Introduction

Much of extant biodiversity may have arisen from a small number of adaptive radiations occurring on large spatiotemporal scales [1]–[3]. Under the niche-filling model of adaptive radiation, ecological opportunities arise from key innovations, the extinction of competitors, or geographic dispersal [1],[4],[5]. These cause rapid evolutionary rates in ecologically relevant traits, as diverging lineages exploit distinct resources. Rates of trait evolution then decelerate as niches become saturated, a pattern that has been formalised as the “early burst” model (e.g., [6],[7]).

Most phylogenetic studies of adaptive radiations focus on small scales such as island radiations and other recently diverging clades, including Anolis lizards, cichlid fishes, and geospizine finches [2],[6],[8]–[10]. Detailed study of these model systems has demonstrated the importance of ecological and functional divergence as drivers of speciation early in adaptive radiations (e.g., [11],[12]). Surprisingly though, early burst patterns of trait evolution receive only limited support from model comparison approaches for these and other adaptive radiations occurring in geographically restricted areas and on short timescales (<50 million years [Ma]; most <10 Ma) [6] (but see [13],[14]).

Studies of morphological evolution on longer timescales, unfolding over 100 Ma or more, are central to establishing whether niche-filling or early burst patterns of trait evolution are important evolutionary phenomena on large phylogenetic scales. A small number of recent studies quantified patterns of trait evolution on large scales using neontological phylogenies. For example, diversification rates and morphological rates are positively correlated in actinopterygians [15] (∼400 Ma); rapid rates of both morphological and molecular evolution occur on deep, Cambrian, nodes of the arthropod tree of life [16] (∼540 Ma); and the early evolution of placental mammals was characterised by rapid rates of diversification [17] (100–65 Ma) and perhaps body size evolution [18] (but see [19]).

However, even the largest neontological studies [15]–[18],[20],[21] are limited to explaining the rise of important extant groups. A more complete characterisation of macroevolutionary processes on long timescales should also explain the ascent and demise of important extinct groups (e.g., [22]), which in fact represent most of life's diversity. Substantial evidence for the dynamics of past adaptive radiations might have been erased from the neontological archive, and macroevolutionary models for extinct or declining/depauperate clades may be tested most effectively using deep time data from the fossil record [23],[24].

Palaeontologists often quantify patterns of morphological radiation using time series of disparity (e.g., [25],[26]). However, few phylogenetic studies including fossil data have attempted to explain patterns of morphological radiation in large clades on timescales >100 Ma, and most have individually targeted either the roots of exceptional modern clades such as birds or mammals (e.g., [19],[27],[28]) or extinct/depauperate clades (e.g., [29]–[31]; studies based on discrete characters). Thus, patterns of morphological evolution in major extinct clades, and their links to successful modern clades, are not well understood.

Non-avian dinosaurs are an iconic group of terrestrial animals. They were abundant and ecologically diverse for most of the Mesozoic, and included extremely large-bodied taxa that challenge our understanding of size limits in terrestrial animals [32]. The first dinosaurs appeared more than 230 Ma ago in the Triassic Period, as small-bodied (10–60 kg), bipedal, generalists. By the Early Jurassic (circa 200 Ma), they dominated terrestrial ecosystems in terms of species richness [33],[34], and Cretaceous dinosaurs (145–66 Ma) had body masses spanning more than seven orders of magnitude (Figure 1A). Non-avian dinosaurs became extinct at the catastrophic Cretaceous/Paleogene (K/Pg) boundary event, at or near the peak of their diversity [35],[36]. In contrast, extant dinosaurs (neornithine birds) comprise around 10,000 species and result from one of the most important large-scale adaptive radiations of the Cenozoic [3],[21].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Dinosaur body masses.

(A) Dinosaur body mass through time (the full set of mass estimates is given in Dataset S1). (B) Box-and-whisker plot showing median (dark line), hinges (box range), and ranges (whiskers) of body masses for major dinosaur groups. Outliers (circles) include the iguanodontians Mochlodon vorosi (31 kg), Elrhazosaurus, and Valdosaurus (both 48 kg), the sauropods Europasaurus (1,050 kg) and Magyarosaurus (746 kg), and the flightless avialan Gargantuavis (180 kg).

https://doi.org/10.1371/journal.pbio.1001853.g001

The proposed drivers of early dinosaur diversification are controversial. Although various causal factors have been suggested to underlie a presumed adaptive radiation, few studies have tested the predictions of niche-filling models, and these have yielded equivocal results. An upright, bipedal gait, rapid growth, and possible endothermy have been proposed as key innovations of Triassic dinosaurs (reviewed by [34]), and mass extinctions during the Triassic/Jurassic boundary interval removed competing clades, perhaps leading to ecological release and rapid rates of body size evolution in Early Jurassic dinosaurs [37] (but see [34]). However, quantitative studies using body size proxies [34] and discrete morphological characters [33] have found only weak support for the niche-filling model during early dinosaur evolution, instead favouring gradualistic evolutionary rates. These studies focussed on the Late Triassic–Early Jurassic, so it is unclear whether Early Jurassic dinosaur evolution differed from later intervals (consistent with radiation following a mass extinction), or how the Middle Jurassic–Cretaceous radiation of birds and their proximate relatives relates to overall patterns of dinosaur diversification.

We used phylogenetic comparative methods [6],[14],[38],[39] to analyse rates of dinosaur body mass evolution (Materials and Methods; Appendix S1). For this study, we compiled a large dataset of dinosaur body masses (441 taxa; Dataset S1) using the accurate scaling relationship of limb robustness (shaft circumference) derived from extant tetrapods [40] (Appendix S1; Dataset S1). Body mass affects all aspects of organismal biology and ecology (e.g., [41],[42]), including that of dinosaurs (e.g., [43]–[45]). Because of its relationship with animal energetics and first-order ecology, understanding the evolution of body mass is fundamental to identifying the macroevolutionary processes underlying biodiversity seen in both ancient and modern biotas. Therefore, by studying body mass evolution, we assess the broad pattern of niche filling in the assembly of dinosaur diversity through 170 Ma of the Mesozoic.

In many hypotheses of adaptive radiation, ecological speciation is an important process generating both morphological and taxonomic diversity (e.g., [2]; but see [46]), according to which ecological differentiation is essentially simultaneous with lineage splitting [12]. In consequence, many large-scale studies of adaptive radiation have focussed on diversification rates (e.g., [17],[21],[47]). A correlation between diversification rates and morphological rates is consistent with adaptive radiation (e.g., [15]). However, even when this can be demonstrated, the occurrence of ecological speciation is difficult (perhaps impossible) to test in clades even only a few Ma old [48]. Methods for estimating diversification rates on non-ultrametric trees (e.g., those including deep time data) have recently become available [49]. However, these methods require accurate estimates of sampling probability during discrete time intervals, and it is not clear that it is possible to obtain such estimates from the dinosaur fossil record, which contains many taxa known only from single occurrences. Therefore, our study focuses on the predictions of niche-filling models of morphological evolution during adaptive radiation, as done in some previous studies (e.g., [6],[13]).

Results

Most of the earliest dinosaurs weighed 10–35 kg (Figure 1); Herrerasaurus was exceptionally large at 260 kg. Maximum body masses increased rapidly to 1,000–10,000 kg in sauropodomorphs, with especially high masses in early sauropods such as Antetonitrus (5,600 kg; Norian, Late Triassic) and Vulcanodon (9,800 kg; Early Jurassic), whereas minimum body masses of 1–4 kg were attained by Late Triassic ornithischians and theropods (Figure 1). Jurassic Heterodontosauridae (∼0.7 kg [50]), Middle Jurassic and younger Paraves (e.g., Epidexipteryx, 0.4 kg; Anchiornis, 0.7 kg), and Cretaceous Avialae (birds: 13–16 g to 190 kg [51]) extended this lower body size limit (Table 1). Archaeopteryx weighed 0.99 kg (the largest, subadult specimen [52]) and the Cretaceous sauropod Argentinosaurus weighed approximately 90,000 kg (Table 1). Our full set of mass estimates is available in Dataset S1 and a summary is presented in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Estimated masses in kilograms of smaller- and larger-bodied adult representatives of major dinosaur groups, given to two significant figures. The standard error of all mass estimates is 0.135 log₁₀(kg) [40].

https://doi.org/10.1371/journal.pbio.1001853.t001

Our node height tests indicate that evolutionary rate estimates at phylogenetic nodes (standardised phylogenetically independent contrasts [39]) vary inversely with log-transformed stratigraphic age for most phylogenies (Figure 2). This relationship is significant (based on robust regression [14],[53]) for most phylogenies of non-maniraptoran dinosaurs, and for ornithischians and non-maniraptoran theropods when analysed separately (Figure 2B). This result is weakened, and becomes non-significant, when Triassic nodes are excluded (Figure S1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Node height test for early burst of rates of dinosaur body mass evolution.

(A) Nodal evolutionary rate estimates (standardised independent contrasts [39],[89]) versus node age for data excluding (dashed lowess line) and including (solid lowess line) Maniraptora. (B–C) Box-and-whisker plots detailing results of: (B) robust regression of evolutionary rate on node age: slope (upper plot) and p-value (lower plot); (C) robust regression of evolutionary rate on nodal body mass: slope (upper plot) and p-value (lower plot). In (B–C) dashed lines occur at zero (upper plots) and 0.05 (lower plots: threshold for statistical significance). 1 = Dinosauria; 2 = Ornithischia; 3 = Sauropodomorpha; 4 = Theropoda; and 5 = Maniraptora.

https://doi.org/10.1371/journal.pbio.1001853.g002

Declining evolutionary rates through time are not found in any analyses including maniraptorans. Indeed, when maniraptorans are added to analyses of Dinosauria, a burst of high nodal rate estimates is evident in lowess lines spanning the Middle Jurassic–Early Cretaceous interval of maniraptoran diversification (Figure 2A). Maniraptorans have a weakly positive (non-significant) relationship between evolutionary rates and body mass, and do not show diminishing evolutionary rates through time (Figure 2B–C). This contrasts with non-maniraptoran dinosaurs, in which evolutionary rates vary inversely with body mass (Figure 2C).

Maximum-likelihood models [6],[38] were fitted to phylogenies calibrated to stratigraphy using the “equal” and “mbl” (minimum branch length) methods (see Materials and Methods), and complement the results of our node height tests in showing support for early burst models only in analyses excluding Maniraptora (Table 2; Figure S2). Note, however, that the maximum-likelihood method has less statistical power to detect early burst patterns than does the node height test when even a small number of lineages escape from the overall pattern of declining rates through time [14]. Two models that predict saturation of trait variance through a clade's history were commonly supported in our analyses: the early burst model of exponentially declining evolutionary rates through time, and the Ornstein–Uhlenbeck (OU) model of attraction to a “trait optimum” value. Other models (e.g., Brownian motion, stasis) had negligible AICc weights in all or most (directional trend model) analyses (AICc is Akaike's information criterion for finite sample sizes).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Summary of maximum-likelihood model-fitting approaches, AICc weights (see also Figure S2), and parameter values provided in the form “median (minimum–maximum)” over a set of 60 time-calibrated phylogenies (for AICc weights) or for those phylogenies in which the model received an AICc weight greater than 0.3 (the number of which is given in the column “Number”).

https://doi.org/10.1371/journal.pbio.1001853.t002

Early burst models received high AICc weights for analyses of ornithischians, non-maniraptoran theropods, and non-maniraptoran dinosaurs when using the “equal” branch length calibration method (Table 2; Figure S2). Early burst models had comparable AICc weights to Ornstein–Uhlenbeck models for sauropodomorphs when using the “equal” branch length calibration method, and for ornithischians and non-maniraptoran theropods when using the “mbl” method. Early burst models had generally lower AICc weights for non-maniraptoran dinosaurs and for sauropodomorphs when using the “mbl” branch length calibration method (Table 2; Figure S2). Support from some phylogenies for Ornstein–Uhlenbeck models of attraction to a large body size optimum from small ancestral body sizes [54],[55] in ornithischians [56], non-maniraptoran theropods, and especially sauropodomorphs and non-maniraptoran dinosaurs (Table 2; Figure S2), suggests the occurrence of Cope's rule in dinosaurs. All phylogenies provide strong support for this pattern in maniraptorans (Table 2).

Exceptionally high rates at individual nodes in our phylogenies were identified as down-weighted datapoints in robust regression analyses [14],[53]. Five sets of exceptional nodes in the Triassic–Early Jurassic represent rapid evolutionary shifts from primitive masses around 10–35 kg to large body masses in derived sauropodomorphs (>1,000 kg), armoured ornithischians (Thyreophora; Figure 1B) and theropods (Herrerasaurus, and derived taxa such as Liliensternus (84 kg) and Dilophosaurus (350 kg)), and to smaller body sizes in heterodontosaurid ornithischians (Figure 3; Table 3). Rapid body size changes were rare in later ornithischians and sauropodomorphs, which each show only one exceptional Jurassic node, marking the origin of body sizes greater than 1,000 kg in derived iguanodontians, and of island dwarfism in the sauropod Europasaurus [57]. By contrast, up to six exceptional Jurassic nodes occur in theropod evolution, with especially high contrasts at the origins of body sizes exceeding 750 kg in Tetanurae, and marking phylogenetically nested size reductions on the line leading to birds: in Coelurosauria (e.g., Ornitholestes, 14 kg; Zuolong, 88 kg) and in Paraves, which originated at very small body masses around 1 kg [58].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Dinosaur phylogeny showing nodes with exceptional rates of body size evolution.

Exceptional nodes are numbered and indicated by green filled circles with diameter proportional to their down-weighting in robust regression analyses (Appendix S1). Details of these nodes are given in Table 2. The sizes of shapes at tree tips are proportional to log₁₀(mass), and silhouettes are indicative of approximate relative size within some clades. The result from one tree calibrated to stratigraphy by imposing a minimum branch duration of 1 Ma is shown; other trees and calibration methods retrieve similar results. Silhouettes used were either previously available under Public Domain or with permission from the artists. Non-avialan dinosaur silhouettes used with thanks to the artist, Scott Hartman. Avialan silhouettes are modified from work by Nobumichi Tamura, and /Archaeopteryx/ from Mike Keesey.

https://doi.org/10.1371/journal.pbio.1001853.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Details of body size changes at exceptional nodes indicated in Figure 3.

https://doi.org/10.1371/journal.pbio.1001853.t003

The contrast between theropods and other dinosaurs is even greater in the Cretaceous, when no exceptional nodes occur in Sauropodomorpha, and only two in Ornithischia: at the origins of large-bodied Ceratopsidae and island dwarf rhabdodontid iguanodontians (e.g., Mochlodon [59]). At least nine shifts occurred during the same interval of theropod evolution, including seven in maniraptorans (Figure 3; Table 3).

Discussion

Materials and Methods

We used phylogenetic comparative methods to analyse rates of dinosaur body mass evolution [6],[14],[38],[39] (Appendix S1). Body mass, accompanied by qualitative observations (Table 3), was used as a general ecological descriptor. Body mass was estimated for all dinosaurs for which appropriate data were available (441 taxa; Dataset S1) using the empirical scaling relationship of limb robustness (stylopodial circumference) with body mass, derived from extant tetrapods [40] (Appendix S1). We analysed log₁₀-transformed data (excluding juveniles), which represent proportional changes in body mass.

Stylopodial shaft circumferences are infrequently reported in the literature, so many were taken from our own measurements, or were calculated from shaft diameters (Appendix S1). Previous large datasets of dinosaurian masses were based on substantially less accurate methods, using the relationship between linear measurements (e.g., limb bone lengths) and volumetric models of extinct dinosaurs ([84]–[86]; reviewed by [40]).

Quantitative macroevolutionary models were tested on composite trees compiled from recent, taxon-rich cladograms of major dinosaur groups (Appendix S1; Figure S4, Figure S5, Figure S6, Figure S7). Phylogenetic uncertainty was reflected by analysing alternative topologies and randomly resolved polytomies (Appendix S1). Tip heights and branch durations were stratigraphically calibrated, and zero-length branches were “smoothed” using two methods: (1) by sharing duration equally with preceding non-zero length branches (the “equal” method [87]); and (2) by imposing a minimum branch length of 1 Ma (the “mbl” method [88]).

We used maximum-likelihood model comparison [6],[38] and “node height” test [14],[39] methods (Appendix S1) to test the prediction of the niche-filling hypothesis: that rates of morphological evolution diminish exponentially through time after an adaptive radiation [1],[2],[4]. The node height test treats standardised independent contrasts [89] as nodal estimates of evolutionary rate [39] and tests for systematic deviations from a uniform rate Brownian model, using regression against log-transformed geological age (robust regression [14],[53]). We also regressed standardised contrasts against nodal body mass estimates (a proxy for generation time and other biological processes that might influence evolutionary rates). As well as testing for a “background” model of declining evolutionary rates through time, robust regression identifies and down-weights single nodes deviating substantially from the overall pattern [14],[53]. These nodes represent substantial, temporally rapid, niche-shift events [14], following the macroecological principle that organisms in different body size classes inhabit different niches and have different energetic requirements [41]. We used lowess lines to visualise non-linear rate variation with time and body mass.

Exponentially declining rates of evolution through time, predicted by the niche-filling model of adaptive radiation [1]–[3], were also tested by comparing the fit of an early burst model [6],[7] with other commonly used models: Brownian motion, directional evolution (“trend”), the Ornstein–Uhlenbeck model of evolution attracted to an optimum value, and stasis (“white noise”) [38],[56],[90] (Appendix S1). Explicit mathematical models of trait evolution on our phylogenies were fitted using the R packages GEIGER version 1.99–3 [91] and OUwie version 1.33 [55] (for Ornstein–Uhlenbeck (OU) models only), and compared using AICc [92],[93]. Unlike GEIGER, OUwie allows estimation of a trait optimum (θ) that is distinct from the root value (Z₀) in OU models. Values from GEIGER and OUwie are directly comparable: identical log likelihood, AICc, and parameter estimates are obtained for test datasets when fitting models implemented in both packages (Brownian motion in all instances; and OU models when θ = Z₀ for ultrametric trees); although note that comparable standard error values entered to the OUwie function of OUwie 1.33 are the square of those entered to the fitContinuous function of Geiger 1.99–3. The algorithm used to fit OU models in GEIGER 1.99–3 is inappropriate for non-ultrametric trees (personal communication, Graham Slater to R. Benson, December 2013). This problem is specific to OU models implemented by GEIGER 1.99–3, and does not affect the other models that we tested. GEIGER 1.99–3 fits models of trait evolution using independent contrasts, after rescaling the branch lengths of the phylogenetic tree according to the model considered [7]. For all models, except the OU model in the case of non-ultrametric trees, the covariance between two taxa i and j can be written as a function of the path length s_ij shared between the two taxa (e.g., [6],[7]). The tree can thus easily be rescaled by applying this function to the height of each node before computing independent contrasts. In the case of the OU model, the covariance between two taxa i and j is a function of both the shared (pre-divergence) portion of their phylogenetic history and the non-shared (post-divergence) portion [54]. In the case of an ultrametric tree, the non-shared portion can also be written as a function of s_ij (it is simply the total height T of the tree, minus s_ij [90],[94]), and the corresponding scaling function can be applied to the tree (this is what is performed in GEIGER 1.99.3). However, in the case of a non-ultrametric tree, the post-divergence portion of the covariance cannot be written as a function of s_ij, so there is no straightforward scaling function to apply. Instead, it is necessary to fit the model by maximum likelihood after computing the variance–covariance matrix. This is what is implemented in OUwie, and now in GEIGER 2.0 (personal communication, Josef Uyeda to R. Benson, January 2014).

Our data and analytical scripts are available at DRYAD [95].

Supporting Information

Acknowledgments

We thank Graham Slater, Josef Uyeda, Rich FitzJohn, and Jeremy Beaulieu for discussion. We thank Ronan Allain, John Bird, Stephen Brusatte, Sérgio Cabreira, Jonah Choiniere, Serjoscha Evers, Maria Malabarba, Octavio Mateus, Jay Nair, Attila Õsi, Dennis Parsons, Diego Pol, Stephen Poropat, Toru Sekiya, Virginia Tidwell, Peggy Vincent, John Whitlock, and Thomas Williamson for sharing measurements.