Introduction

The face is one of the most distinct features of morphological variation in the human population, and is closely associated with personal identity1. This variation arises from a combination of visible soft tissue characteristics such as skin pigmentation, thickness, and hair distribution, and deeper structural elements, including facial musculature, fat distribution, and skeletal morphology2,3. Among these, the facial skeleton plays a dominant role in shaping overall facial form.

Among primates, humans exhibit unusually high intra-species variability in cranial and facial morphology4,5,6. The craniofacial skeleton is a complex, modular structure7, whose morphogenesis is regulated by interactions among genetic, epigenetic, and environmental influences3,8. Understanding the relative contributions of these factors remains a central goal in the study of facial development9,10,11,12. Large-scale genome-wide association studies (GWAS) have identified both shared and population-specific genetic loci associated with facial morphology across global populations13,14,15,16, highlighting the polygenic and globally variable architecture of facial traits. Heritability studies utilizing monozygotic and dizygotic twin pairs report moderate-to-high heritability but may be inflated by shared environments, while SNP-based methods yield more moderate estimates of heritability11,17,18,19. Early pedigree-based work including analyses of the Hallstatt skulls, found lower and heterogeneous values for linear traits, while a later geometric morphometric study of the same series demonstrated that cranial shape as a whole is highly heritable and exhibits pervasive genetic integration across facial, vault, and basicranial regions20,21. Overall, variation in reported heritability partly reflects methodological differences such as twin versus pedigree studies, distances versus full-landmark configurations, or SNP-based estimates, but all converge on the view that craniofacial morphology is polygenic and developmentally integrated. Beyond DNA sequence, population-level differences may also arise from epigenetic mechanisms, such as DNA methylation, histone modification, and non-coding RNAs, which regulate gene expression and mediate how genetic architecture responds to developmental and environmental inputs22,23.

Long-term evolutionary processes may also have contributed to population-level differences in craniofacial morphology. These include genetic drift, founder effects24,25,26; and local adaptation in response to ecological pressures such as climate27,28,29, which has been linked to nasal and midfacial variation relevant to thermoregulation and respiration. Diet-related mechanical loading and subsistence strategies may shape facial form via masticatory stresses3,30,31,32. In addition, cultural practices and social dynamics4,33, such as subsistence strategies, and sexual selection, may have also influenced craniofacial variation via both developmental plasticity and heritable selection pressures3,4. Plasticity refers to the ability of developing tissues to respond to environmental inputs, and can contribute to systematic inter‑population differences when environmental exposures consistently vary between groups across generations, potentially leading to epigenetic effects or genetic accommodation34,35,36.

While such mechanisms may help to provide explanations for adult phenotypic divergence, they do not address a more fundamental developmental question: at what point during ontogeny do population-level differences in facial morphology first emerge? Most studies to-date, rely on postnatal datasets from children, adolescents, or adults27,37,38,39,40,41,42,43, limiting insight into the early developmental origins of craniofacial diversity. Given that early facial morphogenesis is shaped by tightly orchestrated molecular and cellular processes, it remains unclear whether population differences arise only later through environmental exposures, or whether they are already evident at the prenatal period. Supporting the latter, Collado et al.44 demonstrated that fetal nasal bone length already shows significant ethnic variation as early as 11–14 weeks of gestation, underscoring that population-level morphological differences can be detectable in the first trimester.

Facial morphogenesis during very early mammalian embryonic development follows a conserved patterning process across species, with divergence in morphology becoming more pronounced after the initial 3–4 weeks of gestation45. Multispecies morphometric analyses demonstrate that distinct facial shapes are already discernible among amniotes, even before the completion of facial prominence fusion, highlighting shared developmental blueprints with early, sharp divergence in patterning trajectories45.

In the human context, Vidarsdóttir et al.46, hypothesized that population-level variation may originate prenatally, though they did not directly study fetal specimens. Katsube et al.47 reported that crucial facial development occurs during the early fetal period and is important for midfacial morphogenesis. Complementing these findings, Sardi et al.48 showed that during late prenatal to early postnatal stages (32–47 weeks), craniofacial growth follows distinct allometric trajectories: the neurocranium undergoes marked changes driven by brain growth, while the face remains relatively isometric. These results suggest that many aspects of facial form are already established before birth, supporting the possibility that population differences emerge during fetal development.

Here, we apply geometric morphometric (GM) analyses to three-dimensional µCT scans of human fetal specimens from Japanese (JP) and North American (NAP) populations focusing on the second trimester of gestation. Although our sample sizes are limited and the collections differ in several aspects, such as period of acquisition and documentation of ancestry, it is exceptionally rare to access such large, well-preserved collections of fetal craniofacial material. These specimens, therefore, offer a unique opportunity to investigate human-specific, early developmental processes. Our analyses reveal discernible, population-level differences in facial morphology at these early prenatal stages, prior to birth.

Results

Overall growth pattern of fetal specimen

Landmark configurations belonging to the JP and NAP samples (Fig. 1; Table 1) were collectively subjected to a principal component analysis (PCA). The scatter plots for the upper face, midface and mandibular landmarks, failed to show a discernably distinct pattern of distribution along PCs 1 and 2 or 1 and 3 (Supplemental Fig. 2). We performed a multivariate linear regression of facial shape on log centroid size of the face to assess the effects of allometry on facial shape. The contribution of size on shape of the upper face, midface and mandible was significant, with approximately 31%, 15% and 10% shape variation explained for each region, respectively. The residuals from the regression were used to compute another PCA (Fig. 2). Once again, the scatter plots for the upper face, midface and mandibular landmarks, failed to show a discernably distinct pattern of distribution along the tested PCs (Fig. 2).

Fig. 1
figure 1

Landmarks (purple: upper face, red: midface, blue: mandible) and semi-landmarks (green) annotated on the specimen. From left to right, external, internal, inferior and right lateral views.

Full size image
Table 1 Specimen information of Japanese (JP) and North American (NAP) populations.
Full size table
Fig. 2
figure 2

Principal component plots (left column: PC1 v/s 2, right column: PC1 v/s PC3) of the upper face (top), midface (middle) and mandible (bottom) utilizing residuals from a multivariate linear regression of facial shape on log centroid size. Specimens from JP and NAP collections are colored in blue and red, respectively. Plots depict distribution of specimen along PC1 (x – axis) and PC2 or PC3 (y-axis) with 80% confidence ellipses for each population. The first three components explained 40.28%, 31.82%, and 17.93% of shape variance in the upper face; 38.38%, 20.10%, and 12.10% in the midface; and 29.23%, 15.01%, and 10.19% in the mandible.

Full size image

Population-level differences in fetal facial morphology

Population-level differences between the JP and NAP landmark configurations of the upper face, midface, and mandible were evaluated using Procrustes distances calculated from regression residuals. Significant shape differences between JP and NAP specimen were detected in the upper face and midface (Table 2, p = 0.001), while differences in the mandible did not reach statistical significance (Table 2, p = 0.075). Notably, we found no influence of sex on these differences (Supplemental Table 2).

Table 2 Summary of group differences in facial shape by region (Procrustes ANOVA, *p < 0.05).
Full size table

To visualize the population-level differences, linear discriminant analysis (LDA) was performed on the upper face and midface landmark configurations. The superior orbital rim of the NAP specimens had a more rounded shape and retruded position when compared to the JP specimen (Fig. 3, top panel). Within the midface, the landmarks annotated on the maxilla appear to be more laterally positioned in the JP specimens. Similarly, the landmarks on the zygomatic bone and arch are different in their lateral extent (Fig. 3, bottom panel, frontal and inferior views). The inferior orbital margin on the JP specimen was positioned more laterally and anteriorly, compared to the NAP specimen. Overall, the JP specimen displayed a wider zygoma, with an orbital margin which is more projected, in the superior, anterior and lateral directions, compared to the NAP specimens (Fig. 3). Links to the interactive 3D models of the reconstructed craniofacial morphology depicting LDA results of the upper face and the midface regions are provided (see Supplemental Data - Links).

Fig. 3
figure 3

Landmark shifts along the LD axis (left to right: viewed from anterior, superior, inferior and right lateral aspects). Landmarks from JP and NAP collections are colored in blue and red, respectively. A standard deviation (SD) of 2 was utilized for visualization. Links to the interactive 3D models of the reconstructed craniofacial morphology depicting LDA results of the upper face and the midface regions are provided (see Supplemental Data - Links).

Full size image

Facial growth allometry in the second trimester

Next, a multivariate analysis of variance (MANOVA) was conducted for the upper face and midface to investigate whether population differences were also present in the growth trajectories of these two regions. However, no significant differences were noted between the NAP and JP populations. (Supplemental Table 3).

Hence, we combined the landmark configurations from both datasets to visualize the overall allometric vectors of growth during the studied time period (Fig. 4). Notably, the lengths of the growth vectors which depict unit standard deviation, were much larger in the mandible compared to the upper face and midface (Fig. 4, purple arrows), indicating greater variability in this region. Links to interactive 3D models of the reconstructed craniofacial morphology depicting growth allometry of the midface and the mandibular regions are provided (Supplemental Data - links).

Fig. 4
figure 4

Mean fetal skull showing overall allometric growth vectors (purple arrows) of all specimens (combined JP and NAP) at the individual landmarks and semi landmarks annotated on the (A) upper face, midface and (B) mandibular regions. The length of the arrows depicts unit standard deviation (SD=1 in panel A and 0.25 in panel B). Links to the interactive 3D models of the reconstructed craniofacial morphology depicting growth allometry of the midface and the mandibular regions are provided (see Supplemental Data - Links).

Full size image

The superior orbital rim displays an overall anterior-inferior-medial growth direction while the inferior rim located on the zygomatic bone shows an anterior-superior-medial growth direction, with an overall reduction in bony orbit size relative to the rest of the growing facial skeleton (Fig. 3A). With respect to the landmarks on the maxillary, zygomatic, palatal and squamous temporal bones, there was an overall relatively anterior and medial displacement during growth. The nasal bone displayed a significant elongation anterior-posteriorly, with an overall relatively inferior displacement with growth. Contrastingly, the elongating vomer, displays a relative superior displacement, when compared to the overall configuration of landmarks. Together, these results indicate relatively greater growth in the anterior-posterior dimension, compared to the lateral and vertical dimensions.

In the mandible (Fig. 3B), the landmarks along the anterior border of coronoid process show a posterior-lateral displacement while those at the tip of the coronoid displace inferior-medially relative to the rest of the mandible. The posterior border, extending interiorly from the developing condylar head, displays a posterior-medial displacement while the angle of the mandible displays a lateral expansion. The inferior border of the mandible displays a similar lateral expansion, with an increasingly anterior component to the growth vectors noted as we move anteriorly from the angle to the symphysis region. Interestingly, the displacement shifts dramatically to the medial direction, at the junction between the tooth crypts of the developing canines and first primary molars. Together, these changes denote a transition from a V-shaped to a more U-shaped inferior mandibular border.

Discussion

By the fetal stage of human development, the major steps of craniofacial organogenesis have already been completed, and bony ossification of the facial skeleton is well underway. Muscle development is also advanced and fetal movements have begun49. Our previous study evaluating JP specimens demonstrated that after the 11th week of gestation, the midface demonstrates a more rapid anterior–posterior growth trajectory compared to the mandible, stabilizing into a more balanced, orthognathic relationship around the 19th week of gestation47. Additionally, it was demonstrated that the lateral parts of the facial skeleton including the zygomatic areas, transform significantly during this period and correlate strongly with masseter muscle development47. A similar trend towards a significant antero-posterior expansion of the midface region, with a relative narrowing of the periorbital and nasomaxillary structures was also detected here (Fig. 4, A). The mandible also demonstrated a discernable change in overall morphology, with a medio-lateral twist of the posterior and anterior margins of the ramus, and transition from a V- to U-shaped inferior border, through ontogeny (Fig. 4, B).

Interestingly, our results indicate that subtle yet significant facial morphological difference already exist between the JP and NAP specimens during the 2nd trimester (Fig. 3). The zygomatic bone in JP specimens was positioned slightly more anteriorly and laterally compared to NAP. This aligns with previous studies showing that adult individuals of East Asian ancestry typically exhibit zygomatic bones that are more anteriorly and laterally projected relative to those of European ancestry, contributing to a broader and flatter midfacial contour27,41,42,43. These skeletal characteristics contribute to population-level differences in orbital depth, facial convexity, and periorbital prominence. The inter-population differences observed in our fetal sample mirror adult-like zygomatic positioning (Fig. 3), suggesting that the early ossifying regions around the orbit already exhibit population-specific morphological patterns at this early stage of development. These patterns differ from those reported by Xing et al.50, likely because their analysis was limited to the orbital morphology, whereas our study considered the entire facial skeleton. This broader anatomical context influences the centroid position and consequently influenced comparative outcomes in morphology.

Vidarsdóttir et al.46 suggested that population-level differences in facial morphology arise through three interrelated ontogenetic mechanisms: (1) emergence of population-specific facial features very early in development, possibly even prenatally; (2) population-specific differences in ontogenetic trajectories, including allometric vectors that diverge significantly in some groups; and (3) variation in ontogenetic scaling, whereby differences in the extent or truncation of growth trajectories further shape adult morphology. Importantly, these authors observed that early morphological differences are not necessarily predictive of later growth trajectories, emphasizing the complexity and plasticity of craniofacial development. In line with these suggestions, Evteev et al.51 showed that midfacial growth trajectories diverge across populations from birth through early childhood, while Freidline et al.52 demonstrated distinct allometric pathways between different ethnic groups (Khoisan vs. European). Together, these findings align with our data, reinforcing the notion that population-specific facial morphology is established prenatally, beginning at the second trimester, and then continues to diverge through growth-related processes.

Surprisingly, our data did not show population level shape differences in the mandibular landmarks between JP and NAP specimen (Table 2). Based on previous studies, and our own data (Fig. 4), the mandible undergoes significant ontogenetic changes during the fetal stages studied here47. Significant population-level shape variations have been identified in studies analyzing adult mandibular morphology as well as sexual dimorphism53. An absence of a statistically quantifiable shape differences between populations may suggest that mandibular morphology is more influenced by function54, and hence difference may become more evident later in life, likely after puberty. Other possible explanations are that high inter-individual variability in mandibular shape likely obscures subtle inter-population shape differences, or that the within-population variation in mandibular shape is less structured compared to the upper face and midface regions.

The mechanisms of specific population-level differences in facial form remain incompletely understood. As noted, genetics, epigenetics (gene regulation), functional dynamics, environmental exposures, climate, and cultural or social variables, are all factors thought to contribute, at varying degrees4,11,22,23,28,55,56,57. In a prior study, we found the superior orbital margin in fetal specimen to be highly symmetrical, with minimal fluctuating asymmetry noted in the inferior margin58. This early morphological symmetry, combined with the population-level differences identified in the current study, supports the role of inherited genetic and epigenetic influences, at these developmental stages. Although fetal muscle activity is already present at the developmental stage examined49, its contribution to bone growth through functional loading is likely minimal.

In conclusion, our findings suggest that by the 2nd trimester of gestation, subtle, yet significant shape differences can be noted in the developing upper and midfacial skeleton amongst individual of Japanese and North American ancestry. While these shape variations are small and limited to certain regions of the orbital rim and lateral areas of the zygomatic bone, they point to an ontogenically early divergence in facial morphogenesis in genealogically distinct populations.

Methods

Specimen

Twenty-two conceptuses with crown rump length (CRL) ranging from 105 to 198 mm (gestational age ~ 13–20 weeks) from the Congenital Anomaly Research Center at Kyoto University (referred to as the JP – Japanese population) and 20 specimens with CRL ranging from 99 to 181 mm (gestational age ~ 12–19 weeks) from University of British Columbia (referred to as the NAP- North American population) were selected for the study. Use of these specimens for the purpose of craniofacial morphological analyses59 has been approved by the Ethics Committee at Kyoto University Graduate School and Faculty of Medicine, Kyoto, Japan (R0316, R0347, and R0989) and University of British Columbia, Vancouver, Canada (H08-02576, and H22-02933).

The specimens used in this study were collected several decades ago. For the NAP population, the specimens were collected in the 1980’s, while the JP specimens were collected beginning around the 1960’s to-date. Hence, our samples were collected from slightly different time periods, which could introduce biases due to changes in environmental factors, healthcare practices, and population demographics over time. Detailed population ancestry of the NAP population is unknown, since this information was not recorded during sample collection and is difficult to deduce because Western Canada experienced moderate immigration during that time frame. However, census-based demographic data indicate that in 1971, approximately 93.5% of the Vancouver Census Metropolitan Area population was of European ancestry, with small proportions of East Asian (5.2%) and Indigenous groups (0.6%, First Nations, Métis, and Inuit)60. Comparatively, the JP specimens are all exclusively of Japanese ancestry. If any of the NAP specimen have an East Asian ancestry, it may have inadvertently diluted out ability of find robust morphological differences in the evaluated geometries. These labels are used solely to reflect broad population ancestry for the purposes of comparing developmental trajectories, without any implication of racial typologies.

All conceptuses were selected based on absence of any distinct congenital anomalies of the face. Since some of the specimens presented artificial deformities due to storage compression one side of the face, only one side of the face was used in this study.

The average differences in crown-rump length (CRL, Table 1) between the two sample sets were not statistically significant (unpaired t test: p > 0.05), indicating that the two population samples are comparable with regards to their overall size and ages. We found a strong, positive correlation between the CRL and the centroid sizes for each specimen (NAP and JP combined, Supplemental Fig. 1), indicating that there were no significant proportional differences between the sizes of the head and the corresponding body lengths (CRL).

Imaging and segmentation

Computed Microtomography (µCT) imaging was performed at 120 kVp, 180–200 µA using a 0.5 mm aluminum filter at a range of 38–66 microns using TOSCANER-30,000 µFD-ZⅡ(Toshiba, Tokyo, Japan) for JP specimens and 70 kVp, 100 µA using a 0.5 mm aluminum filter at a range of 50–51.3 microns using Scanco Medical µCT100 (Scanco, Switzerland) for NAP specimens (Supplemental Table 1). Scans were reconstructed into 3D volumes and digitally segmented using the Amira® software (version 6.0.1; Visualization Sciences, Berlin, Germany) to isolate select paired craniofacial bones. Threshold based semi-automated segmentation was accomplished by selecting voxels with an intensity exceeding a specified minimum masking value (3500 for maxilla and palatal bones, 3300 for the zygomatic, frontal and temporal bones, and 4000 for the nasal bone, vomer and mandible). To maintain consistency of segmentations, the same values were utilized for each bone across all specimen. Subsequently, surface meshes were generated for the segmented bone and utilized for landmark annotation.

Landmark annotation and error measurements

Landmarks were placed manually on the 3D surface files of each specimen using Checkpoint™ (Stratovan Corporation, Davis, CA, USA). A combination of the multiplanar views (boundary) and 3D rendering of the surface meshes was utilized for visualization and landmark placement (Fig. 1). Surface normal to landmarks were visible during annotation to aid in precise localization. The list of landmarks and their descriptions can be found in a previous publication assessing asymmetry in the fetal skeleton58. Note that we only utilized landmarks on one side of each specimen. This is because in a few specimens, storage conditions had resulted in slight distortions on one side of the face, which may impact assessment of shape. Hence, only the non-affected sides were used. All landmarks were mirrored appropriately to represent the right side for GM analysis. In a previous report using the JP specimen, we analyzed craniofacial bone asymmetry at these early developmental timepoints. While some significant asymmetries were detected, the majority were confined to lateral areas of the frontal bone58. Moreover, most asymmetry was fluctuating (i.e. without a side predilection), supporting our utilization of only one side of the facial bones. Repeated landmarks were placed on all specimens by the same investigator (SRB) ~ 1 week apart to ensure good repeatability. Intra-investigator error was calculating by measuring the Euclidean distance between the Cartesian coordinates from the original and repeat landmarking attempts for all specimen. The mean errors for some specimen have been reported previously58, and were recalculated here for the NAP specimen, and found to be 0.189 mm on average (range 0.075–0.402 mm – Supplemental Table 1).

Geometric morphometrics

Geometric morphometric (GM) analyses utilize landmark configurations retaining geometric information, allowing for comprehensive quantification of complex developmental events during growth, and detection of subtle variations in morphology61,62. Landmark X, Y, Z coordinate data were exported to MATLAB 9.0.1 (Mathworks, Natick, MA, USA) and R 3.4.1 using the library Morpho for subsequent analyses. Sliding semi-landmarks were placed upon selected curvatures using package Morpho in R. The purpose of semi landmarks is to better represent morphological structures where individual homologous landmarks are difficult to define, however, where ridges or curve maxima can be identified between two landmarks63. The fit of semi landmarks are optimized by minimizing the bending energy between them, along the selected curves (inferior border of the orbit and mandible).

In a previous study by Katsube et al.47, ontogenetic allometry during fetal facial skeleton development was investigated, using a non-linear weighted average analysis, in specimen from the JP collection. Uniquely, potential mandibular positional artifacts were corrected by applying a set of transformations to all mandibular landmarks based upon a hinge axis defined by the bi-condylar points. Since we only used one side of the facial skeleton here, we were unable to perform such a correction. Consequently, the mandibular and facial landmarks have been analyzed as separate geometries, to avoid any confounding errors imposed by an open-mouth position. Additionally, the superior orbital margin, a component of the frontal bone, was designated as a separate region due to its anatomical proximity to the developing brain and orbit, and the potential influence of these structures on its growth. Since the primary question asked here was whether populations differed in craniofacial skeletal shape at early developmental time points, we focused our attention on the three regions, i.e. the upper face, midface and the mandible, individually.

Generalized Procrustes analysis was applied separately on the upper face, the midface and the mandible to ensure that the landmark coordinates were translated, scaled, and rotated to the best superimposition. To summarize the variations in the craniofacial skeleton, principal component analysis (PCA) was performed. Multivariate linear regression analyses were performed between facial shape and log-transformed centroid size for the upper face, midface, and mandible to assess the statistical significance of growth allometry. Since the influence of allometry was significant for regions, principal component analysis (PCA) was re-conducted using regression residuals. To evaluate population differences, Procrustes distances based on these residuals were tested using permutation tests with 1000 iterations. The threshold of significance was set at P = 0.05. To visualize population-level differences in shape, linear discriminant analysis (LDA) was performed. In addition, multivariate analysis of variance (MANOVA) was conducted for the upper face and midface to investigate whether population differences were also present in growth trajectories. The allometric shape (size-related shape variation) (AS) vector58,64, which indicates the ‘growth vector’ in our study, was calculated from the multivariate regression (ordinary least squares, OLS) of principal components (PCs) on log-transformed centroid size, using only those PCs that showed significant size-related variation. Since no statistically significant difference between the populations were noted (see Results), we combined the specimen for determining the allometric shape vector. Although each region (upper face, midface and mandible) was analyzed independently, the visualization of the upper face and midface landmarks are shown together. For illustrations, the length of the arrows depicts unit standard deviation (SD). A SD of 1 was used for the upper face and midface vectors, while an SD of 0.25 was used for the vectors superimposed on mandibular landmarks.