ANGIOSPERM DIVERGENCE TIMES: THE EFFECT OF GENES, CODON POSITIONS, AND TIME CONSTRAINTS

Transcription

1 Evolution, 59(8), 2005, pp ANGIOSPERM DIVERGENCE TIMES: THE EFFECT OF GENES, CODON POSITIONS, AND TIME CONSTRAINTS SUSANA A. MAGALLÓN 1,2 AND MICHAEL J. SANDERSON 3,4 1 Departamento de Botánica, Instituto de Biología, Universidad Nacional Autónoma de México, 3er Circuito de Ciudad Universitaria, A.P , México D.F , Mexico 2 s.magallon@ibiologia.unam.mx 3 Section of Evolution and Ecology, One Shields Avenue, University of California, Davis, California mjsanderson@ucdavis.edu Abstract. An understanding of the evolution of modern terrestrial ecosystems requires an understanding of the dynamics associated with angiosperm evolution, including the timing of their origin and diversification into their extraordinary present-day diversity. Molecular estimates of angiosperm age have varied widely, and many substantially predate the Early Cretaceous fossil appearance of the group. In this study, the effect of different genes, codon positions, and chronological constraints on node ages are examined on divergence time estimates across seed plants, with a special focus on angiosperms. Penalized likelihood was used to estimate divergence times on a phylogenetic hypothesis for seed plants derived from Bayesian analysis, with branch lengths estimated with maximum likelihood. The plastid genes atpb, psaa, psbb, and rbcl were used individually and in combination, using first and second, third, and the three codon positions, including and excluding age constraints on 20 nodes derived from a critical examination of the land-plant fossil record. The optimal level of rate smoothing according to each unconstrained and constrained dataset was obtained with penalized likelihood. Tests for a molecular clock revealed significantly unclocklike rates in all datasets. Addition of fossil constraints resulted in even greater departures from constancy. Consistently with significant deviations from a clock, estimated optimal smoothing values were low, but a strict correlation between rate heterogeneity and optimal smoothing value was not found. Age estimates for nodes across the phylogeny varied, sometimes substantially, with gene and codon position. Nevertheless, estimates based on the four concatenated genes are very similar to the mean of the four individual gene estimates. For any given node, unconstrained age estimates are more variable than constrained estimates and are frequently younger than well-substantiated fossil members of the clade. Constrained estimates of ages of clades are older than unconstrained estimates and oldest fossil representatives, sometimes substantially so. Angiosperm age estimates decreased as rate smoothing increased. Whereas the range of unconstrained angiosperm age estimates spans the fossil age of the clade, the range of constrained estimates is narrower (and older) than the earliest angiosperm fossils. Results unambiguously indicate the relevance of constraints in reducing the variability of ages derived from different partitions of the data and diminishing the effect of the smoothing parameter. Constrained optimizations of divergence times and substitution rates across the phylogeny suggest appreciably different evolutionary dynamics for angiosperms and for gymnosperms. Whereas the gymnosperm crown group originated shortly after the origin of seed plants, a long time elapsed before the origin of crown group angiosperms. Although absolute age estimates of angiosperms and angiosperm clades are older than their earliest fossils, the estimated pace of phylogenetic diversification largely agrees with the rapid appearance of angiosperm lineages in stratigraphic sequences. Key words. parameter. Chronogram, fossils, gymnosperms, penalized likelihood, rate heterogeneity, seed plants, smoothing Angiosperms (flowering plants) are one of the principal components of modern biota and play crucial roles in terrestrial ecosystems. Their relatively late appearance in the fossil record, especially in comparison with other major groups of seed plants, suggests that the group diversified relatively recently. However, it has been argued (e.g., Axelrod 1952, 1970) that angiosperms are much older than what the fossil record suggests, being a cryptic component of ancient floras. An accurate estimate of angiosperm age is crucial to achieve an understanding of their early evolution, that is, ancient origin with cryptic early history versus relatively recent origin and subsequent rapid diversification. It is useful to distinguish between the time of divergence of the angiosperm lineage from its living sister taxon (the stem group age) from the time of the oldest divergence that gave rise to a living angiosperm species (the crown group age; Doyle and Donoghue 1993; Magallón and Sanderson 2001). Recent and, by and large, consistent hypothesis about the phylogeny of living seed plant clades (e.g., Qiu et al. 1999; Chaw et al. 2000; Gugerli et al. 2001; D. Soltis et al. 2002; Magallón 2005 The Society for the Study of Evolution. All rights reserved. Received September 16, Accepted May 18, and Sanderson 2002) suggests that the difference between angiosperm crown group and stem group ages is large. The angiosperm crown group age, hereafter referred to simply as the age of angiosperms, is the focus of this study. The oldest unequivocal angiosperm fossils are dispersed monosulcate pollen grains with a thin, nonlaminated inner wall and a reticulate-columellate outer wall, found in Valanginian to Hauterivian (Early Cretaceous) sediments from geographically distant areas (Brenner 1984, 1996; Trevisan 1988; Brenner and Bickoff 1992; Hughes 1994). Slightly younger sediments (Barremian) contain a considerably greater diversity and abundance of angiosperm pollen (e.g., Doyle 1969, 2000; Doyle and Hickey 1976; Doyle and Hotton 1991; Friis et al. 1994a, 1999) including tricolpate grains, which signal the existence of eudicot angiosperms. Barremian to Aptian sediments contain a wealth of angiosperm reproductive remains that document an extensive taxonomic and structural diversity among which phylogenetically early diverging angiosperm lineages have been identified (e.g., Nymphaeales, Illiciales, Chloranthaceae; Friis et al. 1994a, 1999, 2001; Fru-

2 1654 S. A. MAGALLÓN AND M. J. SANDERSON min and Friis 1999; Gandolfo et al. 2004). The majority of these early angiosperms, while exhibiting affinities with magnoliids and perhaps with monocots, do not belong to existing groups (Friis et al. 1999, 2000, 2005). Younger Early Cretaceous sediments contain reproductive remains of living magnoliid lineages (e.g., Drinnan et al. 1990; Crane et al. 1994; Friis et al. 1994a,b), as well as early diverging eudicots (e.g., Upchurch 1984; Drinnan et al. 1991; Crane et al. 1993). It is in younger Late Cretaceous sediments where the oldest representatives of lineages that originated after a substantial amount of phylogenetic branching can be found. Molecular-clock estimates of angiosperm age provide substantially different perspectives. Some point at dramatically older ages, hundreds of million years older than the earliest angiosperm fossils, implying that angiosperm diversification into modern groups took place even as early as the Lower Carboniferous, million years ago (Ramshaw et al. 1972; Martin et al. 1989). Initial estimates were based on a few molecular data derived from one or a few genes (e.g., Ramshaw et al. 1972; Martin et al. 1989, 1993) or a small taxonomic sample (e.g., Wolfe et al. 1989; Martin et al. 1993; Laroche et al. 1995; Goremykin et al. 1997). In several of these studies, angiosperms were represented by taxa with particularly high rates of molecular substitution (e.g., grasses), which, in the light of later evidence (Sanderson and Doyle 2001), were found to introduce significant biases in divergence time estimation. Moreover, these initial studies relied on the assumption of approximately constant (clocklike) rates of molecular substitution among lineages and were sometimes calibrated with divergence times estimated for distantly related taxa. Recent work has documented the inherent complexity of estimating divergence times and identified numerous stochastic and deterministic sources of error and bias, including, among many others, the stochastic nature of the process of molecular substitution and the presence of different substitution rates in different lineages (Sanderson 1998; Sanderson and Doyle 2001; Arbogast et al. 2002; P. Soltis et al. 2002; Bromham and Penny 2003; Sanderson et al. 2004). Methods and strategies to account for recognized sources of error have been devised. Possibly the most relevant progress consists of the development of analytical methods that account for rate variation among lineages in several different ways (e.g., Sanderson 1997, 2002; Huelsenbeck et al. 2000; Thorne and Kishino 2002; Yoder and Yang 2000; Yang 2004). Available methods aim toward greater realism by including protocols to incorporate absolute temporal information (e.g., Sanderson 1997, 2002; Kishino et al. 2001) or taking advantage of shared divergence times among different molecular partitions (e.g., different genes; Thorne and Kishino 2002; Yang 2004). Several recent studies that investigated angiosperm age implemented some of these novel methodological approaches. Sanderson (1997) developed a nonparametric method (nonparametric rate smoothing, NPRS) that estimates rates and times by smoothing rate variation among neighboring branches on a tree, inducing an autocorrelation in rates. The concept of temporal autocorrelation is consistent with the idea that rates of substitution among closely related lineages are likely to be similar (Gillespie 1991). Temporal constraints on ages of nodes can be straightforwardly incorporated into NPRS estimation of rates and times (Sanderson 1997, 2003). Sanderson (1997) used nucleotide sequences of the plastid gene rbcl for a sample of 36 land-plant taxa to compare ages estimated according to a molecular clock and NPRS and to evaluate the effect of including fossil-derived constraints on the ages of nodes. The NPRS estimate of angiosperm age was approximately 175 million years when constraints were not included, and as old as 200 million years when constraints were added. Wikström et al. (2001) used NPRS to estimate the age of angiosperms and angiosperm clades using data (plastid rbcl and atpb and 18S nrdna for 567 species) and a phylogenetic hypothesis (one of 8000 most parsimonious trees) obtained by Soltis et al. (1999). The phylogeny was calibrated by fixing the age of the divergence of Fagales and Cucurbitaceae with the fossil record of Fagales. Resulting estimates were compared with first fossil occurrences, but fossil ages were not used to constrain nodes. The age of angiosperms was estimated between 158 and 179 million years. P. Soltis et al. (2002) used NPRS to evaluate the effect of different genes, different branch lengths, and alternative calibration points in estimating divergence times for lineages of tracheophytes (vascular plants). The study was based on the data (plastid rbcl, atpb, and rps4 and 18S nrdna sequences for 35 taxa across land plants) and phylogenetic tree obtained by Pryer et al. (2001). Angiosperms were represented by only two terminal taxa belonging to the magnoliid grade. Well-substantiated fossil information was used to calibrate the phylogeny at alternative nodes; however, constraints on ages of nodes were not used. The majority of discussed age estimates (P. Soltis et al. 2002, tables 4, 8) are derived from trees calibrated at the angiosperm node with its age fixed at 125 million years. The results documented substantial differences for node ages according to different genes, different estimates of branch lengths, and, in some cases, alternative calibration points. When nodes other than angiosperms were used as calibration points, in at least two cases (calibration at lycophytes and at gymnosperms), the NPRS age of angiosperms agreed closely with the fossil age of the group (P. Soltis et al. 2002, p. 4435). Thorne et al. (1998; see also Kishino et al. 2001; Thorne and Kishino 2002) developed an explicit model of rate variation implemented in a Bayesian framework. The initial model was tested with the same vascular plant data used by Sanderson (1997), except for using amino acid instead of nucleotide sequences (Thorne et al. 1998). Because independent absolute temporal information was not used, divergence times were measured as relative time units. The age of angiosperms was estimated as about one-half of the age of the divergence between lycophytes and euphyllophytes (vascular plants excluding lycophytes), which, according to the fossil record, took place at least by the Upper Silurian, 420 million years ago (Kenrick and Crane 1997). The age of angiosperms would thus correspond to approximately 210 million years. Sanderson (2002) developed a semiparametric penalized likelihood (PL) approach, which combines a parameter-rich model that optimizes rates of substitution on branches with a numerical penalty that minimizes rapid rate changes. A smoothing parameter regulates the relative contribution of the parametric model and the numerical penalty into esti-

3 ANGIOSPERM DIVERGENCE TIMES 1655 mating rates and divergence times. Depending on its magnitude, a smaller or greater level of rate heterogeneity is allowed. Selection of the optimal smoothing magnitude according to the degree of rate heterogeneity contained in the data is guided by a data-driven cross validation procedure. As implemented in the program r8s (Sanderson 2003), penalized likelihood permits temporal constraints on the ages of nodes during estimation of rates and divergence times. Schneider et al. (2004) compared penalized likelihood-estimated timing of radiation of polypodiaceous ferns and angiosperms. The timing of angiosperm radiation was based on a dataset of rbcl, atpb, and 18S sequences for 95 taxa, modified from Soltis et al. (1999) and Soltis et al. (2000). A Bayesian phylogenetic hypothesis was calibrated at the tracheophyte node, and 30 nodes were constrained with fossilderived minimum ages. The angiosperm crown node received two different treatments: its age was fixed at 132 million years, the oldest fossil occurrence of the group, or the fossil age was used as a minimum age constraint. In the first case, estimated ages of internal angiosperm nodes were approximately the same as the corresponding fossil ages. In the second case, estimates were much older than fossil first occurrences including the age of angiosperms, which was calculated as approximately 250 million years. In this study, we estimate the age of nodes across a phylogeny of seed plants, with a special focus on the angiosperms. We evaluate the effect of different genes, gene combinations, and codon position partitions, as well as the implementation of chronological constraints on nodes. Particular care was taken to correct for factors known to cause biases in age estimation, when these can be accounted for. The taxon sample consists of critically selected placeholders for clades across tracheophytes and was particularly designed to investigate the age of angiosperms. Selection criteria included the choice of taxa with different habits (woody and herbaceous) and generation times (annuals and perennials) to represent possible different substitution regimes, as well as representatives that span the root node of clades. The data consist of the nucleotide sequences of four highly conserved protein-coding plastid genes that exhibit rates of molecular substitution appropriate to investigate phylogenetic relationships and divergence times at deep levels in the plant phylogeny. Recent studies have revealed the complexity involved in resolving relationships among major clades of spermatophytes (seed plants; e.g., Chaw et al. 2000; Magallón and Sanderson 2002; Rydin et al. 2002; D. Soltis et al. 2002). We use an estimate of phylogeny congruent with the most recurrent results in these analyses. Maximum likelihood was used to estimate branch lengths according to different genes and codon positions. To estimate divergence times, we used penalized likelihood implementing optimal smoothing magnitudes and incorporating chronological constraints on nodes. The temporal calibration of the topology and the chronological constraints on nodes are based on a critical assessment of the fossil record across land plants. This study has three specific aims: (1) to evaluate the effect of different genes and gene combinations, codon positions, and the use of age constraints on the degree of rate heterogeneity found in the data, on the magnitude of the smoothing parameter estimated through penalized likelihood cross validation, and on the estimated age of angiosperms; (2) to evaluate the effect of the magnitude of the smoothing parameter on estimates of angiosperm age; and (3) to obtain an interval of ages for strongly supported nodes in the phylogeny derived from point estimates from different genes, compare it with point estimates derived from gene combinations, and evaluate the effect of age constraints on the size of intervals. The results not only provide information about the effect of different factors in estimating ages with penalized likelihood, but also provide estimates of the age of major seed plant clades, as well as a glimpse of evolutionary dynamics across seed plants. MATERIALS AND METHODS Phylogeny Reconstruction Taxonomic sample The taxonomic sample consists of representatives of the lineages of land plants, with a special focus on seed plants, and was targeted specially to investigate the age of angiosperms. Each seed plant clade is represented by taxa that span the root node of well-supported monophyletic groups recognized in previous studies (e.g., Chaw et al. 1997; Pryer et al. 2001; D. Soltis et al. 2000, 2002; Magallón and Sanderson 2002; Rydin et al. 2002). The taxonomic sample includes 63 terminal taxa representing angiosperms (31 terminal taxa represented by 32 genera), conifers (15 species in 13 genera), gnetophytes (three species), Ginkgo (one species), cycads (three species), ferns (seven species), lycophytes (one species), bryophytes (one species), and charophytes (one species). Chara, the placeholder for chlorophytes, was the designated outgroup in phylogeny estimation. A list of the genera and species in the taxonomic sample is provided in Supplementary Table 1 (available online only at dx.doi.org/ /04-565/1.s1). Genes and molecular methods The nucleotide sequences of four highly conserved proteincoding plastid genes were used as primary data for phylogeny estimation. The sequences of atpb and rbcl were added to the psaa and psbb dataset used in the seed plant phylogeny study of Magallón and Sanderson (2002). Whereas rbcl and atpb have been used extensively in plant phylogeny estimation (e.g., Chase et al. 1993; Lewis et al. 1997; Savolainen et al. 2000; D. Soltis et al. 2002), the use of psaa and psbb has been limited (e.g., Graham and Olmstead 2000; Sanderson et al. 2000). The four genes exhibit a high percentage of amino-acid similarity and a relatively low nucleotide substitution rate, with psbb being the most conserved of the four, followed by psaa, and rbcl and atpb exhibiting very similar measures of sequence conservation (Olmstead and Palmer 1994). Primers and protocols for amplification of psaa and psbb are provided in Magallón and Sanderson (2002). Most atpb and rbcl sequences were obtained from GenBank. Sequences of the same species as in the psaa and psbb dataset were preferred, but when unavailable, the sequence of a different species within the same genus was used. When sequences for a genus were unavailable in GenBank, we obtained these

4 1656 S. A. MAGALLÓN AND M. J. SANDERSON sequences using the same species as in the psaa and psbb dataset, for a total of 19 new atpb sequences (four angiosperms, 12 conifers, two gnetophytes, one cycad), and two new rbcl sequences (one angiosperm, one conifer). In only one instance was a taxon represented by two different genera (Capparaceae: Isomeris arborea for psaa and psbb, and Cleome hassleriana for atpb and rbcl). The GenBank/EMBL accession numbers of included sequences are provided in Supplementary Table 1 (available online). New atpb and rbcl sequences were obtained from the same DNAs or from the same biological individuals represented in the psaa and psbb study (Magallón and Sanderson 2002). Gene amplification was achieved through polymerase chain reaction (PCR), using the thermocycler cycling parameters described by Hoot et al. (1995). PCR primers were rbcl1f, rbcl1460r, atpb2f, the last two modified from Savolainen et al. (2000), and the newly designed atpb1450r. The use of atpb2f did not result in successful PCR amplification for some conifers, and was therefore substituted with rbcl1r (corresponding to the reverse and complement of rbcl1f), which, in combination with atpb1450r, amplified atpb plus the atpb rbcl intergenic spacer. This method was not universally successful among conifers, and in some cases, only a partial atpb sequence was obtained. It was not possible to amplify atpb for Pinus parviflora and Encephalartos lebomboensis. Sequences of PCR and sequencing primers are shown in Supplementary Table 2 (available online only at dx.doi.org/ / s1). Purified PCR products were sequenced directly by automated fluorescent dye methods on an ABI model 377 sequencer (Applied Biosystems, Inc., Foster City, CA) at the University of California, Davis, Division of Biological Sciences Sequencing Facility. For each taxon, partial sequences obtained from reactions using different primers were assembled and edited using Sequencher 4.0 (GeneCodes, Ann Arbor, MI). The high conservation of rbcl and atpb permitted manual sequence alignment. Parsimony analyses Datasets for each gene (assigning equal weight to all codon positions) were used to perform separate parsimony analyses. Each parsimony analysis consisted of a heuristic search with 100 random addition of sequences and TBR branch swapping, saving all minimal trees found in each swap. A nonparametric bootstrap, consisting of 100 bootstrap replicates with full heuristic searches equal to those in the phylogeny search, except for including only 10 random sequence additions, provided information about clade support. Analyses were conducted using PAUP* 4.0b10 (Swofford 2002). The most parsimonious trees resulting from each analysis were summarized as strict consensus trees, including bootstrap support (results available upon request). A visual examination of the strict consensus trees derived from individual gene datasets revealed no strongly supported differences ( 85%). We therefore concatenated the sequences to create a combined four-gene dataset, which was used for phylogeny reconstruction with Bayesian inference. Bayesian analyses Bayesian estimation of phylogeny was conducted with MrBayes 3.0B4 (Huelsenbeck 2000; Huelsenbeck and Ronquist 2001). To examine Markov chain Monte Carlo (MCMC) convergence and the effect of a slightly suboptimal model, Bayesian inference of phylogeny was performed in two steps. In the first step, a GTR G model was applied uniformly to the four-gene dataset. A 250,000 generation MCMC was started from a random tree with one cold and three incrementally heated chains, sampling every 100th tree. The negative natural logarithm of the likelihood ( ln L) score of sampled trees was plotted against their generation number to determine the approximate generation number at which the MCMC chain achieved stationarity. After checking that they corresponded to the stable phase of the chain (see results), 2000 sampled trees corresponding to generations 50,001 to 250,000 were used to construct a 50% majority rule consensus tree in which the percentage of times a clade is represented among the sampled trees is proportional to its posterior probability (PP). The second step of Bayesian phylogeny inference consisted of a search that differed from the first one by using a GTR I G model, the one identified by the Akaike information criterion in Modeltest (Posada and Crandall 1998) as the one with best fit to the data, which was applied uniformly to the four-gene dataset, and by implementing a generation MCMC. As in the previous step, every 100th tree was sampled, and the ln L score of the 10,000 sampled trees was plotted against generation number. Sampled trees corresponding to generations 200,001 to 1,000,000 (8000 trees), which correspond to the stable phase of the chain, were used to construct a 50% majority-rule consensus. The majority-rule consensus trees obtained from the two Bayesian searches were compared to detect differences in the topology, and in the PPs associated with clades. A single topology selected at random among the 2000 retained trees from the first Bayesian search was used as a working phylogenetic hypothesis to estimate divergence times (Fig. 1). Obtaining phylograms Divergence Time Estimation Length of branches on the selected topology were estimated according to each of four independent genes and two gene combinations (the two photosystem genes [PS] and the four genes [FOUR]), using first and second, third, and all codon positions. Chara was excluded from branch length estimation because the sequences of two genes were unavailable (Supplementary Table 1, available online). The atpb sequences of Encephalartos and P. parviflora were also missing, therefore, to avoid unreliable estimates due to missing data, both species were pruned from the topology before branch length estimation with the atpb and FOUR datasets. In all cases, branch lengths were estimated with maximum likelihood with GTR G, using PAUP* for Macintosh. Zerolength terminal branches were removed from phylograms using the PRUNE command in r8s (Sanderson 2003). Although not always necessary, this limits occasional degenerate solutions that can arise in the numerical optimizations when attempting to smooth rate variation in a region of the tree in which both rates and time durations are zero.

5 ANGIOSPERM DIVERGENCE TIMES 1657 FIG. 1. Phylogenetic hypothesis. This topology was obtained from Bayesian analysis of the concatenated sequences of atpb, psaa, psbb, and rbcl, including all codon positions. The tree was selected at random from the sampled trees of the Markov chain Monte Carlo. The tree has a ln L of , a rate matrix of A-C 1.80, A-G 5.70, A-T 0.70, C-G 1.01, C-T 7.57, and shape of the gamma distribution to explain among-site rate variation Numbers above branches represent posterior probabilities (PP). Numbered nodes are supported by 0.95 PP, and correspond to numbers in the Appendix (available online). Numbers inside circles correspond to node number; numbers inside rectangles are the node number, and the age in million years used to constrain the node. Constraints were implemented as minimum ages except for node 1 (tracheophytes), the calibration node, whose age was fixed, and node 44 (eudicots), constrained with a maximal age.

6 1658 S. A. MAGALLÓN AND M. J. SANDERSON Calibration and constraints The age assigned to a calibration point has a determinant influence on the ages estimated across the tree (Sanderson and Doyle 2001; P. Soltis et al. 2002; Near et al. 2005). It has been empirically determined that using different calibration points has a major impact on ages estimated across the tree (P. Soltis et al. 2002). In this study, the tracheophyte crown group node (eutracheophyte node, Kenrick and Crane 1997), corresponding to the divergence of lycophytes (Huperzia) and euphyllophytes (all other vascular plants) was selected as the calibration point, and fixed at an absolute age of 419 million years. The eutracheophyte node was selected as calibration point because its members uniquely share a trait that is decay resistant and abundantly produced (tracheids with thick, decay-resistant walls and pitlets between thickenings or within pits; Kenrick and Crane 1997), allowing for unequivocal identification and the possibility of a small time lapse between the origin of the trait and its preservation in the fossil record (Magallón 2004). The calibration age was derived from the stratigraphic position of the oldest fossils that belong to either of the two branches derived from the eutracheophyte node. The oldest unequivocal representatives of the lycophyte branch are Zosterophyllophya and Baragwanathia, both from the Upper Silurian (Ludlow, ca. 419 million years; Tims and Chambers 1984; Garrat and Rickards 1987; Hueber 1992; Kenrick and Crane 1997). The oldest unequivocal representatives of the euphyllophyte branch are younger, from the Lower Devonian (discussion in Supplementary Table 3, available online only at / s1). To the best of our knowledge, the age assigned to the eutracheophyte node represents one of the most reliable fossil-derived ages that can be assigned across tracheophytes. An alternative calibration point is the crown eudicot node, within angiosperms, given the possibility of unequivocal identification of its members in the fossil record by an abundantly produced and decay-resistant trait that is unique to the clade (i.e., tricolpate pollen). Little is known about the effect of the position of the calibration point in the phylogeny, however, the eudicot node was not chosen as calibration point because it is separated by many branching events from the root of the tree and because of its relative closeness to the angiosperm node. The ages of 20 additional nodes were constrained with well-substantiated fossil ages. Nineteen nodes were constrained with minimal ages (i.e., estimated ages are not allowed to be younger than the constraint), and one, crown group eudicots, with a maximal age (i.e., estimated ages are not allowed to be older than the constraint). Fossils used to assign dates to calibration and constraint nodes are discussed in Supplementary Table 3 (available online). Fossil stratigraphic occurrences were transformed into absolute ages by using the midpoint between the boundaries of the narrowest formally recognized stratigraphic interval to which each belongs. Absolute ages of stratigraphic boundaries were obtained from Haq and van Eysinga (1998). Divergence times were estimated without applying (unconstrained analyses; F ), and applying constraints on the ages of nodes derived from the fossil record (constrained analyses; F ). Estimating divergence times Each unconstrained and constrained phylogram was tested for rate constancy using the 2 test of Langley and Fitch (1974; Sanderson 1998) implemented in r8s v1.60 for Unix. Divergence times were estimated using penalized likelihood (PL; Sanderson 2002). In PL, a smoothing parameter ( ) regulates the relative contribution of a fully parametric model that optimizes rates on branches with a numerical penalty against pronounced rate change among branches. Greater accuracy is expected by assigning a value to that correctly reflects the level of rate heterogeneity in the data. The optimal magnitude of can be estimated through a cross-validation procedure consisting of pruning terminal branches of the phylogram, estimating model parameters with PL (including ) on the remaining branches, and predicting the length of the pruned branches. The cross-validation procedure can be performed on unconstrained or constrained phylograms. The cross-validation procedure and divergence time estimations were conducted with r8s v1.60 (Sanderson 2003) on all unconstrained and constrained phylograms, for a total of 36 divergence time reconstructions. Cross-validations were performed by obtaining the treewide prediction error associated with smoothing values in an interval from log to log , in increments of opt is the associated with the lowest 2 cross-validation score, reflecting the smallest estimated error between observed and predicted parameters. Divergence times were estimated with PL using a TN algorithm with bound constraints, implementing the opt estimated in the previous step. To prevent finishing a search at a local optimum, each divergence time estimation consisted of five initial starts with three perturbed restarts, with perturbations of magnitude 0.05 in random directions. Estimating the length of branches derived from the root node of a tree is problematic because of the uncertain placement of the root along the branch linking the outgroup from the ingroup. An empirical solution is to exclude from divergence time reconstruction the outgroup used during branch length estimation this way, the root node is shifted to the next node within the original phylogram and the lengths of the branches derived from it are accurate (Sanderson 2002). In this study, Marchantia, the sister to the remaining taxonomic sample during branch length estimation, was excluded from the molecular-clock tests, from opt estimations, and divergence time reconstructions. The divergence between Huperzia and all other vascular plants (i.e., the tracheophyte node) became the new root of the tree. This node is also the calibration point. Magnitude of smoothing parameter ( ) versus age The effect of the magnitude of on estimated ages was evaluated by screening point estimates of the age of angiosperms across a range of magnitudes. This effect was evaluated on unconstrained and constrained phylograms of the four individual genes and the two gene combinations, using the three codon positions. In each case, ages were reconstructed according to magnitudes ranging from log to 7.5, at increments of 0.5. Ages were estimated using the TN algorithm in r8s. Each search consisted of three initial

7 ANGIOSPERM DIVERGENCE TIMES 1659 starts with three perturbed restarts and perturbations of magnitude 0.05 in random directions. RESULTS AND DISCUSSION Phylogeny Reconstruction Data and phylogenetic analysis The dataset used to estimate phylogeny includes 6387 bp. Insertions and deletions are absent, except for a single 3-bp insertion near the 5 end of the psaa sequence of Zea, Ephedra, and Welwitschia (its presence in Gnetum is unknown, because the corresponding gene fragment could not be sequenced). Visual sequence alignments were unproblematic. The database is available in TreeBASE or from the authors. The trees progressively sampled in the two independent Bayesian searches (one using a GTR G model for 250,000 MCMC generations and the other using a GTR I G model for generations) stabilized around constant ln L scores by generation 30,000. In the first search, sampled trees corresponding to generations 1 50,000 (500 trees) were excluded as the burn-in phase of the chain, and the remaining 2000 sampled trees, which correspond to the stable phase of the chain, were used to construct a 50% majorityrule consensus tree. In the second search, sampled trees corresponding to generations 1 200,000 (2000 trees) were discarded, and the remaining 8000 trees, well inside the stable phase of the chain, were summarized as a 50% majority-rule consensus. The consensus trees derived from the two searches are equal, and the PP values associated with clades in the two trees are very similar. Therefore, in this case, the independent Bayesian searches, which started from different random trees and used slightly different models and different MCMC lengths, converged at equal consensus topologies and very similar PPs associated with clades. A randomly selected topology of the 2000 trees sampled in the first Bayesian search was used as a working phylogenetic hypothesis to estimate divergence times (Fig. 1). Seed plant relationships Relationships among and within seed plant clades found in this study are congruent with independent analyses (e.g., Qiu et al. 1999; Bowe et al. 2000; Chaw et al. 2000; Nickrent et al. 2000; Sanderson et al. 2000; D. Soltis et al. 2000, 2002; Gugerli et al. 2001; Rydin et al. 2002). Our results support the monophyly of embryophytes (land plants), tracheophytes, euphyllophytes, spermatophytes, and angiosperms. Within tracheophytes, the monilophytes, a clade that includes eusporangiate and leptosporangiate ferns, Equisetum, and Psilotum, are the sister taxon to spermatophytes, in agreement with previous results (Pryer et al. 2001). The deepest phylogenetic split within spermatophytes separates living gymnosperms and angiosperms. Cycads are the sister to all other gymnosperms, and Ginkgo is the sister to the expanded conifer clade, which includes the gnetophytes. As in other studies (e.g., Qiu et al. 1999; Bowe et al. 2000; Chaw et al. 2000; Nickrent et al. 2000; Gugerli et al. 2001; Magallón and Sanderson 2002; D. Soltis et al. 2002), the gnetophytes appear as sister to Pinaceae, and together, they constitute the sister to a clade that includes all other conifers. Relationships within Pinaceae conflict with those found by Wang et al. (2000). Relationships among angiosperms are congruent with strongly supported results in independent analyses (e.g., Qiu et al. 1999; D. Soltis et al. 2000; Zanis et al. 2003). Amborella, Nymphaeales, and Austrobaileyales are the basal grade in angiosperms. The monocots are the sister to a weakly supported clade that includes the remaining angiosperms. Within this clade, a eumagnoliid clade is the sister to Chloranthaceae plus eudicots, but this relationship is also weak. Ranunculids are the sister group to all other eudicots, and a core eudicot clade is strongly supported (Fig. 1). As previously recognized, resolving phylogenetic relationships among major clades of seed plants has proven to be a complex problem, due in part to the substantial proportion of extinct diversity that cannot be accounted for in molecular studies and to the difficulty of accurately judging character homology in morphological analyses. Although a full resolution of relationships among major seed plant clades is pending, in this study we use a phylogenetic working hypothesis that reflects, to the best of our knowledge, the most recurrent relationships obtained from explicit analyses of molecular characters. Phylograms The likelihood scores, nucleotide substitution rate matrix, and shape of the gamma distribution estimated for each phylogram according to different genes, gene combinations and codon position partitions are shown in Supplementary Table 4 (available online only at A visual examination of phylograms (available upon request) indicates that long branches consistently subtend Huperzia, all monilophyte terminal nodes, the gnetophyte crown node, the three gnetophyte terminal nodes, the angiosperm crown node, and within angiosperms, the node corresponding to Poaceae, the grass family. In several phylograms, the branch subtending Pinaceae is also long. Short branches consistently subtend the gymnosperm crown node, and in many phylograms, the branch subtending euphyllophytes and many internal conifer branches. Many internal angiosperm branches are also consistently short, in particular, the branches immediately above the angiosperm crown node (discussed below). Internal and terminal zero-length branches are more frequent in phylograms estimated using first and second codon positions, as expected, given their overall slower substitution rate. Also as expected, increasing the amount of data by combining two or more genes decreased the number of zero-length branches. Zero-length terminal branches were removed from each phylogram before testing the molecular clock and estimating divergence times (Supplementary Table 5, available online only at Effect of Genes, Codon Positions, and Constraints on Rate Heterogeneity, Optimal Smoothing, and Age of Angiosperms and Seed Plants Rate heterogeneity The Langley and Fitch (LF) test applied to each unconstrained and constrained phylogram (after excluding Mar-

8 1660 S. A. MAGALLÓN AND M. J. SANDERSON chantia) in all cases indicated significant departures from rate constancy (Table 1). The 2 values can be used as a proxy for departure from a clock model that is more or less comparable across partitions because the degrees of freedom are about the same (Table 1). Although all tests indicated significantly unclocklike rates, the least unclocklike datasets are those for first and second positions. The more clocklike behavior of first and second codon positions with respect to third positions has been observed in other datasets and results from the greater influence of variations in rates among lineages on synonymous substitutions than on nonsynonymous substitutions (Gillespie 1991). Comparisons among genes indicate that psbb substitution rates are the least unclocklike. Combined datasets with two or more genes (PS and FOUR datasets) are less clocklike than individual datasets. Finally, fossil constraints increased rate heterogeneity. Magnitude of optimal smoothing The range of smoothing values used in cross validations to estimate the optimal magnitude of the smoothing parameter for each unconstrained and constrained phylogram in all cases resulted in a range of initially decreasing and subsequently increasing treewide prediction errors, thus allowing unambiguous identification of the smoothing magnitude ( opt ) associated with lowest treewide prediction error. Selected optimal smoothing values are shown in Table 1. The magnitude of the smoothing parameter is related to the degree of rate heterogeneity in the data, thus, it would be expected that the magnitude of optimal smoothing for phylograms with low LF molecular clock test P-values (see above) was also small and vice versa. In our data, phylograms with the lowest LF P-values received low optimal smoothing scores, however, a strict correlation between the two values is not found (Table 1). The rbcl and psbb phylograms had the highest optimal smoothing magnitudes, as indicated by gene-to-gene comparisons in which codon positions and inclusion or exclusion of constraints were constant. Phylograms based on combined genes (PS and FOUR) had lower optimal smoothing magnitudes than those based on the corresponding individual genes. Consistently with being the least unclocklike, phylograms based on first and second positions usually obtained higher optimal smoothing values than phylograms based on third or all positions. Very often, the optimal smoothing magnitude for third and all codon position phylograms is the same (Table 1). A surprising result emerged from the comparison of optimal smoothing magnitudes on phylograms that differ only in the exclusion/inclusion of constraints on ages of nodes. Optimal smoothing magnitudes of constrained phylograms are higher than those of their unconstrained counterparts (except in one case in which optimal smoothing values were equal), a result inconsistent with the greater rate heterogeneity found in constrained phylograms (LF test; Table 1). An initial interpretation would suggest that constrained datasets are more clocklike than their unconstrained counterparts, in other words, that adding constraints from the fossil record results in more clocklike rates. This interpretation is contrary to the results of the LF molecular clock test, which indicate TABLE 1. Effect of genes, codon positions, and constraints. The effect of using different genes (atpb, psaa, psbb, rbcl) or gene combinations (PS, FOUR), different codon positions (first and second, third, all), and including (F ) or excluding (F ) fossil-derived age constraints on nodes is evaluated on measures of rate heterogeneity ( 2, P), magnitude of optimal smoothing parameter ( opt ), and estimated ages for seed plants and angiosperms (in million years). Constrained (F ) Unconstrained (F ) Age Age Spermatophytes Angiosperms Angiosperms 2 df P opt Spermatophytes 2 df P opt atpb atpb atpb psaa psaa psaa psbb psbb psbb rbcl rbcl rbcl PS PS PS FOUR FOUR , FOUR ,

9 ANGIOSPERM DIVERGENCE TIMES 1661 TABLE 2. Cross-validation scores obtained with the molecular clock and penalized likelihood. All cross-validations were performed on constrained phylograms. The cross-validation using the molecular clock was implemented through the Langley and Fitch (LF) method, which produces a single value for the chi-squared error. The cross-validation using penalized likelihood (PL) produced a series of estimated chi-squared errors, corresponding to the indicated range of values of (see Materials and Methods). The smallest PL chi-squared error and its associated are shown. This is the opt chosen to conduct age and rate estimations. In all cases, PL chi-squared errors are smaller than LF errors. Dataset LF 2 error Minimal 2 error atpb123 F psaa123 F psbb123 F rbcl123 F PS123 F FOUR123 F greater departures from constancy in constrained phylograms (expressed as higher 2 and lower P-values) than in their unconstrained counterparts (Table 1). An alternative interpretation is that implementing the constraints in our phylograms forces such strong departures from rate constancy, that the PL optimization finds it easier to predict pruned branch lengths during the cross validation procedure by using more clocklike rates. To check if better estimates were obtained by imposing a molecular clock on cross validations, we compared the 2 error obtained by imposing a molecular clock on the constrained phylograms (for each gene and gene combination, using the three codon positions) with the observed minimal 2 obtained with PL. Cross-validations imposing the molecular clock were performed using the LF method with a TN algorithm in r8s, and the resulting 2 errors are shown in Table 2. In all comparisons, the minimal 2 errors obtained with PL are smaller than the 2 error obtained when the molecular clock was imposed, indicating that, according to an objective criterion of predictive accuracy, PL cross-validations provide better estimates of the optimal smoothing parameter for each phylogram than cross-validations with the molecular clock. Estimates of the age of angiosperms and seed plants Point estimates of the age of angiosperms are quite variable depending on the gene, codon position, and the implementation of constraints (Table 1). The gene that most frequently provided the youngest estimates for the angiosperm node is psaa (five of six comparisons; Table 1), with the youngest age being million years (psaa123f ). The atpb gene most frequently provided the oldest estimates (four of six comparisons; Table 1), however the oldest and second oldest estimates resulted from rbcl12f and rbcl12f ( and million years, respectively). Point estimates derived from combined genes (PS and FOUR) usually fall within the interval of the ages derived from the corresponding individual genes. The exceptions are the ages derived from unconstrained combined photosystem genes (i.e., PS12F, PL opt PS3F and PS123F ), which are younger than independent psaa and psbb ages (Table 1). The results do not point clearly at a particular codon position partition as most frequently providing the oldest or the youngest estimates. The oldest point estimates of angiosperm age were obtained from first and second position phylograms (rbcl12f and rbcl12f ; ages shown above). Age estimates derived from the three codon positions are usually intermediate between first and second and third position estimates (Table 1). The use of first and second positions resulted in the greatest difference between the youngest and oldest estimated ages (differences of and million years, from unconstrained and constrained phylograms, respectively), whereas the narrowest difference was obtained from the three codon positions without imposing constraints (30.49 million years). Use of constraints on ages of nodes usually results in older angiosperm ages (16 of 18 comparisons; Table 1). Angiosperm ages estimated without constraints range from to million years, with only one estimate (PS12F ) being younger than 132 million years, the age of the oldest unequivocal angiosperm fossils, thus being inconsistent with the fossil record. Constrained angiosperm age estimates range from to million years. Whereas constrained estimates are older than the unconstrained ones, the difference between youngest and oldest ages derived from constrained phylograms is narrower than the difference in unconstrained phylograms ( and million years, respectively; Table 1). The interval of unconstrained angiosperm ages brackets the age of the earliest angiosperm fossils, but constrained age estimates are at least 44 million years older. The interval of estimates for the age of angiosperms is very large, possibly as a result of the distant placement of nodes that can effectively constrain their age the closest is the euphyllophyte node, which is constrained with a very ancient minimum age. One way to reduce the variability in age would be to use the ages of angiosperm stem lineage representatives to impose maximum bounds to the age of the angiosperm crown group. Unfortunately, several practical complications arise, first and foremost, the fact that a clear understanding of phylogenetic relationships among living and extinct seed plant clades, which take into account the consistent fraction of molecular-based phylogenetic results, is currently lacking (although some relevant efforts have been undertaken, e.g., Doyle 2001). Also lacking is an unambiguous identification of the extinct sister group of angiosperms and of other angiosperm stem lineage representatives. The age of seed plants (the spermatophyte crown node) was used as an additional indicator of the effect of genes, codon positions, and use of constraints. These estimates should be considered less conclusive than those for angiosperms, because taxonomic sampling was not designed to address the seed plants in particular. Point estimates of seed plant age vary according to different genes, codon positions, and the use of constraints, however, variability is not as large as for angiosperms (Table 1), possibly as a result of the closeness of the former to constrained nodes (i.e., euphyllophytes and gymnosperms) and to the calibration point (Fig. 1). As for the angiosperms, atpb provided the oldest age estimates (six of six comparisons), with the oldest estimate