Introduction
We now know that the appearance of the limb does not signal the conquest of land by vertebrates, because the first stegocephalians (i.e. limbed vertebrates) were probably aquatic. This idea is not new because Romer [45,46] thought that the first stegocephalians were largely aquatic (see the glossary for the phylogenetic definitions of the taxon names adopted here [9], [16], [17], [24] and [26]). However, it received a spectacular confirmation with the discovery of osteological characters that strongly suggest the presence of internal gills in the Devonian stegocephalians Ichthyostega and Acanthostega [12,13]. The shape of the stapes of Ichthyostega, apparently optimized for underwater hearing, corroborates this idea [12]. The presence of grooves for the cephalic portion of the lateral-line organ in numerous Carboniferous and Permian limbed vertebrates suggests that several stegocephalians remained aquatic a few dozen million years after the appearance of the limb [35,48]. Thus, the limb probably appeared in aquatic vertebrates and may have been used for walking on dry land much later. Its initial function remains unknown, but it may have been used for locomotion in an aquatic environment, like the fins of lungfishes, that are used in an underwater walk ([21], and personal observation).
It is difficult to determine when some stegocephalians became terrestrial because the absence of grooves for the lateral-line organ does not prove that this organ was absent. Indeed, several aquatic lissamphibians (Siren, proteids, Amphiuma, pipids, Telmatobius, etc.) possess this organ (in the form of superficial neuromasts), but this organ leaves no trace on their skull ([14], personal observation). The morphology of Palaeozoic stegocephalians seldom gives unambiguous proofs about their lifestyle because it includes few forms that are clearly adapted only to an aquatic lifestyle, contrary to amniotes, in which a return to an aquatic lifestyle is often associated with a transformation of epipodials into swimming paddles (e.g. ichthyosaurs, plesiosaurs, mosasaurs, pinnipeds, cetaceans, etc.) The degree of ossification has been used to infer the lifestyle of extinct vertebrates [53] because the epiphyses of long bones, hands and feet are often poorly ossified in aquatic forms, whereas terrestrial tetrapods have a well-ossified endoskeleton. However, this criterion is not always sufficient because young individuals of terrestrial vertebrates also have a poorly ossified skeleton. Histological characters, such as the presence of calcified cartilage, are often associated with an aquatic lifestyle, but they may also reflect neoteny. Indeed, neoteny is often associated with an aquatic lifestyle and with the presence of calcified cartilage in lissamphibians.
For these reasons, we chose to work on bone microanatomy. Several studies have suggested that the return to an aquatic lifestyle in tetrapods is often accompanied by microanatomical changes in the skeleton. In tetrapods that live in shallow water and that recently returned to an aquatic lifestyle, the long bones (Fig. 1A and B) are generally more compact than those of their terrestrial relatives [36,43]. In tetrapods that live in deep water and that swim actively, the skeleton tends to become lighter, but the medullar region is often occupied by spongy bone (Fig. 1D). This adaptation is found in Neogene cetaceans [6], in most ichthyosaurs [4,38], in plesiosaurs [52], etc. In contrast, long bones of terrestrial vertebrates usually have compact cortical bone and a large medullar cavity (Fig. 1C). Thus, microanatomical data are among the potentially most useful to infer the lifestyle of early stegocephalians [43].
Most previous studies that used bone microanatomical data to assess the lifestyle of extinct vertebrates have relied on a visual comparison of cross-sections. This technique works when there are obvious adaptations to an aquatic lifestyle, but in numerous other cases, it is difficult to use and requires a thorough knowledge of the diversity and distribution of osteological characters. This technique was used to infer the lifestyle of early stegocephalians [43], sauropsids [42], synapsids [41], but it gave better results in Mesozoic amniotes, such as champsosaurs [8], ichthyosaurs [6], placodonts [5], plesiosaurs [52], and crocodilians [22], in which the adaptations to an aquatic lifestyle are more obvious. Since the bone structure of early stegocephalians is difficult to interpret, we have developed quantitative methods that should enable us to extract as much palaeoecological information from long-bone microstructure as possible.
Material and methods
Girondot and Laurin [19] have developed a mathematical model and software (named Bone Profiler) that can express the quantity and distribution of bone tissue in a cross-section of a long bone. It may be useful to summarize this model here: it includes four parameters (Fig. 2) that correspond approximately to the compactness in the medulla (Min), in the superficial portion of the cortex (Max), the relative size of the medullar region (P), and the width of the transition zone between the cortical and medullar regions (S). The compactness of the bone C can be expressed as a function of the distance d from the centre of the section:
C
(
d
)
=
1
1
+
e
1
S
(
P
-
d
)
(
M
a
x
-
M
i
n
)
+
M
i
n
\[
\mathit{C}\left(\mathit{d}\right)=\frac{1}{1+{\mathit{e}}^{\frac{1}{\mathit{S}}(\mathit{P}-\mathit{d})}}(\mathit{M}\mathit{a}\mathit{x}-\mathit{M}\mathit{i}\mathit{n})+\mathit{M}\mathit{i}\mathit{n}
\]
where P is the relative distance between the centre of the bone and the inflexion point, where the compactness gradient is the highest. Thus, P is the distance between the centre and the inflexion point divided by the radius of the bone. P has a simple relationship with the cordico-diaphyseal index (CDI) that has often been used in comparative osteology [11]; it is 1-CDI. S is the slope-1 at point P. Max is the upper asymptote. Min is the lower asymptote. It is usually 0 in terrestrial vertebrates that have a free medullar cavity (Fig. 1C), but it may be positive in aquatic taxa that possess a medullar spongiosa (Fig. 1A, B, and D).
To gather these data, we developed the software Bone Profiler [19], which can read an image (picture or drawing) of a bone section and gather data on the compactness in 51 concentric zones and 60 sectors (each of which measures 6° in width), which together delimit 3060 areas on which the compactness is measured (Fig. 3). These data are used to establish 60 compactness profiles on a complete section, but incomplete sections can also be used, and this is particularly useful to analyse fossil bones, which are often fragmentary or damaged. Average and standard deviation of each variable are calculated, outliers (i.e. profiles for which at least one parameter is more than four standard deviations from the average) are removed and this procedure is repeated as many times as necessary. Thus, data on the average and variability of the parameters are gathered.
We are currently gathering data on a large number of extant tetrapods to produce a model to infer the lifestyle of early stegocephalians. The analysis summarized here bears on the femora of 105 individuals belonging to 46 lissamphibian species (23 anuran species and 23 urodele species). Lissamphibians are particularly suitable for such an analysis, because they encompass a broad range of lifestyles; from an initially amphibious lifestyle [51], several taxa became terrestrial or aquatic (Fig. 4). A tree that incorporates divergence times between all included species has been produced from formerly published phylogenies [30]. Divergence times were estimated from the geological age of the oldest fossil that belongs to each clade and from molecular studies (Fig. 4).
Several tests were performed to determine which characters included ecological information. The best known is the concentrated changes test described by Maddison [37]. This test rests on the premise that if no correlation exists between two characters, the distribution of the dependent character has no impact on the frequency or direction of changes in the dependent character. It requires a completely resolved phylogeny, does not use branch length information, and can be used only for discrete data. Thus, it was necessary to convert our continuous variables into discrete ones, and this was the least satisfactory part of this test.
Permuted multiple linear regressions are more appropriate for our purpose because this test can be used on datasets that include a mix of continuous and discrete characters. This test was described by Legendre et al. [34], who used it to determine which characters were correlated with brain evolution in marsupials. Legendre et al. [34] used a simple taxonomy in which marsupials were classified into three orders, and that amounts to using a poorly resolved phylogeny that includes three polytomies and no branch length data. However, a completely resolved phylogeny incorporating branch-length information can be used, as we have done. The variables are then converted into distance matrices, character by character. For quantitative characters, the distance is simply the absolute value of the difference between taxa (the Euclidean distance), taken pair by pair (Fig. 5). For discrete, binary characters (such as lifestyle, that was coded as 0 for aquatic taxa and 1 for amphibious or terrestrial taxa), the difference was 1 when the lifestyles differed, and 0 when they were identical. Finally, the phylogenetic distance between taxa is simply the time that has elapsed since their lineages diverged (Fig. 5). A multiple linear regression is then performed on all the matrices, and this yields regression coefficients between the dependent variable (lifestyle) and each independent variable (axial length, femoral compactness, phylogenetic position). The values of the dependent variable are then randomly permuted and the multiple regression analysis is repeated. This yields a new set of regression coefficients. A large number of analyses on randomized data sets is performed (typically, at least 999, but in our study, we have performed 9999) to generate a null distribution of regression coefficients. The relationship is deemed to be statistically significant if the initial regression coefficient between two variables is greater than at least 95% of the regression coefficients obtained from randomized datasets. This test does not require the data to be normally distributed, and this is critical here because the lifestyle was coded as a binary character. In classical regressions, the dependent variable must have a normal distribution (and must therefore be continuous or possess enough states to approximate a normal distribution). The permutations remove this constraint. This method is less known than the phylogenetically independent contrasts [15], but it is flexible and has recently been used to study trends in body size evolution in early stegocephalians [25].
Using phylogenetically weighted logistic regressions, we have produced inference models for the lifestyle of stegocephalians that require only data on body size and the compactness profile. The phylogenetic weight attributes equal weights to two sister-taxa, even if they are represented by very uneven numbers of species. These models were produced by minimizing the difference between the observed lifestyle (on extant amphibians) and the lifestyle inferred by the model on the basis of compactness profile (parameters bci
with i ranging from 1 to 4), section diameter (parameter bc5
) and log-transformed body length data, used as a covariable.
h
=
1
1
+
e
(
C
0
+
∑
i
=
1
5
(
C
i
1
b
c
i
C
i
2
ln
presacral
length
)
)
\[
\mathit{h}=\frac{1}{1+{\mathit{e}}^{\left({\mathit{C}}_{0}+\sum _{\mathit{i}=1}^{5}\left({\mathit{C}}_{\mathit{i}1}{{\mathit{b}\mathit{c}}_{\mathit{i}}}^{{\mathit{C}}_{\mathit{i}2}\text{ln}\mathit{presacral}\mathit{length}}\right)\right)}}
\]
Eleven parameters (C0
, C1i
and C2i
with i from 1 to 5) were estimated to minimize the weighted least-square sum (SSQ) of the difference between observed and inferred habitat over all individuals. This model was simplified by removing some constants.
One of the resulting models was then used to infer the lifestyle of four stegocephalians (Fig. 6) that represent four major clades (Temnospondyli, Embolomeri, Amphibia, and Amniota).
Results
The concentrated changes test shows that the femoral compactness and axial length (measured from the atlas to the sacral vertebrae in urodeles, and between the atlas and the posterior end of the urostyle in anurans) contain ecological information [30]. Compared to terrestrial or amphibious lissamphibians, aquatic amphibians are on average larger (p = 0.01) and their femur is more compact (p = 0.042).
The permuted multiple regressions confirm that axial length (p = 0.0026) and compactness profile parameter P (p = 0.0225) are correlated with the habitat.
Two inference models were produced using phylogenetically weighted logistic regressions [30]. Both yield inferences comprised between 0 and 1 (0 for aquatic and 1 for amphibious, since the differences between terrestrial and amphibious species was not statistically significant). The first model is based on 46 species; it requires estimating 11 constants and yields a correct lifestyle in 85% of the 105 individuals and 82% of the 46 species. We suspected that the relatively high error rate resulted from great differences in proportions (and perhaps, in the mode of locomotion) between anurans and urodeles. Furthermore, we believed that anurans could mislead us because their body proportions are very different from those of the early stegocephalians whose lifestyle we wanted to infer. Thus, we produced a second model based on the 23 species of urodeles included in this study. This model requires estimating only four constants and requires information on compactness profile parameter P, maximal diameter of the bone section, and axial length. It estimates correctly the lifestyle of 39 of the 41 individuals and of 21 of the 23 species.
Applying the second model to four early stegocephalians yields results generally congruent with the lifestyle that had previously been attributed to these taxa.
The small temnospondyl Doleserpeton from the Lower Permian of North America has generally been considered to be closely related to lissamphibians [2,3,47], but most recent phylogenetic analyses suggest that it is a stem-tetrapod [1,23,28,29,51]. Doleserpeton is usually considered terrestrial [3]; this interpretation is confirmed by the logistic inference model.
Chroniosaurus is an Upper Permian embolomere from Russia [20]. It is one of the last embolomeres, since that group appeared in the Lower Carboniferous, was most diverse in the Upper Carboniferous, and became extinct in the Triassic. Golubev [20] suggested that Chroniosaurus was aquatic. Our logistic model infers an amphibious or terrestrial lifestyle. This is not necessarily a mistake because Chroniosaurus possessed well-ossified dorsal scutes that were articulated to each other and to the neural spines [20]. This may be an adaptation to strengthen the axial skeleton to better support the weight of the animal on dry land, and this suggests an amphibious or a terrestrial lifestyle.
The Lower Permian amphibian Cardiocephalus is generally considered terrestrial because of its morphology and of its presence in the fossiliferous locality of Fort Sill, that seems to preserve a terrestrial community [39,50]. The inference model yields an amphibious to terrestrial lifestyle for this taxon.
The small parareptile Phaanthosaurus, from the Triassic of Russia, is probably closely related to turtles, like all parareptiles [27,32,33,40]. Our model confirms the terrestrial lifestyle usually attributed to Phaanthosaurus [32].
Discussion
The greater compactness of the femur of aquatic lissamphibians confirms results of several previous studies on various stegocephalian taxa [5,8,22,36,38,42,44,52], but seems to conflict with the results of the two other studies on the relationship between skeletal mass and compactness and lifestyle in anurans [10,31]. Indeed, Castanet and Caetano [10] found that two amphibious ranid species (Rana perezi et Rana esculenta) have a lower relative skeletal mass and a smaller cortico-diaphyseal index (i.e., a greater parameter P and a lower compactness) than two more terrestrial congeneric species (Rana temporaria et Rana dalmatina). Leclair et al. [31] found that in seven species of anurans (Bufo americanus, five species of Rana and Xenopus laevis), the proportion of the corporal mass represented by the skeleton is lower in aquatic to amphibious taxa than in terrestrial taxa. This contradiction is more apparent than real, because we have not studied the relative mass of the skeleton, that is not necessarily tightly correlated with skeletal compactness. The observed difference in the cortico-diaphyseal index [10] may be partly explained by a size effect, because there is an interaction between parameter P (and hence, the cortico-diaphyseal index) and body size [30 (table 3)], and the amphibious ranids studied by Castanet and Caetano [10] were larger than the terrestrial ones [10 (table 1)]. However, there is no real contradiction with our results because we have found no statistically significant differences between amphibious and terrestrial lissamphibian species, but only between aquatic and other (amphibian and terrestrial) species. The sample studied by Castanet and Caetano [10] included no aquatic species. Hence, their study [10] found a difference where we found none, perhaps because they had a greater intraspecific sample (from 31 to 48 individuals per species, whereas ours included only one to 21 individuals par species, with an average of only 2.3 individuals per species).
The results obtained on the femur of lissamphibians suggest that bone microstructural data can be used to reconstruct the history of the conquest of land by vertebrates. The inference models that we produced cannot be used to infer the lifestyle of mid-sized to large stegocephalians because they are based on lissamphibians, that are relatively small animals, and body size includes ecological information. Thus, bone microstructural data on extant taxa of larger body size have to be gathered to complete this research program [18]. We are also studying the bone microstructure of Eusthenopteron, a primitively aquatic sarcopterygian from the Devonian, that will provide a valuable basis for comparison with potentially primitively aquatic stegocephalians, and that will be used to produce better inference models. Finally, other bones (e.g., humerus, radius and tibia) are being studied [18], because we do not know which will yield the most reliable ecological signal.