Direct evidence that late Neanderthal occupation precedes a technological shift in southwestern Italy

Abstract Objectives During the middle‐to‐upper Paleolithic transition (50,000 and 40,000 years ago), interaction between Neanderthals and Homo sapiens varied across Europe. In southern Italy, the association between Homo sapiens fossils and non‐Mousterian material culture, as well as the mode and tempo of Neanderthal demise, are still vividly debated. In this research, we focus on the study of two human teeth by using 3D geometric morphometric approaches for a reliable taxonomical attribution as well as obtaining new radiometric dates on the archeological sequence. Material and Methods This work presents two lower deciduous molars uncovered at Roccia San Sebastiano (Mondragone‐Caserta, Italy), stratigraphically associated with Mousterian (RSS1) and Uluzzian (RSS2) artifacts. To obtain a probabilistic attribution of the two RSS teeth to each reference taxa group composed of Neanderthals and Homo sapiens, we performed and compared the performance of three supervised learning algorithms (flexible discriminant analysis, multiadaptive regression splines, and random forest) on both crown and cervical outlines obtained by virtual morphometric methods. Results We show that RSS1, whose Mousterian context appears more recent than 44,800–44,230 cal BP, can be attributed to a Neanderthal, while RSS2, found in an Uluzzian context that we dated to 42,640–42,380 cal BP, is attributed to Homo sapiens. Discussion This site yields the most recent direct evidence for a Neanderthal presence in southern Italy and confirms a later shift to upper Paleolithic technology in southwestern Italy compared to the earliest Uluzzian evidence at Grotta del Cavallo (Puglia, Italy).

The Italian Peninsula plays a key role in the study of human evolution due to its geographical position (at the centre of the Mediterranean), broad environmental diversity ) (See supplementary "Environmental setting"), and richness of archeological evidence (Benazzi, Viola, et al., 2011;Hoffecker, 2009;Marciani et al., 2020). However, only a handful of Neanderthals and Homo sapiens remains dated between $50 and 40 ka cal BP have been found in Italy (Benazzi et al., 2014Benazzi, Douka, et al., 2011;Fabbri et al., 2016;Moroni et al., 2018;Romandini et al., 2020;Zanchetta et al., 2018), thereby preventing a comprehensive overview on the relationship between these two species. The multi-layered site of Roccia San Sebastiano (Mondragone-Caserta, Southern Italy) (Figure 1) is of great interest because of (1) its location which confirms a dense prehistoric occupation on the Tyrrhenian side of Italy; (2) the richness and variety of its assemblages; and (3) the continuous settlement spanning from the middle-to-the upper Paleolithic (i.e., it attests several technocomplexes: Gravettian, Aurignacian Dufour, Uluzzian and Mousterian).
Here, we provide evidence of two left second lower deciduous molars (hereafter called RSS1 and RSS2) coming from the Roccia San Sebastiano that were discovered in the Late Mousterian (RSS1) and the Uluzzian (RSS2) deposits. We run state-of-the-art attribution methods on two new human fossils and obtain new radiometric dates on the archeological sequence of RSS in order to (1) disentangle potential association with different human species and (2) ascertain the timing of middle-to-upper Paleolithic technological shift to compare to other areas in Italy (e.g., Apulia).  (Belluomini et al., 2007;Collina et al., 2008; (Figure 1). The cave (about 12 meters in length and 3 m in depth) is divided into two distinct parts: the outer portion named the rock shelter and the internal portion (whose dimensions

| Archeological setting
have not yet been determined). The exploration of the chrono-cultural sequence of the entire stratigraphic deposit is provided by three trenches (E14-E15, F14, and E16) (see supplementary material).
The archeological sequence (Table S1) dug in trench E14-15 at Roccia San Sebastiano can be divided into three main units (labeled Units 1 to 3) (See supplementary "The site: stratigraphic sequence").

| MicroCT scan and digital reconstruction
High-resolution microCT images of the RSS1 Ldm2 were obtained with the XALT scanner (Panetta et al., 2012) at the Institute of Clinical Physiology CNR, Pisa (Italy). The tooth was scanned at 50 kVp, 2 mm Al filtration, 960 projection over 360 , 0.9 mAs/projection for a total scan time of 50 min per sample. The tomographic images were reconstructed using a parallelized version of a modified Feldkamp algorithm (Feldkamp et al., 1984) with included raw data filtration and correction procedures for various unwanted effects and artifacts. The final reconstructed volume consisted of an array of 950 Â 950 Â 770 isometric voxels, each with a side length of 13.8 μm.
MicroCT images of RSS2 Ldm2 were acquired at the Department of Physics and Astronomy of the University of Bologna. The tooth was scanned at 130 kVp, 0.1 mm Fe filtration, 900 projections over 360 , 1.32 mAs/projection for a total scan time of 279 min per sample. The tomographic images were reconstructed using the modified Feldkamp algorithm (Feldkamp et al., 1984) with embedded compensation for mechanical misalignments and raw data pre-correction for beam-heardening. The final reconstructed volume consisted of an array of 950 Â 950 Â 770 cubic voxels, each with a side length of 13.8 μm.
The microCT images of the original samples were virtually segmented using Avizo 9.2 software (Thermo Fisher Scientific, Waltham, Massachusetts, USA). The segmented enamel caps and virtually filled dentins were converted to meshes using the Geomagic Design X (3D Systems Software, Rock Hill, South Carolina, USA), a 3D metrology software.  (Table 1)

| Morphological description and metric comparison
Morphological descriptions and morphometric analyses of the teeth were undertaken on the original specimens and also on the digital models. For the study of morphological traits of deciduous teeth, nonmetric dental traits were assessed following standards outlines adapted from Arizona State University Dental Anthropology System, (ASUDAS) (Turner et al., 1991) and deciduous dental morphology (Garralda et al., 2020;Hanihara, 1956), which was devised for H. sapiens dentition (Harvati et al., 2015;Riga et al., 2018). Occlusal wear stage (Oxilia et al., 2015Oxilia, Menghi Sartorio, et al., 2021) was scored based on Molnar (1971). Age at death was estimated by evaluating tooth formation and root resorption information based on recent H. sapiens time ranges (Moorrees et al., 1963).

| Morphometric analysis
Orientation of each tooth was performed by using the best-fit plane computed at the cervical line (i.e., the cervical plane that best fits a spline curve digitized at the cervical line) parallel to the xy-plane of the Cartesian coordinate system (Been et al., 2017;Benazzi et al., 2013;Bocherens & Drucker, 2006;Fiorenza et al., 2018) by using the Geomagic Design X 3D metrology software (3D Systems Software, Rock Hill, South Carolina, USA). The teeth were then rotated around the z-axis in order to comply with the following criteria: the mesiodistal fissure parallel to the x-axis and the lingual fissure parallel to the yaxis. Finally, the mesiodistal (MD) and buccolingual (BL) dimensions (the size of the bounding box enclosing the crown and cervical outlines) of the RSS1 and RSS2 were analyzed. The measurements were then compared to 99 lower deciduous second molars including Neanderthals (N; n = 34), early Homo sapiens (EHS; n = 8) and recent Homo sapiens (RHS; n = 57), collected from the scientific literature (Hershkovitz et al., 2011) following comparable and previously published protocols (Benazzi et al., 2013;Harvati et al., 2015;Margherita et al., 2016Margherita et al., , 2017. The cervical line of each tooth crown was digitized ( Figure S1a) with a 'spline curve' in Geomagic Design X software (3D Systems Software, Rock Hill, South Carolina, US) and the best-fit plane of the cervical line (here, cervical plane) was computed ( Figure S1a) Note: The calibrated radiocarbon dates are provided using IntCal20 (Reimer et al., 2020) into the OxCal v. 4.4 program (Ramsey, 2009 silhouette of the oriented crown outline was then projected onto the cervical plane. The fractured or worn areas of original crown outline were corrected (gray color) by using as reference the buccolingual contour extent. In Rhino v. 4.0 (Robert McNeel & Associates, Seattle, WA), both crown outlines and cervical outlines were centered superimposing the centroids of their area ( Figure S1b). All the outlines were represented by 16 pseudo-landmarks obtained by equiangularly spaced radial vectors out of the centroid. The first radius is directed buccally and parallel to the y axis of the Cartesian coordinate system ) ( Figure S1c). Finally, size information was oriented and centered with a uniform scaling of the pseudolandmark configurations to unit centroid size ( Figure S1d).
The chosen shape variables (crown and cervical outlines) were then projected into the shape-space obtained from a principal component analysis (PCA) of the comparative sample (Table 2)  The lateral enamel thickness was computed for the region of the tooth included between the following planes ( Figure S1e) Toussaint et al., 2010): (1) "cervical plane"; (2) "cutting plane" (a plane parallel to the cervical plane), which passed through the lowest point of the enamel-dentine junction (EDJ) in the mid-occlusal basin , ultimately cutting the occlusal portion of the tooth (called "cutting plane" in Figure S1e).
From this portion of the crown, three measurements were collected: the lateral enamel volume (mm 3 ; Figure S1f), the lateral dentine plus pulp volume (LDPV in mm 3 ; corresponding to the yellow portion in Figure S1f), and the EDJ lateral surface (mm 2 ; Figure S1g), which do not consider the fractured side of the dental crown (i.e. distal side of RSS 1) .
These measurements were used for the computation of both the cubic root of LDPV (this index is scale-free) (Martin, 1985;Olejniczak et al., 2008). Finally, LAET and LRET of RSS1 and RSS2 were compared to the comparison sample (N = 10; MH = 18) published in Benazzi et al.2012 (Table 3).

| Statistical analysis
We calculated pairwise Euclidean distances on n-1 PCA coordinates for all individuals and computed a permutational multivariate analysis of variance (PERMANOVA) to assess the presence of significant differences between the three examined groups (N, UPHS and RHS) by considering the variability expressed by all principal components (PCs) at once. We avoided problems related to multiple testing by treating results with Bonferroni correction (Table S2 and S3). To obtain a probabilistic attribution of the two RSS teeth to each of the three reference taxa groups (N, UPHS, RHS) we performed and compared the performance of three supervised learning algorithms on both crown and cervical outlines, sincewhen taken individuallythey may lead to different taxonomic attributions (Harvati et al., 2015;Zanchetta et al., 2018). More specifically, we selected the first eight principal components (PCs) accounting for 92% and 95% of crown and cervix data total variance, respectively Jolliffe, 2002;Oxilia et al., 2018;Sorrentino et al., 2020). We tested for normality (Shapiro-Wilk test) and homogeneity of variances across groups (Fligner-Killeen tests) on each PC. Since we found violations of both assumptions in crown data (Tables S4 and S5) and heteroscedasticity in cervix data (Tables S6 and S7) Supplementary Information). We also used a multiadaptive regression splines (MARS) model (Friedman, 1991;Ruppert, 2004).
This algorithm identifies the value intervals that the best discriminate between groups by iteratively running linear regressions for each group and finding the predictor points that minimize within-group total error (knots). These points are then used to link individual linear functions into the final model (Friedman, 1991;Hastie et al., 2009).
We controlled for overfitting of the models using generalized crossvalidation (GCV), a stepwise process which assesses the ratio between the goodness of fit of the model and the number of parameters (knots; Supplementary Information). We then tested the performance of FDA and MARS against a Random Forest (RF) classifier (Liaw & Wiener, 2002

| RSS1
RSS1 is a worn (wear stage 5) (Molnar, 1971) lower left second deciduous molar (Ldm2) with a complete crown and less than a quarter of the root preserved ( Figure 2a). More specifically, root resorption is at a Res3/4 stage (Moorrees et al., 1963) suggesting that the tooth had been lost antemortem, at an age ranging from 9 to 12 years old. in Neanderthal but also a low frequency in H. sapiens molars deciduous (Bailey, 2017). On the buccal side a protostylid with a positive free apex is evident and developed close to the buccal groove (Turner et al., 1991). Interproximal facets are clear both mesially (length = 4.29 mm; height = 2.26 mm) and distally (length = 3.69 mm; height = 2.17 mm), but their size is underestimated due to occlusal wear and, for the distal side, the formation of a post-depositional fracture. The tooth crown has an MD diameter of 9.63 mm (minimum estimation due to the interproximal wear) and a BL diameter of 9.03 mm. At the cervix the MD diameter is 8.25 mm and the BL diameter is 7.58 mm (see Supplementary Information).

| RSS2
RSS2 is a worn (wear stage 5) (Molnar, 1971) Ldm2 with a complete crown, the entire root preserved and an open apical foramen ( Figure   2b). Specifically, the root suggests that the tooth was lost post mortem at an age ranging from 4 (tooth in occlusion) to 6 years (absence of distal interproximal wear facet) (Moorrees et al., 1963).

| RSS1
No difference exists in the distributions of either the crown or cervical outlines between UPHS and RHS (Tables S2 and S3). We therefore grouped all Homo sapiens specimens in our reference sample into a single class (MH) and measure the probability of RSS1 and RSS2 of being attributed to either MH or Neanderthals (N). After repeated 10-fold cross-validation, FDA is the best performing algorithm for both crown and cervix data (Accuracy crown = 0.94; Accuracy cervix = 0.95), followed by MARS (Accuracy crown = 0.93; Accuracy cervix = 0.93) and RF (Accuracy crown = 0.92; Accuracy cervix = 0.9).
As far as RSS1 is concerned, crown and cervical outlines of both the original and the restored tooth fall within the Neanderthal range of variability (Figure 3). All classification algorithms coherently assign  Table S8). Taxonomic attribution is also supported by lateral enamel thickness analysis, where RSS1 lateral average and relative enamel thickness (LAET and LRET, respectively) are closer to Neanderthal values than to the H. sapiens mean (Table S9).
No difference exists in the distributions of either the crown or cervical outlines between EHS and RHS (Tables S2 and S3). Both crown and cervical outlines of RSS1 fall within the Neanderthal range of variability (Figure 3).

| RSS2
Turning to RSS2, both crown and cervical outlines fall within the H. sapiens range of variability (Figure 3), and are consistently attributed to MH by all classification algorithms (Figure 4, TableS8). Lateral Relative Enamel Thickness (LRET) for RSS2 is considerably higher than mean values for both MH and N (Table S10) (Williams et al., 2018) and allow RSS1 to be assigned to this human group. This result is confirmed by morphometric analysis of coronal diameters (Table S12) and outlines (crown and cervical) that corroborate the taxonomic attribution, as well as by the posterior probabilities obtained by all the supervised classification algorithms used in this work (Figure 4, Table S8).
Moreover, the layer in which RSS1 was found appears more recent than spit t39 that we dated to 44,810-44,230 cal BP (68.3% probability). These results allowed us to confirm the most recent evidence in Southern Italy.
In this study we also provide the oldest date for the Uluzzian techno-complex in the Tyrrhenian side of the Italian Peninsula  (Table S13) outline characterized by bucco-distal narrowing, straighter lingual side, and a complex morphology in the occlusal aspect of the EDJ (i.e. crests in the midocclusal basin, but absence of MTC) (Figure 2b). Indeed, as confirmed by all state of the art classification algorithms used, we can state that this tooth belongs to Homo sapiens ( Figure 4, Table S8).

| Techno-cultural observations
Understanding the dynamics of contact between species is crucial, especially investigating at the middle-to-upper Paleolithic transition.
Several questions are indeed still unsolved regarding the mode and tempo of Neanderthals extinction and the following success of Homo  (Peresani et al., 2019), Fumane (Peresani et al., 2016), La Fabbrica (Villa et al., 2018), Castelcivita Gambassini, 1997), Cala (Benini et al., 1997); Uluzzo C (Silvestrini et al., 2021), and Cavallo (Fabbri et al., 2016;Moroni et al., 2013). However, only a few human remains dated between 50 and 40 ka have been discovered in Italy (Buzi et al., 2021) and it is notable that the association between one fossil in one context does not imply that one techno-complex is exclusively related with one or the other species. To date, all the Neanderthal human remains are constrained between 50 and 45 ka cal BP, as the most recent being an incisor from Cavallo cave dated to 45 ka ago (Moroni et al., 2018). At Cavallo the stratigraphic sequence suggests Homo sapiens were already present in southern Europe at least since 45-43 ka cal BP (Cavallo Bleft dP3, and Cavallo C -left dP4) (Benazzi, Douka, et al., 2011) and spread to other parts of Italy at least by $41-40 ka cal BP as documented in Fumane cave (Fumane 2right di 2 ) and Bombrini (Bombrini tooth -left di 2 ) .
At RSS we show that human remain RSS1 belong to a Neanderthal, and RSS2 belongs to Homo sapiens. RSS1 (the Neanderthal specimen) was discovered in association with Mousterian materials. The layer where RSS2 (Homo sapiens) was discovered contains all the main features of the Uluzzian lithic assemblage    Sarti & Martini, 2020 (Douka et al., 2014;Wood et al., 2012). This supports a later expansion of Uluzzian groups from the core area in Apulia , where the oldest Uluzzian settlement was found at Grotta del Cavallo, (ca. 45-43 ka cal BP) (Benazzi, Douka, et al., 2011;Moroni et al., 2018) to the area of Roccia San Sebastiano ( Figure S2). Present results confirm that Southern Italy, and Roccia San Sebastiano in particular, is a key region to disentangle the biocultural dynamics of the two human groups. viviane Slon: Formal analysis (equal); investigation (equal); methodology (equal). Marcello Piperno: Data curation (lead); resources (lead); writingoriginal draft (equal); writingreview and editing (equal).

ACKNOWLEDGMENTS
This work is dedicated to the memory of the late Marcello Piperno.