This study examined 43 accessions of M. dubia maintained in an ex situ germplasm bank at the Experimental Field “El Dorado” (03°57′17″S, 73°24′55″W, 112 m.a.s.l), located at km 25.5 of the Iquitos-Nauta highway in Loreto, Peru. The site experiences typical Amazonian conditions: tropical humid climate with a mean annual temperature of 26.83°C, relative humidity of 89.5%, and average monthly rainfall of 272.5 mm. Established in 1988, this 36-year-old germplasm collection captures natural genetic diversity of eight major hydrographic basins in the Loreto region: Amazonas, Curaray, Itaya, Nanay, Napo, Putumayo, Tigre, and Ucayali ( Figure 1 ). These basins were strategically selected to represent both the accessibility constraints of remote Amazonian locations and the distribution patterns of major M. dubia populations. Within these basins, the 43 wild accessions were systematically collected using a stratified sampling approach designed to capture maximum ecological diversity, targeting populations from contrasting microhabitats including várzea (seasonally flooded forests), terra firme margins, oxbow lakes, and tributary confluences, thereby ensuring representation of the full spectrum of environmental conditions that shape the species’ adaptive variation across its natural range.

Owing to the lack of effective vegetative propagation techniques at that time, a seed-based germplasm bank approach was implemented following protocols for forest seed handling (Willian, 1985). Healthy, mature plants were identified in 43 natural populations from which fully ripe fruits were harvested. The seeds were extracted, washed, air-dried, and germinated in shaded beds. After two months, the resulting seedlings were transferred to unshaded seedbeds with 10 cm spacing between plants and rows. Seedlings were transplanted at a height of 50 cm to field positions with 4×4 m spacing. This seed-based approach was necessary because of the lack of effective vegetative propagation methods. A germination rate of 100% was achieved through optimal seed handling and immediate sowing after extraction, which is consistent with the recalcitrant seed behavior of M. dubia. Each accession comprises 10 individuals maintained under uniform management conditions. Vegetative (agamic) propagation methods such as air layering and central cleft grafting have been developed and are now employed to produce uniform, high-quality planting materials. The complete passport data are provided in Supplementary Table S1 .

The timing of phenotypic evaluation is critical to ensure accurate descriptor measurements and data reliability. After 35 years of establishment, the germplasm collection reached full maturity with stable phenotypic expression, enabling comprehensive characterization during the 2023–2024 fruiting season. The evaluation period from December 2023 to March 2024 was strategically selected to coincide with peak flowering and fruiting when morphological and agronomic descriptors achieved their most representative expression under the local climatic conditions of the Peruvian Amazon.

A systematic sampling approach was implemented to balance statistical rigor with practical constraints of germplasm evaluation. The sampling design followed established protocols for tropical fruit germplasm characterization (Hernández-Delgado et al., 2018; Morales et al., 2019), selecting five individuals from ten plants maintained per accession for detailed evaluation. The 50% sampling intensity was determined through power analysis and validated against practical constraints including: (1) spatial heterogeneity within accession blocks requiring stratified selection, (2) potential edge effects minimized by selecting from central positions, (3) phenological synchrony challenges addressed through multiple observation periods, and (4) resource optimization for comprehensive descriptor measurements. This approach balanced statistical rigor with logistical feasibility while ensuring adequate representation of within-accession variability.

Quality control measures were integrated throughout the sampling process to ensure data reliability and minimize potential bias. Plant selection within each accession followed predetermined criteria designed to minimize edge effects and microenvironmental variation, with individuals chosen from central positions within accession blocks representing the full range of observed phenotypic expression. The selection criteria prioritized plants exhibiting typical morphological characteristics, good health status, and consistent vigor, with standardized parameters applied uniformly across all accessions. The extended establishment period of the collection represents a particular strength of this study, as three decades of cultivation under consistent management conditions likely stabilized phenotypic expression, allowing a clearer manifestation of genetic potential while reducing environmental noise in the dataset.

Data collection encompassed the leaves, flowers, fruits, and seeds of the selected plants across all 43 accessions, following standardized descriptors for M. dubia (Imán et al., 2022). We evaluated twenty-three quantitative descriptors and six qualitative descriptors ( Supplementary Table S2 ). For each plant, there were 10 fully expanded leaves from the middle canopy third, 10 flowers at anthesis, 10 mature fruits at commercial ripeness, and all seeds from the sampled fruits.

Measurements followed standardized protocols with calibrated instruments: linear dimensions (leaf length/width, petiole length, flower parts, fruit size, and seed size) using rulers and digital Vernier calipers (Mitutoyo^® 6″ × 150 mm); mass (fruit weight, pulp weight, and seed weight) using analytical balances (precision ± 0.01 g); sugar content (°Brix) using a digital refractometer (HANNA^® HI96801 calibrated with purified water); and pH using a digital potentiometer (Bonajay^® PH-009 III, calibrated with buffers at pH 4.0 and 7.0). All the instruments were calibrated before each measurement session. To ensure data reliability, all measurements were conducted by a trained four-person team following standardized operating procedures with inter-observer calibration sessions. Cross-validation protocols included 10% re-measurement by different observers, achieving >95% concordance for quantitative traits and 100% agreement for qualitative descriptors, thereby minimizing measurement bias across the multi-month evaluation period. For all quantitative data, the measurements were averaged across ten samples per plant to obtain representative values. Qualitative data (leaf shape, apex shape, base shape, base angle, leaf margin, and the presence of anthocyanins) were recorded as modal categories. Supplementary Figure S1 illustrates the key morphological features measured during characterization, including leaf length, petiole length, style length, flower characteristics, presence of anthocyanin pigmentation, fruit dimensions, sugar content, and seed parameters.

We employed a three-tiered statistical framework that systematically progressed through increasing levels of complexity ( Supplementary Figure S2 ). The analysis began with univariate methods to examine individual descriptor distributions and assess variability patterns. These initial findings informed our bivariate assessments, which revealed associations and correlations between descriptors. Finally, multivariate techniques identified underlying patterns and group structures within the germplasm collection. This hierarchical approach ensured that each analytical level built upon insights from the previous stage, ultimately providing a comprehensive understanding of diversity patterns across multiple dimensions. All statistical procedures were performed using Python (version 3.11.6) selected for its extensive scientific libraries enabling reproducible workflows and handling of large multidimensional datasets, while the Julius computational platform (https://julius.ai/) provided cloud-based processing power essential for computationally intensive multivariate analyses. This combination ensured both methodological transparency and computational efficiency for analyzing 23 quantitative and 6 qualitative descriptors across 215 individual plants.

The M. dubia germplasm bank exhibited distinct patterns of phenotypic variability that aligned with our hypothesis regarding differential variation in descriptors. The descriptive statistics revealed a moderate overall variability (mean CV = 17.4%) across all descriptors, with descriptors falling into three distinct categories ( Table 1 , Figure 2 ): low variability (CV = 3.24%), moderate variability (CV from 10.31 to 19.93%), and high variability (CV from 20.07 to 23.76%).

A clear hierarchical pattern of variability emerged across the 23 quantitative descriptors, with reproductive structures demonstrating substantially greater variation than vegetative traits ( Figure 2 ). Seed-related descriptors exhibited the highest variability levels, all exceeding the 20% coefficient of variation threshold, with total seed weight showing maximum variation at 23.76%. Fruit mass components displayed similarly high variability, including fruit weight (20.82%) and pulp weight (20.63%). Among vegetative traits, petiole length emerged as a notable exception with high variability (20.82%), while among floral descriptors, pedicel length showed exceptional variation (21.42%), distinguishing these two structural traits from their respective categories.

In contrast, the remaining vegetative and floral descriptors demonstrated more conservative variation patterns. Leaf dimensions maintained moderate variability around 10%, indicating morphological stability in foliar traits. Floral measurements, excluding pedicel length, remained within the moderate range, suggesting relatively stable flower morphology across the germplasm collection. Biochemical parameters revealed the most striking contrast in variability patterns: while sugar content showed moderate variation (13.55%), pH exhibited remarkable stability across all accessions with the lowest coefficient of variation of any measured trait (3.24%). This hierarchical distribution of variability from high in reproductive structures to low in biochemical traits characterized the phenotypic diversity pattern within the germplasm bank.

These patterns of differential variability between descriptor categories confirm our first hypothesis that the wild M. dubia germplasm exhibits moderate overall phenotypic variability with greater reproductive structure variation than vegetative descriptors.

Qualitative descriptor assessment across the germplasm bank revealed limited polymorphism, with most traits showing restricted variation. Leaf shape exhibited the highest degree of morphological uniformity, with lanceolate forms comprising 69.3% of the 215 assessed plants ( Figure 3 ). The remaining accessions displayed ovate (20.93%) and elliptical (9.77%) leaf shapes. Additional qualitative traits demonstrated similarly limited diversity: base angle was predominantly obtuse (60.47%), base shape was evenly distributed between rounded (50.70%) and attenuated (41.86%) forms, and base symmetry was primarily symmetrical (66.98%).

Apex shape represented the most variable qualitative descriptor, displaying a more even distribution across four morphological categories. Caudate forms were most frequent at 46.51% of accessions, followed by attenuate (30.23%), acuminate (18.14%), and cuspidate (5.12%) variants. Anthocyanin pigmentation occurred in 79.53% of all assessed plants, while 20.47% lacked this trait. These frequency distributions across the six qualitative descriptors characterized the morphological diversity patterns within the ex situ collection, with apex shape alone showing balanced representation across multiple trait states.

Shannon diversity analysis of qualitative descriptors revealed limited polymorphism across the germplasm bank. The overall mean diversity index of H = 0.823 fell below the moderate diversity threshold of 1.0, with individual descriptors displaying a hierarchical pattern of variability ( Figure 4 ). The diversity values ranged from 0.51 to 1.18, demonstrating substantial variation in polymorphism levels among different morphological traits. Only the apex shape surpassed the moderate diversity threshold with H = 1.18, distinguishing it as the most variable qualitative descriptor in the germplasm bank.

The remaining five descriptors exhibited progressively lower diversity values, forming a clear gradient of morphological variability. Base angle (H = 0.91) and base shape (H = 0.90) approached but did not reach the moderate diversity threshold, followed by leaf shape with H = 0.81. The lowest diversity values were recorded for base symmetry (H = 0.63) and presence of anthocyanins (H = 0.51), with the latter showing the most limited variation among all assessed qualitative traits. This hierarchical distribution of diversity indices characterized the morphological variability structure within the collection.

Complementing the collection-level patterns, within-accession analysis demonstrated considerable heterogeneity for most descriptors. Apex shape maintained its distinction as the most variable trait, with 97.7% of accessions (42 of 43) containing multiple morphological variants ( Supplementary Figure S6 ; Supplementary Table S3 ). Similar within-accession polymorphism was observed for other foliar descriptors, with approximately 75% of accessions exhibiting variation in base shape, base angle, and leaf shape. In contrast, anthocyanin presence showed within-accession variation in only 44.2% of accessions, consistent with its low collection-level diversity.

Spearman correlation analysis revealed distinct patterns of trait associations across the 23 quantitative descriptors, with correlations ranging from strongly negative to strongly positive relationships ( Figure 5 ). The correlation matrix displayed clear clustering of functionally related traits, with the strongest associations occurring within morphological categories rather than between them. Three primary correlation patterns emerged: tightly correlated fruit mass components, inverse relationships between seed number and seed size, and largely independent variation among vegetative, floral, and reproductive structures.

Fruit mass components formed the most cohesive correlation cluster in the dataset. Fruit weight, pulp weight, and fruit width demonstrated exceptionally strong positive associations with correlation coefficients exceeding 0.90, indicating these traits vary in concert. Shell weight and total seed weight showed moderate to strong positive correlations with fruit mass traits, ranging from 0.55 to 0.73. In contrast, biochemical parameters displayed minimal associations with morphological traits, with pH showing particularly weak correlations across all descriptors.

A notable inverse relationship characterized seed traits, revealing a fundamental trade-off in reproductive allocation. Seed number exhibited consistent negative correlations with all individual seed dimensions, with coefficients ranging from -0.42 to -0.48. Conversely, individual seed measurements showed strong positive intercorrelations among themselves, with values between 0.67 and 0.80 for length, width, and thickness combinations. This pattern extended to individual seed weight, which correlated positively with seed dimensions but negatively with seed count.

Vegetative and floral descriptors demonstrated largely independent variation from reproductive traits. Within floral structures, petal and sepal dimensions formed tightly correlated pairs, with coefficients reaching 0.96 for petal measurements and 0.93 for sepal traits. Vegetative descriptors showed weaker intercorrelations, with leaf dimensions displaying only moderate associations. Cross-category correlations between vegetative and reproductive structures remained consistently low, with most coefficients below 0.30, indicating these trait categories vary independently across the germplasm collection.

Association analysis among qualitative descriptors revealed structured relationships that inform efficient phenotyping strategies. Base angle and base shape showed the strongest interdependence (V = 0.667, p < 0.001), indicating these traits are largely redundant for characterization purposes ( Figure 6 ). This strong association suggests measuring one trait could predict the other with reasonable accuracy, streamlining data collection protocols.

Conversely, several several descriptor pairs showed complete independence, providing additional dimensions for accession characterization. The absence of association between leaf shape and base symmetry (V = 0.000) or between base symmetry and anthocyanin presence indicates these traits vary independently, potentially reflecting different genetic control mechanisms or adaptive responses. Understanding these association patterns enables more efficient germplasm characterization by focusing on non-redundant traits that maximize discriminatory power.

Hierarchical cluster analysis successfully classified the germplasm collection into four distinct phenetic groups, providing a framework for understanding morphological diversity patterns ( Figure 7 ). The optimal number of phenetic groups within the M. dubia ex situ germplasm bank was determined through the elbow method ( Supplementary Figure S3 ). At this classification level, the sum of the squared distances within clusters was 634, with the curve showing a notable change in slope that designated four as the most efficient number of clusters for the classification of accessions ( Supplementary Table S5 ).

The phenogram constructed using Ward’s modified hierarchical method and Rogers-Tanimoto distance measure ( Figure 7 ; Supplementary Table S4 ) confirmed the four-group classification structure. The dendrogram exhibited clear differentiation among phenetic groups when sectioned at the 0.55 distance threshold (red dashed line). The cophenetic correlation coefficient (r = 0.677; Supplementary Table S4 ) confirms the robustness of our clustering approach, indicating a good representation of the original distance matrix by the clustering structure, and supporting the validity of the phenetic relationships portrayed in the dendrogram.

The 43 M. dubia accessions were unevenly distributed across the four phenetic groups ( Supplementary Table S5 ). Phenetic groups 1 and 4 each contained 11 accessions, whereas groups 2 and 4 comprised 10 accessions. In phenetic group 1, accessions PER1000421–PER1000409 formed a distinct branch that was separated at approximately 0.35 distance units. Group 2 included accessions PER1000388 through PER1000425, which diverged from the remaining groups by approximately 0.45 distance units. Group 3 contained accessions PER1000405–PER1000423, branching at approximately 0.40 distance units. Finally, group 4 included accessions PER1000396 through PER1000408, which formed the most distinct branch, separated by a 0.48 distance level from the other clusters.

The dendrogram revealed distinct patterns of internal cohesion within each phenetic group. Group 1 showed the highest within-group heterogeneity, with multiple sub-clusters branching at distances ranging from 0.15 to 0.45. Group 3 displayed intermediate cohesion with most accessions clustering below 0.35 distance units, suggesting moderate phenotypic uniformity. Group 2 exhibited a bifurcated structure with two main sub-clusters separating at approximately 0.32 distance. Group 4 demonstrated the tightest internal cohesion, with most branching occurring below 0.30 distance units, indicating greater phenotypic similarity among its members. Notably, no single accession appeared as an outlier, confirming that all 43 accessions were successfully integrated into coherent phenetic groups at the established threshold. This absence of outliers supports the comprehensive nature of the morphological variation captured within the four groups and validates the classification system.

Principal component analysis (PCA) of the 23 quantitative descriptors revealed a structured pattern of phenotypic variation, with the first seven components explaining 82.9% of total variance ( Table 2 ; Supplementary Figure S4 ). The analysis identified distinct groupings of morphological descriptors that contributed differentially to accession differentiation within the germplasm bank. Three primary axes of variation emerged, each characterized by specific descriptor loadings that captured different aspects of morphological diversity.

PC1, with an eigenvalue of 5.987, was predominantly characterized by negative loadings from fruit-related descriptors. The five variables with the strongest negative contributions to PC1 were fruit width (-0.388), seed length (-0.373), pulp weight (-0.366), fruit weight (-0.357), and seed width (-0.344). Hydrogen potential (pH) exhibited the strongest positive association with PC1 (0.160), although this contribution was relatively modest. These loading patterns indicated that PC1 primarily captured variations in fruit size and dimensional attributes.

PC2, with an eigenvalue of 4.032, showed distinctly different descriptor associations than PC1. The most substantial negative loadings on PC2 were from floral descriptors, specifically petal width (-0.389), petal length (-0.382), stamen length (-0.373), style length (-0.355), and hydrogen potential (-0.311). Conversely, vegetative descriptors, including leaf length (0.300), petiole length (0.302), and leaf width (0.298), provided the strongest positive contributions to this component. This pattern indicated that PC2 primarily differentiated accessions based on the contrast between reproductive and vegetative structures.

PC3, with an eigenvalue of 3.115, captured variation patterns that were not accounted for by the first two components. The strongest negative contributions to PC3 were from seed number (-0.506), total seed weight (-0.334), fruit weight (-0.244), and pulp weight (-0.205). In contrast, seed thickness (0.354), weight (0.257), and width (0.223) showed the highest positive loadings. This component appears to reflect an inverse relationship between seed count and individual seed dimensions within the fruits.

The remaining four components (PC4-PC7) each explained between 4.3% and 9.0% of variance, with more specific trait associations. Sepal descriptors dominated PC4, while PC5 was characterized by vegetative traits and sugar content. PC6 showed strong association with pedicel length, and PC7 was primarily influenced by fruit length. Together, these seven components provided a comprehensive representation of the morphological variation present in the germplasm collection.

Quantitative loading patterns across these principal components revealed that fruit mass-related measurements contributed more substantially to phenotypic differentiation than spatial dimensions. Fruit width exhibited the highest absolute loading value on any single component (-0.388 on PC1), followed closely by seed length (-0.373 on PC1), highlighting its prominent role in distinguishing between accessions in the germplasm collection.

A three-dimensional biplot visualization ( Figure 8 ; Supplementary Video S1 ) provided a comprehensive representation of the relationships between morphological descriptors and accession distributions across the four phenetic groups identified in the M. dubia germplasm bank. This multivariate projection, encompassing the first three principal components that collectively explained 57.1% of the total phenotypic variation, revealed clear patterns of descriptor associations and group separation in multidimensional space. The spatial distribution of the 43 accessions, color-coded by phenetic group assignment, demonstrated distinct clustering patterns that corresponded to specific descriptor combinations, as indicated by the direction and magnitude of descriptor vectors. Figure 8 reveals the three-dimensional relationships among accessions, with phenetic group 3 accessions (red points) occupying a distinct region in the negative PC1 space, pulled in that direction by fruit width and pulp weight vectors. Their separation from other groups along this commercially valuable axis, combined with intermediate positions on PC2 and PC3, indicates these accessions have optimized fruit traits without compromising plant vigor or seed quality.

Examination of the descriptor vector patterns within the 3D coordinate system revealed structured associations between morphologically and functionally related descriptors. Fruit-related descriptors (fruit width, fruit weight, and pulp weight) showed a strong parallel orientation, projecting primarily along the negative PC1 axis (-0.388, -0.357, -0.366) and forming a tight cluster of vectors. Seed dimension descriptors (seed length, width, and thickness) were similarly oriented along the negative PC1 axis but diverged along PC3, with seed thickness showing a stronger positive association with PC3 (0.354) compared to seed number, which projected strongly negatively along PC3 (-0.506). Floral descriptors exhibited a distinct pattern, clustering primarily along the negative PC2 axis, with petal dimensions and stamen length forming a coherent group separate from the fruit and seed descriptors.

Phenetic group 1 (yellow points) occupied a region characterized by positive values along PC1 and showed considerable dispersion along both the PC2 and PC3 axes. This group demonstrated an inverse relationship with fruit-related descriptors, displaying generally smaller fruits with a lower weight (8.85 g) and reduced pulp content (6.68 g). Accessions in this group, including PER-399, PER-410, and PER-413, exhibited positive associations with hydrogen potential (pH) and Brix degrees, suggesting higher sugar content and acidity levels compared to the other groups. These accessions showed intermediate to small seed dimensions, but maintained relatively normal floral characteristics.

Phenetic group 2 (green points) displayed a more centralized distribution in multivariate space, with moderate dispersion predominantly along PC3. These accessions, including PER-389, PER-402, and PER-419, exhibited intermediate values for most fruit-related descriptors, which were neither strongly positive nor negative along PC1. The defining descriptors of this group were notably smaller sepal dimensions (sepal width 0.23 cm, sepal length, 0.19 cm). Despite having intermediate fruit weights (9.49 g), the proportional relationship between pulp weight (7.28 g) and total fruit weight suggested higher pulp efficiency, with the pulp percentage exceeding that of groups with larger absolute fruit dimensions.

Phenetic group 3 (red points) occupied the negative sector of PC1, with considerable spread along both PC2 and PC3 axes. Accessions in this group, including PER-411, PER-417, and PER-424, demonstrated strong positive associations with the commercially desirable descriptors. These included wider fruits (2.97 cm), higher fruit weights (11.51 g), greater pulp content (8.72 g), and notably larger seed dimensions (seed length 1.60 cm, seed width 1.20 cm, seed thickness 0.37 cm). Interestingly, this group showed an inverse relationship with leaf length (7.95 cm), which was significantly shorter than that of the other groups.

Phenetic group 4 (purple points) occupied a region characterized by negative values along PC1 and PC2, with moderate dispersion along PC3. This group, containing accessions PER-391, PER-396, and PER-404, exhibited the largest leaf dimensions (leaf length 8.89 cm, leaf width, 3.77 cm) among all groups. Similar to Group 3, these accessions produced larger and heavier fruits (11.46 g) with substantial pulp weights (8.47 g) but were distinguished by having significantly more seeds per fruit and greater total seed weight (1.64 g) than the other groups. A notable characteristic of this cluster was its consistently lower pH value (2.80), representing the most acidic fruit in the collection.

Comparative analysis revealed distinctive trade-offs between commercially important descriptors across the four phenetic groups. Groups 3 and 4 contained the largest and heaviest fruits, respectively. Group 3 exhibited superior seed descriptors (larger and thicker individual seeds), whereas group 4 produced more seeds per fruit. Although groups 1 and 2 displayed smaller absolute fruit dimensions, group 2 demonstrated the highest pulp-to-fruit weight ratio, suggesting differential efficiency in resource allocation across the germplasm bank. These multidimensional relationships were further elucidated through an interactive three-dimensional representation provided in the Supplementary Material ( Supplementary Video S1 ), allowing a more comprehensive visualization of phenotypic associations and accession distributions within the multivariate morphological space.

Multivariate Analysis of Variance (MANOVA) revealed highly significant differences among the four phenetic groups across all the measured quantitative variables ( Table 3 ). All four multivariate test statistics demonstrated significant differentiation among the clusters (p<0.0001). The Wilks’ Lambda test yielded a value of 0.01 (F=2.9, df=69,52), the Pillai’s trace was 2.34 (F=2.91, df=69,57), the Lawley-Hotelling trace showed a value of 12.7 (F=2.88, df=69,47), and Roy’s largest root was 7.09 (F=5.85, df=23,19). These consistently significant results across all four multivariate criteria confirmed robust differentiation among the phenetic groups.

Post-hoc analysis using Hotelling’s T² test identified three statistically distinct groups among the four phenetic clusters. Phenetic group 4 was classified independently as group A, whereas phenetic group 3 was assigned to group B, indicating significant multivariate differentiation between these two clusters. In contrast, phenetic groups 1 and 2 were both categorized as group C, demonstrating that these two clusters were not significantly different from each other in the multivariate space despite being separated in the hierarchical cluster analysis. The clear statistical separation of phenetic groups 3 and 4 from each other and from groups 1 and 2 provided quantitative confirmation of the distinctiveness of the clusters identified through hierarchical classification.

Analysis of Variance revealed significant differences among the four phenetic groups for 17 of the measured quantitative descriptors ( Table 4 ). These differences were statistically significant, with ten descriptors showing highly significant differences (p<0.001), five showing very significant differences (p<0.01), and two showing significant differences (p<0.05). The Scott-Knott test distinguished specific groupings among the four phenetic groups for each descriptor.

For vegetative and flower descriptors, distinct patterns emerged across the phenetic groups. pH values showed groups 1, 2, and 3 forming a statistically homogeneous group (2.87-2.91) that differed significantly from group 4 (2.80). Leaf length measurements differentiated group 3 (7.95 cm) from the other three groups, which formed a statistically uniform group (8.47-8.89 cm). Leaf width separated groups 1 and 3 (3.43-3.46 cm) from groups 2 and 4 (3.63-3.77 cm). Phenetic groups 1 and 3 consistently formed a single statistical group with higher sepal width, sepal length, petal width, petal length, and stamen length, whereas phenetic groups 2 and 4 generally displayed lower values.

The fruit-related descriptors demonstrated particularly pronounced differences between phenetic groups. Fruit width measurements separated the clusters into three distinct groups: groups 3 and 4 with the highest values (2.88-2.97 cm), group 2 had intermediate values (2.46 cm), and group 1 had the lowest values (2.25 cm). Fruit weight followed a similar pattern, with groups 3 and 4 forming one statistical group with higher weights (11.46-11.51 g) that differed significantly from groups 1 and 2 (8.85-9.49 g). Shell weight, pulp weight, and total seed weight showed similar grouping patterns, with groups 3 and 4 consistently forming a statistically homogeneous group with higher values than those of groups 1 and 2.

The seed descriptors displayed more complex differentiation patterns across phenetic groups. Seed length showed the highest degree of discrimination, with each cluster forming a statistically distinct group, except for groups 1 and 2, which were grouped together. Seed width separated the clusters into three groups: group 3 with the highest width (1.20 cm), groups 2 and 4 with intermediate values (1.07-1.10 cm), and group 1 with the lowest values (0.95 cm). Seed thickness differentiated group 3 (0.37 cm) from the other three groups (0.32-0.34 cm), while seed weight separated groups 3 and 4 (0.57-0.61 g) from group 1 and 2 (0.50-0.54 g).

The spatial distribution of the four phenetic groups across the Loreto region revealed a heterogeneous pattern, with no clear geographic structure ( Figure 9 ). Phenetic group 1 (yellow) was predominantly found in areas near Iquitos, with several accessions clustered around the central part of the region. Phenetic group 2 (green) showed a dispersed distribution, with representatives found in the northwest, northeast, and central sections of the study area. Phenetic group 3 (red), which contained accessions with superior fruit characteristics, displayed a widespread distribution across the region, with representatives found along multiple river systems, including areas north of Iquitos and in central river networks. Phenetic group 4 (purple) was concentrated in two main areas: the central zone around Iquitos and the southern portions of the Loreto department.

Statistical analysis using the Mantel test revealed no significant correlation between phenotypic variability and geographical distance (r = 0.0639, p = 0.4034), as shown in Figure 9 and Supplementary Figure S5 , confirming our hypothesis that local environmental factors, rather than regional adaptations, drive the phenotypic patterns. The low correlation coefficient and non-significant p-value demonstrated that accessions with similar phenotypic characteristics were not necessarily geographically proximate. The random distribution pattern was consistent across all eight hydrographic basins in the germplasm collection, with representatives of multiple phenetic groups often found within the same river system. This random spatial arrangement was particularly evident in the central region, where accessions belonging to all four phenetic groups occurred in relatively close proximity, despite their phenotypic differences.

The distribution map also revealed that phenetic group membership transcended collection site boundaries, with accessions from the same phenetic group often originating from distant, disconnected water systems. This pattern was evident for all four phenetic groups, with no exclusive association between any single phenetic group and a particular geographic region or hydrographic basin within the Loreto Department. The most phenotypically distinct groups (3 and 4) showed overlapping geographical distributions with intermediate groups (1 and 2), further emphasizing the lack of geographic structuring in the observed phenotypic variation.

Statistical evaluation of commercially important fruit descriptors across the 12 selected accessions confirmed our third hypothesis that distinct phenotypic groups would harbor promising germplasm with superior trait combinations ( Figure 10 ). The comprehensive analysis encompassed six key parameters: fruit weight, pulp weight, total seed weight, pulp percentage, seed number per accession, and seed count in high pulp accessions (>75% pulp). While parametric tests revealed no significant differences among accessions for fruit mass components (p > 0.05), non-parametric analyses identified significant variation in seed-related traits with direct implications for commercial selection.

Fruit weight distribution showed considerable uniformity across evaluated accessions, with no significant differences detected (ANOVA F=1.3613, p=0.2298; Figure 10A ). Values ranged from 8.5 to 14.5 g, with median weights clustering between 10–13 g. The red dashed line at 11 g represents the overall mean, revealing that accessions PER1000411 and PER1000416 displayed the highest median fruit weights (approximately 13 g), while PER1000383 showed the lowest (approximately 10 g). This statistical homogeneity suggests fruit weight alone is insufficient as a primary selection criterion.

Pulp weight patterns mirrored total fruit weight, showing no significant differences among accessions (ANOVA F=1.4846, p=0.1775; Figure 10B ). Values ranged from 6 to 11 g, with median values between 7.5-10.5 g. Accessions PER1000411 and PER1000416 maintained superior performance with the highest median pulp weights (approximately 10 g), while PER1000383 exhibited the lowest. The strong correlation between fruit and pulp weight (ρ=0.95) was evident in these parallel patterns.

Total seed weight showed marginally non-significant variation (ANOVA F=1.8307, p=0.083; Figure 10C ), ranging from 1.0 to 2.3 g. Despite the lack of statistical significance, PER1000416 displayed notably higher total seed weight (median >2.0 g), while PER1000407 and PER1000411 showed the lowest values (median <1.4 g), suggesting differential seed production strategies among accessions.

Pulp percentage, a critical commercial parameter, showed no significant differences (ANOVA F=1.4903, p=0.1753; Figure 10D ), with values ranging from 68% to 80%. Accessions PER1000385, PER1000417, and PER1000418 displayed the highest median pulp percentages (78-80%), while PER1000424 showed the lowest (approximately 72%). The red dashed line at 75% indicates the threshold for high-quality fruit, with most accessions exceeding this benchmark.

Seed number per accession revealed the first significant differentiation (Kruskal-Wallis H=18.4387, p=0.048; Figure 10E ). Post-hoc analysis clearly separated accessions into three groups: PER1000411 and PER1000407 with the lowest seed counts (group ‘a’, median 2.0-2.2 seeds), intermediate accessions (group ‘ab’), and PER1000385 and PER1000405 with the highest seed numbers (group ‘b’, median 3.5-4.0 seeds).

The most discriminating analysis emerged when examining seed number in high-pulp accessions (>75% pulp content). This subset analysis revealed highly significant differences (Kruskal-Wallis H=13.8412, p=0.0315; Figure 10F ), effectively separating accessions into distinct commercial categories. PER1000411 demonstrated the optimal combination (group ‘ab’) with consistently low seed numbers (2.0-2.2) while maintaining pulp percentage above 75%. PER1000385 showed the highest seed count (group ‘a’, 4.0 seeds) despite high pulp percentage, while PER1000416 occupied an intermediate position (group ‘ab’).

These results confirmed our hypothesis and identified three superior accessions from phenetic group 3 with complementary trait profiles. Accession PER1000411 emerged as the most promising overall, combining high fruit weight, substantial pulp weight, minimal seed number, and a superior pulp percentage. Accession PER1000416 excelled in absolute fruit and pulp weight but carried a moderate seed load. Accession PER1000417 balanced the moderate fruit size with an excellent pulp percentage and acceptable seed count. The concentration of these elite materials within phenetic group 3, despite the limited statistical differentiation in individual traits, validates the multivariate clustering approach for identifying synergistic trait combinations that would be overlooked by univariate selection alone.

Phenotypic characterization of the M. dubia ex situ germplasm bank reveals distinct variability patterns that provide significant insights into the evolutionary history and domestication status of the species. The differential pattern between quantitative and qualitative descriptors characteristic of species with limited domestication history emerges clearly from our analysis ( Table 1 ; Figures 2 , 4 ). This distinction proves particularly informative because the restricted diversity of qualitative descriptors prevented clear morphological differentiation between accessions. In contrast, quantitative descriptors, especially those related to fruit mass, facilitated the identification of distinct phenetic groups. The absence of a correlation between these groups and geographic origin suggests that local environmental conditions, rather than regional adaptations, are the primary drivers shaping the observed phenotypic diversity patterns.

Understanding these variability patterns requires examination of the specific environmental pressures that have shaped trait evolution in M. dubia. The higher variability in reproductive descriptors compared to vegetative descriptors likely reflects specific environmental pressures in riparian habitats (Haghpanah et al., 2024; Seleiman et al., 2021). During the annual flooding cycle (December-May), M. dubia trees remain partially submerged for up to 4–5 months, creating selection for larger, more buoyant fruits that enhance water-mediated dispersal. The negative correlation between seed number and size suggests resource allocation trade-offs optimized for variable flooding intensities, fewer large seeds for long-distance dispersal versus numerous small seeds for local colonization. Additionally, the remarkably low pH variability indicates a strong stabilizing selection for fruit acidity, potentially maintaining vitamin C stability under humid conditions (Peters and Vasquez, 1987; Ferreira et al., 2010).

While our findings reveal novel aspects of M. dubia diversity, they also complement and extend previous research on this species. These findings align with previous research while offering a more comprehensive assessment of the variability patterns of M. dubia in relation to its evolutionary trajectory. Our study extends beyond the single-trait analyses of Imán et al. (2011) by providing a multivariate characterization of 23 traits. Our observations corroborate the findings of Bardales-Lozano et al. (2016), who reported higher variability in fruit morphological descriptors than vegetative descriptors in Brazilian populations, but our comprehensive dataset revealed that this pattern is consistent across the species range. Chagas et al. (2015) identified substantial variations in fruit descriptors across wild populations in the northern Amazon Basin, with fruit mass and pulp content displaying particularly high coefficients of variation. Our analysis uniquely demonstrated that this variation clusters into distinct phenetic groups rather than continuous gradients. This consistency across independent studies conducted in different regions of the Amazon Basin strengthens the conclusion that fruit-related variability represents an intrinsic characteristic of M. dubia populations, shaped by natural selection rather than human intervention.

The patterns observed in M. dubia become particularly meaningful when contrasted with variability patterns in domesticated crops. The variability pattern observed in M. dubia differs markedly from that typically documented in highly domesticated crops, providing evidence for its limited historical domestication. In contrast to M. dubia, intensively selected crops display pronounced morphological variation across multiple trait categories, resulting from human-mediated selection (Pickersgill et al., 1976; Meyer and Purugganan, 2013; Pickersgill, 2016). M. dubia exhibits a variability pattern that more closely resembles that of semi-domesticated or recently domesticated species (Clement et al., 2010). This limited domestication status has important implications for breeding: the high natural variability in fruit traits provides excellent selection opportunities without the genetic bottlenecks typical of domesticated crops, whereas the absence of domestication syndrome allows for rapid genetic gains through simple mass selection.

This limited domestication status creates specific opportunities for targeted improvement efforts. The absence of classic domestication syndrome features, such as the loss of seed dormancy, uniform fruit size, or reduced chemical defenses, indicates specific opportunities for improvement. Future domestication efforts should aim to (1) extend the harvest season through selection for asynchronous ripening, (2) develop thornless varieties for easier harvesting, (3) reduce seed number while maintaining fruit size, and (4) enhance post-harvest fruit stability. Understanding these targets allows breeders to accelerate domestication while preserving valuable wild characteristics, such as flood adaptation and high vitamin C content (Stetter, 2020; Iqbal et al., 2020).

Molecular genetic evidence remarkably aligns with our phenotypic observations. Recent genomic investigations using microsatellite markers have revealed higher diversity within populations than between populations, indicating that M. dubia maintains considerable heterozygosity while exhibiting limited differentiation between geographically separated populations (Castro et al., 2024). This pattern aligns remarkably well with our phenotypic findings of low population-level differentiation, combined with moderate individual variability. Similar conclusions were reached by Šmíd et al. (2017), whose microsatellite analysis confirmed this genetic structure, and by Nunes et al. (2017), who used ISSR markers to demonstrate that the M. dubia population structure exhibits high within-population variation (73.8%) and comparatively low between-population variation (26.2%).

Such a genetic organization is characteristic of outcrossing species that have experienced minimal artificial selection pressure, which is a key indicator of early stage domestication. Genetic evidence presents a cohesive picture of a species maintaining high heterozygosity that contributes to its adaptive potential and phenotypic plasticity, while notably lacking the selective sweeps and reduced diversity commonly observed in genomic regions of crops subjected to intense artificial selection. This congruence between the genetic and phenotypic data strongly reinforces our hypothesis that M. dubia remains in the preliminary stages of domestication, with its current genetic and phenotypic characteristics predominantly shaped by natural evolutionary processes rather than sustained human-mediated selection.

However, some apparent discrepancies merit further explanation. Although neutral molecular markers show limited population differentiation (Šmíd et al., 2017), our phenotypic data revealed four distinct clusters. This reflects different evolutionary forces: neutral markers capture demographic history and gene flow, whereas phenotypic traits respond to local selection pressures. The phenetic groups we identified likely represent adaptive peaks for different microhabitat conditions within the relatively uniform genetic background maintained by gene flow.

These insights into population structure and genetic organization have direct implications for conservation and breeding strategies. The moderate genetic uniformity observed necessitates strategic conservation approaches: (1) prioritizing accessions from different phenetic groups to maintain allelic diversity, (2) implementing complementary in situ conservation in diverse microhabitats to capture ongoing adaptive evolution, and (3) utilizing the retained heterozygosity for rapid genetic gains through recurrent selection without the constraints of narrow genetic bases typical of highly domesticated crops. This conservation strategy should incorporate both static preservation in germplasm banks and dynamic conservation through on-farm management, allowing continued evolution in response to changing environmental conditions.

Multivariate analysis provided valuable insights into the structure of the phenotypic diversity within the germplasm bank ( Figures 7 , 8 ; Supplementary Video S1 ). The identification of four phenetic groups primarily differentiated by fruit-related descriptors has important implications for breeding strategies, as it indicates that selection of fruit descriptors can effectively target the primary axes of variation within the species.

Building on previous phenetic studies in M. dubia, our analysis reveals both consistencies and novel insights. Our results extend those of Chagas et al. (2015), who identified five groups when analyzing fruit descriptors in wild M. dubia populations in northern Brazil. While previous studies have identified fruit descriptors as the most discriminating variables in Peruvian cultivated populations (Pinedo, 2017; Šmíd et al., 2017), our comprehensive analysis uniquely demonstrates that these traits form coherent clusters suitable for targeted breeding. The consistency across these independent studies suggests that fruit descriptors are evolutionarily and agronomically significant in M. dubia, likely reflecting both natural selection for reproductive success, and incipient human selection for improved fruit quality.

The geographic distribution of these phenetic groups reveals important patterns about the forces shaping diversity in this species. The lack of correlation between phenetic groups and geographic origins ( Supplementary Figure S3 ), indicates that phenotypic variation was not structured by regional adaptation. This finding differs from the conclusions of Bardales-Lozano et al. (2016), who documented a significant relationship between fruit descriptors and geographic distribution in Roraima, Brazil. In contrast, the present study aligns with broader research highlighting the complex spatial distribution patterns of Amazonian plant diversity influenced by multiple factors, including river dynamics and historical biogeography (de Oliveira et al., 2014; Maestri and Duarte, 2020; Thom et al., 2020). These findings suggest that ecological and biogeographical interactions in the Amazon are convoluted and are often shaped by habitat heterogeneity, which plays a crucial role in species richness and distribution (Palmer et al., 2010).

These complex ecological interactions are increasingly understood through advances in landscape genetics. For example, recent studies using these approaches have significantly enhanced our understanding of the complex interplay among genetic variation, environmental conditions, and phenotypic descriptors in Amazonian tree species. Luize et al. (2024) and Lemke et al. (2012) convincingly demonstrated that local microhabitat conditions, particularly soil characteristics, nutrient availability, light exposure, and flooding regimes, exert stronger selective pressures on riparian species than broader regional factors. These findings provide substantial evidence that fine-scale environmental heterogeneity, rather than geographic distance or regional adaptations, primarily drives the observed patterns of phenotypic diversity in these ecosystems.

The mechanistic basis for these environmental adaptations in Amazonian trees is becoming clearer through molecular approaches. Landscape genomics has proven to be instrumental in identifying specific genetic loci associated with adaptive descriptors in diverse tree populations (Ćalić et al., 2015). A compelling example comes from research on Himatanthus sucuuba, where populations in flooded versus non-flooded environments display significant genetic and phenotypic differentiation, indicating that hydrological regimes function as powerful selective forces that drive local adaptation (Ferreira et al., 2010). These studies revealed that genetic variation underlying phenotypic descriptors is often structured in response to fine-scale environmental gradients rather than broader geographic patterns.

The implications of these findings extend beyond academic understanding to practical applications in conservation and germplasm collection. The absence of geographic structuring in phenotypic variation has profound implications. It suggests that microenvironmental heterogeneity within sites exceeds macrogeographic differentiation, supporting a conservation strategy focused on ecological gradients rather than geographic representation. This pattern, consistent with high gene flow in river-connected populations, indicates that future climate adaptation may depend more on standing variation within populations than on region-specific ecotypes. For germplasm collection strategies, this finding suggests that sampling multiple microhabitats within a single location may capture more functional diversity than extensive geographic sampling across uniform habitats.

Having established that phenetic groups represent coherent clusters of variation driven by environmental factors, the practical application of these findings becomes paramount. The identification of elite materials becomes particularly relevant given that phenetic group 3 consistently demonstrated superior commercial traits across multiple analyses. From an applied perspective, translating our understanding of diversity patterns into actionable breeding targets represents a critical step in the domestication continuum of M. dubia.

The superiority of phenetic group 3 accessions extends previous findings while providing new insights for breeding strategies. The identification of elite germplasm extends beyond previous single-location evaluations by providing multi-trait selection indices. For example, Alves et al. (2013) demonstrated that phenotypic selection based on fruit quality descriptors in tropical fruits leads to significant genetic gains in subsequent breeding generations, suggesting similar potential for M. dubia. Similarly, Bardales-Lozano et al. (2016) reported the successful development of improved M. dubia varieties through selection based on phenotypic characterization of germplasm collections.

Beyond identifying superior accessions, understanding their complementary trait profiles enables strategic breeding approaches. The distinct descriptor combinations observed in the identified promising accessions suggest specific breeding strategies tailored to different market demands and processing requirements. Crossing these complementary accessions could pyramid favorable alleles while maintaining the genetic diversity essential for long-term improvement. This strategy aligns with modern approaches to crop improvement that balance genetic gain with genetic conservation, as advocated by Tanksley and McCouch (1997) and Liu et al. (2023) for underutilized species.

The identification of promising germplasm accessions with superior trait combinations extends beyond immediate breeding applications to encompass broader socioeconomic considerations. The foundation provided by these elite materials enables development of improved varieties that could substantially increase the productivity and profitability of M. dubia cultivation. Penn (2006) demonstrated that improved M. dubia varieties could generate 2–3 times the income of traditional crops in the region while creating sustainable livelihoods that reduce pressure on forest resources. Furthermore, the commercial development of M. dubia can create value chains that benefit rural communities through processing and value addition (Pinedo Panduro et al., 2004; Pinedo, 2009).

The economic foundation for this transformation is already evident in current production systems. The economic potential of M. dubia in the Peruvian Amazon is substantial, with a current production of 14,226 tons, generating approximately $28 million annually (MIDAGRI, 2025), demonstrating significant promise for improving rural livelihoods. Blare and Donovan (2018) demonstrated that the cultivation and value addition of high-value native fruits increased rural household income while simultaneously requiring less land and contributing to the conservation of forest cover. By adopting improved varieties derived from our identified accessions, smallholder farmers can significantly enhance yield and quality, thereby positioning themselves to meet the rising international demand for this superfruit (Fracassetti et al., 2013; García-Chacón et al., 2023).

Beyond primary production, the development of value chains creates multiplier effects throughout rural economies. Establishing sustainable M. dubia production systems generates employment throughout the value chain from cultivation to processing and marketing. This addresses poverty reduction and rural development goals, particularly for indigenous communities that maintain traditional knowledge of M. dubia management (Blare and Donovan, 2018). Value-added processing of M. dubia fruit into products such as freeze-dried powder, juices, and functional foods could multiply economic returns in rural communities five-fold compared to fresh fruit sales (Conceição et al., 2019; Do et al., 2021). Additionally, byproducts, such as seed oil for cosmetics, reduce waste and provide additional income streams (Azevedo et al., 2019; Conceição et al., 2019).

Ultimately, the conservation and utilization of characterized germplasm represents a model for sustainable development in the Amazon. The ex situ conservation approach exemplified by the INIA germplasm bank complements in situ conservation efforts by preserving genetic diversity, while enabling characterization and utilization. This dual approach operationalizes sustainable development principles by balancing immediate economic needs with long-term conservation goals. As climate change threatens Amazonian ecosystems, maintaining both ex situ collections and developing climate-resilient production systems based on characterized germplasms are critical (Clement et al., 2010).

While our study provides valuable insights into the phenotypic diversity of M. dubia, several limitations should be acknowledged. First, reliance on morphological data alone provides an incomplete picture of the genetic diversity. Phenotypic descriptors are influenced by both genetic and environmental factors, and the observed variability may not fully reflect the underlying genetic diversity. Future studies should integrate molecular marker data to comprehensively assess genetic diversity patterns and population structures (Govindaraj et al., 2015).

Second, the ex situ germplasm bank, which is representative of the Loreto region, encompasses a relatively limited geographic area compared with the total distribution range of the species. M. dubia populations from other Amazonian countries, including Brazil, Colombia, and Venezuela, may harbor additional diversity not captured in the current study (Šmíd et al., 2017; Castro et al., 2018). Investigations have reported significant genetic differentiation between populations of Amazonian species from different watersheds, suggesting that broader geographic sampling may reveal additional patterns of diversity (Clement et al., 2010).

Third, our study focused primarily on morphological descriptors, with a limited assessment of biochemical characteristics beyond basic measurements of sugar content and pH. Recent studies have demonstrated significant variations in the vitamin C content and antioxidant capacity among M. dubia genotypes (Castro Gómez et al., 2013; Freitas et al., 2016; Grigio et al., 2016), highlighting the importance of such evaluations. Given the commercial importance of M. dubia fruit as a source of vitamin C and other bioactive compounds, future phytochemical profiling studies are essential.

Finally, phenotypic characterization was conducted in a single environment, limiting the assessment of genotype-by-environment interactions and phenotypic plasticity. Such interactions are crucial for understanding how different genotypes respond to varying environmental conditions, which can significantly affect descriptor expression and stability (Montesinos-López et al., 2016; Nguyen et al., 2023). Conducting multi-environment trials would provide more comprehensive insights into the stability of observed descriptors across diverse growing conditions (i.e., flooding gradients, soil types, and climate zones), facilitating the identification of resilient genotypes that can adapt to varying agricultural environments (Sakazaki et al., 2022).

The advancement of M. dubia improvement requires immediate deployment of genomic technologies to accelerate breeding and conservation efforts. Whole-genome sequencing, successfully implemented in other tropical fruit species (Alves et al., 2024; Chakrabarty and Külheim, 2025), will establish the foundation for developing SNP markers and enabling marker-assisted selection for commercially important traits. Genotyping-by-sequencing approaches, proven effective for determining genetic diversity and population structure in Myrtaceae species (Klápště et al., 2021; Chen et al., 2025; Zolkafli et al., 2025; Leal et al., 2025). Most transformatively, CRISPR/Cas-based genome editing technologies targeting the L-galactose pathway for vitamin C biosynthesis (Castro et al., 2023; Mall et al., 2024; Redondo-López et al., 2025), could dramatically accelerate genetic gains. These molecular tools will support development of mapping populations from crosses between phenetic groups, particularly utilizing superior accessions from group 3, enabling quantitative trait loci identification and implementation of genomic selection protocols that could reduce breeding cycles from 8–10 years to 4–5 years (Kumar et al., 2020). Integration of speed breeding techniques using controlled environments (Watson et al., 2018) and multi-omics approaches combining genomics, transcriptomics, and metabolomics (Kumar et al., 2017) will further accelerate understanding of the genetic basis underlying the exceptional nutritional properties of M. dubia.

To fully exploit the genetic potential revealed by genomic approaches, integrative multi-omics strategies must elucidate the complex biological pathways underlying the exceptional nutritional properties of M. dubia. Integration of multi-omic approaches would significantly deepen our understanding of the genetic and biochemical bases underlying the exceptional properties of M. dubia. Integration of genomics with transcriptomics, proteomics, and metabolomics will reveal how genetic variation translates into the biochemical diversity of bioactive compounds. High-throughput metabolomic profiling across the germplasm collection will identify relationships between genetic markers and phytochemical composition, following successful models in other species (Pott et al., 2020; Saldanha et al., 2020; Colantonio et al., 2022; Manickavasagam et al., 2024). RNA-Seq analyses across developmental stages can identify key regulatory networks controlling fruit development and quality descriptors, building on approaches successfully employed with Benincasa hispida (Du et al., 2024), Citrus sinensis (Feng et al., 2019), and Fragaria chiloensis (Gaete-Eastman et al., 2022). These molecular insights will guide targeted breeding for enhanced nutritional value and optimized agronomic performance (Ceccarelli, 2015).

Additionally, conservation strategies must evolve beyond traditional ex situ maintenance to encompass dynamic approaches that preserve adaptive potential while supporting phenotypic evaluation. Establishment of multilocational collections across environmental gradients will capture genotype-by-environment interactions, complemented by cryopreservation protocols for long-term security (Panis et al., 2020) and on-farm conservation networks that maintain traditional varieties while preserving indigenous knowledge (Galluzzi et al., 2010). Supporting these efforts, next-generation phenotyping technologies (Tao et al., 2022; Awada et al., 2024) including UAV-based remote sensing (Ampatzidis et al., 2020) and near-infrared spectroscopy for non-destructive fruit quality assessment (Walsh et al., 2020) will enable rapid evaluation of large germplasm collections. Multi-environment phenotyping networks across the Amazon basin, focusing on climate resilience traits such as drought and flooding tolerance (Oliveira et al., 2019), will identify germplasm capable of maintaining productivity under future climate scenarios while incorporating farmer preferences through participatory selection approaches (Ceccarelli and Grando, 2020).

The ultimate success of M. dubia as a driver of sustainable Amazonian development requires integration of biological research with production system optimization and socioeconomic considerations. Development of agroforestry systems that integrate M. dubia with complementary species (Miller and Nair, 2006), combined with research on pollinator ecology (Campbell et al., 2018) and post-harvest technologies (Neves et al., 2015), will address current production constraints. Innovation platforms uniting researchers, farmers, processors, and policymakers will ensure research relevance and facilitate technology adoption (Schut et al., 2016), while decision support tools incorporating genetic, environmental, and market information will guide variety deployment. Long-term studies quantifying ecosystem services from M. dubia cultivation, including carbon sequestration and biodiversity conservation (Reed et al., 2019), will demonstrate broader societal benefits beyond direct economic returns. These integrated research efforts will establish M. dubia as a model for sustainable utilization of underutilized Amazonian species, creating a replicable framework that simultaneously advances conservation and development objectives throughout tropical regions globally.