Introduction

Winter squash (Cucurbita moschata D.) is one of the most socioeconomically and nutritionally important vegetables in the Cucurbita genus, cultivated worldwide. It stands out for the nutritional value of its fruits, characterized by a pronounced content of bioactive components such as β and α-carotene1,2. These precursors have the highest provitamin A activity3 and show high antioxidant activity4. Considered one of the main sources of β and α-carotene among the vegetables consumed in Brazil5, C. moschata is one of the species prioritized in programs aimed at vitamin A biofortification and circumventing vitamin A deficiencies6.

Studies also point to the potential of using C. moschata seeds in human food production, emphasizing their high levels of unsaturated fatty acids (UFA) and bioactive components. The seed oil of C. moschata consists of approximately 70% UFA and a high content of monounsaturated fatty acid (MUFA) such as oleic acid7,8. Such characteristics make this oil an excellent option in human nutrition, particularly in replacing lipid sources harmful to health, such as those with a predominance of saturated fatty acids. It has also been demonstrated that C. moschata seed oil contains a high content of vitamin E components, such as α- and γ-tocopherol, and carotenoids; which are bioactive components widely known for their beneficial effects on human health9,10.

C. moschata, a cosmopolitan vegetable crop, is cultivated in a wide geographic range, from tropical to temperate regions11. Archaeological evidence points to the presence and food use of this vegetable in Latin America, specifically Colombia and Ecuador, for more than 7000 years12,13. Its cultivation spread throughout Latin America, mostly to countries such as Argentina, Peru, and Brazil11,14. In Brazil, C. moschata cultivation is widespread, covering different edaphoclimatic conditions and production systems15, mainly family-based production. Linked to this, studies have highlighted the remarkable variability in the agronomic characteristics, resistance against phytopathogens, and chemical–nutritional traits of fruits and seeds of the C. moschata germplasm found in Brazil16,17,18.

The Vegetable Germplasm Bank of the Federal University of Viçosa (BGH-UFV) was founded in 1966. Since then, it has carried out germplasm collection for a period of more than five decades, covering different geographic regions of Brazil19. Currently, this bank maintains a collection of around 350 accessions of C. moschata, which represents a substantial sample of the Brazilian germplasm and is one of the largest collections of this species in the country19,20.

The usefulness of plant germplasms conserved in banks depends on the quantity and quality of information associated with it, corroborating the efforts aimed at its proper evaluation. However, common restrictions in the evaluation of germplasms kept in banks, such as financial limitations and the lack of human resources, generally limit this evaluation21. The application of computational intelligence, and more specifically of artificial neural networks (ANNs), is a promising tool for evaluating and managing plant germplasms conserved in banks22,23. This tool has aroused interest because it can map non-linear systems, extracting the particularities of these systems from information such as measurements, samples, or patterns. Interest in applying ANNs also stems from their ability to adapt through experience, learning ability, generalization ability, and fault tolerance. The fact that their implementation is not linked to the experimentation process and the nature of the data set, allowing them to circumvent limitations often associated with multivariate analyzes, is another advantage in using ANNs24.

The emphasis placed on the conservation of plant germplasm in banks has resulted in the establishment of extensive collections. On the other hand, it has been highlighted that the use optimization of a collection is inversely related to its size25. Therefore, with a view to improving the use, accessibility, and conservation of accessions maintained in germplasm banks, Frankel26 proposed the concept of the core collection. Brown et al.27 Defined the term core collection as a set of accessions chosen from a germplasm collection to represent the maximum genetic variability of the initial collection, with minimum redundancy. The establishment of core collections consists of a procedure widely used in collections of plant germplasm, covering different species and based on different methodologies28,29,30. Cucurbita moschata crop is characterized by vigorous growth, demanding large areas and intense labor for the agro-morphological evaluation of its germplasm. With this, the implementation of ANNs represents a promising approach for optimizing the evaluation and management of this vegetable germplasm.

Given the above, this present study evaluated the agronomic and chemical–nutritional aspects of 91 accessions of C. moschata maintained at the BGH-UFV. The study analyzed the variability of the germplasm, and established a core collection based on multivariate approaches and the implementation of ANNs, aiming at optimizing the use and management of this germplasm.

Materials and methods

Origin of germplasm and conduct of the experiment

This work initially comprised the agro-morphological assessment of part of the C. moschata collection maintained at the BGH-UFV, including 91 accessions from different regions of Brazil, mostly landraces collected from family-based properties (Table 1). This germplasm was previously collected from several collections, as detailed by (Silva 2001)19.

Table 1 Origin of part of the C. moschata accessions kept in the Vegetable Germplasm Bank of the Federal University of Viçosa.
Full size table

Agro-morphological evaluations

The agro-morphological evaluation was carried out in a field experiment conducted from January to July 2016 at the Experimental Unit of the Department of Agronomy at UFV-“Horta Velha” (20° 4524″ S, 42° 5045″ W; altitude, 648.74 m). The soil in the experimental area is classified as dystrophic Red Yellow Oxisol with a flat topography, and the climate in the region is Cwb, with an average annual temperature of 19.4 °C and annual precipitation of approximately 1200 mm.

The evaluation of the genotypes comprised vegetative traits, production, and chemical–nutritional aspects of fruits, seeds, and seed oil. Details about the agro-morphological descriptors used in the evaluation of germplasm are provided in Supplementary Table 2. The genotypes were also evaluated for multi-categorical traits and details about these traits are provided in Supplementary Table 3. The accessions were evaluated together with four commercial cultivars used as controls: the hybrids Tetsukabuto and Jabras (C. moschata × C. maxima) and the cultivars Jacarezinho and Maranhão. These genotypes were evaluated using Federer's augmented block design31, with five replications for each control. The four controls were randomly distributed in each block, and the accessions were randomly distributed among all the blocks in equal numbers. The experiment was established using a spacing of 3 × 3 m between plants and rows, with five plants per plot. The production and transplanting of seedlings and the cultural treatments were carried out in accordance with the local recommendations for the crop32.

The agro-morphological evaluations were carried out on the three central plants of each plot, using three fruits per plant. The carotenoid content was estimated based on the analysis of colorimetric parameters of fruit pulp, using a manual tristimulus colorimeter (Color Reader CR-10; Konica Minolta, Tokyo, Japan). This assessment was performed as detailed by18, according to the equations proposed by33, described below:

$$C=\sqrt{{a}^{2}+{b}^{2}}$$
$$TC = {6},{1226} + {1},{71}0{6}^*{\text{a}}$$
$$L = - {6}.{3743} + 0.{2818}^*{\text{C}}$$

where C corresponds to the saturation or chroma of fruit pulp; a and b correspond to the contribution of red and yellow to the color of fruit pulp (dimensionless), respectively; TC corresponds to the total content of carotenoids, and L corresponds to the lutein content of fruit pulp, both expressed in μg g−1 of fresh fruit mass.

The seed oil content (SOC) was determined using an extractor (ANKOM XT15, ANKON, Macedon, United States), according to a standard method from the Association of Official Analytical Chemists (AOAC), described by34. The extraction of seed oil was carried out using mechanical pressing, according to the methodology by18, and the fatty acid profile was analyzed using gas chromatography (GC). GC was performed using the GC-17A gas chromatograph (Shimadzu Corporation, Kyoto, Japan), equipped with an automatic insertion platform, flame ionization detector, and a Carbowax capillary column (30 m × 0.25 nm). Chromatography was performed under injection and detection temperatures of 230 and 250 °C, respectively. Column operation started at 200 °C, with an increase of 3 °C·min−1, until reaching a temperature of 225 °C. Nitrogen was used as a carrier gas with a flow rate of 1.3 L·min−1, and the concentration of each methyl ester was determined as a percentage of the relative peak area.

Implementation of restricted maximum likelihood procedures and the best linear unbiased prediction for analysis of agromophological data

Agromophological data were analysed from restricted maximum likelihood (REML) procedures and the best linear unbiased prediction (BLUP). This analysis was carried out using the R program and the package lme435. The genotypic values of accessions (BLUP) and controls (BLUES) were obtained from the BLUP, while the estimates of variance components were obtained from the REML, based on the following model:

$$y = {\text{W}}b + {\text{X}}a + {\text{Z}}t + e$$

where, y representes the phenotypic data vector, b representes the vector of blocks effect ssumed to be random, a representes the vector of accessions effect assumed to be random, t representes the vector of controls effect assumed to be fixed, and e represents the error vector. The letters W, X and Z represents the incidence matrices of parameters b, a, and t, respectively, with the data vector y. Both multivariate and ANNs analysis were carried out using the estimates of BLUP and BLUES.

The estimates of variance components comprised only the genotypic variance (σ2g). Heritability was obtained based on the following estimator: h2 = 1 − (Pev/σ2g), where Pev represents the prediction of error variance36.

Analysis of genetic variability using multivariate approaches and artificial neural networks

Multivariate analysis included the grouping of genotypes and the distribution of accessions in relation to principal components. Multivariate analysis of variability was carried out using both quantitative and multi-categorical information. For quantitative data, the distance matrix between genotypes was obtained using the standardized average Euclidean distance, from the estimates of BLUPs and BLUES. For multi-categorical data, the distance matrix was obtained from the arithmetic complement of the simple coincidence index. These matrices were then summed, resulting in a single distance matrix. For the sum of the matrices, they were standardized and each one received an equal weight in the summation procedure. The choice of the grouping method was based on cophenetic correlation, opting for the grouping that provided highest cophenetic correlation coefficient; while the determination of number of groups to be formed in clustering was based on the methodology proposed by37. Multivariate analysis were performed with the help of Matlab38 and the Genes software39.

A principal component analysis was implemented in order to identify the distribution of accessions in relation to the principal components. This analysis considered the data of quantitative and multi-categorical traits, according to the methodology of40; and was implemented with the help of Matlab38.

The analysis of the genetic variability organization through neural networks was carried out using Kohonen self-organizing maps (SOM). For this, different two-dimensional hexagonal topological maps were tested in which the N units (neurons) were allocated considering the number of rows and columns, ranging from 1 to 7. This procedure was based on the understanding that defining the topological map and, consequently, the number of neurons and parameters should be based on the researcher's experience, and trial and error methods41. Next, the selection of the best network architecture from 2000.00 training sessions for each of the combinations was carried out. The defined network topology had a hexagonal neighborhood. Network analysis was performed with the help of Matlab38 and the Genes software39.

Core collections were established from the random sampling of accessions from the full collection using sampling intensities of 10, 15, 20, and 25%. Thus, 9, 14, 18, and 23 accessions were sampled from the full collection to form the core collections with sampling intensities of 10, 15, 20, and 25%, respectively. The sampling of accessions for the establishment of the core collection was random and with no replacement. The validation of core collections was carried out from the comparison with the complete collection, based on the parameters obtained for the agro-morphological characteristics such as mean and variance29. Means and phenotypic variances of variables in the complete collection and nuclear collections were estimated with the aid of the Genes software39.

Collection and use of any plant materials statement

The authors declare that the plant collection and use was carried in accordance with all the relevant guidelines.

Results

Phenotypic range and heritability of traits

Based on the distribution analysis of traits, we observed a high phenotypic range for fruit production traits and chemical–nutritional aspects of fruit pulp and seed oil (Fig. 1). These amplitudes were especially higher for the productivity of fruits (PF), the total carotenoid content of fruit pulp (TC), and the oleic and linoleic fatty acid contents. Associated to this, these traits expressed significant genotypic variances and heritability estimates ranging from high to very high (Fig. 1).

Figure 1
figure 1

Frequency distribution of characteristics associated with fruit production and chemical–nutritional aspects of fruit pulp and seed oil. DDF, Accumulated degree days for flowering; NFP, Number of fruits per plant; PF, Productivity of fruits; TC, Total carotenoid content of fruit pulp; PS, Productivity of seeds; SOP, Seed oil productivity; LAC, Linoleic acid content, and OAC, oleic acid content.

Full size image

Most of the traits expressed greater amplitude between the accessions compared with the hybrids or lines used as controls.

Clustering of genotypes and principal components analysis from multivariate approach

The unweighted pair-group method using arithmetic averages (UPGMA) grouping method provided one of the highest cophenetic correlation indexes (> 0.7) and was adopted for the grouping of genotypes. Analysis of the variability using the multivariate approach showed that accessions and controls were grouped into seven groups. Groups 1 and 2 were the largest groups consisting of 33 and 37 genotypes, respectively (Fig. 2). Group 7 contained only the BGH-6749 genotype and was the smallest group.

Figure 2
figure 2

Grouping of accessions and controls based on a multivariate approach.

Full size image

Group 7 had the lowest number of accumulated degree days for flowering (DDF), followed by groups 6 and 4. These groups also had the lowest averages for DDF. Group 5 contained the genotype with the highest productivity of fruits (44.67 t. ha−1) and was the group with the highest average productivity of fruits (21.74 t. ha−1).

Groups 2 and 3 contained the genotypes with the highest average for total carotenoid content in fruit pulp, with contents of 187.21 and 181.17 μg g−1 of fresh mass, respectively. Group 4 contained the genotype with the highest content of oleic fatty acid in the oil (40.18%) and was also the group with the highest average for this characteristic (26.28%). Group 4 also contained the genotype with the lowest linoleic fatty acid content.

Figure 3 demonstrates the distribution of accessions in relation to the first two principal components (PC), emphasizing the analysis of genotype variability based on the multivariate approach. PC analysis highlighted accessions BGH-5456A, BGH-1992, and BGH-291 as those with the highest loads in the first PC. The first PC explained 80.6% of the total variation of genotypes in relation to agro-morphological characteristics, and the second PC explained 16.7%.

Figure 3
figure 3

Genetic variability of the 91 accessions of C. moschata kept in BGH-UFV from principal components (multivariate approach), showing the dispersion of genotypes in relation to the first two principal components.

Full size image

Organization of the accession’s variability from ANNs

Figure 4 shows the variability of accessions from ANN and SOM. It is observed that each neuron concentrated a similar number of accessions and controls, demonstrating an equitable concentration of these genotypes in the neurons (Fig. 4A,B, Table 2). ANNs analysis provided information about the genetic distance between the accessions and controls in each neuron. A tendency for genotypes with greater genetic distance to concentrate in the extreme neurons was observed (Fig. 4C).

Figure 4
figure 4

Kohonen’s self-organizing map demonstrating the concentration and genetic distances of genotypes in neurons. Distribution of genotypes in neurons (A,B) and genetic distance between the genotypes of each neuron (C). In Fig. 4A, the lighter color denotes greater number of accessions per neuron, while the darker color denotes smaller number of accessions per neuron. The lighter color denotes a greater distance between the genotypes in the neuron, while the darker color denotes a smaller distance in Fig. 4C.

Full size image
Table 2 Concentration of genotypes in neurons from Kohonen’s self-organizing map, as shown in Fig. 4A,B.
Full size table

Establishment and validation of the core collection

Table 3 shows the list of accessions of each core collection obtained from the different sampling intensities. Validation of the core collections was performed by comparing the mean and variance of the complete collection and the mean and variance of each core collection.

Table 3 List of accessions in core collections formed from different sampling intensities.
Full size table

In general, the core collection obtained under a sampling intensity of 15% (15% CC) presented a mean and variance closer to those of the complete collection. The means and variances for degree days accumulated for flowering (DDF), number of fruits per plant (NFP), mass of seeds per fruit (MSF), productivity of seeds (PS), and SOP characteristics using 15% CC were very close to those of the complete collection (Table 4).

Table 4 Means and variances of agro-morphological traits in the different core collections.
Full size table

Discussion

The high phenotypic ranges observed in this study for traits such as fruit productivity, total carotenoid content of fruit pulp, and oleic and linoleic fatty acid levels are in line with the genetic variability observed in previous studies of the C. moschata germplasm16,17,18,42.

Accessions BGH-5455A and BGH-5598A expressed the highest carotenoid contents with 187.21 and 181.17 μg g−1 of fresh pulp mass, respectively. This result is much higher than those reported in previous studies2,43. For example, the study involving the characterization of 55 accessions of C. moschata, also maintained by the BGH-UFV, reported a total content of carotenoids in the fruit pulp not exceeding 118.70 μg g−1 of fresh pulp mass44. On the other hand, when evaluating the C. moschata germplasm from Northeastern Brazil, Carvalho et al.1 reported averages of up to 404.98 μg g−1. The differences observed in the total content of carotenoids in fruit pulp between the present and previous studies may be mainly associated with the genetic aspects of the germplasm evaluated in each study. Studies with C. moschata generally reported high levels of carotenoids in fruit pulp1,45, particularly β- and α-carotene. These components are known for their important biological functions, such as provitamin A46 and antioxidant activity4.

Accessions BGH-5456A, BGH-3333A, BGH-5361A, and BGH-5472A expressed the highest levels of oleic acid. The emphasis on the analysis of the fatty acid profile of C. moschata aims at exploring the potential of this vegetable as an oleaginous crop. Consisting of approximately 75% of UFA and with a high content of MUFA such as oleic acid7,8, the oil from C. moschata seeds is an excellent substitute for lipid sources with high levels of saturated fatty acids, harmful to human health. Corroborating this, studies demonstrate the association between the consumption of lipid sources composed predominantly of saturated fatty acids and the high risk of cardiometabolic pathologies, particularly cardiovascular diseases and type II diabetes mellitus47,48. This has encouraged the replacement of saturated lipids in human food with UFA, with a particular focus on vegetable oils—the main source of UFA in the human diet.

Using multivariate analyzes and ANNs highlighted the high variability of C. moschata accessions. Clustering using a radial dendrogram allowed the identification of groups with the most promising averages in terms of accumulated DDF, PF, TC, and fatty acid profile (Fig. 2). The analysis of variability using PC corroborated the accession grouping pattern using the dendrogram, highlighting accessions BGH-5456A, BGH-1992, and BGH-291 as the most divergent (Fig. 3).

The analysis of the organization of the accession’s variability from ANNs corroborated the variability observed from the multivariate approach. This was confirmed by the concentration of a similar number of accessions along the neurons (Fig. 4A,B). This demonstrates that the adopted network architecture, consisting of seven columns and seven rows, efficiently organized the variability of the genotypes. Similar to the present study, a series of studies with Kohonen SOM also defined their topology randomly or by trial and error22,49,50. With this, it is assumed that the method to find the best architecture should be established judiciously. This is because different results can be obtained each time a SOM is used, given that networks have random synaptic weights at the beginning of training22.

Analysis using ANNs identified a tendency for genotypes with greater genetic distance to concentrate in the most extreme neurons (Fig. 4C), information that will support the establishment and validation of the core collections. Thus, genotypes concentrated in the extreme neurons express greater genetic distance. ANNs analysis enabled the organization of the genotypes into closer groups than those obtained from the radial dendrogram grouping (Fig. 2), proving to be more efficient in identifying similarity patterns and in organizing the proximity of genotypes between groups. Close to this, Santos et al.22 also used the SOM technique as an alternative method to assess genetic diversity in rice breeding programs. However, it should be noted that there is the possibility of greater variation in the allocation of genotypes in neurons as the number of neurons increase 51.

The variability observed among the genotypes of C. moschata in the present study is in line with previous studies with this species, characterized by high genetic variability, reflected, at first, in the variation of morphological aspects of plants and fruits. Studies have highlighted the variability of the Brazilian germplasm of C. moschata16,17,18, possibly a result of the adaptation of this germplasm to a wide ecological range found in the country, consisting of different edaphoclimatic conditions15. In addition, the occurrence of natural hybridization between populations also contributes to the variability in the germplasm of this vegetable18.

When establishing core collections, they must be evaluated regarding their ability to maintain the existing variability in the complete collection29,30. The averages and variances of agro-morphological characteristics of 15% CC were closest to the averages and variances of the complete collection, particularly in relation to DDF, NFP, MSF, PS, and SOP (Table 4). The 15% CC variances tended to be higher than the complete collection variances for most traits, which indicates that with this sampling intensity, the core collection effectively preserved the complete collection's genetic variability.

The validation of nuclear collections can be carried out using different approaches, such as the analysis of the amplitude coincidence index30,52, and are based on parameters analysis such as mean, variance, and amplitude29,53,54. For example, when proposing the establishment of a core collection based on the US Department of Agriculture soybean germplasm collection, Oliveira et al.29 emphasized the analysis of mean, variance, and amplitude observed in core collections as an approach for their validation. In this sense, Frankel55 highlighted that the sampling strategy is efficient when the core collection retains at least 80% of the original amplitude for a trait.

The establishment of a core collection aims to maintain the greatest possible variability from a minimum number of accessions, thus providing greater efficiency in identifying useful genetic diversity by breeders and other scientists. Given this, it is assumed that the 15% CC was effective since it presented means and variances very close to those of the complete collection and a number of accessions considerably lower than the full collection29,56.

According to27, establishing a core collection provides advantages for both collection curators and breeders. With the proposal of a core collection, two hierarchical levels are established, namely the core collection and the complete collection. From this, the curators can prioritize conservation activities such as germination and regeneration tests, in the core collections, in addition to concentrating efforts in the characterization and evaluation of the accessions of these collections. For breeders, evaluations of core collections often become less onerous due to the smaller number of accessions in these collections.

With the present study, the agro-morphological characterization of the collection of C. moschata maintained at the BGH-UFV approaches its conclusion18,44,57. Constituting a substantial sample of the Brazilian germplasm of C. moschata and one of the largest collections of this species in the country20, the characterization of this collection has covered the evaluation of an extensive set of characteristics, including the analysis of resistance against important phytopathogens of the crop, fruit and seed productivity; as well as chemical–nutritional aspects of fruits, seeds and seed oil18,44,57. Previous studies with the collection of C. moschata at the BGH-UFV allowed the identification of promising accessions as sources of genes for genetic improvement of this species.

The implementation of ANNs in the present study proved to be a useful tool to base the establishment of core collections, allowing a clearer distinction of the formed groups compared to the multivariate approach. Implementing ANNs for analyzing the organization of germplasm variability initially brings the advantage of mapping even trends or performances that do not follow linear behaviors58. Additionally, multivariate approaches bring disadvantages such as their association with the experimentation process and the nature of the data set. Therefore, a series of factors related to how the experimentation is conducted can compromise the efficiency of these analyses. For example, different genetic distance indices might be recommended for analyzing the diversity of a set of genotypes, depending on the statistical design in which they were evaluated. The Euclidean distance index, for example, is indicated for cases in which samples under evaluation have not been evaluated with repetition59, and in this case, the multivariate analysis does not include environmental errors that possibly have influenced the average results of samples. On the other hand, if there was repetition, the Mahalanobis distance is recommended59, which allows environmental errors to be contemplated in the multivariate analysis. The use of distance measures, such as the Euclidean ones, is restricted to quantitative data and recommended for cases in which there is no correlation between the variables, that is, for cases in which the variables are independent.

C. moschata crop presents characteristics that make the evaluation of its germplasm challenging. This species is characterized by branches with vigorous growth and long internodes32,60, which requires an extensive area for the evaluation of a reduced number of accessions, making the process costly. On the other hand, as already explained, the fruits and seeds of C. moschata express high nutritional value. Its fruits are characterized by a high content of carotenoids such as β- and α-carotene1,61, components with high provitamin A and antioxidant function4,46. Moreover, the seed oil of C. moschata consists of approximately 75% of UFA and has a high content of MUFA such as oleic acid7,8, components that are beneficial to human health. The establishment of the core collection proposed in the present study will be crucial to optimize the evaluation and use of promising accessions from this collection, especially for characteristics of high chemical–nutritional importance, such as the carotenoid profile of fruit pulp and the fatty acid profile of seed oil. The core collection could also be used as a source of alleles for genetic improvement programs of C. moschata and other cucurbits.

Conclusion

The accessions of C. moschata expressed a considerable phenotypic range for productivity of fruits, total carotenoid content of fruit pulp, and oleic and linoleic fatty acid contents, which enabled the identification of promising accessions for use as a source of genes for genetic improvement of these traits.

Multivariate analyzes and the approach using ANNs highlighted the high variability of C. moschata accessions evaluated in this study. The variability organization of accessions from ANNs corroborated the variability of accessions observed from the multivariate approach. This demonstrates that the network architecture adopted efficiently organized the genotype variability. ANNs were able to organize the genotypes into closer groups than those obtained from the radial dendrogram grouping, proving to be more efficient in identifying similarity patterns and in organizing the proximity of genotypes between groups. This information was fundamental to supporting the core collections' establishment and validation.

The averages and variances of agro-morphological traits using 15% CC were those closest to the averages and variances of the complete collection, particularly in relation to DDF, NFP, MSF, PS, and SOP, demonstrating that this core collection was efficient in maintaining the variability of accessions. Establishing the 15% CC will be crucial to optimize the evaluation and use of promising accessions from this collection, especially for traits of high chemical–nutritional importance, such as the carotenoid profile of fruit pulp and the fatty acid profile of seed oil.