useful than the mean for a character, is a representative, single value in values. which to capture the variation of a species (Table 2). In the present study, most methods reconstruct When appropriate, MAX/MIN is useful because it ancestral character states that necessitate three dif- provides the minimum change necessary between ferent rates of evolution to be hypothesized at some species with very different values for a character. nodes (Fig. 3). At nodes where these different rates Although MEAN and MAX/MIN provide useful are observed, the two descendant species tend to information on the evolution of a continuous charac- greatly differ in corolla tube length, which results in ter, they are only single point values and not a range; intermediate values reconstructed at these nodes. If therefore, the reconstructed ancestral values will this is the case, the rate of evolution that gives rise not represent the morphological variation of extant to the ancestor differs from that which gives rise to species (Herrera, 2009) or hypothesized ancestors each descendant (Fig. 3). Therefore, three different (Swiderski et al., 1998). Additionally, some studies rates of evolution must be hypothesized: one for the (Crisp & Weston, 1987; Pimentel & Riggins, 1987) ancestor and one for each of the two descendants have suggested that the mean of a continuous char- (Fig. 3). By contrast to most methods, the use of acter has no meaning in a phylogenetic sense. From a TNT and linear parsimony results in only two rates practical point, species can have the same mean for a of evolution hypothesized: one for the ancestral character despite having different ranges, or vice- species and one descendant species (often the one versa. The mean may be the most appropriate single with the shorter corolla tube) and another for the point to encompass the variation of a character, other descendant species (Fig. 3A). Hypothesizing although this single value only can provide limited two, rather than three, different rates of evolution is information concerning patterns of evolution. more parsimonious. Fewer rates of evolution arise To capture all of the variation a species exhibits, because the value of the ancestral species is not two methods were employed: one for software that intermediate between the values of the descendants, accepts point data and another for software that rather it is similar to one of the two descendant accepts range data. If only point values are accepted, species. the use of MAX and MIN allows for putative limits to

© 2012 The Linnean Society of London, Biological Journal of the Linnean Society, 2012, ••, ••–•• EVOLUTION OF COROLLA TUBE LENGTH 13 be placed on the ancestral values. The use of MIN, diate length are not known currently (Johnston, 1952, MEAN, MAX, and MAX/MIN (if appropriate) provides 1954; Cohen, 2011), although it is possible that inter- information on the mean and range of species, as well mediate forms have existed and are now extinct. as the minimum change between smaller and larger Alternatively, it is possible that some species with values. Because most software packages only accept long, hummingbird-pollinated corolla tubes originated single point values, not a range of data, the use of without progression through intermediate floral these four scoring schemes is a practical approach to forms, thus allowing them to adapt quickly to a par- encompass the morphological variation in extant and ticular niche. ancestral species. Lithospermum colonized the New World, at the If a range of values can be input, RANGE and ±1/2 most recent, approximately 8.4 Mya, which is RANGE would be the most appropriate manners in approximately the same time as hummingbirds began which to represent the variation of a continuous char- to diversify in North America (Bleiweiss, 1998). acter. The inclusion of all the variation for each termi- Shortly after individuals of Lithospermum arrived nal is the most critical manner in which to investigate and began to diversify in the New World, the ances- patterns of evolution (Swiderski et al., 1998). For tors of two species pairs – Lithospermum flavum example, if speciation occurs at the edges of ranges, Sessé & Moc. + Lithospermum notatum (I. M. Johnst.) which often include extreme morphological values J. I. Cohen and Lithospermum exsertum (D. Don) J. I. (Weber & Schmid, 1998; Jonas & Geber, 1999; Lowry, Cohen + Lithospermum johnstonii J. I. Cohen (Fig. 1) 2012), then the most critical test in which to investi- – developed long, tubular corollas (Fig. 2). Given the gate the evolution of a morphological character should long, yellow, green or orange corolla tubes of these include all potential values, not just those that are ancestors (Cohen, 2011) (Table 3), the flowers possess most common (i.e. mean ± SD). Even though RANGE the features of those pollinated by hummingbirds and ±1/2 RANGE were each only investigated in one (Richards, 1997). By contrast to these two ancestral software package, these two approaches capture all of species, two other species, L. leonotis and L. mac- the variation that exists in the extant species. There- romeria (Figs 1A, B, respectively) (Grant & Grant, fore, these two coding schemes also provide the great- 1970; Boyd, 2002; Cohen, 2011), developed long, est opportunity in which to hypothesize the variation tubular corollas more recently, approximately within that was present in ancestral species. Butler & Losos the past 1–1.5 Myr, at minimum. Consequently, (1997) suggest that a range of values at a node creates species of Lithospermum that bear hummingbird- ambiguity in the reconstruction of ancestral character pollinated flowers arose at two different times: in the states. Although this comprises one way of viewing the shadow of hummingbird diversification in North issue, a range of values at a node also can be viewed in America, as well as more recently. Some of these a positive light. A range of values accounts for the species, such as the ancestor of L. flavum and L. natural variation in hypothesized ancestors, and this notatum, likely exploited a new and novel group of is the same manner in which we accommodate the organisms (i.e. hummingbirds) in North America, morphological variation present in extant species whereas others, including L. leonotis and L. macrome- (Johnston, 1952; Alford, 2008). ria, may have developed long, tubular corollas in ecosystems that already housed other hummingbird- pollinated species. EVOLUTION OF COROLLA TUBE LENGTH Although corolla tube length has frequently IN LITHOSPERMUM increased in Lithospermum, ancestral character state In Lithospermum, corolla tube length is a character reconstructions provide evidence that corolla tube with much interspecific variation (Fig. 2), with shifts length also has decreased in some lineages (Fig. 2; in length often developing later in the evolution of the Table 3). For example, the corolla tubes of the sister genus (Fig. 3; Table 4; d > 1 and k < 1) (Pagel, 1999). species L. matamorense and L. californicum range in These shifts in corolla tube length may be correlated length from 1–2 mm and 10–15 mm, respectively, with changes in particular pollination syndromes. For although the corolla tube of the ancestor of these example, at least six origins of corolla tubes > 20 mm two species is hypothesized to be approximately in length are resolved (Fig. 2, triangles), and flowers 6.4–18.3 mm long, depending on the method of analy- with corolla tubes of this length are putatively polli- sis (Fig 2; Table 3). Therefore, corolla tube length nated by hummingbirds (Grant & Grant, 1970; Boyd, decreased in L. matamorense at least three-fold (and 2002). Of these six origins, three involve extant sister potentially up to 18-fold) from the ancestor. The corol- species that bear corolla tubes that differ in length by las of other species, such as L. calcicola B. L. Rob. and at least 15 mm, with two of these species pairs involv- L. latifolium Michx., also arose via a decrease in ing even greater differences in length (at least 21 mm) corolla tube length (Fig. 2). Because of the close prox- (Figs 1, 2). Species that bear corolla tubes of interme- imity of the anthers and stigmas in the flowers of the

© 2012 The Linnean Society of London, Biological Journal of the Linnean Society, 2012, ••, ••–•• 14 J. I. COHEN species with short corolla tubes, the breeding system Boyd AE. 2002. Morphological analysis of sky island popu- may be primarily selfing, which might have origi- lations of Macromeria viridiflora (Boraginaceae). Systematic nated as a result of its advantages in particular Botany 27: 116–126. environments (Armbruster et al., 2002; Kalisz & Butler MA, Losos JB. 1997. Testing for unequal amounts of Vogler, 2003). evolution in a continuous character on different branches of In Lithospermum, the evolution of corolla tube a phylogenetic tree using linear and squared-change parsi- length best fits a pattern of Brownian motion. This mony: an example using lesser Antillean Anolis lizards. is consistent with the observation that in the genus Evolution 51: 1623–1635. Cohen JI. 2011. A phylogenetic analysis of morphological and there is no uniform directional trend for corolla tube molecular characters of Lithospermum L. (Boraginaceae) length. Increases and decreases in corolla tube and related taxa: evolutionary relationships and character length are scattered among various clades of the evolution. Cladistics 27: 559–580. phylogeny (Fig. 3), with six and four substantial Cohen JI, Davis JI. 2009. Nomenclatural changes in Lithos- increases and decreases, respectively (Table 3), permum (Boraginaceae) and related taxa following a reas- resolved. In clades that include species with corolla sessment of phylogenetic relationships. Brittonia 61: 101– > tubes 20 cm, the lack of species with intermediate 111. corolla tube lengths suggests that changes in corolla Cohen JI, Davis JI. 2012. Molecular phylogenetics, molecu- tube length tend to occur in bursts, not gradually. lar evolution, and patterns of clade support in Lithosper- Although gradual directional trends in corolla tube mum (Boraginaceae) and related taxa. Systematic Botany length in Lithospermum are not observed, it is 37: 490–506. interesting to note multiple independent origins of Crisp M, Weston P. 1987. Cladistics and legume systemat- both long and short corolla tubes. Additionally, the ics, with an analysis of the Bossiaeeae, Brongniartieae and shortest corolla tubes are, in general, present in the Mirbelieae. In: Stirton CH, ed. Advances in legume system- Old World species, with numerous increases in atics, Part 3. London: Kew Publishing, 65–130. corolla tube length resolved among members of the Cronquist A, Holmgren AH, Holmgren NH, Reveal JL, New World. Holmgren PK. 1984. Intermountain flora, Vol. 4. The Asteridae except the Asteraceae. New York, NY: New York Botanical Garden Press. ACKNOWLEDGEMENTS Davis JI, Soreng RJ. 2010. Migration of endpoints of two The author would like to thank C. D. Kellogg, Mac genes relative to boundaries between regions of the plastid genome in the grass family (Poaceae). American Journal of H. Alford, Michael A. Sundue, John A. Allen, and Botany 97: 874–892. four anonymous reviewers for their very useful com- Drummond AJ, Rambaut A. 2007. BEAST: Bayesian evo- ments on this manuscript. Andrew Meade provided lutionary analysis by sampling trees. BMC Evolutionary helpful information on the analysis of continuous Biology 7: 214. characters in BayesTraits. Thomas Couvreur pro- Edgar RC. 2004. MUSCLE: multiple sequence alignment vided insight on the creation of data matrices for with high accuracy and high throughput. Nucleic Acids BEAST. Mark Gabel graciously confirmed the age of Research 32: 1792–1797. Lithospermum dakotense. Farris JS. 1970. Methods for computing Wagner trees. Sys- tematic Zoology 19: 83–92. REFERENCES Farris JS. 1990. Phenetics in camouflage. Cladistics 6: 91–100. Alford MH. 2008. Revision of Neosprucea (Salicaceae). Farris JS, Källersjö M, Kluge AG, Bult C. 1995. Testing Systematic Botany Monographs 85: 1–62. significance of incongruence. Cladistics 10: 315–319. Al-Shehbaz IA. 1991. The genera of Boraginaceae of Felsenstein J. 1985. Phylogenies and the comparative the southeastern United States. Journal of the Arnold method. American Naturalist 125: 1–15. Arboretum, Supplementary Series 1: 1–169. Felsenstein J. 1988. Phylogenies and quantitative characters. Archie J. 1985. Methods for coding variable morphological Annual Review of Ecology and Systematics 19: 445–471. features for numerical taxonomic analysis. Systematic Ferrero V, Arroyo J, Vargas P, Thompson JD, Navarro Zoology 34: 326–345. L. 2009. Evolutionary transitions of style polymorphisms in Armbruster WS, Mulder CPH, Baldwin BG, Kalisz S, Lithodora (Boraginaceae). Perspectives in Plant Ecology, Wessa B, Nute H. 2002. Comparative analysis of late floral Evolution and Systematics 11: 111–125. development and mating-system evolution in tribe Col- Gabel ML. 1987. A fossil Lithospermum (Boraginaceae) from linsieae (Scrophulariaceae s.l.). American Journal of Botany the tertiary of South Dakota. American Journal of Botany 89: 37–49. 74: 1690–1693. Bleiweiss R. 1998. Tempo and mode of hummingbird Garcia-Cruz J, Sosa V. 2006. Coding quantitative character evolution. Biological Journal of the Linnean Society 65: data for phylogenetic analysis: a comparison of five 63–76. methods. Systematic Botany 31: 302–309.

Gift N, Stevens PF. 1997. Vagaries in the delimitation of Martins EP. 1999. Estimation of ancestral states of continu- character states in quantitative variation – an experimental ous characters: a computer simulation study. Systematic study. Systematic Botany 46: 112–125. Biology 48: 642–650. Gleason HA, Cronquist A. 1991. Manual of vascular plants Martins EP. 2004. COMPARE, Version 4.6b. Computer of Northeastern United States and adjacent Canada, 2nd programs for the statistical analysis of comparative data. edn. New York, NY: New York Botanical Garden Press. Available at: http://compare.bio.indiana.edu Goloboff P, Farris JS, Nixon KC. 2008. TNT, a free Martins EP, Hansen TF. 1997. Phylogenies and the com- program for phylogenetic analysis. Cladistics 24: 774–786. parative methods: a general approach to incorporating phy- Goloboff PA, Mattoni CI, Quinteros AS. 2006. Continuous logenetic information into the analysis of interspecific data. characters analyzed as such. Cladistics 22: 589–601. American Naturalist 194: 646–667. Graham SW, Reeves PA, Burns ACE, Olmstead RG. 2000. Mickevich M, Johnson M. 1976. Congruence between Microstructural changes in noncoding chloroplast DNA: morphological and allozyme data in evolutionary inference Interpretation, evolution, and utility of indels and inver- and character evolution. Systematic Zoology 25: 260– sions in basal angiosperm phylogenetic inference. Interna- 270. tional Journal of Plant Sciences 161: S83–S96. Nixon KC. 2002. WinClada, Version 1.7. Available at: http:// Grant V, Grant KA. 1970. A hummingbird-pollinated species www.cladistics.com/about_winc.htm of Boraginaceae in the Arizona flora. Proceedings of the Ochoterena H. 2009. Homology in coding and non-coding National Academy of Sciences of the United States of DNA sequences: a parsimony perspective. Plant Systematics America 66: 917–919. and Evolution 282: 151–168. Hall TA. 1999. BioEdit: a user-friendly biological sequence Pagel M. 1997. Inferring evolutionary processes from alignment editor and analysis program for Windows 95/98/ phylogenies. Zoologica Scripta 26: 331- 348. NT. Nucleic Acids Symposium Series 41: 95–98. Pagel M. 1999. Inferring the historical patterns of biological Herrera CM. 2009. Multiplicity in unity: plant subindividual evolution. Nature 401: 877–884. variation and interactions with animals. Chicago, IL: Uni- Pagel M, Meade A. 2011. Bayestraits. Available at: http:// versity of Chicago Press. www.evolution.rdg.ac.uk Johnston IM. 1952. Studies in the Boraginaceae, XXIII. A Pimentel R, Riggins R. 1987. The nature of cladistic data. survey of the genus Lithospermum. Journal of the Arnold Cladistics 3: 201–209. Arboretum 33: 299–366. Posada D, Crandall KA. 1998. Modeltest: testing the model Johnston IM. 1954. Studies in the Boraginaceae, XXVI. of DNA substitution. Bioinformatics 14: 817–818. Further revaluations of the genera of the Lithospermeae. Rae TC. 1998. The logical basis for the use of continuous Journal of the Arnold Arboretum 35: 1–81. characters in phylogenetic systematics. Cladistics 14: 221– Jonas CJ, Geber MA. 1999. Variation among populations of 228. Clarkia unguiculata (Onograceae) along altitudinal and lati- Rambaut A, Drummond AJ. 2007. Tracer, Version 1. tudinal gradients. American Journal of Botany 86: 333–343. Available at: http://beast.bio.ed.ac.uk/Tracer Kalisz S, Vogler DW. 2003. Benefits of autonomous selfing Richards AJ. 1997. Plant breeding systems, 2nd edn. London: under unpredictable pollinator environments. Ecology 84: Chapman & Hall. 2928–2942. Samuels ML, Witmer JA. 1999. Statistics for the life Kelley RB. 1993. Lithospermum. In: Hickman JC, ed. The sciences, 2nd edn. Englewood Cliffes, NJ: Prenctice Jepson manual: higher plants of California. Berkeley, CA: Hall. University of California Press, 381–382. Simmons MP, Ochoterena H. 2000. Gaps as characters in Kittelson PM, Handler SD. 2006. Genetic diversity in iso- sequence-based phylogenetic analyses. Systematic Biology lated patches of the tallgrass prairie forb, Lithospermum 49: 369–381. canescens (Boraginaceae). Journal of the Torrey Botanical Skinner MF, Johnson FW. 1984. Tertiary stratigraphy and Society 133: 513–518. the Frick collection of fossil vertebrates from north-central Leins P, Erbar C. 2003. Floral developmental features and Nebraska. Bulletin of the American Museum of Natural molecular data in plant systematics. In: Stuessy TF, Mayer History 178: 215–368. V, Hörandl E, eds. Deep morphology: toward a renaissance Stevens PF. 1991. Character states, morphological variation, of morphology in plant systematics. Ruggell: A. R. G. and phylogenetic analyses: A review. Systematic Botany 16: Ganterner Verlag, 81–106. 553–583. Lowry DB. 2012. Ecotypes and the controversy over stages in Stevens PF. 2000. On characters and character states: do the formation of new species. Biological Journal of the overlapping and non-overlapping variation, morphology, and Linnean Society 6: 241–257. molecules all yield data of the same value? In: Scotland RW, Maddison WP. 1991. Squared-change parsimony reconstruc- Pennington RT, eds. Homology and systematics: coding tions of ancestral states for continuous-valued characters on characters for phylogenetic analysis. London: Taylor and a phylogenetic tree. Systematic Zoology 40: 304–314. Francis Inc, 81–105. Maddison WP, Maddison DR. 2010. Mesquite: a modular Swiderski DL, Zelditch ML, Fink WL. 1998. Why system for evolutionary analysis, Version 2.73. Available at: morphometics is not special: coding quantitative data for http://mesquiteproject.org phylogenetic analysis. Systematic Biology 47: 508–519.

Swofford DL, Maddison WP. 1987. Reconstructing ances- P, eds. Morphology, shape, and phylogenetics. London: tral character states under Wagner parsimony. Mathemati- Taylor & Francis, 247–268. cal Biosciences 87: 199–229. Webster AJ, Purvis A. 2002b. Testing the accuracy of Thomas DC, Weigend M, Hilger HH. 2008. Phylogeny and methods for reconstructing ancestral states of continuous systematics of Lithodora (Boraginaceae – Lithospermeae) characters. Proceedings of the Royal Society of London and its affinities to the monotypic genera Mairetis, Halacsya Series B, Biological Sciences 269: 143–149.

and Paramoltkia based on ITS1 and trnLuaa-sequence data Weigend M, Gottschling M, Selvi F, Hilger HH. 2009. and morphology. Taxon 57: 79–97. Marbleseeds are gromwells – systematics and evolution of Valentine DH, Chater AO. 1972. Boraginaceae. In: Tutin Lithospermum and allies (Boraginaceae tribe Lithosper- TG, Heywood VH, Burges NA, Moore DM, Valentine DH, meae) based on molecular and morphological data. Molecu- Waters SM, Webb DA, eds. Flora Europaea, Vol. 3. Diapen- lar Phylogenetics and Evolution 52: 755–768. siaceae to myoporaceae. London: Cambridge University Weller SG. 1985. The life history of Lithospermum caro- Press, 83–122. liniense, a long-lived herbaceous sand dune species. Ecologi- Weber E, Schmid B. 1998. Latitudinal population differen- cal Monographs 55: 49–67. tiation in two species of Solidago (Boraginaceae) introduced Zhu G, Riedl H, Kamelin RV. 1995. Boraginaceae. In: Wu into Europe. American Journal of Botany 85: 1110–1121. ZY, Raven PH, eds. Flora of China, Vol. 16. Gentianaceae Webster AJ, Purvis A. 2002a. Ancestral states and evolu- through Boraginaceae. Beijing, St Louis: Science Press, and tionary rates of continuous characters. In: Macleod N, Forey Missouri Botanical Garden Press, 329–427.

SUPPORTING INFORMATION Additional Supporting Information may be found in the online version of this article: Table S1. List of species included in analyses, along with collection information and GenBank accession numbers. Table S2. EXCEL table that includes: values reconstructed, at nodes, for corolla tube lengths (A). Node 2 not included because of polytomy at base of tree (Fig. 2) (MP, maximum parsimony; ML, maximum likelihood; BI, Bayesian inference). For discussion of methods and coding schemes, see text. Change in corolla tube length between nodes (B). Negative values are increase in corolla tube length between ancestor and descendant, and positive values are decrease in corolla tube length. Rates of evolution for each branch (C). Negative values are increase in corolla tube length between ancestor and descendant, and positive values are decrease in corolla tube length. Lengths for all branches of phylogenetic tree in Fig. 2 (D). Please note: Wiley-Blackwell are not responsible for the content or functionality of any supporting materials supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the article.