doi: 10.1111/jeb.12268
SHORT COMMUNICATION
Turtle origins: insights from phylogenetic retrofitting and
molecular scaffolds
M. S. Y. LEE*†
*Earth Sciences Section, South Australian Museum, North Terrace, Adelaide, SA, Australia
†School of Earth and Environmental Sciences, University of Adelaide, Adelaide, SA, Australia
Keywords:
Abstract
amphibians and reptiles;
Archosauria;
Bayesian inference;
Diapsida;
molecular scaffold;
morphological evolution;
Parareptilia;
parsimony;
phylogenetics;
Testudines.
Adding new taxa to morphological phylogenetic analyses without substantially revising the set of included characters is a common practice, with
drawbacks (undersampling of relevant characters) and potential benefits
(character selection is not biased by preconceptions over the affinities of the
‘retrofitted’ taxon). Retrofitting turtles (Testudines) and other taxa to recent
reptile phylogenies consistently places turtles with anapsid-grade parareptiles
(especially Eunotosaurus and/or pareiasauromorphs), under both Bayesian
and parsimony analyses. This morphological evidence for turtle–parareptile
affinities appears to contradict the robust genomic evidence that extant (living) turtles are nested within diapsids as sister to extant archosaurs (birds
and crocodilians). However, the morphological data are almost equally consistent with a turtle–archosaur clade: enforcing this molecular scaffold onto
the morphological data does not greatly increase tree length (parsimony) or
reduce likelihood (Bayesian inference). Moreover, under certain analytic
conditions, Eunotosaurus groups with turtles and thus also falls within the
turtle–archosaur clade. This result raises the possibility that turtles could
simultaneously be most closely related to a taxon traditionally considered a
parareptile (Eunotosaurus) and still have archosaurs as their closest extant
sister group.
Introduction
One of the major unresolved issues in vertebrate phylogeny concerns the affinities of turtles (Testudines: tortoises, terrapins and sea turtles). Their highly aberrant
morphology means that many aspects of their anatomy
are difficult to compare with other extinct and extant
vertebrates; consequently, turtles have a disproportionate number of debated homologies in morphological
phylogenetic analyses and are often highly unstable
‘wildcard’ taxa (Lee, 2001; Harris et al., 2007). The
advent of molecular systematics has not supported any
of the morphological hypotheses. Several recent genomic-scale molecular studies have robustly nested turtles
within diapsid reptiles, as sister group to archosaurs
Correspondence: M. S. Y. Lee, Earth Sciences Section, South Australian
Museum, North Terrace, Adelaide 5000, SA, Australia.
Tel.: +61 8 8207 7568; fax: +61 8 8207 7422;
e-mail: [email protected]
(birds and crocodilians) among extant taxa (e.g. Chiari
et al., 2012; Crawford et al., 2012; Wang et al., 2013).
Simultaneously however, the most recent morphological
studies (Carroll, 2013; Lyson et al., 2013) place turtles in
a much more basal position, with extinct anapsid-grade
reptiles and thus outside of diapsids altogether.
Early cladistic analyses united turtles with various
groups of anapsid-grade reptiles, initially captorhinids
(e.g. Gauthier et al., 1988), and two groups of parareptiles: procolophonids (e.g. Reisz & Laurin, 1991; Laurin
& Reisz, 1995) and pareiasaurs (e.g. Lee, 1995). However, the assumption that turtles were anapsid-grade
reptiles was strongly challenged by studies placing
them with sauropterygians, a clade of marine diapsids
that includes placodonts, plesiosaurs (Rieppel &
deBraga, 1996; deBraga & Rieppel, 1997). Even
though the strength of the character support for this
arrangement was debated, other lines of evidence
were interpreted as additional independent support for
the sauropterygian hypothesis. First, discovery of the
ª 2013 THE AUTHOR. J. EVOL. BIOL. 26 (2013) 2729–2738
JOURNAL OF EVOLUTIONARY BIOLOGY ª 2013 EUROPEAN SOCIETY FOR EVOLUTIONARY BIOLOGY
2729
2730
M. S. Y. LEE
most primitive known turtle (Odontochelys) revealed
that it was highly aquatic (Li et al., 2008). However,
primitive turtles apart from Odontochelys are highly terrestrial (Joyce & Gauthier, 2004), leading to suggestions
that turtles might be primarily terrestrial, and Odontochelys thus secondarily aquatic (Reisz & Head, 2008).
Second, molecular data – culminating in genomes –
started to place turtles within diapsids with increasing
support (Hedges, 2012). However, the molecular data
place turtles near archosaurian diapsids, whereas the
morphological analyses that unite turtles with sauropterygians place both groups with lepidosaurian diapsids.
There is little morphological support that turtles are
related to archosaurian diapsids (Rieppel, 2000; but see
Bhullar & Bever, 2009). Nevertheless, morphological
hypotheses placing turtles with anapsid groups (and thus
outside of diapsids altogether) were even more incongruent with the emerging molecular consensus. Based on all
these considerations, turtles are currently widely
accepted to be most closely related to archosaurs among
extant taxa (e.g. Hedges, 2012) and, with less certainty,
to be most closely related to sauropterygians among
extinct taxa (e.g. Li et al., 2008; M€
uller & Tsuji, 2007),
although the above contradiction (sauropterygians are
related to lepidosaurs, not archosaurs) has not been
resolved.
This emerging consensus that turtles were somewhere within diapsids has been recently challenged by
the observation that the anapsid-grade parareptile Eunotosaurus uniquely shares with turtles a suite of unique
morphological novelties (Lyson et al., 2010, 2013;
Carroll, 2013). These include a carapace formed by
expanded ribs, a greatly shortened body (trunk) region
consisting of no more than 10 vertebrae, and reduction
in intercostal musculature. These resemblances have
been known for a century (Watson, 1914), but surprisingly, no quantitative phylogenetic analysis had ever
simultaneously included both Eunotosaurus and turtles.
This was partly because of the assumption that Eunotosaurus was nested within parareptiles and turtles were
nested within diapsids. Analyses of anapsid reptiles
(including Eunotosaurus) thus did not consider turtles
(Modesto, 2000; M€
uller & Tsuji, 2007; Tsuji et al.,
2012), whereas analyses of turtle affinities focused on
diapsid reptiles and did not sample minor anapsid
groups such as Eunotosaurus (deBraga & Rieppel, 1997;
Rieppel & Reisz, 1999; Li et al., 2008). When turtles
were eventually added to analyses of anapsid reptiles,
and Eunotosaurus added to analyses of diapsid reptiles, a
robust Eunotosaurus–turtle grouping resulted on both
occasions (Lyson et al., 2010, 2013). Furthermore, the
Eunotosaurus–turtle clade fell within anapsids, nested
within parareptiles, often close to two other candidate
turtle relatives (procolophonids and pareiasaurs). There
thus seemed to be phylogenetic signal placing turtles
with Eunotosaurus, and within anapsid parareptiles, but
outside of Diapsida.
‘Retrofitting phylogenies’ – adding taxa to existing
data matrices without substantially reevaluating the
character list – is a necessarily common practice.
Researchers rarely have time or resources to construct
fresh morphological character lists when describing
new taxa; typically, character lists from previous analyses are employed largely unchanged, perhaps supplemented by a few extra characters relevant to the
‘retrofitted’ taxon. However, the pros and cons of this
approach (to my knowledge) have been rarely discussed (Lee, 1995): there are obvious limitations, but
perhaps also potential benefits. Major errors would
occur if retrofitted taxon does not belong in the analysis (ingroup) at all, but such problems are unlikely
except in the most poorly known groups or for highly
incomplete taxa. A recent example concerns a fragmentary pterosaur mandible (Buffetaut, 2011), which
emerged as a ‘bird’ when inappropriately included in a
data set focused on avian and nonavian theropods
(Naish et al., 2012). Also, such errors are not directly
related to retrofitting taxa per se; rather, they result
from incorrect ingroup–outgroup assumptions and can
affect any analysis. Another potential problem concerns
the unavoidable subjectivity of many morphological
character states (e.g. ‘large’ vs. ‘small’); adding taxa to a
previous analysis is problematic if different workers
have different interpretations of character states. Ultimately, though, this problem is again not particular to
retrofitted phylogenies, but rather affects all morphological analyses: character states should ideally be
defined so scorings can be replicated by other workers.
An ideal solution would entail precise definitions (e.g.
numerical ratios for quantitative traits) and illustrations
of the states observed and scored in every single terminal taxon (as done by O’Leary et al. 2013).
A limitation more specific to ‘retrofitting’ involves
undersampling of informative characters relevant to the
phylogenetic position of the added taxon. Phylogenetic
data sets typically consider only parsimony-informative
phenotypic traits, where the derived state is present in
two or more terminal taxa (Yeates, 1992). Thus, the
character sampling in such analyses would typically
exclude traits where the derived condition is either
(i) present only in retrofitted taxa (omitted as invariant
in original study) or (ii) present only in the retrofitted
taxa and a single original terminal taxon (omitted as
autapomorphies in original study). Adding taxa to morphological matrices can make turn many previously
invariant or autapomorphic characters into parsimonyinformative characters; such traits would be undersampled if taxa are retrofitted without expanding the pool
of relevant characters.
Paradoxically, the character bias introduced by ‘retrofitted taxa’ is also potentially illuminating. It is difficult
or impossible to objectively sample and subdivide morphological characters, and many morphological studies
have been argued to overly favour characters supporting
ª 2013 THE AUTHOR. J. EVOL. BIOL. 26 (2013) 2729–2738
JOURNAL OF EVOLUTIONARY BIOLOGY ª 2013 EUROPEAN SOCIETY FOR EVOLUTIONARY BIOLOGY
Turtle relationships
one particular phylogenetic hypothesis, for example, by
oversampling or oversplitting such traits. It is notable
that workers with particular interests in procolophonids
(Reisz & Laurin, 1991), pareiasaurs (Lee, 1995) and
sauropterygians (Rieppel & Reisz, 1999) each considered their focal taxon to be related to turtles – raising
the possibility that character choice in these studies
favoured synapomorphies linking turtles with (respectively) procolophonids, pareiasaurs and sauropterygians.
In fact, as all biologists are aware of the basic branching
pattern of the tree of life, it would be difficult to
exclude phylogenetic preconceptions when choosing
characters for any morphological phylogenetic analysis.
This inherent bias might be reduced in retrofitted phylogenies. For instance, the first data set evaluated here
was originally constructed to ascertain the position of
turtles within diapsids (Li et al., 2008), without any
consideration of Eunotosaurus. Thus, it is unlikely that
there could have been any conscious bias favouring
characters uniting Eunotosaurus with any particular taxa
(such as turtles or parareptiles). Similarly, the second
data set was designed to resolve affinities within parareptiles, with turtles being explicitly excluded (Tsuji
et al., 2012). It is thus unlikely that there was any bias
favouring traits uniting turtles with any particular parareptile taxon (such as Eunotosaurus). Thus, affinities of
the retrofitted taxon revealed in such analyses are unlikely to be the result of biased character selection introduced by preconceptions over the ‘correct’ position of
that taxon; rather, the phylogenetic signal for the retrofitted taxa can be considered to have emerged despite
potential undersampling of relevant characters.
Retrofitting taxa to two (relatively) independent data
sets has retrieved a turtle–Eunotosaurus clade (Lyson
et al., 2010, 2013). However, two important and interrelated questions remain. First, what are the relationships of other parareptiles to the Eunotosaurus–turtle
clade, given that many of these parareptiles have been
historically linked to Eunotosaurus and/or turtles? Second, how can the Eunotosaurus–turtle clade be reconciled with the robust molecular evidence that extant
turtles are sister to extant archosaurs? These questions
are investigated here by the evaluation of updated versions of the two ‘retrofitted’ data sets, and enforcing
molecular scaffolds (Springer et al., 2001) to test
whether the morphological data are truly incompatible
with the genomic data.
Materials and methods
Morphological character matrices
Two retrofitted matrices were analysed: Eunotosaurus
was added to a diapsid-focused data set, and turtles
were added to an anapsid-focused data set. Although
these data sets are similar to previous studies (Lyson
et al., 2010, 2013), the changes described below
2731
required new analyses to be performed. Each data set
was analysed both with original character sampling (i.e.
the original character list, constructed without any reference to the retrofitted taxa) and with character
sampling slightly broadened to reflect information
introduced by the new taxa (i.e. inclusion of new relevant characters, but potentially influenced by phylogenetic preconceptions about the retrofitted taxa). It is
notable, however, that most of the new characters
introduced related to similarities between Eunotosaurus
and turtles; there was no similar effort to identify additional relevant characters supporting alternative (e.g.
pareiasaur or procolophonid) affinities.
The diapsid-focused data sets (Diapsid169 and Diapsid189) are based on an amniote data set of 169 characters with a broad sampling of diapsids, which originally
placed turtles within diapsids as sister group to sauropterygians (Li et al., 2008). The (putative) parareptile
Eunotosaurus was added to this data set, and turtle and
diapsid taxon sampling slightly expanded with the addition of the stem-turtle Proganochelys and the armoured
sauropterygian Sinosaurosphargis (Lyson et al., 2010,
2013). The latest version of this retrofitted matrix
(Lyson et al., 2013) was analysed, first including just
the 169 characters in the original diapsid-focused
matrix (Li et al., 2008) and then with an additional 20
newly identified characters bringing the total to 189.
These retrofitted data sets, with the original 169 characters and the expanded 189 characters, are here termed
Diapsid169 and Diapsid189, respectively. The 20 new
characters were based on the 22 identified by Lyson
et al. (2010, 2013). Two of the 22 characters were omitted, and another redefined, due to likely redundancy
(e.g. number of dorsal vertebrae and number of dorsal
rib pairs are here treated as a single character rather
than two). The new list of 20 extra characters, and
changes to the previous set of 22, are listed in Table S1,
and the full matrix is provided in Table S2.
The anapsid-focused data sets (Anapsid136 and Anapsid154) are based on a very recent amniote data set of
136 characters with a broad sampling across anapsids,
especially parareptiles (Tsuji et al., 2012). Turtles were
not included in this matrix; thus, Odontochelys and Proganochelys were added to this data set using codings
taken directly from Lyson et al. (2010). The latter study
scored these basal turtles for these exact characters,
when ‘retrofitting’ them to an earlier version of the
anapsid data set (M€
uller & Tsuji, 2007). The anapsid
matrix was analysed using only the original 136 characters and with the addition of 18 newly identified characters (bringing the total to 154). These new characters
represented 18 of the 20 added to the diapsid data sets
(the other 2 were already in the original anapsid
matrix). These 18 characters were scored for most terminal taxa by Lyson et al. (2010, 2013); however, some
lesser-known parareptile taxa in the anapsid matrix had
to be coded for these characters using the primary
ª 2013 THE AUTHOR. J. EVOL. BIOL. 26 (2013) 2729–2738
JOURNAL OF EVOLUTIONARY BIOLOGY ª 2013 EUROPEAN SOCIETY FOR EVOLUTIONARY BIOLOGY
2732
M. S. Y. LEE
literature and personal observations. The list of 18 extra
characters, with anatomical sources, is provided in
Table S1, and the full matrix is provided in Table S2.
Phylogenetic analyses
Each of the two data sets with original and expanded
character sampling (Diapsid169, Diapsid189, Anapsid136,
Anapsid154) was analysed using both Bayesian inference and maximum parsimony, to test whether the
results were sensitive to analytic methods. Seymouriamorphs were treated as the most distant outgroup (sister to all other taxa). To facilitate comparisons
with previous analyses of earlier versions of these data
sets, all multistate characters were treated as unordered.
However, essentially identical results were obtained
when multistate characters were treated as ordered (if
they formed clear linear morphoclines), an approach
that has been argued to improve phylogenetic accuracy
(Wiens, 2001; Grand et al., 2013). The Bayesian and
parsimony analyses were also repeated for the Diapsid169 and Diapsid189 data sets with molecular scaffolds
(i.e. turtle–archosaur clade) enforced, to test whether
the morphological data were significantly incongruent
with the molecular data (see below).
Bayesian inference used MrBayes 3.2 (Ronquist et al.,
2012), with the Lewis (2001) stochastic morphological
model and correction for sampling of only variable characters. Model selection used stepping-stone analyses (Xie
et al., 2011) to infer Bayes factors (sensu Kass and Raftery
1995), that is, twice the difference in marginal lognLikelihoods (BFKR). There was ‘decisive’ (BFKR > 10) support
for inclusion of the gamma parameter for rate variation
across characters for all data sets (BFKR: Diapsid169 = 15.0, Diapsid189 = 33.8, Anapsid136 = 15.68,
Anapsid154 = 24.22). To ensure convergence (stationarity), each analysis consisted of four replicate runs of
50 million steps, with sampling every 5000 steps. Each
of the four replicate runs consisted of one unheated and
three incrementally heated chains (temperature 0.2).
Convergence was reached well before the burn-in of
10 million (20%): convergence in parameters was diagnosed by essentially identical sampled distributions
across runs (potential scale reduction factor ~ 1); convergence in topology was diagnosed by similar clade (split)
frequencies across runs (standard deviation of clade frequencies < 0.05). A majority-rule consensus tree was
constructed from the combined (post-burn-in) samples
of all four runs. Exact MrBayes settings used are
appended at the end of each datafile. Analyses were run
on the Tizard computer grid at eResearch SA (http://
www.eresearchsa.edu.au/supercomputers).
Parsimony analyses used PAUP* (Swofford, 2003),
with most parsimonious trees (MPTs) found via heuristic searches involving 100 random addition searches
(followed by a strict consensus, if > 1 tree), and bootstrapping involving 1000 replicates (followed by a
majority-rule consensus). In order to prevent any
bootstrap replicate from being trapped on islands with
large numbers of MPTs, nchuck was set to 1000. Exact
PAUP settings used are appended at the end of each
datafile.
Molecular scaffolds
These analyses explicitly address the question: if one
accepts that extant turtles are related to extant archosaurs, as robustly supported by genomic data (see
above), where do the fossils best fit within such a
molecular scaffold (Lee, 2013)? The diapsid data sets
sampled sufficient extant (living) amniote lineages to
investigate such interactions with molecular phylogenies: extant turtles are represented by Testudines,
extant archosaurs by Archosauriformes, extant lepidosaurs by Squamata and Rhynchocephalia and extant
mammals by Cynodontia. Accordingly, the Diapsid169
and Diapsid189 matrices were re-analysed with relationships among extant taxa constrained to the pattern
robustly supported by genomic data: [Cynodontia
(Testudines, Archosauriformes) (Squamata, Rhynchocephalia)]. All other (i.e. fossil) taxa were unconstrained and allowed to be placed by the morphological
data into their optimal position within this backbone
constraint. The Bayesian and parsimony analyses were
repeated with such backbone constraints; all other settings remained unchanged. The nonparametric test of
Templeton (1983) and Bayes factors (Kass & Raftery,
1995; see above) were used to test whether enforcing
the molecular backbone constraints resulted in significantly poorer fit to the morphological data.
The anapsid data sets did not sample any extant diapsid lineages (i.e. no archosaurs or lepidosaurs), so a
molecular scaffold enforcing an archosaur–turtle sistergroup relationship could not be applied.
Results
The discussion focuses on the unordered analyses, to
facilitate comparison with recent studies that have all
used this approach (e.g. Lyson et al., 2010, 2013; Tsuji
et al., 2012). Ordered analyses yielded very similar (often
identical) trees in all analyses. In all the diapsid analyses,
after ordering the characters, the Bayesian consensus
trees remained identical and the parsimony consensus
trees remained fully consistent (no conflicting clades;
differences only consisted of one clade in the ordered
analysis being unresolved in the unordered analysis or
vice versa). In all the anapsid analyses, after ordering the
characters, the Bayesian trees either remained identical
(Anapsid154) or differed in only a single branch
(Anapsid136); the parsimony consensus trees remained
identical. Unless noted otherwise, all discussion and
metrics refer to the unordered analyses; relevant differences in the ordered analyses are specifically noted.
ª 2013 THE AUTHOR. J. EVOL. BIOL. 26 (2013) 2729–2738
JOURNAL OF EVOLUTIONARY BIOLOGY ª 2013 EUROPEAN SOCIETY FOR EVOLUTIONARY BIOLOGY
Turtle relationships
The results of all (unordered) analyses are summarized in Figs 1 and 2 (topology) and Table 1 (tree statistics). Trees and support values for individual analyses
are discussed below and depicted in Figs S1–S3. In the
Diapsid169 and Diapsid189 data sets (Fig. 1), the position
of turtles, and to a lesser extent Eunotosaurus and sauropterygians, was the most unstable across analyses;
interrelationships between other taxa remained essentially unchanged. Retrofitting Eunotosaurus and other
taxa to the original diapsid-focused matrix, without
increasing character sampling (Diapsid169), did not alter
the original basic topology: a turtle–sauropterygian
clade was obtained in all analyses (without or with a
molecular scaffold). However, adding 20 additional
characters (Diapsid189) tended to result in a turtle–Eunotosaurus clade within parareptiles. In both anapsidfocused data sets (Anapsid136, Anapsid154), retrofitting
of turtles always resulted in them being placed somewhere within parareptiles (Fig. 2). In three analyses,
turtles grouped with (or within) pareiasauromorphs; in
the other analyses, they grouped with Eunotosaurus.
Turtles never grouped with diapsids.
Diapsid169 (diapsid matrix with retrofitted
Eunotosaurus)
Bayesian inference and parsimony gave almost identical
results (Fig. S1a), placing turtles with sauropterygians
and lepidosauromorphs (Bayesian PP = 0.99, parsimony
bootstrap = 28%). In particular, turtles grouped with
the armoured sauropterygian Sinosaurosphargis, but with
poor support (PP = 0.73, bs = 44%). Eunotosaurus was
placed within a monophyletic Parareptilia (PP = 0.99,
bs = 26%). Thus, addition of Eunotosaurus and additional diapsids did not greatly change topology from the
original study (Li et al., 2008).
Diapsid169 with molecular scaffold (constraining
turtles to be sister to archosaurs among extant
taxa)
Bayesian inference and parsimony gave almost identical
results (Fig. S1b), again retrieving a turtle–sauropterygian
clade but moving it from lepidosauromorphs to
archosauromorphs (pp = 0.54, bs = 22%), in accordance
with the molecular scaffold. Eunotosaurus remained
within a monophyletic Parareptilia (pp = 0.94, bs =
30%). Enforcing the scaffold increased tree length by
only two steps (P > 0.48) and fractionally decreased the
-LnL by 0.71 (BFKR = 1.42).
Diapsid189 (diapsid matrix with retrofitted
Eunotosaurus and 20 new informative characters)
Bayesian inference and parsimony gave similar results
(Fig. S2a), uniting turtles with Eunotosaurus (pp = 0.95,
bs = 69%), placing both within a monophyletic
2733
Parareptilia (pp = 0.95, bs = 29%). Bayesian inference
placed the Eunotosaurus–turtle clade basally within
parareptiles, above millerettids, but parsimony united
this clade with pareiasaurs. Sauropterygians remained
with lepidosauromorphs (pp = 0.94, bs = 22%), as per
the original analysis (Li et al., 2008). Thus, addition of
extra characters as well as taxa was required to break
up the turtle–sauropterygian clade found in the original
study.
Diapsid189 with molecular scaffold (constraining
turtles to be sister to archosaurs among extant
taxa)
Bayesian inference and parsimony gave conflicting
results. Bayesian inference (Fig. S2b) placed turtles
alone with archosaurs (pp = 0.53; pp is < 1.0 because
fossil turtles and fossil archosauromorph groups were
left unconstrained). Neither sauropterygians nor Eunotosaurus fell with turtles and archosaurs: sauropterygians
grouped with lepidosauromorphs (pp = 0.65) and Eunotosaurus fell within a monophyletic Parareptilia
(pp = 0.99). In contrast, parsimony (Fig. S2c) united
turtles, Eunotosaurus and sauropterygians (bs = 57%)
and placed this clade with archosauromorphs
(bs = 22%). Enforcing the scaffold increased tree length
by only six steps (P > 0.41) and increased the -LnL by
4.9 (BFKR = 9.8; a ‘strong’ but not ‘decisive’ decrease in
fit [sensu Kass and Raftery (1995)].
Anapsid136 (anapsid matrix with retrofitted turtles)
Bayesian inference and parsimony gave similar results
(Fig. S3a), either grouping turtles as sister to pareiasaurs and nycteroleterids (Bayesian; pp = 0.61) or to
pareiasaurs alone (parsimony; bs = 54); the latter topology was also retrieved under Bayesian inference when
characters were ordered (pp = 0.55). Eunotosaurus
grouped with millerettids (pp = 0.34, bs = 56%) and
together formed the basal parareptile clade, not particularly close to turtles.
Anapsid154 (anapsid matrix with retrofitted turtles
and 18 additional characters)
Bayesian inference and parsimony gave conflicting
results. Bayesian inference (Fig. S3b) now grouped turtles with Eunotosaurus (Bayesian; pp = 0.79) and placed
this clade with millerettids basally within parareptiles.
Parsimony continued to unite turtles with pareiasaurs
(Fig. S3c), but the newly added characters reduced the
support for this clade (bs = 32%).
Discussion
The variable position of turtles (and to a lesser extent
Eunotosaurus) across these analyses contrasts with the
ª 2013 THE AUTHOR. J. EVOL. BIOL. 26 (2013) 2729–2738
JOURNAL OF EVOLUTIONARY BIOLOGY ª 2013 EUROPEAN SOCIETY FOR EVOLUTIONARY BIOLOGY
2734
M. S. Y. LEE
(a)
Millerettidae
Ankyromorpha
Archosauromorpha
Lepidosauromorpha
(pareiasaurs,
procolophonoids etc)
(crocodylians, dinosaurs,
birds etc)
(squamates,
rhynchocephalians etc)
1
2
Parareptilia
Stem
3
a
psid
Dia
Diapsida
(crown)
(b)
1
2
3
Fig. 1 The phylogenetic results when Eunotosaurus is retrofitted to an analysis focused on diapsid reptiles and turtles (matrix modified from
Li et al. (2008) and Lyson et al. (2013)). (a) Schematic phylogeny showing three alternative positions for turtles, Eunotosaurus and
sauropterygians: with ankyromorph anapsids, with archosauromorph diapsids and with lepidosauromorph diapsids. (b) Table showing the
actual position(s) of turtles, Eunotosaurus and sauropterygians in 8 (2 9 2 9 2) different analyses: the original (Diapsid169) and expanded
(Diapsid189) data set, under Bayesian inference (BI) and maximum parsimony (MP), without and with a molecular scaffold. Eunotosaurus
falls with parareptiles in most analyses, but the position of turtles and sauropterygians is more variable.
ª 2013 THE AUTHOR. J. EVOL. BIOL. 26 (2013) 2729–2738
JOURNAL OF EVOLUTIONARY BIOLOGY ª 2013 EUROPEAN SOCIETY FOR EVOLUTIONARY BIOLOGY
Turtle relationships
2735
Millerettidae
Procolophonoidea
(a)
Pareiasauromorpha
Diapsida
(pareiasaurs,
nycteroleters etc)
(archosauromorphs
lepidosauromorphs etc)
1
Diapsida
2
3
Ankyramorpha
Parareptilia
(b)
1
2
3
Fig. 2 The phylogenetic results when turtles are retrofitted to an analysis focused on parareptiles (matrix modified from Tsuji et al., 2012
and Lyson et al., 2010). (a) Schematic phylogeny showing three alternative positions for turtles and Eunotosaurus: with millerettid
parareptiles, with pareiasauromorph (ankyromorph) parareptiles and with diapsids. (b) Table showing the actual position(s) of turtles and
Eunotosaurus and sauropterygians in four different analyses: the original (Anapsid136) and expanded (Anapsid154) data set, under Bayesian
inference (BI) and maximum parsimony (MP). There was insufficient sampling of diapsids to employ a molecular scaffold. Eunotosaurus
falls with parareptiles in all analyses; the position of turtles varies, but they always fall within parareptiles.
relatively stable position of the remaining amniote taxa,
including other potential relatives of turtles. Sauropterygians always nest within diapsids, and procolophonids
and pareiasaurs always nest within parareptiles. These
analyses therefore highlight the particularly unstable
position of turtles within amniote phylogeny (Harris
ª 2013 THE AUTHOR. J. EVOL. BIOL. 26 (2013) 2729–2738
JOURNAL OF EVOLUTIONARY BIOLOGY ª 2013 EUROPEAN SOCIETY FOR EVOLUTIONARY BIOLOGY
2736
M. S. Y. LEE
Table 1 Tree statistics from analyses of all data sets. Note that
adding a molecular scaffold to either the Diapsid169 or the
Diapsid189 (constraining turtles to be sister to archosaurs among
extant taxa) does not greatly decrease the fit of the morphological
data: the Bayesian harmonic mean – LnL, and the parsimony tree
length barely increase; these changes are not statistically
significant (see main text).
Data set
analysed
Bayesian inference
(harmonic mean – LnL)
Parsimony
(length & number
of trees)
Diapsid169
Diapsid169
(with scaffold)
Diapsid189
Diapsid189
(with scaffold)
Anapsid136
Anapsid154
2529.87
2529.16
869; 15 trees
871; 15 trees
2693.00
2697.92
906; 2 trees
912; 5 trees
1685.58
1763.28
497; 4 trees
520; 4 trees
et al., 2007) and also some instability of Eunotosaurus.
There is often weak support for any particular hypothesis of turtle affinities, but certain important patterns are
revealed.
First, the hypothesis that turtles are related to parareptiles, first proposed cladistically over two decades
ago (Reisz & Laurin, 1991) and recently revived
(Carroll, 2013; Lyson et al., 2013), is at least as strongly
supported as the currently dominant assumption of
diapsid affinities (e.g. Li et al., 2008; Ruta et al., 2011).
Turtles nest within parareptiles in all analyses of the
anapsid-focused data set and in some analyses of the
diapsid-focused data set (i.e. analyses of Diapsid189
without a molecular scaffold). Thus, the suggestion that
parareptile affinities of turtles can now be abandoned
(Tsuji et al., 2012) might be premature. Different analytic approaches also yield slightly different topologies.
Bayesian inference retrieves a wide range of positions
for turtles (varying across all reptiles in the diapsidfocused data sets and varying across parareptiles in the
anapsid-focused data sets). Parsimony, in contrast,
tends to specifically support a position for turtles near
pareiasauromorphs. Parsimony analyses of the Anapsid136, Anapsid154 data sets retrieve turtles and pareiasaurs as sister groups (to the exclusion of all other taxa,
including Eunotosaurus), whereas maximum parsimony
(MP) analysis of the Diapsid189 matrix places pareiasaurs as sister to the Eunotosaurus–turtle clade (see Fig.
S3 in Lyson et al., 2013). In the parsimony analysis of
the full anapsid-focused data set (Anapsid154), pareiasaurs and turtles are united by 8 unambiguous (optimization-independent) synapomorphies, of which the
most notable are frontal excluded from orbit (character
2), 20 or fewer presacrals (89), supinator process indistinct (108), reduced manual (113) and pedal (128) phalangeal formula, and dermal armour (130; present in
turtles as peripheral bones, regardless of the homologies
of other shell elements). In the parsimony analysis of
the full diapsid-focused data set (Diapsid189), pareiasaurs are united with turtles and Eunotosaurus via 7
unambiguous synapomorphies, including choana (character 1) and suborbital foramen (74) positioned medially, wide frontal (26), 20 or fewer presacrals (97),
chevrons articulating with 1 centrum, rather than
intercentral (112), enclosed ectepicondylar foramen
(127).
Second, the molecular scaffold analyses (enforcing
turtles+archosaurs) reveal less conflict between the
genomic and morphological data than often proposed
(e.g. Lyson et al., 2010; Hedges, 2012). Enforcing a turtle–archosaur clade to the Diapsid169 and Diapsid189
matrices barely decreases fit in either parsimony or
Bayesian analyses (Table 1); thus, the morphological
data cannot refute (and thus is consistent with) the
genomic evidence. However, enforcing this turtle–
archosaur clade does result in a highly unstable position
for Eunotosaurus. Eunotosaurus remains with parareptiles
in most ‘scaffolded’ analyses, but one analysis places
Eunotosaurus as sister to turtles within the turtle–archosaur
clade. This raises the possibility that turtles could be
most closely related to a taxon traditionally considered a
‘parareptile’ and also simultaneously have archosaurs as
their extant sister group. This result in turn raises the
possibility that other reptiles traditionally considered
‘parareptiles’ (e.g. millerettids, pareiasaurs and procolophonids) could also occupy a similar position, if new
studies uncover sufficient synapomorphies with turtles
and/or Eunotosaurus. The anapsid analyses consistently
reveal a signal linking turtles with pareiasaurs and procolophonids, and addition of these characters to the
diapsid data sets might shift the affinities of these taxa.
Furthermore, many parareptiles have at least a lower
temporal fenestra (Modesto et al., 2009), which might be
primitive for parareptiles or crown amniotes (Pi~
neiro
et al., 2012). It is worth highlighting that if turtles nest
within archosauromorphs, they must have secondarily
closed both upper and lower temporal fenestrae at some
stage. This could have happened prior to, coincident
with, or subsequent to, the acquisition of the shell; currently known basal turtle fossils (and Eunotosaurus) have
both anapsid skulls and (most) shell characters and thus
do not elucidate the relative order of character acquisition. If the skull openings closed before the shell
evolved, this would entail the existence of unarmoured
reptiles with varying reductions in skull fenestration
along the turtle stem (and thus within Archosauromorpha).
Retrofitting taxa to existing morphological data
matrices is a necessarily common practice in systematics, which can be problematic (undersampling of relevant characters) as well as illuminating (character
selection cannot be biased by preconceptions about the
added taxon). Adding Eunotosaurus to diapsid analyses,
ª 2013 THE AUTHOR. J. EVOL. BIOL. 26 (2013) 2729–2738
JOURNAL OF EVOLUTIONARY BIOLOGY ª 2013 EUROPEAN SOCIETY FOR EVOLUTIONARY BIOLOGY
Turtle relationships
and turtles to anapsid analyses, suggests that there is
phylogenetic signal uniting turtles with different groups
of parareptiles, which is revealed by even an apparently
disinterested survey of the character evidence. The original diapsid matrix (Diapsid169) did not consider Eunotosaurus: it is thus unlikely that preconceptions about
the position of Eunotosaurus biased the selection and
coding of the 169 characters used in this study. The
placement of Eunotosaurus within parareptiles in all
analyses of Diapsid169 matrix presumably reflects genuine phylogenetic signal rather than biased character
selection. In that matrix, however, turtles remain
embedded within diapsids: addition of extra characters
shared by turtles and Eunotosaurus (Diapsid189) moves
turtles to group with Eunotosaurus inside parareptiles.
Similarly, the original anapsid-focused matrix (Anapsid136) was aimed at resolving the relationships
between parareptiles, and explicitly ignored turtles
(Tsuji et al., 2012). Nevertheless, when turtles were
added to this data set, they emerged within parareptiles, under both Bayesian inference and parsimony
(see also Lyson et al., 2010). Within parareptiles, however, turtles group with ankyromorphs (pareiasaurs and
procolophonids), rather than Eunotosaurus. Addition of
extra characters (Anapsid154) causes turtles to group
with Eunotosaurus under Bayesian inference, but turtles
still group with pareiasaurs and procolophonids under
parsimony. The persistent grouping of turtles with pareiasaurs and procolophonids is notable, given that added
characters (mostly similarities between Eunotosaurus
and turtles) are potentially biased against this relationship. However, the support for all these hypotheses is
relatively weak (Figs S1–S3), as measured by posterior
probabilities (Bayesian inference) or bootstrap
(parsimony).
Retrofitting phylogenies therefore does not conclusively resolve the position of turtles with respect to
their putative fossil relatives. However, it reveals that
the morphological data cannot refute the genomic evidence that turtles are most closely related to archosaurs
among extant taxa – regardless of the exact position of
Eunotosaurus, pareiasaurs and procolophonids. It also
suggests that there are sufficient putative synapomorphies uniting turtles with disparate groups of parareptiles to warrant continuing investigation (Carroll, 2013;
Lyson et al., 2013). The position of these taxa with
respect to turtles and archosaurs remains to be
resolved; one intriguing possibility is that some parareptiles could be nested within diapsids, as sister to turtles
within a turtle–archosaur clade.
Acknowledgments
I thank Sean Reilly and e-Research SA (www.
eresearchsa.edu.au/supercomputers) for grid computing resources and assistance, T. Lyson for discussion
and original matrices from Lyson et al. (2013) and
2737
M. Laurin and an anonymous reviewer for constructive
comments. This research was funded by the
Australian Research Council and The Environment
Institute (University of Adelaide).
References
Bhullar, B.-A.S. & Bever, G.S. 2009. An archosaur-like laterosphenoid in early turtles (Reptilia: Pantestudines). Breviora
518: 1–11.
deBraga, M. & Rieppel, O. 1997. Reptile phylogeny and the
interrelationships of turtles. Zool. J. Linn. Soc. 120: 281–354.
Buffetaut, E. 2011. Samrukia nessovi, from the Late Cretaceous
of Kazakhstan: a large pterosaur, not a giant bird. Ann.
Paleontol. 97: 133–138.
Carroll, R.L. 2013. Problems of the ancestry of turtles. In: Morphology and Evolution of Turtles (D.B. Brinkman, P.A. Holroyd
& J.D. Gardner, eds), pp. 19–36. Springer, Dordrecht.
Chiari, Y., Cahais, V., Galtier, N. & Delsuc, F. 2012. Phylogenomic analyses support the position of turtles as the sister
group of birds and crocodiles (Archosauria). BMC Biol. 10:
e65.
Crawford, N.G., Faircloth, B.C., McCormack, J.E., Brumfield,
R.T., Winker, K. & Glenn, T.C. 2012. More than 1000 ultraconserved elements provide evidence that turtles are the sister group of archosaurs. Biol. Lett. 8: 783–786.
Gauthier, J., Kluge, A.G. & Rowe, T. 1988. The early evolution
of the Amniota. In: The phylogeny and classification of the tetrapods, Volume 1: amphibians, reptiles, birds (M.J. Benton, ed.),
pp. 103–155. Clarendon Press, Oxford.
Grand, A., Corvez, A., Duque Velez, L.M. & Laurin, M. 2013.
Phylogenetic inference using discrete characters: performance of ordered and unordered parsimony and of threeitem statements. Biol. J. Linn. Soc. doi: 10.1111/bij.12159.
Harris, S.R., Pisani, D., Gower, D.J. & Wilkinson, M. 2007.
Investigating stagnation in morphological phylogenetics
using consensus data. Syst. Biol. 56: 125–129.
Hedges, S.B. 2012. Amniote phylogeny and the position of turtles. BMC Biol. 10: e64.
Joyce, W.G. & Gauthier, J.A. 2004. Palaeoecology of Triassic
stem turtles sheds new light on turtle origins. Proc. R. Soc.
Lond. B 271: 1–5.
Kass, R.E. & Raftery, A. E. 1995. Bayes Factors. J. Amer. Stat.
Assoc. 90: 773–795.
Laurin, M. & Reisz, R.R. 1995. A reevaluation of early amniote
phylogeny. Zool. J. Linn. Soc. 113: 165–223.
Lee, M.S.Y. 1995. Historical burden in systematics and the
interrelationships of ‘parareptiles’. Biol. Rev. 70: 459–547.
Lee, M.S.Y. 2001. Reptile relationships turn turtle. Nature 389:
245.
Lee, M.S.Y. 2013. Turtles in transition. Curr. Biol. 23:
R513–R515.
Lewis, P.O. 2001. A likelihood approach to estimating phylogeny from discrete morphological character data. Syst. Biol.
50: 913–925.
Li, C., Wu, X.-C., Rieppel, O., Wang, L.-T. & Zhao, L.J. 2008.
An ancestral turtle from the Late Triassic of southwestern
China. Nature 456: 497–501.
Lyson, T.R., Bever, G.S., Bhullar, B.-A.S., Joyce, W.G. &
Gauthier, J.A. 2010. Transitional fossils and the origin of turtles. Biol. Lett. 6: 830–833.
ª 2013 THE AUTHOR. J. EVOL. BIOL. 26 (2013) 2729–2738
JOURNAL OF EVOLUTIONARY BIOLOGY ª 2013 EUROPEAN SOCIETY FOR EVOLUTIONARY BIOLOGY
2738
M. S. Y. LEE
Lyson, T.R., Bever, G.S., Scheyer, T.M., Hsiang, A.Y. &
Gauthier, J.A. 2013. Evolutionary origin of the turtle shell.
Curr. Biol. 23: 1113–1119.
Modesto, S.P. 2000. Eunotosaurus africanus and the Gondwanan
ancestry of anapsid reptiles. Palaeont. Africana 36: 15–20.
Modesto, S.P., Scott, D.M. & Reisz, R.R. 2009. A new parareptile with temporal fenestration from the Middle Permian of
South Africa. Can. J. Earth Sci. 46: 9–20.
M€
uller, J. & Tsuji, L.A. 2007. Impedance-matching hearing in
Paleozoic reptiles: evidence of advanced sensory perception
at an early stage of amniote evolution. PLoS ONE 2: e889.
Naish, D., Dyke, G., Cau, A., Escuillie, F. & Godefroit, P. 2012.
A gigantic bird from the Upper Cretaceous of Central Asia.
Biol. Lett. 8: 97–100.
O’Leary, M.A., Bloch, J.I., Flynn, J.J., Gaudin, T.J. Giallombardo, A., & Giannini, N.P. et al. 2013. The placental mammal ancestor and the post-K-Pg radiation of placentals.
Science 339: 662–667.
Pi~
neiro, G., Ferigolo, J., Ramos, A. & Laurin, M. 2012. Cranial
morphology of the Early Permian mesosaurid Mesosaurus tenuidens and the evolution of the lower temporal fenestration
reassessed. C. R. Palevol. 11: 379–391.
Reisz, R.R. & Head, J.J. 2008. Turtle origins out to sea. Nature
456: 450–451.
Reisz, R.R. & Laurin, M. 1991. Owenetta and the origin of turtles. Nature 349: 324–326.
Rieppel, O. 2000. Turtles as diapsid reptiles. Zool. Scr. 29:
199–212.
Rieppel, O. & deBraga, M. 1996. Turtles as diapsid reptiles.
Nature 384: 453–455.
Rieppel, O. & Reisz, R.R. 1999. The origin and evolution of
turtles. Annu. Rev. Ecol. Syst. 30: 1–22.
Ronquist, F., Teslenko, M., van der Mark, P., Ayres, D.L.,
Darling, A., H€
ohna, S. et al. 2012. MrBayes 3.2: efficient
Bayesian phylogenetic inference and model choice across a
large model space. Syst. Biol. 61: 539–542.
Ruta, M., Cisneros, J.C., Liebrecht, T., Tsuji, L.A. & M€
uller, J.
2011. Amniotes through major biological crises: faunal turnover among parareptiles and the end-Permian mass extinction. Palaeontology 54: 1117–1137.
Springer, M.S., Teeling, E.C., Madsen, O., Stanhope, M.J. & de
Jong, W.W. 2001. Integrated fossil and molecular data
reconstruct bat echolocation. Proc. Natl. Acad. Sci. USA 98:
6241–6246.
Swofford, D.L. 2003. PAUP*: Phylogenetic Analysis Using Parsimony (*and Other Methods). Version 4.0. Sinauer Associates,
Sunderland MA.
Templeton, A.R. 1983. Phylogenetic inference from restriction
endonuclease cleavage site maps with particular reference
to the evolution of humans and apes. Evolution 37: 221–
244.
̈
Tsuji, L.A., Muller, J. & Reisz, R.R. 2012. Anatomy of Emeroleter levis and the phylogeny of the nycteroleter parareptiles.
J. Vert. Paleont. 32: 45–67.
Wang, Z., Pascual-Anaya, J., Zadissa, A., Li, W., Niimura, Y.,
Huang, Z. et al. 2013. The draft genomes of soft-shell turtle
and green sea turtle yield insights into the development and
evolution of the turtle-specific body plan. Nat. Genet. 45:
701–706.
Watson, D.M.S. 1914. Eunotosaurus africanus (Seeley) and the
ancestors of the Chelonia. Proc. Zool. Soc. London 11:
1011–1020.
Wiens, J.J. 2001. Character analysis in morphological phylogenetics: problems and solutions. Syst. Biol. 50: 689–699.
Xie, W., Lewis, P.O., Fan, Y., Kuo, L. & Chen, M.-H. 2011.
Improving marginal likelihood estimation for Bayesian phylogenetic model selection. Syst. Biol. 60: 150–160.
Yeates, D.K. 1992. Why remove autapomorphies? Cladistics 8:
387–389.
Supporting information
Additional Supporting Information may be found in the
online version of this article:
Figure S1 Phylogeny and support values from Bayesian
and parsimony analyses of the Diapsid169 data set.
Figure S2 Phylogeny and support values from Bayesian
and parsimony analyses of the Diapsid189 data set.
Figure S3 Phylogeny and support values from Bayesian
and parsimony analyses of the Anapsid136 and Anapsid154 data sets.
Table S1 Discussion and list of characters included in
Diapsid189 and Anapsid154 data sets, and sources of anatomical information on these characters.
Table S2 Diapsid189 and Anapsid154 data sets, with
MrBayes and PAUP* commands. (Diapsid169 and Anapsid136 data sets can be derived from these by excluding
characters).
Received 18 July 2013; accepted 15 September 2013
ª 2013 THE AUTHOR. J. EVOL. BIOL. 26 (2013) 2729–2738
JOURNAL OF EVOLUTIONARY BIOLOGY ª 2013 EUROPEAN SOCIETY FOR EVOLUTIONARY BIOLOGY