Skip to content Skip to navigation

Article: A new phylogenetic hypothesis of turtles with implications for the timing and number of evolutionary transitions to marine lifestyles in the group

Palaeontology - Volume 62 Part 1  - Cover
Publication: Palaeontology
Volume: 62
Part: 1
Publication Date: January 2019
Page(s): 93 134
Author(s): Serjoscha W. Evers, and Roger B. J. Benson
Addition Information

How to Cite

EVERS, S.W., BENSON, R.B.J. 2019. A new phylogenetic hypothesis of turtles with implications for the timing and number of evolutionary transitions to marine lifestyles in the group. Palaeontology, 62, 1, 93-134. DOI: /doi/10.1111/pala.12384

Author Information

  • Serjoscha W. Evers - Department of Earth Sciences University of Oxford South Parks Road Oxford OX1 3AN UK
  • Roger B. J. Benson - Department of Earth Sciences University of Oxford South Parks Road Oxford OX1 3AN UK

Publication History

  • Issue published online: 27 December 2018
  • Manuscript Accepted: 23 May 2018
  • Manuscript Received: 21 February 2018

Funded By

NERC studentship to SWE on the DTP Environmental Research. Grant Number: NE/L0021612/1
European Community Research Infrastructure Action

Online Version Hosted By

Wiley Online Library (Free Access)
Get Article: Wiley Online Library [Free Access]


Evolutionary transitions to marine habitats occurred frequently among Mesozoic reptiles. Only one such clade survives to the present: sea turtles (Chelonioidea). Other marine turtles originated during the Mesozoic, but uncertain affinities of key fossils have obscured the number of transitions to marine life, and the timing of the origin of marine adaptation in chelonioids. Phylogenetic studies support either a highly‐inclusive chelonioid total‐group including fossil marine clades from the Jurassic and Cretaceous (e.g. protostegids, thalassochelydians, sandownids) or a less inclusive chelonioid total‐group excluding those clades. Under this paradigm, these clades belong outside Cryptodira, and represent at least one additional evolutionary transition to marine life in turtles. We present a new phylogenetic hypothesis informed by high resolution computed tomographic data of living and fossil taxa. Besides a well‐supported Chelonioidea, which includes protostegids, we recover a previously unknown clade of stem‐group turtles, Angolachelonia, which includes the Late Jurassic thalassochelydians, and the Cretaceous–Palaeogene sandownids. Accounting for the Triassic Odontochelys, our results indicate three independent evolutionary transitions to marine life in non‐pleurodiran turtles (plus an additional two‐three in pleurodires). Among all independent origins of marine habits, a pelagic ecology only evolved once, among chelonioids. All turtle groups that independently invaded marine habitats in the Jurassic–Cretaceous (chelonioids, angolachelonians, bothremydid pleurodires) survived the Cretaceous–Palaeogene mass extinction event. This highlights extensive survival of marine turtles compared to other marine reptiles. Furthermore, deeply‐nested clades such as chelonioids are found by the middle Early Cretaceous, suggesting a rapid diversification of crown‐group turtles during the Early Cretaceous.

Turtles are a major group of reptiles and show high ecological diversity, inhabiting marine, freshwater and terrestrial environments. The ancestral ecology for crown‐turtles is inferred to be aquatic (Joyce & Gauthier 2004) and there is evidence for variation between terrestrial and aquatic habitats among stem‐group turtles (Joyce & Gauthier 2004; Scheyer & Sander 2007; Li et al. 2008; Reisz & Head 2008; Scheyer et al. 2014; Cerda et al. 2016; Joyce 2017). As an ecologically diverse extant group with an excellent fossil record, turtles provide information on the evolutionary dynamics of ecological transitions in vertebrates, and in particular the secondary reversion to marine life. This transition is not well understood in reptiles, despite its occurrence in at least 18 lineages among many higher groups (mesosaurs, thalattosaurs, ichthyosaurs, sauropterygians, crocodiliforms, rhynchocephalians, squamates, turtles) (Motani 2009; Benson 2013; Bardet et al. 2014). In part, this lack of understanding results from the extinction of most marine reptile groups by the end of the Mesozoic, or mid‐Cenozoic.

Within turtles alone, multiple lineages independently adapted to marine life from as early as the Triassic (Gaffney et al. 2006; Li et al. 2008; Mateus et al. 2009; Anquetin et al. 2015; Ferreira et al. 2015). Of these, only chelonioids (sea turtles) survive to the present. Previous phylogenetic hypotheses imply at least four and as many as eight independent transitions to marine life in turtles, throughout their entire evolutionary history (e.g. Joyce 2007; Li et al. 2008; Mateus et al. 2009; Rabi et al. 2012; Anquetin et al. 2015). However, deep uncertainty regarding the phylogenetic affinities of several extinct groups renders the precise number highly uncertain. This lack of consensus on turtle phylogeny also affects our understanding of wider macroevolutionary dynamics, such as the mode and rate of morphological evolution during ecological transitions.

Extant turtles provide genetic (= molecular) data that help to constrain their phylogenetic relationships and allow the estimation of divergence ages for subclades (e.g. Joyce et al. 2013). Morphological and molecular phylogenies have resulted in different hypotheses for the interrelationships of extant turtles (for instance regarding the monophyly and content of Chelydroidea: Shaffer et al. 1997; Parham et al. 2006; Joyce et al. 2013; Joyce 2016). Although morphological evidence has subsequently been found to support some of the propositions of molecular studies (e.g. Knauss et al. 2011 for the example of Chelydroidea) not all conflicting relationships among crown‐turtles have yet been reconciled. At present, it is common for morphological studies of turtle phylogeny to use a backbone constraint informed by the results of molecular studies (Danilov & Parham 2008; Cadena & Parham 2015; Zhou & Rabi 2015).

Nevertheless, analysis of morphological data remains important, because the placement of fossil taxa has to be established for a reliable tree calibration in molecular clock analyses (Near et al. 2005; Parham et al. 2012; Joyce et al. 2013). Additionally, the consideration of fossils is fundamental to our understanding of the patterns of morphological and ecological evolution during deep divergences, and in entirely extinct groups. This is maybe best documented by the evolutionary transition from fish to tetrapods (e.g. Ahlberg & Milner 1994; Ahlberg et al. 1996; Carroll 1996; Ahlberg & Johanson 1998; Long & Gordon 2004; Clack 2009) and by the dinosaurian origins of birds (Ostrom 1975; Gauthier 1986).

Several Mesozoic turtle groups exemplify the difficulties associated with robustly fitting fossil taxa into the turtle tree. These include the marine thalassochelydians, sandownids and protostegids, as well as the freshwater xinjiangchelyids, sinemydids and macrobaenids. Most of these groups have been alternatively proposed to be stem‐group turtles, stem‐group cryptodires, or within the crown‐group of cryptodires as stem‐ or crown‐group chelonioids, although this is in part caused by the instability of the position of pleurodires across studies (Sterli 2010). Because the placements of these groups influence reconstructions of morphological changes during deep evolutionary divergences, they have implications for the relative phylogenetic placements of other clades. Furthermore, different phylogenetic hypotheses give rise to distinct conclusions regarding: (1) the number of times that marine ecologies have evolved within turtles; (2) the timing of secondarily marine adaptation in the chelonioid total‐group; and (3) inferences of character state changes and polarities, which influence hypotheses of morphological evolution on the stem‐group of cryptodires, and among turtles more widely. Clearly, resolving the phylogeny of key fossil groups can have substantive implications for our understanding of the evolution of turtles as a whole. We present a new phylogenetic analysis that helps to clarify these issues.

Independent origins of marine life in turtles

Adaptation to marine life evolved independently in stem‐group turtles and in both cryptodires and pleurodires. The stem‐group turtle Odontochelys semitestacea from the middle Late Triassic (220 Ma) is generally interpreted as the oldest marine turtle based on its depositional environment (Li et al. 2008; Reisz & Head 2008; Lyson et al. 2010). Furthermore, Odontochelys has limb proportions characteristic of aquatic turtles, contrasting with other Triassic stem‐turtles (Joyce & Gauthier 2004; Li et al. 2008).

Among pleurodires, marine taxa have been identified in two phylogenetically distinct lineages, documenting at least two independent instances of adaptation to marine life: the Cretaceous–Oligocene (and possibly Miocene) Bothremydidae (Gaffney et al. 2006; Rabi et al. 2012), and the Eocene–Pleistocene Stereogyina (Sánchez‐Villagra et al. 2000; Winkler & Sánchez‐Villagra 2006; Gaffney et al. 2011; Ferreira et al. 2015). Bothremydid phylogeny suggests that there could be two independent origins of marine habits in that group, as the marine Taphrosphyini and the marine Bothremydina do not form a monophyletic group with respect to freshwater bothremydids (Gaffney et al. 2006; Rabi et al. 2012).

The number of independent lineages of marine turtles within the total group of cryptodires is less well understood. Chelonioidea, which includes all extant sea turtles, represents one unambiguous evolutionary origin of secondarily marine ecologies within the crown of cryptodires. Chelonioids form a clade of highly marine‐adapted forms that are characterized by the presence of flippers, salt glands, incipient to well‐developed secondary palates, and a range of modifications of the shell, all of which represent adaptations to a marine lifestyle, and are observed in at least some fossil chelonioids (Mlynarski 1976; Zangerl 1980; Hirayama 1994). Undisputed marine stem‐group chelonioids are also known from that time (Toxochelys spp.: Coniacian–Campanian, Nicholls 1988; Joyce et al. 2013) but the phylogenetic position of some putative crown‐group chelonioids from the Cretaceous (e.g. Allopleuron hofmanni from the late Maastrichtian; Gray 1831) have been uncertain. The only non‐marine turtles that have repeatedly been recovered as stem‐group chelonioids are various ‘macrobaenids’ (usually including Judithemys sukhanovi from the Campanian of Canada; Parham & Hutchinson 2003) and ‘sinemydids’ (Brinkman & Wu 1999; Danilov & Parham 2008; Sterli 2010; Pérez‐García 2012; Cadena & Parham 2015) but most authors argue against this (e.g. Zhou & Rabi 2015).

Other marine turtles have been recognized, especially from the Late Jurassic and Early Cretaceous. Depending on their phylogenetic position, these groups, the protostegids, thalassochelydians and sandownids, could represent additional, independent origins of marine adaptation (Fig. 1). Protostegids are known from the Barremian–Campanian (e.g. Zangerl 1953; Collins 1970; Hirayama 1994, 1998; Hooks 1998; Kear & Lee 2006; Cadena & Parham 2015). Thalassochelydians include ‘plesiochelyids’, ‘eurysternids’, ‘thalassemydids’, Solnhofia parsoni and Jurassichelon oleronensis from the Late Jurassic (e.g. Gaffney 1975a, b; Rieppel 1980; Joyce 2000; Anquetin & Joyce 2014; Anquetin et al. 2014a, b, 2015, 2017). Sandownids are known from the Cretaceous and Palaeogene (Aptian–Danian; Meylan et al. 2000; Mateus et al. 2009; Cadena 2015). Many distinct phylogenetic hypotheses have been proposed, not only for these groups, but also for their members. In the following, we provide brief summaries of all relevant groups, including evidence for their inferred ecologies and previously proposed phylogenetic relationships.

Figure 1 Open in figure viewerPowerPoint Simplified phylogenetic hypotheses of selected previous studies highlighting different phylogenetic positions that have been proposed for marine, non‐pleurodiran turtles. A, Joyce (2007). B, Danilov & Parham (2008). C, Sterli (2010). D, Anquetin (2012). E, Tong & Meylan (2013). F, Sterli & de la Fuente (2013). G, Anquetin et al. (2015). H, Cadena & Parham (2015) and Cadena (2015). I, Zhou & Rabi (2015). Note that the colours denote proposed clades, following Anquetin et al. (2017) for Thalassochelydia, Tong & Meylan (2013) for Sandownidae, Cadena & Parham (2015) for Protostegidae, and Rabi et al. (2014) for Sinemydidae/Macrobaenidae. Note that ‘Chelonioidea’ is used to denote the crown‐group herein.

Non‐pleurodiran marine turtles from the Mesozoic

Protostegids are a group of Cretaceous marine turtles with a cosmopolitan distribution (e.g. Zangerl 1953; Collins 1970; Hirayama 1998; Kear & Lee 2006; Cadena & Parham 2015). They include the gigantic forms Archelon ischryos (Wieland 1896) and Protostega gigas (Cope 1871a) from the Campanian of North America, as well as early representatives such as Santanachelys gaffneyi from the late Aptian or early Albian of Brazil (Hirayama 1998) and Desmatochelys padillai from the late Barremian or early Aptian of Colombia (Cadena & Parham 2015). These taxa have been found in a monophyletic group in most phylogenetic studies so far (Hirayama 1994, 1998; Hooks 1998; Kear & Lee 2006; Cadena & Parham 2015; but see Gentry et al. 2018). They share unambiguous adaptations to a pelagic marine lifestyle with chelonioids, including the modification of hands into flippers and the reduction of the shell, and also an enlarged fenestra interorbitale indicative of salt glands (e.g. Wieland 1896, 1906; Case 1897; Hirayama 1998). Many studies have found protostegids on the stem‐lineage of Dermochelys (the extant leatherback turtle; Hirayama 1994; Kear & Lee 2006; Bardet et al. 2013; Cadena & Parham 2015; Gentry et al. 2018) with a few exceptions (Joyce 2007; Anquetin 2012; and studies expanding these matrices; see below).

Thalassochelydians (sensu Anquetin et al. 2017) are an assemblage of Late Jurassic turtles from central and western Europe (Anquetin & Chapman 2016; Anquetin et al. 2017), and possibly Argentina (Fernández & de la Fuente 1988; de la Fuente & Fernández 2011). They are exclusively found in marine depositional environments, with ‘plesiochelyids’ and ‘thalassemydids’ found in sediments indicating open‐marine carbonate platform environments, and ‘eurysternids’ such as Solnhofia parsoni coming from more marginal marine settings (Anquetin et al. 2017). Plesiochelyids have been found as a monophyletic group by almost all phylogenetic studies (e.g. Sterli 2010; Sterli & de la Fuente 2013; Anquetin et al. 2015; but see Anquetin 2012) but the monophyly of the more inclusive Thalassochelydia (including Solnhofia parsoni, Jurassichelon oleronensis and plesiochelyids) has only recently been demonstrated (Anquetin et al. 2015; but see below). The marine habits of these taxa are supported by their oxygen isotope signatures, which indicate brackish environments for ‘eurysternids’ and fully marine environments for ‘plesiochelyids’ (Billon‐Bruyat et al. 2005). Unlike chelonioids and protostegids, thalassochelydian limbs do not take the form of flippers, although their manus is notably elongated, indicating adaptation to aquatic locomotion (Anquetin et al. 2017). Furthermore, an expanded fenestra interorbitale indicates the presence of salt glands, and the presence of shell fontanelles provides morphological evidence consistent with a marine lifestyle (Anquetin et al. 2017). The shell bone histology is also consistent with an aquatic lifestyle (Scheyer et al. 2014).

Sandownids are a secondarily marine group of putative stem‐group cryptodires currently known from four taxa ranging in age from the Aptian (Sandownia harrisi; Meylan et al. 2000) or possibly late Barremian (Leyvachelys cipadi; Cadena 2015) to the Paleocene (Brachyopsemys tingitana; Tong & Meylan 2013). These taxa have been recovered as a clade by all phylogenetic studies that included more than one of them (Mateus et al. 2009; Tong & Meylan 2013; Cadena 2015). Their fossils come from shallow marine deposits, and distinctive morphological adaptations such as their extensive secondary palate indicative of a durophagous diet have been interpreted as marine adaptations (Mateus et al. 2009; Tong & Meylan 2013; Cadena 2015). Informative shell and postcranial material is only known from Leyvachelys cipadi. Leyvachelys lacks features found in pelagic marine turtles (cheloniids, dermochelyids and protostegids) as it has an unflattened and unexpanded humerus and metacarpal, a strongly ossified contact between the plastron and carapace, and lacks shell fontanelles (Cadena 2015). Sandownids are therefore thought to have lived in littoral habitats (Cadena 2015).

Previous phylogenetic work

Extant chelonioids include cheloniids (hard‐shelled sea turtles) and the leatherback sea turtle, Dermochelys coriacea, the only dermochelyid. Molecular phylogenies find chelonioids within Cryptodira, as the sister group to Chelydroidea (Kinosternoidea + Chelydridae; e.g. Crawford et al. 2015; Pereira et al. 2017) in a clade called Americhelydia (Joyce et al. 2013). The number of independent origins of marine adaptation among eucryptodiran lineages depends on the relationships of fossil marine taxa with respect to chelonioids. A summary of phylogenetic hypotheses recovered in previous studies is presented in Figure 1.

Protostegids have frequently been placed on the stem of Dermochelyidae (Hirayama 1994, Kear & Lee 2006; Tong & Meylan 2013; Bardet et al. 2013; Cadena & Parham 2015; Gentry et al. 2018; see Fig. 1E, H). However, Joyce (2007) and some subsequent analyses (Danilov & Parham 2006, 2008; Sterli & de la Fuente 2011, 2013; Anquetin 2012) found protostegids to be basal eucryptodires on the cryptodiran stem (Fig. 1A, B, F, G), or on the turtle stem (Sterli 2010; Sterli & de la Fuente 2011; Fig. 1C). Either position would imply that protostegids evolved marine habits independently of chelonioids, and have typically been found by analysis of the Early Cretaceous Santanachelys as the sole representative of protostegids in a global matrix. When recovered outside Chelonioidea, Santanachelys has been found in a clade with Solnhofia (Danilov & Parham 2006), or Solnhofia and Jurassichelon to the exclusion of plesiochelyids (Joyce 2007; Pérez‐García et al. 2012; Sterli & de la Fuente 2013; Sterli et al. 2013), or a clade including Solnhofia, Jurassichelon and plesiochelyids (Anquetin 2012).

Many previous studies have found the taxa included in Thalassochelydia (sensu Anquetin et al. 2017) to be polyphyletic or paraphyletic (Danilov & Parham 2006, 2008; Joyce 2007; Mateus et al. 2009; Anquetin 2012; Pérez‐García et al. 2012; Sterli & de la Fuente 2013; Sterli et al. 2013; Cadena 2015; Cadena & Parham 2015; see Fig. 1A–F, H, I). This is in part because of the uncertainty that surrounds the positions of Jurassichelon (‘Thalassemysmoseri in studies previous to the revision provided by Pérez‐García 2015) and Solnhofia, which have been variously grouped with non‐thalassochelydian taxa such as the early protostegid Santanachelys (Danilov & Parham 2006; Joyce 2007; Pérez‐García et al. 2012; Sterli & de la Fuente 2013; Sterli et al. 2013) but also sandownids (e.g. Mateus et al. 2009). Nevertheless, some concept of either one or more clades of Late Jurassic marine turtles has been present in many phylogenetic studies. Three major hypotheses have been presented concerning the positions of these turtles (regardless of their monophyly), as either: (1) stem‐group chelonioids, implying an inclusive clade of marine cryptodires that originated in the Late Jurassic; (2) stem‐group cryptodires, and therefore representing a distinct, independent origin of marine life from that in chelonioids; (3) stem‐group turtles, also representing a distinct origin of marine life. It is clear therefore that thalassochelydians share some traits with chelonioids, and others with stem‐group turtles or stem‐group cryptodires. Indeed, when constituents of Thalassochelydia have been found as stem‐group chelonioids, groups that are otherwise widely regarded as stem‐group cryptodires or stem‐group turtles (xinjiangchelyids, sinemydids; Zhou & Rabi 2015) have sometimes also been found on the stem‐group of chelonioids (Cadena & Parham 2015). However, a sinemydid–chelonioid relationship has also been found in some studies irrespective of the position of thalassochelydians (Brinkman & Wu 1999; Danilov & Parham 2008; Sterli 2010; Pérez‐García 2012). Clearly, resolving the phylogenetic affinities of Thalassochelydia is of great importance for our knowledge of turtle evolution.

Little consensus exists about the relative placement of sandownids (Fig. 1A, E, G, H). Specifically, the relationships of sandownids to ‘eurysternids’, plesiochelyids and chelonioids are unclear. The first reported sandownid, Sandownia harrisi, was originally interpreted as a basal member of Trionychia (softshell turtles; Meylan et al. 2000). However, later studies found little support for this placement (e.g. Joyce 2007) and the patterning of the carapace of Levyachelys is also inconsistent with this hypothesis (Cadena 2015). Sandownia (commonly the only sandownid taxon included in global phylogenetic studies) was tentatively placed on the stem of Cryptodira as a basal eucryptodire by Joyce (2007). Mateus et al. (2009) found sandownids to be sister to Solnhofia in a group called Angolachelonia, which formed part of a marine grade including Jurassichelon, Portlandemys and Plesiochelys on the stem of Cryptodira. Tong & Meylan (2013) and Anquetin et al. (2015) recovered sandownids as the sister‐taxon of chelonioids, and Solnhofia and Jurassichelon as eucryptodires. Finally, Cadena (2015) recovered sandownids as the sister to xinjianchelyids + sinemydids as stem‐group chelonioids, forming a large panchelonioid clade including Jurassichelon and Solnhofia, but excluding plesiochelyids, which were found as stem‐group turtles (Cadena 2015).

Material and method

Character dataset

For our character matrix, we used information from three primary sources: Cadena (2015), Zhou & Rabi (2015), and Anquetin et al. (2015). All of these matrices are modified versions of previous studies and can be traced back to Joyce (2007), which represents the first global phylogenetic analysis of Mesozoic turtles. All three source matrices are global in scope but have different foci. Cadena's (2015) matrix is a modified version of the work of Cadena & Parham (2015) and includes character information specific to sandownids. The matrix of Cadena & Parham (2015) is a combination of global matrices (Joyce 2007; Anquetin 2012; Sterli 2008; Sterli & de la Fuente 2013) but importantly includes characters from matrices that were designed to specifically test chelonioid relationships (Hirayama 1994; Hirayama 1998; Kear & Lee 2006; Parham & Pyenson 2010). The Zhou & Rabi (2015) matrix is similarly based on several previous studies, which were focused on the relationships of stem‐group turtles and stem‐group cryptodires (Joyce 2007; Sterli 2008; Sterli & de la Fuente 2013; Rabi et al. 2013, 2014; Zhou et al. 2014). Finally, the analysis of Anquetin et al. (2015) was specifically focused on thalassochelydian relationships. By combining the character lists of these studies, we attempt to capture all morphological variation documented so far among non‐pleurodiran turtles. The 257 characters of Cadena (2015) were reconciled with the slightly different character definitions given in the other two studies, and were further modified by our own observations. This resulted in a phylogenetic data matrix of 345 characters and 80 taxa. In total, 70 characters of Cadena & Parham (2015) were modified (see Coding strategy) and modified versions of 13 characters used by Anquetin (2012), Anquetin et al. (2015), Cadena (2015), Havlik et al. (2014) and Zhou & Rabi (2015) were added. We also added 58 new cranial and mandibular characters. We re‐scored all cranial and mandibular characters from both Cadena & Parham (2015) and other sources, for all taxa in our matrix. Our scorings are based on first hand examination of specimens (44 taxa), examination of CT scans for 59 taxa, photographs of specimens not seen first hand by the authors (shared with SWE by Dr Walter Joyce), and published descriptions (data sources in Table 1). An illustrated character list with discussions of character modifications and additions, and justifications for modified scorings is available in Evers & Benson (2018b, appendix S1).

Table 1. Sources for scorings in our character–taxon matrix Taxon Specimen(s) used for character scoring References used for character scoring Personal observation of taxon CT Allaeochelys libyca BSPG 1991 II 130 Havlik et al. (2014) Yes Yes Allopleuron hoffmanni NHMUK R4291; NHMM 000001 Mulder (2003) Yes Yes Annemys sp. IVPP V18106 Brinkman et al. (2013) No Yes Annemys levensis PIN 4636‐4‐2 Sukhanov (2000); Rabi et al. (2014) No No Annemys latiens PIN 4636‐6‐1 Rabi et al. (2014) No No Apalone spinifera emoryi FMNH 22178 No additional references used Yes Yes Araripemys barretoi AMNH 30778 Meylan (1996); Gaffney et al. (2011) Yes Yes Archelon ischyros YPM 3000 Wieland (1896, 1900, 1902) Yes No Argillochelys cuneiceps NHMUK R38955; NHMUK R41636 No additional references used Yes Yes Arundelemys dardeni USNM 497740 Lipka et al. (2006) No Yes Australochelys africanus BP/1/4933 Gaffney & Kitching (1995) Yes Yes Baptemys wyomingensis YPM 374, 3758; DMNH 511 Meylan & Gaffney (1989); Knauss (2014) No No Bouliachelys suteri QMF 31669 Kear & Lee (2006) No Yes Brachyopsemys tingitana AMNH 30001; AMNH 30612 Tong & Meylan (2013) Yes Yes Caretta caretta NHMUK 1938.1.9.1 No additional references used Yes Yes Carettochelys insculpta NHMUK 1903.7.10.1 No additional references used Yes Yes Chelodina oblonga NHMUK 64.12.22 No additional references used Yes Yes Chelodina longicollis a a Species indicated by asterisk were used only for postcranial scores. Unknown (specimen number not given in Cadena & Parham 2015) Cadena & Parham (2015) No No Chelonia mydas NHMUK 1969.776 No additional references used Yes Yes Chelonoidis sp. SMF 67582 No additional references used No Yes Chelonoidis chilensis a a Species indicated by asterisk were used only for postcranial scores. Unknown (specimen number not given in Cadena & Parham 2015) Cadena & Parham (2015) No No Chelus fimbriatus NHMUK No additional references used Yes Yes Chelydra serpentina SMF 32846 No additional references used No Yes Chrysemys picta NHMUK No additional references used Yes Yes Chubutemys copelloi MPEF‐PV1236 Gaffney et al. (2007); Sterli et al. (2015) No No Dermatemys mawii SMF 59463 No additional references used No Yes Dermochelys coriacea FMNH 171756; UMZC R3031 No additional references used Yes Yes Desmatochelys lowii KUVP 1200 No additional references used No Yes Desmatochelys padillai FCG‐CBP 01, 13, 15, 39, 40; UCMP 38345, 382456 Cadena & Parham (2015) No No Dracochelys bicuspis IVPP V4075; IVPP V12091 Gaffney & Ye (1992); Brinkman (2001) No No Eileanchelys waldmanni NMS.G.2004.31.15; NMS.G.2004.31.16a‐f Anquetin (2010) Yes Yes Elseya dentata NHMUK No additional references used Yes Yes Emarginachelys cretacea KUVP 23488 Whetstone (1978) No No Emys orbicularis WGJ1987a No additional references used No Yes Eochelone brabantica NHMUK R37213 Casier (1968) Yes Yes Eosphargis breineri FUM‐N‐1450 Nielsen (1959) Yes No Eretmochelys imbricata FMNH 22242 No additional references used Yes Yes Eubaena cephalica DMNH 96004; AMNH 4948, 2602, 2604, 2606; YPM 1785 Gaffney (1972, 1979); Gaffney (1982a, b) No No Geoclemys hamiltoni NHMUK No additional references used Yes Yes Glyptops plicatulus AMNH 336; YPM 1784, 4717, 5821 Gaffney (1979) No No Gopherus polyphemus FMNH 211815 No additional references used Yes Yes Judithemys sukhanovi TMP 87.2.1 Parham & Hutchison (2003) No No Jurassichelon oleronensis PIMUZ AIII 514 Rieppel (1980) Yes Yes Kallokibotion bajazidi NHMUK R4918; NHMUK R4921; NHMUK R4925 Gaffney & Meylan (1992) Yes Yes Kayentachelys aprix TMM 4370‐2, 43653‐1; MCZ 8914‐8917; MNA V1558, V2664 Sterli & Joyce (2007); Gaffney & Jenkins (2010) No Yes Kinosternon suburum hippocrepis FMNH 211711 No additional references used Yes Yes Kinosternon flavescens a a Species indicated by asterisk were used only for postcranial scores. Unknown (specimen number not given in Cadena & Parham 2015) Cadena & Parham (2015) No No Kirgizemys dmitrievi ZIN PH7/15 Danilov et al. (2006) No No Kirgizemys hoburensis PIN 3334‐4; PIN 3334‐35; PIN 3334‐36 Sukhanov (2000) No No Lepidochelys kempii M009/08 Jones et al. (2012) No Yes Lepidochelys olivacea SMNS 11070 No additional references used Yes Yes Levyachelys cipadi FCG‐CBP‐71; SMU 75377, 72852, 74982, 75327; FWMSH 93B‐17 Vineyard (2009); Cadena (2015) No No Lissemys punctata SMF 74141 No additional references used No Yes Macrochelys temminckii FMNH 22111 No additional references used Yes Yes Meiolania platyceps NHMUK R682; specimens listed in Gaffney (1983) Gaffney (1983) Yes Yes Natator depressus R112123 Jones et al. (2012) No Yes Notochelone costata NHMUK R9590 No additional references used Yes Yes Ocepechelon bouyai OCPDEK/GE516 Bardet et al. (2013) No No Ordosemys sp. IVPP V12092 Brinkman & Wu (1999) No Yes Pelodiscus sinensis IW576‐2 No additional references used No Yes Pelomedusa subrufa SMF 70504 No additional references used Yes Yes Phrynops geoffroanus SMF 45470 No additional references used No Yes Platysternon megacephalum SMF 69684 No additional references used No Yes Plesiochelys planicpes OUMNH J1582 Gaffney (1975a, 1976) Yes Yes Plesiochelys etalloni MNB 435 Gaffney (1975a, 1976) Yes Yes Pleurosternon bullockii UMZC T1041 Evans & Kemp (1975) No Yes Podocnemis unifilis FMNH 45657 No additional references used Yes Yes Podocnemis expansa a a Species indicated by asterisk were used only for postcranial scores. Unknown (specimen number not given in Cadena & Parham 2015) Cadena & Parham (2015) No No Portlandemys mcdowelli NHMUK R2914 No additional references used Yes Yes Proganochelys quenstedti SMNS 15759, 16980 Gaffney (1990) Yes No Protostega gigas AMNH 1502, 1503; CMNH 1421; FMNH PR2, P27385 Case (1897); Hay (1908) Yes No Puppigerus camperi NHMUK R14375 Moody (1974) Yes Yes Rhinochelys pulchriceps CAMSM B55775, B55783, B55776; NHMUK R43980, R2226, R35197 No additional references used Yes Yes Sandownia harrisi MIWG 3480 Meylan et al. (2000) Yes Yes Sinemys gamera IVPP V9532‐11 Brinkman & Peng (1993) No No Sinemys lens IVPP V9533‐1; IVPP V9533‐3 Brinkman & Peng (1993) No No Solnhofia parsoni TM 4023 Gaffney (1975b) Yes Yes Staurotypus salvinii NHMUK 1879.1.7.5 No additional references used Yes Yes Staurotypus triporcatus a a Species indicated by asterisk were used only for postcranial scores. Unknown (specimen number not given in Cadena & Parham 2015) Cadena & Parham (2015) No No Sternotherus minor FMNH 211696 Bever (2009) Yes Yes Sternotherus odoratus a a Species indicated by asterisk were used only for postcranial scores. Unknown (specimen number not given in Cadena & Parham 2015) Cadena & Parham (2015) No No Testudo marginata FMNH 51672 No additional references used Yes Yes Testudo hermanni a a Species indicated by asterisk were used only for postcranial scores. Unknown (specimen number not given in Cadena & Parham 2015) Cadena & Parham (2015) No No Toxochelys sp. AMNH1496, 5118; FMNH UR3; FMNH PR219, PR648 Hay (1896); Nicholls (1988) Yes Yes Xinjiangchelys radiplicatoides IVPP V9539‐1 Brinkman et al. (2013) No Yes Xinjiangchelys wusu PMOL‐SGPA0100‐1; PMOL‐SGPA0100‐3 Rabi et al. (2013) No No
  • a Species indicated by asterisk were used only for postcranial scores.
  • Bold specimen numbers indicate specimens for which CT scans were available.
  • Postcranial scorings were taken from Cadena & Parham (2015).

Coding strategy

Phylogenetic analyses of morphology are used to infer relationships from a set of comparative anatomical observations. The concept of homology is central to the delimitation of these observations as characters and character states. Homology is the hypothesis that similar attributes shared by different species are equivalent to one another due to common ancestry, and therefore reflect their phylogenetic relationships (Platnick 1979; Eldredge & Cracraft 1980; de Pinna 1991; see Patterson (1982) for a disambiguation and extensive discussion of homology). Hypotheses of homology are established based on topographic and ontogenetic correspondence and on compositional similarity of traits (Patterson 1982; Rieppel 1988; de Pinna 1991; Hawkins et al. 1997). These initial (‘primary’; de Pinna 1991) hypotheses of homology are embodied by the characters and character states, and are tested against one another by means of congruence in parsimony analysis, which maximizes the number of initial homology propositions that are upheld as true homology rather than the result of convergent evolution (homoplasy; Patterson 1982; de Pinna 1991).

Character construction therefore represents the proposition of initial hypotheses of homology based on observations of similarity. Character states are different forms that a homologous trait (i.e. the character) can take among organisms, and are equivalent transformations of one another on a single hierarchical level (the level of the character) (Platnick 1979; de Pinna 1991; Hawkins et al. 1997). Characters and character states therefore form a hierarchical framework of homology propositions. For example, a specific site in the aligned nucleotide sequences of several organisms is hypothesized to be comparable on the character level, and the (four or five) nucleobases represent possible variation of this character on the character state level (and could therefore not represent separate characters themselves) (Pleijel 1995). Construction of morphological characters is potentially more subjective. Nevertheless, we have taken various steps to ensure that our cladistics characters for turtles, both new, and modified from previous analyses, reflect homology (described below).

Character states should also be independent of each other (e.g. Pimental & Riggins 1987; Strong & Lipscomb 1999). Non‐independence of character states is problematic because it results in a single observation of similarity being coded twice, and therefore the character state is arbitrarily given a greater weight than others (Wilkinson 1995; Strong & Lipscomb 1999). Logical dependence between characters exists when the score of a taxon for one character directly affects the scoring of the same taxon in another character (e.g. coding the absence and zero size of a feature; Wilkinson 1995), and this can be readily avoided. Biological dependence is a hypothesis that character states of different characters are non‐independent due to functional or developmental linkage (Wilkinson 1995). Information on the underlying biological cause for such a linkage if often not available, which makes is hard to fully avoid biological independence in character construction.

Many morphological characters used in previous studies of turtle phylogeny combine complex observations across multiple hierarchical levels of similarity, or lump together independently varying character states. We re‐coded characters so that independently varying aspects of morphological variation are coded in separate characters. For example, character 10 of Cadena & Parham (2015) was defined as:

  • Jugals, jugal–squamosal contact: 0 = present; 1 = absent, contact between postorbital and quadratojugal present.

In this character definition, the absence of a jugal–squamosal character is assumed to be linked to the presence of a postorbital–quadratojugal contact. The character includes the assumption that a postorbital–quadratojugal character must be present when a jugal–squamosal contact is absent. However, these contacts vary independently from one another, as taxa exist (trionychids) in which a postorbital–quadratojugal contact is absent, but a jugal–squamosal contact is also absent. The approach taken here is to break up this character into two separate characters, which each code for one of the contacts (our chs 25 and 42).

We also re‐coded characters so that observations across multiple hierarchical levels of similarity are coded in separate characters. For example, we modified character 60 of Cadena & Parham (2015), which was phrased as:

  • Epipterygoid: 0 = present, rod like; 1 = present, laminar; 2 = absent.

The comparative anatomical observations underlying this and similar characters include both the absence of a certain feature, and variation in the shape of that feature when present (the ‘no tail; red tail; blue tail’ problem; Maddison 1993; Smith 1994; Hawkins et al. 1997; Brazeau 2011). Coding these observations into a single character, as done in most turtle studies (e.g. Joyce 2007; Cadena & Parham 2015; but see Anquetin 2012) assumes that the condition of having no epipterygoid is on the same hierarchical level as the observed shape conditions of taxa with epipterygoids. This disregards the homology implied by the presence of an epipterygoid in various taxa, irrespective of its of shape (Hawkins et al. 1997), resulting in loss of information.

To reflect the hierarchical levels of similarity of these observations of the epipterygoid, and other characters in our analysis, we used a hierarchical coding strategy. For the above example, this reformulation results in two separate characters. One character codes for the presence versus absence of the feature (e.g. epipterygoid; our ch. 87), and the second codes for the various shapes that the feature can take (e.g. epipterygoid shape; our ch. 88) (Hawkins et al. 1997; Brazeau 2011; see Anquetin 2012 for a turtle study using the same approach). Logical dependence of the character states of the second character on the first is avoided by scoring the second character as inapplicable (‘–’) when the first character is scored as ‘absent’. Inapplicability does not introduce logical dependency (contra Strong & Lipscomb 1999) because it does not represent a separate character state, and transitions to ‘inapplicable’ do not count towards the length of a cladogram. However, inapplicable scores can be problematic because inapplicability is treated the same as missing data in software implementations of parsimony analysis (e.g. Platnick et al. 1991; Strong & Lipscomb 1999; Brazeau et al. 2017). This can introduce illogical character state reconstructions during the optimization process (e.g. Platnick et al. 1991; Maddison 1993). This problem relates to computational issues that may be solved in future generations of software (e.g. Brazeau et al. 2017), and problems with inapplicable character states are herein taken to be less problematic than the homology issues of compound characters discussed above (Strong & Lipscomb 1999; Hawkins et al. 1997).

Some of our characters appear superficially to be hierarchical compound characters (as in Cadena & Parham 2015, ch. 60; above), because they include an absence state as well as several other non‐absence states. However, these characters are measurements such as the size of a structure, for which absence equates to zero. It is therefore appropriate to treat their character states as homologous with each other at the same hierarchical level. For example, our character 7 (ch. 7 of Cadena & Parham 2015):

  • Prefrontal, dorsal prefrontal exposure: 0 = present, large; 1 = present, reduced; 2 = absent or near absent.

Our approach also differs from some recent analyses of turtles (e.g. Anquetin 2012) that exclusively used binary characters. Instead, we coded instances for which multiple conditions are observed for a character as multistate characters, reflecting the hypothesis that these alternative conditions represent transformations of a single underlying feature. For example, our character 297 (identical to ch. 210 of Cadena & Parham (2015)):

  • Caudals, anterior caudal centra: 0 = amphicoelous; 1 = procoelous or platycoelous; 2 = opisthocoelous.

The alternative approach of only using binary characters (e.g. Anquetin (2012) for turtles) requires the construction of a series of characters, or ignoring observations to artificially create binary characters, or lumping a set of distinct character states into a single character state (as done in the above example by Anquetin (2012): ch. 160, in which the conditions of procoely and opisthocoely are lumped into a single character state). Anquetin (2012) justified the usage of binary characters by stating that this approach minimizes the amount of ‘a priori assumptions about homology’ (Anquetin 2012, p. 5). However, this avoidance of homology proposition misapplies the similarity criterion used to establish primary hypotheses of homology, which are always a priori hypotheses (de Pinna 1991; Hawkins et al. 1997). Instead of avoiding assumptions of homology, the approach of Anquetin (2012) generates unexpected hypotheses of homology (e.g. that procoely and opisthocoely are interchangeable, and homologous with each other on a low hierarchical level).

Taxon sampling

Our operational taxonomic units (OTUs) were chosen specifically to test the phylogenetic relationships of non‐pleurodiran marine turtles, and we preferentially used taxa for which we had CT data. Our matrix includes many chelonioids, thalassochelydians, protostegids and sandownids (Table 1). To test the placement of these groups in a global framework, we also included a thorough sample of stem‐group turtles, including Triassic and Early Jurassic species, as well as paracryptodires and meiolaniforms. We further included a range of xinjiangchelyids, sinemydids and macrobaenids, which have been hypothesized to be stem‐group cryptodires (e.g. Zhou & Rabi 2015). We sampled extant representatives of modern clades including pleurodires, and all families of turtles. We included early fossil representatives of extant families when CT data was available to us or when sufficiently detailed anatomical descriptions were available in the literature. Of particular importance, two fossil chelydroids, Emarginachelys cretacea and Baptemys wyomingensis, were included because chelydroids (kinosternoids + chelydrids) are the extant sister‐group to chelonioids (Crawford et al. 2015) and fossil representatives of this group can thus influence character optimizations for the total group of Americhelydia (chelydroids + chelonioids).

Most of our OTUs represent modern or fossil taxa on the species level, but some OTUs are supraspecific: the extant taxa Chelodina, Chelonoidis, Kinosternon, Podocnemis, Staurotypus, Sternotherus and Testudo. This was done because our CT scans, which were used to re‐score the cranial and mandibular characters, belonged to different species from those used by Cadena & Parham (2015) (Table 1), and we retained their postcranial characters. We chose to score a supraspecific OTU for Toxochelys because this allowed inclusion of data from specimens that have been assigned to different species, and thus score a more complete OTU. Species assignments regarding specimens of Toxochelys are currently uncertain (Zangerl 1953; Nicholls 1988; Brinkman et al. 2006), particularly because the holotype specimens of valid species are relatively incomplete and lack cranial material. Pending a taxonomic revision of Toxochelys material, we consider it best to refer most Toxochelys material to Toxochelys sp., as done by some other authors (e.g. Brinkman et al. 2006).

The protostegid Rhinochelys pulchriceps was scored primarily from CT scans of the holotype specimen CAMSM B55775. However, we also used data from CT scans of other specimens, which have been referred to different species of Rhinochelys (Collins 1970; see Table 1 for specimens used). These added information on the mandible (R. cantabrigiensis: CAMSM B55783; R. elegans: CAMSM B55776). It is likely that all British specimens of Rhinochelys represent a single species (SWE, unpublished data), and we therefore consider all material in Collins (1970) to be Rhinochelys pulchriceps.

CT data and 3D models used in this study

We used high‐resolution X‐ray computed‐tomography (CT) data for as many taxa as possible (summarized in Table 1). To support our anatomical observations and opinions stated throughout this work, we deposited several CT datasets and 3D models generated through manual segmentation in the software Mimics (versions 16.0–19.0; in MorphoSource (Evers & Benson 2018a). The data represent a sample of fossil and extant turtle specimens that have been extensively used in character construction and illustration of characters (see Evers & Benson 2018b, appendix S1). These are: Apalone spinifera emroyi (FMNH 222178), Dermochelys coriacea (UMZC R3031), Elseya dentata (NHMUK, Gopherus polyphemus (FMNH 211815), Lepidochelys olivacea (SMNS 11070), Phrynops geoffroanus (SMF 45470), Plesiochelys planiceps (OUMNH J1582), Argillochelys cuneiceps (NHMUK R38955), Sandownia harrisi (MIWG 3480), Solnhofia parsoni (TM 4023), Sternotherus minor (FMNH 211696) and Rhinochelys pulchriceps (CAMSM B55783; referred to Rhinochelys cantabrigiensis by Collins (1970), but see Taxon Sampling, above). Because scanning parameters varied between all scans, details are given in the online repository with the data (Evers & Benson 2018a) and are not summarized here.

Similarity matrix

To illustrate the structure of our dataset, we produced heatmaps of similarity matrices calculated from our character–taxon matrix in R (R Core Team 2016). Polymorphisms and missing data (‘?’) were treated as inapplicable data. Similarity between two taxa was calculated as the number of identical scoring matches between these taxa divided by the number of comparable scores between them (e.g. Sneath & Sokal 1973; Foote 1994; Wagner 1997; see Benson & Druckenmiller 2014 for a similar code). The resulting values are represented in a symmetric similarity matrix. Similarity matrices were calculated for the total dataset, and partitions of that dataset representing cranial and mandibular characters, shell characters, limb characters and axial + girdle characters. For the anatomical partitions of the total dataset, taxa with no anatomical information for the respective partition, as well as taxa with 90% or more missing data were excluded. Each similarity matrix was coloured by assigning a spectrum of colours to the similarity values of each matrix using the package gclus (Hurley 2012). The R code for calculating similarity and producing the heatmaps is available in Evers & Benson (2018b).

Parsimony analysis

Tree searches were performed in TNT 1.5 for Windows (Goloboff et al. 2008; Goloboff & Catalano 2016). For all analyses, we used a molecular backbone constraint following the topology of Pereira et al. (2017) for extant taxa. All fossil taxa were left unconstrained and were thus allowed to fall anywhere within the constrained topology of extant taxa. Proganochelys quenstedti was set as the outgroup. In TNT, new technology algorithms were used with default settings, and tree drifting (Goloboff 1999) and parsimony rachet (Nixon 1999) enabled. The initial level of the driven search was set to 30, and the number of times the minimum tree length should be obtained was set to 30. This procedure does not require specifying a number of replicates to be performed (Goloboff 1999; Giribet 2007). The most parsimonious trees (MPTs) of this analysis were then subjected to a final round of tree bisection and reconnection (TBR). The resulting MPTs were used to construct a strict consensus tree.

Characters were treated as unordered and equally weighted in our initial analysis. We also executed a second run in which several characters were ordered, for comparison (chs 7, 14, 18, 34, 44, 67, 76, 79, 90, 93, 94, 103, 107, 123, 130, 131, 138, 142, 147, 205, 210, 217, 248, 253, 292, 305, 325, 339, 340, 344). Ordering character states represent a priori hypotheses that transitions between character states only occur in a specific sequence (e.g. Wilkinson 1992, 1995). However, prior knowledge about the evolutionary sequence of character state transitions is often missing or otherwise not substantiated by evidence. Therefore, we only ordered multistate characters in which one character state was directly intermediate in form, size, or number of two other states (for instance ch. 340: size of ulnare vs intermedium: 0 = ulnare smaller than intermedium; 1 = ulnare nearly as large as intermedium; 2 = ulnare much larger than intermedium). If clear intermediates are not evident, the states were left unordered (for instance ch. 260: medial contact of extragulars: 0 = absent; 1 = contacting each other anterior to gulars; 2 = contacting each other posterior to gulars).

Bremer support was calculated in TNT, using the absolute Bremer support setting and TBR branch swapping on existing trees to generate suboptimal trees. Character optimization was performed in PAUP* for Macintosh (Swofford 2002) because TNT only returns unambiguous synapomorphies (i.e. those in which ACCTRAN and DELTRAN agree in determining a character state change) whereas the optimization criterion can be specified in PAUP*. The optimization was computed on a fully resolved tree. This tree was chosen from the MPTs on the basis of high congruence with a 50% majority rule consensus tree. The most important differences between the tree selected for optimization and our consensus topology are that paracryptodires are monophyletic and that the species of Sinemys form a clade that is sister group to all remaining sinemydids (see Evers & Benson 2018b, appendix 2, fig. S2.1). The tree topology chosen for the optimization, as well as unambiguous, ACCTRAN, and DELTRAN optimizations for all characters as well as all nodes are given in Evers & Benson (2018b, appendices S1–2).

To illustrate implications of our results for the age of crown‐group chelonioids (see Discussion, below), we produced a time‐scaled tree using age‐ranges of taxa derived from the literature and the PalaeoDB (, summarized in Table 2. Commands from the strap (Bell & Lloyd 2014) and paleotree (Bapst 2012) packages in R were used for time‐scaling and visualization. We arbitrarily used a minimum branch length argument of 1 myr for zero‐branch lengths in the timePaleoPhy command of the paleotree package (Bapst 2012). Because some lineages of turtles have older fossil representatives than the oldest fossils for those groups included in our study, we included minimum constraints for three nodes of our time scaled tree. First, we used the top of the Oxfordian (157.3 Ma) as a minimum constraint for Pleurodira based on Caribemys oxfordiensis (de la Fuente & Iturralde‐Vinent 2001; Joyce et al. 2013). Secondly, we used the base of the Cretaceous (145.0 Ma) as a minimum constraint for Cryptodira based on the total‐group trionychian Sinaspideretes wimani. Although the type locality of Sinaspideretes is not exactly known (Tong et al. 2014) it is likely that the material as well as referred material is from the Shangshaximiao Formation (Danilov & Parham 2006; Tong et al. 2014), which is probably of Late Jurassic but possibly even of Middle Jurassic age (Xing et al. 2013). Using the base of the Cretaceous as a minimum age is thus a conservative constraint, following Joyce et al. (2013). Lastly, again following Joyce et al. (2013), we use the top of the Campanian (72.1 Ma) as a minimum constraint for Chelydroidea based on a peripheral fragment that can be unambiguously identified as belonging to the total‐group of Chelydridae, which was described by Brinkman & Rodriguez de la Rosa (2006).

Table 2. Stratigraphic ranges for taxa used for our time‐calibrated tree (Ma) Taxon First appaearance datum (FAD) Last appearance datum (LAD) Source Allopleuron hofmanni 72.1 66.0 Janssen et al. (2011) Annemys latiens 163.5 145.5 Rabi et al. (2010) Annemys levensis 157.3 145.5 Rabi et al. (2010) Annemys sp. 167.7 145.0 PalaeoDB Araripemys barretoi 112.6 109.0 PalaeoDB Archelon ischyros 83.6 72.1 Hirayama (1997) Argillochelys cuneiceps 56.0 47.8 Weems & Brown (2017) Arundelemys dardeni 122.46 112.6 PalaeoDB Australochelys africanus 201.6 189.6 PalaeoDB Baptemys wyomingensis 55.8 40.4 PalaeoDB Bouliachelys suteri 105.3 99.7 PalaeoDB Brachyopsemys tingitana 66.043 61.7 PalaeoDB Chubutemys copelloi 125.45 112.6 PalaeoDB Desmatochelys lowii 93.9 89.8 Hirayama (1997) Desmatochelys padillai 125.0 120.0 Cadena & Parham (2015) Dracochelys biscuspis 125.0 100.5 Rabi et al. (2010) Eileanchelys waldmanni 168.3 166.1 Joyce et al. (2016b) Eubaena cephalica 70.6 66.043 PalaeoDB Emarginachelys cretacea 70.6 66.043 PalaeoDB Eochelone brabantica 47.8 41.2 Weems & Brown (2017) Eosphargis breineri 56.0 47.8 Nielsen (1959) Glyptops plicatulus 150.8 145.5 PalaeoDB Judithemys sukhanovi 84.9 70.6 PalaeoDB Jurassichelon oleronensis 152.1 145.0 Rieppel (1980); Anquetin et al. (2017) Kallokibotion bajazidi 72.1 66.0 Joyce et al. (2016b) Kayentachelys aprix 199.3 182.7 Joyce et al. (2016b) Kirgizemys dmitrievi 130.0 112.6 PalaeoDB Kirgizemys hoburensis 125.0 100.5 Rabi et al. (2010) Leyvachelys cipadi 130.0 109.0 PalaeoDB Notochelone costata 109.0 99.7 PalaeoDB Ocepechelon bouyai 70.6 66.043 PalaeoDB Ordosemys sp. 129.4 113.0 Rabi et al. (2010) Plesiochelys etalloni 157.3 150.0 Anquetin et al. (2017) Plesiochelys planiceps 152.1 145.0 Anquetin et al. (2017) Pleurosternon bullockii 145.2 140.2 PalaeoDB Portlandemys mcdowelli 152.1 145.0 Anquetin et al. (2017) Proganochelys quenstedti 208.5 201.3 Joyce et al. (2016b) Protostega gigas 85.8 70.6 PalaeoDB Puppigerus camperi 56.0 41.2 Weems & Brown (2017) Rhinochelys pulchriceps 105.3 102.0 PalaeoDB for FAD; LAD following Collins (1970) Sandownia harrisi 125.45 122.46 PalaeoDB Santanachelys gaffneyi 112.6 109.0 PalaeoDB Sinemys gamera 129.4 100.5 Rabi et al. (2010) Sinemys lens 163.5 145.0 Rabi et al. (2010) Solnhofia parsoni 155.7 145.5 PalaeoDB; Joyce (2000) Toxochelys latiremis 85.8 70.6 PalaeoDB; Weems & Brown (2017) Xinjiangchelys radiplicatoides 168.3 152.1 Rabi et al. (2010) Xinjiangchelys wusu 163.5 152.1 PalaeoDB

Testing the effect of ‘marine’ characters and advanced protostegids

To test if our phylogenetic results were influenced by the inclusion of characters that might be functionally linked to a marine lifestyle, we performed an additional analysis in which characters that were identified as being indicative of a marine lifestyle were deleted. Furthermore, to test if our phylogenetic results (specifically, the recovery of protostegids being crown‐group chelonioids, see Results below) were driven by the inclusion of advanced taxa with a highly specialized body plan with clear adaptations to pelagic life, as suggested by some authors (e.g. Cadena & Parham 2015), we performed a series of analyses in which only certain protostegids with the most plesiomorphic characters were included in the analyses. In the first of these ‘least specialized taxa analyses’, we only kept Santanachelys gaffneyi as the only representative of protostegids, as done in all studies that previously found protostegids outside Chelonioidea (e.g. Joyce 2007; Danilov & Parham 2008; Anquetin 2012; Sterli & de la Fuente 2013). In our second analysis (least specialized taxa analysis 2), we additionally kept Rhinochelys pulchriceps, for which we had the most detailed cranial information. In a third analysis (least specialized taxa analysis 3), we kept all protostegids but Archelon ischyros and Protostega gigas, which are commonly described as ‘advanced protostegids’ (e.g. Kear & Lee 2006).

Testing alternative hypotheses

To test if our topology was significantly different in tree length compared to alternative hypotheses regarding the placement of key taxa, we used Templeton's non‐parametric test (Templeton 1983) in PAUP* for Macintosh (Swofford 2002). Templeton's test uses the differences in counts of character state transitions for each character implied by a pair of trees. Resulting differences are sorted in ascending order and fractional rank scores are assigned to them. Rank scores for characters that have fewer transitions in the suboptimal tree are summed to produce a test statistic, which is subjected to a Wilcoxon's signed rank test. The default for PAUP* is a two‐tailed test (Goldman et al. 2000). However, because we use Templeton's test with one tree being known to be optimal (the null hypothesis is that constrained topologies fit the data as well as the optimal, i.e. shortest, tree) a one‐tailed test is appropriate (Templeton 1983; Goldman et al. 2000; Hipp et al. 2004). Although p‐values of one‐tailed tests are less conservative, their use has been advocated because they are generally close to the exact probability values for this test (Felsenstein 1985; Lee 2000) and p‐values reported herein were corrected accordingly (see Goldman et al. 2000). Alternative hypotheses were chosen based on phylogenetic results presented in previous studies. Our alternative hypotheses test the position of plesiochelyids, thalassochelydians, sandownids, xinjiangchelyids, sinemydids, protostegids as stem‐group cryptodires and stem‐group chelonioids, respectively, and the position of protostegids as stem‐group turtles. The alternative topologies were recovered by running parsimony analyses including one OTU relevant to each hypothesis into the molecular backbone constraint, while all other fossil OTUs were kept as floaters. For hypotheses relating to species‐rich groups or groups that have not consistently been found to be monophyletic among previous studies, such as sinemydids or thalassochelydians, several constraints using different OTUs were used. Because Templeton's test uses character changes across a tree in a pairwise comparison, individual trees had to be selected from the available MPTs of the original and constraint analyses. To do this, we computed symmetric differences (Robinson & Foulds 1981) between all MPTs from each constraint analysis and all MPTs of our original analysis using the command treedist in PAUP*, which returns a pairwise symmetric differences matrix. The least and most symmetric tree pairs between original and constraint analyses were selected from the matrix for comparison. If several tree pairs matched the least or most symmetry criterion, we randomly chose a tree pair from the symmetric difference matrix for the Templeton's tests (the procedure, including PAUP* and R code, are provided in Evers & Benson (2018b) along with all MPTs and consensus trees generated for alternative hypotheses).


Similarity matrix

Similarity heatmaps show high or low similarity of character states. Most clades show relatively low between‐group similarity, and relatively high within‐group similarity (Fig. 2), congruent with our phylogenetic topology. Relatively high between‐group similarity of some clades that are not close relatives of one another (according to our phylogenetic results) provide evidence of homoplasy, and possibly indicate convergent evolution.

Figure 2 Open in figure viewerPowerPoint Heatmap of symmetric pairwise‐taxon similarity matrix using our full character dataset (n = 345). The phylogenetic tree to the left is identical to our strict consensus tree using unordered character data. Dotted lines delimit turtle subclades for readability. Abbreviations: AC, Angolachelonia; BS, basal stem‐group turtles; CH, Cheloniidae; CY, Chelydroidea; D, Dermochelyidae; M, Meiolaniformes; PC, Paracryptodira; PL, Pleurodira; PR, Protostegidae; SM; Sinemydidae/Macrobaenidae; T, Toxochelys sp.; TE, Testudinoidea; TR, Trionychia, XI, Xinjiangchelyidae. Colour online.

Some of the highest between‐group similarity is seen between protostegids on one hand, and sinemydids, xinjiangchelyids and angolachelonians (i.e. thalassochelydians + sandownids; details are given below) on the other hand (Fig. 2). This pattern has been noted based on qualitative comparisons before (Cadena & Parham 2015), and is according to our matrix driven predominantly by cranial and, to a lesser extent, axial + girdle characters (Fig. 3A, C). Our phylogenetic results strongly support the hypothesis that protostegids are crown‐group chelonioids, deeply nested within Cryptodira. This pattern of similarity therefore indicates a relatively high amount of convergence of the protostegid skull morphology with that of angolachelonians, sinemydids and xinjiangchelyids. In accordance with our topology, protostegid skull morphology is most similar to the skull morphology of cheloniids. However, unlike in the axial + girdle or shell partitions, cranial similarity between protostegids and cheloniids never exceeds moderate values. Paired with the moderate similarity of cranial morphology of protostegids with non‐chelonioid groups, the relatively strong differences of protostegid skulls to cheloniid skulls reveal a high amount of potential character conflict. This might explain the finding of protostegids outside Chelonioidea in some previous studies (e.g. Joyce 2007; Anquetin 2012). Furthermore, the character similarity between sinemydids, xinjiangchelyids, and angolachelonians can be used as an explanation for the unstable positions of these groups among previous phylogenetic analyses (e.g. Rabi et al. 2013; Cadena & Parham 2015; Zhou & Rabi 2015; see Discussion, below).

Figure 3 Open in figure viewerPowerPoint Heatmaps of symmetric pairwise‐taxon similarity matrices of subsets of our characters data. A, cranial character data (n = 185). B, shell character data (n = 87). C, axial and girdle character data (n = 50). D, limb character data (n = 23). Note that NA values result from comparisons of taxa with
PalAss Go! URL: | Twitter: Share on Twitter | Facebook: Share on Facebook | Google+: Share on Google+