This study analysed 121 seed samples representing 48 distinct varieties or species within the genus Vitis. Detailed information regarding the origin of each sample and its taxonomic classification is provided in Table 1. Of these, seven accessions correspond to Vitis species other than V. vinifera. Twenty-one samples represent wild-type populations, categorized here as sylvestris. These samples were obtained from female vines surveyed between 2003 and 2009 in riparian forests across various locations in France and Spain and subsequently maintained at the IMIDRA germplasm collection. We acknowledge the potential for genomic overlap in these populations; thus, our “sylvestris” category includes both V. vinifera subsp. sylvestris and potential feral individuals (cultivars escaped from cultivation) that exhibit wild-type morphological traits.

Images were captured with a Sony α5100 24-megapixel camera (Sony Corp., Tokyo, Japan) equipped with an AF-S Micro NIKKOR 60 mm f/2.8G ED lens (Nikon Corp., Tokyo, Japan). To minimize geometric distortion and ensure maximum reproducibility, the focal distance of the objective was held constant at an object distance of 17 cm, with all specimens positioned strictly in the center of the frame. Each seed was oriented with its ventral side facing the objective. The longitudinal axis—defined by the line connecting the apex and the stalk base—was precisely aligned with the vertical Y-axis of the image field. This standardized orientation is a critical prerequisite to ensure consistency in subsequent elliptical Fourier descriptors and curvature landmark analyses. Camera settings were configured manually to optimize image quality and depth of field, utilizing a shutter speed of 1/5 s, an aperture of f/18, and an ISO sensitivity of 100. Lighting conditions were maintained between 800 and 1200 lumens with a color temperature of 6400 K. No optical or digital noise-reduction filters were applied during or after image acquisition, as the low ISO values (≤400) guaranteed negligible electronic noise. Furthermore, chromatic correction was unnecessary since the original images were computationally binarized into high-contrast black contours for shape analysis. The raw image dataset is publicly available repository via Zenodo (see Supplementary Materials).

Seed outlines are treated here as closed-plane curves, and, as such, may be approximated using sums of trigonometric functions as truncated Fourier expansions [28,29]. When the parametric components are piecewise linear functions x t = ∑ i = 1 K ∆ x i , y t = ∑ i = 1 K ∆ y i ,  these expansions are: (1) X N t = A 0 + ∑ n = 1 N a n cos 2 n π t T + b n sin 2 n π t T , (2) Y N t = C 0 + ∑ n = 1 N c n cos 2 n π t T + d n sin 2 n π t T , where A 0  and C 0 represent the centroid coordinates, and each addend corresponds to an ellipse (the n-th harmonic) determined by the coefficients a n , b n , c n , d n , which are calculated as:

The coefficients a n , b n , c n , d n of EFT equations can be extracted automatically with the software Momocs [21]. The number N of harmonics determines the degree of precision in the adjustment of the closed curve to the image outline and it may vary depending on the objectives of the study as well as on the characteristics of the image under analysis. While the low-order coefficients ( a i , b i , c i , d i , i = 1 , 2 ) define the basic structure of the curves, the higher-order coefficients ( a i , b i , c i , d i , i > 4 ) are responsible for surface details. In addition, b and c coefficients determine the symmetry of the curve. Curves with bilateral symmetry may have b i , = d i = 0 , i = 1,2 , …

Images (JPG) containing 20–30 seeds were converted into .TPS files, the coordinates for seed outline extracted, and the corresponding coefficients were obtained following the protocol outlined in Momocs v1.4.0 [21]. The coefficients were applied to Mathematica equations to determine the average outlines (Aos) relative to each of the seed populations. In the parametric equations representing the curves, coefficients belonging to N = 6 harmonics were considered for curvature analysis. The values a 1 = 1 , b 1 = c 1 = 0 were fixed, thus leaving a total of 6 × 4 − 3 = 21, varying coefficients. Equations with 21 varying coefficients defined the Aos that served to calculate curvatures. Although it was estimated that N = 7 accounts for 95% of the total of shape variation in Vitis [22,23], and values over it resolve details of the outlines in particular cultivars, for curvature analysis N = 6 results in 10 points well conserved in most cultivars and varieties simplifying the comparisons. In the models N is variable between 7 and 9, giving 4N − 3 coefficients.

Seed outlines were reconstructed using Elliptic Fourier Descriptors (EFD). Curvature analysis was performed on average outlines derived from Fourier coefficients calculated for six harmonics. This level of approximation was selected to achieve an optimal equilibrium between morphological fidelity and computational simplicity. While increasing the number of harmonics improves the resolution of fine-scale irregularities, it also generates a higher density of curvature points that complicates direct inter-varietal comparisons. At six harmonics, the fundamental geometric structure of the Vitis seed—characterized by ten primary curvature extrema—is consistently captured across all accessions, facilitating a standardized comparative framework (see Fig. 1 and Supplementary Materials).

To simplify the geometric model and account for the bilateral symmetry of the seed, the ten detected curvature points were reduced to six representative values: 1.

Unique Points: Points 1 (Apex) and 6 (Base) are unique landmarks. Point 1 consistently represents the maximum positive curvature value across all samples.

The final dataset, comprising the six curvature values, average curvature and curvature ratio (Max to average) and the three morphological indices (A, AR, S), is detailed in Table S1.

The J-Index was employed to quantify the percentage of geometric similarity between a seed’s digital outline and a reference model. This index provides a robust measure of how closely an individual accession adheres to a generalized cultivar or population “archetype.”

The workflow followed a four-step sequence: 1.

Image Capture: Standardized orientation (ventral view, vertical axis).

The dataset for each seed population comprised eleven quantitative variables (see Table S1). Eight variables were derived from the reconstructed average outlines: curvature values at six specific points (P1–P6), the average curvature, and the curvature ratio (maximum to average curvature). The remaining three variables—Area (A), Aspect Ratio (AR), and Solidity (S)—represent the mean values calculated from the individual measurements of 20–30 seeds per population.

All statistical analyses were performed using IBM SPSS Statistics v.28 (SPSS Inc., Chicago, IL, USA) and R v.4.2 [30]. The normality of the data was assessed using the Shapiro-Wilk test; as several variables did not follow a Gaussian distribution, non-parametric methods were employed for all group comparisons.

For analyses involving three or more groups (e.g., comparing the five classification groups), the Kruskal–Wallis H test was used. Significant results were further investigated using the Campbell and Skillings stepwise comparison procedure to identify specific inter-group differences while controlling for Type I error. For comparisons between two independent populations, the Mann–Whitney U test was applied. A Principal Component Analysis (PCA) was conducted in R to visualize the morphometric space and identify the primary variables driving the differentiation between wild and cultivated accessions. Effect sizes for Kruskal–Wallis tests were estimated using epsilon-squared (η²) calculated as: η 2   =   H − k + 1 n − k where H is the Chi-Square value, K is the number of groups and n the total number of samples in the comparison.

Vitis seed images typically present a prominent apical pole (stalk) and a broader basal pole, resulting in a modified elliptical outline (Fig. 3).

Reconstructing these outlines using Elliptic Fourier equations with six harmonics captured ten geometric landmarks, reduced to six primary curvature points (P1–P6) for analysis.

The mean values and measures of dispersion across 121 samples are summarized in Table 2. The highest conservation was observed in average curvature (CV = 6.6%) and P5 (CV = 12.0%). In contrast, P6 (basis) and P4 showed the greatest inter-varietal variability (CV > 40%). Cultivars representing the extreme ends of the morphological spectrum are highlighted in Fig. 4. Notably, P2 (negative curvature) reached its minimum absolute value in V. amurensis (−0.02) and its maximum in ‘Albillo Mayor’ (−2.45), representing a critical landmark for domestication tracking.

Principal Component Analysis (PCA) was performed on the eleven morphometric variables (seed size, aspect ratio, solidity, and curvature metrics) across all 121 samples. The analysis identified three principal components accounting for 79% of the total variance.

PC1 (36.6%): Positively correlated with P2 (negative curvature), solidity, and P4; inversely related to average curvature and P3.

PERMANOVA results confirmed significant differences between the centroids of the pre-defined groups (F = 14.17, p = 0.001). However, low silhouette values (ranging from −0.015 for Muscats to 0.124 for wild species) indicate a phenotypic continuum with substantial morphological overlap between groups.

The PCA biplot (Fig. 5) visualizes the distribution of these groups across four quadrants (Q1–Q4). Q4 is largely occupied by wild species and sylvestris accessions, characterized by higher solidity and P2 values. Domesticated cultivars are distributed across the remaining quadrants: the Hebén group (CG6) predominates in Q1, Muscat types (CG3) in Q2, and the Traminer lineage (CG5) in Q3. Sylvestris samples exhibit an intermediate distribution, overlapping with the Hebén and Traminer groups but notably absent from the Muscat cluster.

While the significant differences in centroids confirm distinct morphological ‘signatures’ for each group, low silhouette values confirm a high degree of overlap—consistent with a common genomic background, and a large variation in the groups (phenotypic continuum). The silhouette values oscillated between −0.015 (Muscat group) and 0.124 (Vitis no vinifera group), thus indicating low separation between the groups. The ellipses of groups sylvestris, Traminer and Hebén are elongated indicating heterogeneity among the samples in these groups. Neverthelelss, there were notable differences among the cultivars offering the possibility to define morphotypes. The samples of Hebén 2020 and 2024 are very close to each other and, in their proximity, there is a good representation of cultivars in the progeny of Hebén (Miguel de Arco, Merseguera, Montúa, Plant Fina, Quigat, Señá, Trepat, Verdejo de Salamanca and Xarello).

The progeny of Savagnin (Traminer) expands to the left in the lower part of the diagram in a region occupied by Chenin Blanc, Gewuertztraminer, Godello, Alfrocheiro, Riesling, Mencía and Pinot Noir, having all these cultivars a related pedigree. Verdejo Blanco, in the progeny of Savagnin and Castellana Blanca, expands to the upper left part of the diagram. In the region between Hebén and Traminer-related cultivars there is a concentration of Muscat-type cultivars, sharing parent-offspring relationships between them, such as Alphonse Lavallee, Cardinal, Imperial, Muscat Hamburg, Muscat a Petits Grains. Afus Ali and Aledo are on top of the diagram with high values of P1 and average curvature.

Those results justify the use of ANOVA to identify specific canalized traits—such as aspect ratio related to seed elongation, solidity or particular variations in curvature points.

Significant differences were identified across the groups using Kruskal–Wallis tests (Table 4).

Group of non-vinifera species (Group 1): Characterized by the highest solidity (968.0), lowest aspect ratio (1.36), and lowest average curvature.

Effect sizes were large for all magnitudes except area (moderate).

The J-Index quantified the morphological fit between seed samples and reference models for Cabernet Sauvignon and Hebén (Fig. 6).

Cabernet Sauvignon Model: Testing against multiple cultivars revealed a significantly higher J-Index for Cabernet Sauvignon (92.4%) compared to Garnacha Tinta (89.9%) and Gewuerztraminer (89.8%) (Table 5).

Hebén Model: This model successfully identified Hebén (92.1%) and its direct progeny, Mollar Cano (92.1%), showing no significant difference between the two. However, it significantly discriminated against the more morphologically distant Muscat Hamburg (88.6%) and Gewuerztraminer (90.9%) (Table 6).

Discriminating between wild and domesticated Vitis seeds has been a central objective of viticultural research for decades [31,32,33,34,35,36]. The transition from traditional measurements to automated image analysis now allows for a high-resolution characterization of both modern cultivars and fossil remains [23,37,38,39]. Our results reinforce the concept of the “wild syndrome,” characterized by rounded seeds with short stalks and high solidity [22,35,36]. A key finding of this study is the geometric relationship between solidity and negative curvature across Vitis species and cultivars. Because solidity represents the ratio of a shape’s area to its convex hull, the presence of negative curvature (inflection zones) at the stalk-chalaza intersection directly reduces solidity values. The observed morphometric gradient—extending from the high solidity and minimal negative curvature of Vitis species outgroups to the low-solidity, elongated forms of the Muscat and Savagnin groups—suggests a progressive shift in seed geometry. While accessions within the Hebén lineage (CG6) occupy an intermediate morphospace, these results should be interpreted as morphometric tendencies rather than definitive taxonomic classifications. These “transitional” patterns likely reflect cultivars that have retained ancestral morphological features while undergoing different selective pressures than more recently bred varieties.

A significant portion of our sylvestris samples were distributed across the PCA plot, overlapping with both non-vinifera species and domesticated cultivars. This high degree of overlap—evidenced by low silhouette values—confirms a phenotypic continuum rather than discrete, isolated clusters. While some populations may represent true relictual V. vinifera subsp. sylvestris, the clustering of certain “wild” samples within domesticated quadrants (Q1 and Q3) suggests the presence of “feral” plants (domesticated escapes) or wild-domestic hybrids [40,41,42]. Because true sylvestris is strictly dioecious, the morphological proximity of some wild samples to hermaphroditic cultivars supports the need for a re-evaluation of European wild Vitis populations [43,44]. However, these patterns are plausible morphometric inferences and require independent molecular validation. High-throughput sequencing is necessary to confirm whether these similarities stem from shared ancestry, recent introgression, or feralization.

The morphometric groups defined here show a notable correspondence with the molecularly defined haplotypes (CG3, CG5, CG6) established by Dong et al. [16]. The relative consistency of the Hebén (CG6) and Savagnin (CG5) progeny clusters suggests that seed geometry can serve as a phenotypic complement to pedigree analysis. For instance, the proximity of ‘Mollar Cano’ to the ‘Hebén’ model aligns with its documented parentage [12,13]. Conversely, instances where progeny deviate from the maternal model, such as ‘Sabro’ or ‘Pedro Ximénez’, likely reflect the geometric influence of an unknown paternal parent or transgressive segregation.

Furthermore, the J-Index analysis of Hebén and Cabernet Sauvignon highlights the stability of these “morphotypes” across diverse geographic origins. This stability suggests that while environmental factors (such as hot summers or water stress) may alter seed size or aspect ratio [45,46,47], the underlying geometric “blueprint” defined by curvature and solidity remains relatively conserved during long periods of vegetative reproduction [48].

Despite the clear trends identified, the substantial overlap in the multivariate space remains a primary limitation. This phenotypic continuum suggests that seed shape is subject to significant variation, likely driven by the complex maternal nature of seed tissues and environmental plasticity. Our study highlights morphometric affinities, but these should not be used as a standalone proxy for genetic identity.

Future research should integrate these geometric markers with independent molecular datasets to isolate the specific genetic factors responsible for seed curvature and solidity. By integrating J-Index modeling with curvature analysis, we provide a non-invasive tool that complements molecular data. This approach offers powerful insights into the phenotypic boundaries of Vitis, provided it is used within a multidisciplinary framework to decode the evolutionary history of the grapevine.