Seeing through the eyes of the sabertooth Thylacosmilus atrox (Metatheria, Sparassodonta)

The evolution of mammalian vision is difficult to study because the actual receptor organs—the eyes—are not preserved in the fossil record. Orbital orientation and size are the traditional proxies for inferring aspects of ocular function, such as stereoscopy. Adaptations for good stereopsis have evolved in living predaceous mammals, and it is reasonable to infer that fossil representatives would follow the same pattern. This applies to the sparassodonts, an extinct group of South American hypercarnivores related to marsupials, with one exception. In the sabertooth Thylacosmilus atrox, the bony orbits were notably divergent, like those of a cow or a horse, and thus radically differing from conditions in any other known mammalian predator. Orbital convergence alone, however, does not determine presence of stereopsis; frontation and verticality of the orbits also play a role. We show that the orbits of Thylacosmilus were frontated and verticalized in a way that favored some degree of stereopsis and compensated for limited convergence in orbital orientation. The forcing function behind these morphological tradeoffs was the extraordinary growth of its rootless canines, which affected skull shape in Thylacosmilus in numerous ways, including relative orbital displacement.


(Colombia), and Forasiepi and Carlini 7named Patagosmilus goini from middle
Miocene Rio Chico locality (Argentina). Anachlysictis is known only from a lower jaw with dentition, a very fragmentary frontal bone, and a partial postcranial skeleton, while Patagosmilus is represented by a partial cranium and a few postcranial elements. Both species have a morphology that is more primitive than that of Thylacosmilus. Derived features shared between Thylacosmilus and Anachlysictis are as follows: large, subvertical symphyseal flange with radiallyoriented lingual bony striations; alveolar and ventral edges of the dentary ramus subparallel and straight; low and poorly developed masseteric crest; poorly inflected angle; low condyle in relation to the alveolar plane; bowed jugal series, and two lower premolars (likely homologous to p2-p3) 74,92 . Derived features shared between Thylacosmilus and Patagosmilus are: short braincase length, bowed postcanine tooth row, two upper premolars (likely homologous to P2-P3), last upper premolar with complex morphology (interpreted as retained DP3 in the adult molar series), and hypertrophied ever-growing upper canine 76 . Phylogenetic analysis has recovered Patagosmilus 42,43,55 or Patagosmilus plus Anachlysictis 92 in the same monophyletic group as Thylacosmilus. For the node-based definition of Thylacosmilidae, we claim the group that includes the common ancestor of Thylacosmilus atrox, Patagosmilus goini, and Anachlysictis gracilis plus all its descendants (e.g., 2 ). Thylacosmilidae is known from the middle Miocene to midearly Pliocene of South America. In addition to the records already considered, one or two other putative thylacosmilids have been discovered in the middle Miocene of La Venta 74,92 , with a much more generalized morphology than other thylacosmilids. This Laventan taxon could either represent a stem or basal thylacosmilid 92 , or alternatively a different sparassodont lineage with incipient and convergent sabertooth architecture 74 . To this record can be added an isolated upper molar from the early Miocene of Patagonia 93 , but this specimen has not yet been included in a phylogenetic framework.
In a recent contribution, Engelman et al. 55 suggested a major systematic modification, reducing Thylacosmilus and its kin to subfamily level 7 (Thylacosmilinae, instead of family Thylacosmilidae), and re-defining the group using a stem-based definition. Under this proposal, the group includes all sparassodonts more closely related to Thylacosmilus atrox than to Proborhyaena gigantea, Borhyaena tuberata, Prothylacynus patagonicus, Lycopsis torresi, Cladosictis patagonica, or Sipalocyon gracilis 55  Thylacosmilus than to any other sparassodont included in those studies. The phylogenetic relationships of the taxa originally named Proborhyaenidae are still unsettled with several alternative propositions according to different authors and data sets: not only they are shown as monophyletic, but also paraphyletic close to thylacosmilids (e.g., 40,[43][44][45][46]55,92,94 ; for discussion see also 2,9,64 ). In view of these uncertainties regarding the arrangement of proborhyaenids, and in order to avoid discrepancies for the sabertooth group, we continue to follow the more conservative node-based definition of Thylacosmilidae.

Supplementary Materials
In this study the orbital region of three crania of the non-marsupial metatherian Thylacosmilus atrox 67  Orbital orientation. To quantify orbital orientation, the orbit is reduced to a plane defined by three landmarks. There are inconsistencies in previous methods regarding the choice and location of the landmarks, which vary according to the orbital anatomy of the group studied and the quality of the material available. To be able to compare our measurements to the data in the literature, we built four orbital planes for each orbit, using landmarks provided by four methods (Table S2).
The landmark coordinates for the 28 specimens studied are given in  120 .
The issue of using only three landmarks to build a plane when studying fossils is not new (e.g., Finarelli and Goswami 118 used 6 landmarks to construct the sagittal plane in order to be able to find at least three of them preserved on a given fossil). Regarding the sagittal plane, we chose not to define three homologous landmarks. Because mammals have bilateral symmetry, any point on the sagittal crest, or the mid-sagittal sutures of the premaxilla, maxilla, nasals, palatine, and basicranium, can be substituted in order to construct the midsagittal plane when one or more of the defined landmarks are not available.
) is the normal vector of the plane and A(xA, yA, zA), B(xB, yB, zB), and C(xC, yC, zC) are three landmarks belonging to the plane. All dihedral angles between two planes are calculated as the angle between the normal vectors of the two planes involved (see Table S1) in the equation: ) is the normal vector of the second plane expressed as For orbital frontation as measured by Casares-Hidalgo et al. 120 , the dorsal plane is built as a plane perpendicular to the sagittal plane 120  Orbit frontation measured by the method of Heesy 117 is calculated as the angle between the direction vector of the nasion-inion ⃗⃗⃗⃗ and the direction vector of the line of intersection between the orbital and sagittal plane (⃗⃗⃗⃗ ). This is found by resolving the system of equations of the two corresponding planes such as Thus ⃗⃗⃗⃗ = ( Orbitolabyrinth angle. The plane of the lateral semicircular canal is built by three landmarks taken at the center of the lumen ( Figure S1): at the exit of the canal from the ampulla [1], at mid-curve of the canal [2], and at the connection of the canal to the vestibule, or, in cases where a secondary crus commune is present (e.g., Didelphis, Figure S1) before its junction to the posterior semicircular canal [3]. The orbitolabyrinth angle is calculated as the mean value of the dihedral angle between the orbital plane and the lateral semicircular canal of the labyrinth on the same cranial side.
The CT scan of the holotype of Thylacosmilus atrox (FMNH P14531) does not allow reconstruction of the semicircular canals at the necessary level of resolution. The canals are however well seen on the CT scan of the paratype Other measurements. Distances between two landmarks are calculated as √( 2 − 1 ) 2 + ( 2 − 1 ) 2 + ( 2 − 1 ) 2 , with (x1, y1, z1) and (x2, y2, z2) corresponding to the coordinates of the two landmarks involved in the distance, following Heesy 117 for the interorbital width and Pilatti and Astúa 119 for skull length and snout length.
The caudal landmark used to measure snout length is taken at the mid-point between the left and right orbitale anterius and its coordinates are calculated as . Snout width is calculated as the distance between the left and right most anterior point on the border on the medial wall of the infraorbital foramen.
Relative rostrum length and rostrum proportional width follow Pilatti and Astúa 119 .
In an attempt to quantify evident deformation, the angle between the sagittal plane and the palatal, frontal, and basal plane is calculated and subtracted from 90° (the expected angle between the respective planes in non-deformed 13 specimens  Although variation between the left and right orbit in fossil specimens is interpreted to come mainly from deformation, some degree of bilateral asymmetry cannot be dismissed. Thylacoleo carnifex is interesting in that regard because the left postorbital bar is complete, whereas the right one has a tiny gap (<2mm). In this case, with no sign of breakage or rugosity, variation can be interpreted either as an artefact of model reconstruction, as deformation, or as true biological variation. Orbit orientation variation at the species level is also present 116,119,121,122 but cannot be taken into account quantitatively because it is impossible to dissociate from deformation in fossil specimens.
14 Virtual eye reconstruction. Virtual eye size for Thylacosmilus atrox and  Table S3). To compare relative eye size, we calculate the ratio between mean eye diameter over skull length, rostrum relative length, and snout width.

Supplementary Discussion of the methods
Vision in fossil taxa is hard to estimate because vision-related soft tissues are lacking and osteological indicia are limited in various ways. For example, although the optic nerve (CNII) has its own foramen (=optic foramen) in almost all eutherians, giving potential information on their visual capacity (see 126  The orbit is traditionally quantified by orbital orientation and orbital size. Orbital orientation is characterized by three angles: convergence, verticality, and frontation. Convergence is the degree to which the left and right orbit are orientated forward 129 . It is calculated as the dihedral angle between the sagittal or mid-sagittal plane and the orbital plane 121 . In the literature, there is confusion between the terms "verticality" (defined by Heesy 116 ) and "frontation" (defined by Cartmill 129 120 ) to quantify the orbit orientation in relation to the posture of the cranium in a general manner (see Table S2). Interestingly, but also adding to the confusion, "frontation" is used by 131 to quantify the angle between the lateral semicircular canal and a plane constructed by the extraocular muscles. Relating orbital orientation to the orientation of the labyrinth has also been undertaken by Simpson and Graf 132 and Graf and Bruken 133 . Unfortunately, these methods cannot be used in the absence of muscles preserved on dry skulls.
To measure orbital orientation in this study, the orbit is reduced to a plane defined by three landmarks. We used four orbital planes for each specimen, following the landmarks provided in 117-120 (see Methods). We do not distinguish better or worse methods because all of them serve the purpose of this study and bring valuable information on various parameters related to mammalian vision. We would however like to highlight some of the issues encountered with each method because they have influence on the interpretation of results. As an example, orbital orientation does not correlate with visual strategy and habitat in one study 120 , but they do in others 130,134 . Similarly, correlation between orbital orientation and encephalization is established in one article 134 , contrasted in another 118 and dismissed in a third 119 . Therefore, interpretation will depend on, and cannot be considered independent of, the method selected.
In The confusion between the two terms "verticality" and "frontation" (see above) make them impossible to compare if they come from two different methods, not only because the orbital orientation is not measured in the same way, but also because the terms refer to different anatomical notions. For example, the orbitotemporal angle of Casares-Hidalgo et al. 120 should correspond in principle to Heesy's supplementary orbitotemporal angle 117 . However, these angles are in fact not intercomparable because the two studies do not use the same landmarks to build the orbital plane.
Furthermore, and as illustrated in Figure 2 and Figures S2-S34, orientation of the orbits will vary significantly depending on the selection of the three landmarks used to define the orbital plane. As a result, between any two methods angles of orbital orientation can differ by as much as 40°. Clearly, measurements that are not based on the same parameters are simply noncomparable (see also 118 ).
Finally, the following issues were encountered with the definition of the landmarks: -Heesy's method 117 : The majority of the landmarks are type II landmarks that do not refer to truly homologous anatomical points 135 .
They are harder to locate and reproduce, whether for one person or two (see 136 ). The intraobserver error can be reduced by utilizing the mean value for measurements of the left and right orbits in all specimens of the same species (when available) (see 117 ). Interobserver error is more complicated to control and we did not generate a dataset to quantify it.
Simply as an example, we noted that orbital convergence of Panthera leo is 60° in our dataset, but 68° in Heesy's 117 . Part of the difference is due to intraspecific variation, as Heesy 117 showed with regard to significant variance between mean values of individuals. In fact, the three Panthera leo specimens from the present study have quite disparate mean values (between 57° and 66°). However, such differences surely also come from interobserver error (see also 137 ). The landmark taken on the ventral suture of the jugal and the maxilla is problematic in terms of its representativeness for orbital orientation. It is not located on the orbital rim in metatherians, but ventral to it at the level of the molar row. Because our intention is to reconstruct the orbital plane, we would expect relevant landmarks to be placed on the orbital rim, or at least within the orbital fossa. In this case the orbital plane constructed with these landmarks deviates from the intuitive plane one could qualitatively estimate, resulting in diverging estimates of its position ( Figure S20, Figure S21, and Table S2). In short, the Finarelli and Goswami method does not render orbital orientation as such, but rather the plane of the rostral part of the zygomatic arch. Thylacosmilus atrox, Thylacoleo carnifex, primates, various felids, herpestids, scandentians, and artiodactyls; e.g., 117,129,134 ). Pilatti and Astúa's method can therefore not be used for these taxa.
Pilatti and Astúa 119 did not find a correlation between verticality and encephalization quotient. However, the plane that they used to calculate verticality is based on the position of the prosthion (most anterior midsagittal point on the alveolar process of the maxilla) and the condyles instead of landmarks directly related to the encephalic cavity.
As a result, they in fact tested the correlation between encephalization and the orientation of the orbit to the horizontal cranial base, rather than the encephalic cavity as such.

Supplementary Results
All results are given in Tables S3 and S4, organized according to the four methods used to quantify orbital orientation ( 117-120 , see Methods). As expected, there are large intermethod differences in the values obtained, which stems from the fact that in each method orbital planes are determined using different landmarks (see Discussion of the methods). In consequence, values cannot be directly compared and must be considered separately.

Orbital orientation measured according to Heesy's method 117 is illustrated
in Figures S2-S7, Figures S12-19 and Figure 34. When orbital convergence is plotted against orbital verticality ( Figure S2 and Figure S3) When orbital convergence is plotted against orbital frontation ( Figure S4 and Figure S5) When orbital convergence is plotted against orbitotemporal angle ( Figure   S6 and Figure S7)

Orbital orientation measured according to Finarelli and Goswami's
method 118 is illustrated in Figure S8, Figures S20-23, and Figure 34.  When orbital convergence is plotted against orbital frontation ( Figure S8), no grouping of clades can be detected. We note in particular the diffuse distribution of canids across the graph. Thylacosmilus is isolated from other sparassodonts.  119 is illustrated in Figure S9, Figures S24-27.   120 is illustrated in Figure S10, Figure S11, Figures S28-34. When orbital frontation is plotted against orbital convergence ( Figure S10), When orbital convergence is plotted against orbitotemporal angle ( Figure   S11), Thylacosmilus is situated far from other taxa. Non-Thylacosmilus sparassodonts are grouped together near pinnipeds, with the exception of Borhyaena tuberata, which is next to some felids, Panthera leo in particular.

Orbital orientation measured according to Casares-Hidalgo et al.'s method
Thylacoleo carnifex stands in the middle of canids with the mustelid Martes foina.
Thylacinus cynocephalus is situated among felids and not far from Sarcophilus harrisii, although the latter is outside of the felid cloud. The other two dasyurids are situated apart from canids, felids and sparassodonts.                                    Tables S3) taken