Selection, genetic parameters, and multi-year stability of cacao yield under agroforestry in Rondônia, Brazil Authors: Caio Márcio Vasconcellos Cordeiro de Almeida, Fernando Luiz de Oliveira Corrêa, Francisco Antônio Neto, Gabriel Mazetti Blasques, Luiz Antônio dos Santos Dias Published: 2026 ABSTRACT AbstractCocoa is a commodity of great global importance and wide acceptance among consumers. Known as the chocolate tree, the cacao is cultivated in developing countries. More than 60% of that production is concentrated Côte d’Ivoire and Ghana and it is declining. This scenario presents a new opportunity for Brazil, the world’s seventh-largest producer. Rondônia, the fourth-largest producing Brazilian state, has several comparative advantages such as proximity to the Asian market, a predominance of a thousand small producers, and the potential to supply the market for fine cocoa. However, there is a need to recommend local clones selected for high performance and stability. To address this, we estimated genetic parameters and genotypic values using mixed models and investigated temporal stability by modeling the genotype × year interaction with covariance structures. Our study involved 30 clones under agroforestry, tested in randomized blocks with three replications and 9-plant plots. Cacao yield was assessed over seven consecutive years of harvests, counted from the third year in the field. Significant clonal effects were observed, and the mean dry bean yield (476 kg ha−1) showed marked amplitude (nearly sevenfold). Satisfactory experimental precision (CV = 37.5%), high values for the coefficient of genetic variation (), the ratio ), and heritability () reflected favorable conditions for selection. Of the ten covariance matrix models evaluated to model the matrix related to the genotype × harvest year (G x A) interaction, the factor analytic with three factors (FA3) provided the best fit. This better fit with FA3 likely arises from the fact that capturing the latent structure of the G × A interaction is not restricted to the time series, allowing it to also accommodate non-autoregressive correlations. The heterogeneous diagonal matrix, associated with FA3, was applied to model the residual covariance matrix. Based on repeatability coefficient’s accuracy, two years of successive harvests were sufficient for selecting clones in advanced breeding cycles, and five years for cultivar release. The genetic correlation matrix between pairs of harvest years reveals that cacao yield remained moderately associated (), with some years being more strongly associated ( than others. There was also a significant change in the ranking of clones over the years of harvests. With the factor analytic selection tools, we selected the five most productive and stable clones (EEOP 33, 63, 65, and 80, and Ipiranga 1), adapted to agroforestry conditions of Rondônia, with a substantial genetic gain of 64.66%. Productive and stable clones guarantee a relatively stable income over the years and ensure the grower’s long-term engagement in the activity.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-56812-y.Keywords: Theobroma cacao L., Mixed models, Temporal stability, Factor analytic, Genetic gain with selectionSubject terms: Biotechnology, Genetics, Plant sciences Introduction Chocolate is one of the most popular commodities globally, enjoying widespread acceptance among consumers. Its primary feedstock is cocoa seeds, referred to as beans after fermentation and drying, which are processed to extract butter, powder, liquor, and cake. These derivatives have diverse culinary applications, in addition to use within the cosmetic and pharmaceutical industries for the formulation of creams, lipsticks, shampoos, and other products. However, the product derived from cocoa beans and best known worldwide is chocolate. Botanically, the term “cocoa” refers to the fruit of the cacao tree (Theobroma cacao L.), a perennial plant native to the Americas belonging to the Malvaceae family. Its cultivation is concentrated in tropical regions, notably between latitudes 8º N and 8º S, within developing countries. Production is predominantly carried out by smallholder farmers operating on areas ranging from two to five hectares1. More than 60% of global cocoa production is concentrated in West Africa, particularly in Côte d’Ivoire and Ghana (https://nipponbusiness.com/cote-divoire-and-ghana-powerhouses-in-cocoa-production-controlling-over-50-of-global-supply/). These countries currently face substantial challenges. According to Nakweya2 “crop failures of these countries have been attributed to climate change and El Niño which triggered erratic rainfall and higher temperatures, stimulating the proliferation of pests and diseases such as black pod disease and cocoa swollen shoot virus disease (CSSVD), transmitted by mealybugs”. As a consequence of these factors, the supply of cocoa has diminished, contributing to a significant increase in market prices in recent years, most notably in 2024, when prices surpassed the record threshold of US$12,000.00 per ton3. This scenario presents a new opportunity for Brazil. The country possesses an established cocoa processing industry with an annual capacity exceeding 275,000 tons of beans. However, domestic production remains insufficient to fully supply this industrial capacity, with the 2024/25 harvest estimated at 180,000 tons of beans4. This scenario compels the Brazilian industry to import cocoa in order to maintain its processing capacity and prevent a decrease in employment opportunities5. However, such imports entail inherent phytosanitary risks, with the potential to introduce pathogen into the beans. Among the most concerning pathogens is Moniliophthora roreri, the causal agent of frosty pod rot or moniliasis, a highly invasive disease capable of causing up to 90% yield losses and still absent from Brazil6. It is important to emphasize that Brazil possesses land suitable for expanding production, a climate favorable to cultivation, a fully integrated supply chain, and validated technologies to increase the average productivity of its plantations7–9. Additionally, the global cocoa market has acted as a catalyst, with annual demand growth averaging 3% over the past 100 years1. The expansion of this commodity in the state of Rondônia adds another strategic advantage, represented by its geographic proximity to the Asian market via the Pacific Ocean, the so-called Bioceanic Route, which offers shorter transit times (up to 17 days less) and lower freight costs (30% less) compared to the Port of Santos10. Rondônia is further distinguished by its agrarian structure, composed predominantly of smallholder farmers, who account for approximately 80% of the state’s 110,000 rural properties. Thousands of these smallholders are settled in agrarian reform projects established by INCRA (National Institute for Colonization and Agrarian Reform)11. This land tenure structure in Rondônia facilitates the organization of the productive base in pursuit of greater added value, particularly through the export of processed cocoa in the forms of butter, powder, liquor, and cake. Over the past five years, annual Brazilian cocoa production has fluctuated around 190,000 tons4, with the states of Pará and Bahia accounting for approximately 95% of this volume, followed by the states of Espírito Santo, Rondônia, and Amazonas. Rondônia, the fourth largest producer nationally, is also the second largest producer in the northern region, with 7,431 hectares under cultivation and a production of 5,053 tons11. This state has distinguished itself in national competitions for the high quality of its beans, demonstrating potential to supply the fine chocolate market. The expansion of Brazilian cacao cultivation has occurred through the use of hybrid varieties, predominantly in Pará, and clonal varieties in other states12,13. In northern Espírito Santo state, as well as in southern Bahia state, monoclonal plantings, mainly with the clones CCN 51 and PS 1319, have become prevalent, despite the inherent risks associated with genetic uniformity that may increase the incidence of pests and diseases, such as moniliasis14. ‘CCN 51’ and ‘PS 1319’ have demonstrated outstanding productivity in commercial monoclonal block plantings in northern Espírito Santo state9, as well as in southern Bahia state15. Notably, ‘CCN 51’ has emerged as the most productive and stable clone in cacao-growing regions of Ecuador, where it was originally selected, as well as in Colombia, Peru, and Brazil16. Genotype-by-environment interaction (G × E) is one of the main challenges faced in plant breeding. It occurs when genotypes respond differently to varying cultivation environments. Thus, genotypes that perform well in one environment may be outperformed by others when tested in different environments17. It is important to note that “environment” is a generic term that may refer to locations, years, or cropping seasons. The presence of G × E interaction is directly linked to the concept of stability, that is, the ability of a genotype to consistently reproduce its performance across multiple environments. Stability itself can be classified as either spatial and/or temporal. In the first case, environments correspond to the locations where genotypes are tested; in the second, to cropping seasons or years of cultivation. In the breeding of perennial species such as cacao and coffee, temporal stability assumes great importance to secure long term productivity18–24, although spatial stability is also evaluated25. The temporal stability of traits is a major issue in perennial crop breeding18–24. In perennial fruit trees such as cocoa and coffee, cumulative yield over the years is one the most common selection criterion, as well as disease resistance26. However, different yield distributions over time can result in identical cumulative yield, since some cultivars are early yielders, with a high yield in the first years, while others are more stable over time20. The temporal stability of yield is also of interest to the cocoa grower, in their pursuit of stable income. Understanding the distribution of yield in consecutive years of harvests is accessed through analyses of repeated measurement data of cocoa trees over time. Such analyses assume the modeling of the structure of cocoa tree data over time. It is important to note that plant breeding is grounded in a quantitative theoretical framework that enables predictions regarding the performance of genotypes evaluated in one or multiple cultivation environments. For this purpose, it is necessary to estimate genetic parameters such as heritabilities, coefficients of variation, genetic variances, selection gains, among others17. The optimal method for estimating genetic parameters and predicting additive and genotypic genetic values is the mixed model approach using REML/BLUP27. It is worth emphasizing that the estimation of genetic parameters in cacao has been occurring28,29 but it needs to be further explored, particularly through the application of mixed model methodologies.30,31. One of the main advantages of using mixed models lies in their flexibility to model complex variance and covariance structures associated with random effects32. In multi-environment trials, modeling G × E interaction through Factor Analytic (FA) structures has become the gold standard, as it allows for a parsimonious representation of the covariance matrix, capturing the main interaction patterns through a reduced number of latent factors33. Based on this approach, Smith & Cullis34 proposed the FA Selection Tools (FAST), a set of metrics derived directly from the FA model, which enable the evaluation and selection of genotypes considering both stability and adaptability simultaneously. Given the opportunity for the expansion of cacao cultivation in Rondônia state and the need to provide producers with productive and reliable cultivars, our research evaluated, through a multi-year trial, the performance of cacao clones under agroforestry. We estimated genetic parameters and genotypic values using mixed models (REML/BLUP) and investigated temporal stability by modeling the genotype × year interaction with covariance structures. Our hypothesis is that the existing genetic variability allows for the recommendation of superior clones and the acceleration of breeding progress in the region, thereby strengthening the sustainability of the production chain. Material and methods Trial site The study was conducted at the “Frederico Afonso” Experimental Station (ESTEX-OP, lat 10º 43′ 42.0″ S, long 62º 14′ 21.5″ W, 280 m alt asl), located in Ouro Preto do Oeste, Rondônia, Brazil. ESTEX-OP is managed by the Executive Committee of the Cocoa Crop Plan (CEPLAC), under the Ministry of Agriculture and Livestock (MAPA). According to the Köppen classification, Rondônia has a tropical Aw climate with a distinct wet season. The climate series for Ouro Preto do Oeste and Ji-Paraná municipalites indicate a hot and humid climate, with a mean annual temperature of 25 ºC, mean annual precipitation of 1900 mm, of which 92% occurs from October to April, and a dry season from May to September35 (Fig. 1). The soil at the experimental site is a Red-Yellow Argisol of medium natural fertility36, with flat topography. Annual rainfall distribution (mm) during the wet (October–April) and dry (May–September) seasons in Ouro Preto do Oeste, RO, Brazil, from 2018 to 202535,72. In our agroforestry system (Suplementary Fig. 1), banana plants (Musa sp.) were established in January 2014 at 3 × 3 m spacing to provide temporary shade, while forest and fruit species were planted at 15 × 15 m to form the permanent overstory. This permanent shade layer consisted of regional species with suitable canopy architecture, including Cedrela fissilis, Eugenia involucrata, Paubrasilia echinata, Caryocar brasiliense, Hymenaea courbaril, Handroanthus heptaphyllus, Carapa guianensis, Inga spp., Talisia esculenta, Garcinia gardneriana, and Annona squamosa. In 2016, stakes of Gliricidia sepium, a fast-growing, nitrogen-rich, repeatedly prunable legume, were added at 6 × 6 m to reinforce shading, and Canavalia ensiformis was sown between rows as a green manure crop in 2016 and 2017. Cacao clones were planted in January 2016 at 3 × 3 m spacing. There are no records reporting some of the species in our agroforestry system as being potential natural hosts of Moniliophthora perniciosa, the causal agent of witches’ broom disease. It is important to note that with the spacing adopted and the types of canopies of the shade trees used, we estimate that our agroecosystem operates with 50 to 60% shading for the cocoa trees. It is also worth noting that cocoa cultivation in Rondônia is implemented by small family farmers (more than 80%), where the use of agroforestry systems is common, differing only in tree composition compared to our trial. Evaluated clones and experimental design The 30 clones planted (Suplementary Table 1) were obtained as cuttings collected from the Active Germplasm Bank at ESTEX-OP and grafted onto rootstocks that had been previously prepared and maintained in a seedling nursery. We used rootstocks formed from seeds of hybrid varieties produced at ESTEX-OP. These hybrids involved crosses between Upper Amazonian x Middle/Lower Amazonian clones, seeking greater heterotic expression. Grafting was performed using the top wedge technique. The selection of these clones was based on preliminary evaluations conducted at ESTEX-OP37,38, on the process of clonal cultivar selection for Rondônia39,40, and on the agronomic performance of clones in Bahia41. Overall, all of these clones were selected for tolerance to witches’ broom disease (WBD) and productivity at different times37–40. At planting, the clonal seedlings were three months old, with plagiotropic branches longer than 40 cm and at least six to eight mature, well-formed leaves. The clones CCN 51 and PS 1319 were used as controls for high productivity and tolerance to witches’ broom disease (WBD), while the clone POUND 12 was the control for susceptibility to WBD. The experimental design adopted was a randomized complete block design, with 30 clones, three replications, and plots consisting of nine plants arranged in three columns and three rows. As a border for the experiment, to avoid microclimatic changes that could alter factors such as light, humidity, and average temperature in these areas, two rows of clonal cacao trees were established in the east–west direction and one row in the north–south direction, using seedlings of the clones CCN 51 and PS 1319. The maintenance of the trial has been carried out according to the technological recommendations for cacao cultivation in the Amazon, which include weed control, pruning and thinning of cacao trees, chemical fertilization, and integrated pest and disease management42. Evaluated traits Phenotypic data were generated from fruit harvests conducted between 2019 and 2025, corresponding to cacao trees aged three to nine years in the trial. We conducted seven to thirteen annual harvests between February and October, peaking from April to August, and evaluated the weight of wet seeds per healthy fruits (WWS) on a per-plot basis. The trait analyzed was dry bean weight per hectare (called cacao yield; in kg ha⁻1), calculated using the following expression: where WWS was divided by the number of plants per plot (in case 9 for full plot, or the number of living trees present), converted to dry seed weight by multiplying by the conversion factor of 40% from wet to dry cocoa, and standardized to 1 hectare by multiplying by 1,111 plants. Individual analyses by harvest year Initially, the cacao yield data were processed for residual analysis and fit to the normal distribution. Subsequently, individual analyses by harvest year (Eq. 1) were conducted. Variance components were estimated using the Restricted Maximum Likelihood (REML)43 algorithm, while genotypic values were predicted using the Best Linear Unbiased Predictors (BLUP)44 algorithm. Thus, considering n phenotypes of v clones (i = 1, 2, …, 30) and b blocks (l = 1, 2, 3), we have: where is the vector of phenotypic observations; is the vector of fixed effects for replication; is the vector of random genotypic effects, with ; and is the vector of random errors, with . and are the incidence matrices associated with the replication and genotypic effects, respectively. Still within the context of individual analyses by harvest year, the significance of the genotypic effect was evaluated using the likelihood ratio test (LRT)45. The test is conducted by comparing the full model, which includes the random genotypic effect, with a reduced model in which this effect is omitted, according to the following expression: where logL and represent the log-likelihoods of the models with and without the genotypic effect, respectively. The test value was compared to the chi-square () distribution with one degree of freedom, at the 5% significance level. From the estimates of variance components obtained from mixed models, the following parameters were estimated: means, experimental coefficient of variation (CV) and genetic coefficient of variation (), the ratio between and CV (), and heritabilities (46, as shown below: where is the residual variance; is the genotypic variance; and is the phenotypic mean of the evaluated trait where PEV is the mean prediction error variance. Joint multi-year analyses by harvests Considering n phenotypes of v clones (i = 1, 2, …, 30), in b blocks (l = 1, 2, 3), and evaluated over m harvest years (j = 1, 2, …, 7), the mixed model used in the joint multi-year analyses by harvests was (Eq. 2): where is the vector of phenotypic observations; is the vector of fixed effects for harvest years; is the vector of fixed effects for replication; is the vector of fixed effects for replication within harvest years; is the vector of random effects for the genotype × harvest year interaction, with ; is the vector of random effects for permanent plots, with ; and is the vector of random errors, with . The matrices and represent, respectively, the genetic and residual covariance matrices between harvest years. , are the incidence matrices associated with the fixed effects of harvest years, replications, and replications within harvest years, respectively. and are the incidence matrices associated with the random effects of genotype × harvest year interaction and permanent plots. Modeling and selection of the covariance matrix A total of 10 covariance matrix models were evaluated to model the matrix related to the genotype × harvest year interaction. The models assessed, as described in Chaves et al.33 and Dessauw et al.18 were: Compound Symmetry (CS), Heterogeneous Diagonal (DIAG), Homogeneous Compound Correlation (CORH), Heterogeneous Compound Correlation (CORGH), First-order Autoregressive (AR1), Heterogeneous Autoregressive (AR1H), Ante-dependence (AD), Factor Analytic with 1 factor (FA1), Factor Analytic with 2 factors (FA2), and Factor Analytic with 3 factors (FA3). For the selection of the best-fitting covariance matrix model, the Akaike Information Criterion (AIC)47 shown below was used: where is the estimate of the maximum point of the residual likelihood function, and p is the number of estimated parameters. We also use the Bayesian Information Criterion (BIC)48 for the same propose as AIC: where k is the number of estimated parameters in the model, and n is the number of independent observational units. Modeling for factor analytic (FA) When selected, the FA structure was used to model the genetic covariance matrix of the genotype × harvest year interaction, . For an FA model of order k, this matrix is expressed as: where is the matrix of factor loadings associated with the k latent factors common to the m harvest years, and is the diagonal matrix containing the specific variances for each harvest year not explained by the common factors. The genotypic effects () were modeled as: where is the vector of factor scores for the v-th genotype, assumed as , e represents the adjustment deviations. To ensure model identifiability, the factor loadings were estimated under lower triangular and column orthonormality constraints, such that . After estimation, the matrix of factor loadings was subjected to an orthogonal rotation based on singular value decomposition (SVD): where e are orthogonal matrices and is the diagonal matrix of singular values. The factor scores were rotated in accordance with this transformation, so that: where is the original vector of scores for the v-th genotype and is the vector of rotated scores. Repeatability analysis across harvest years Based on the estimates of variance components, we calculated the phenotypic variance (), the repeatability coefficient (ρ), and the accuracy of repeatability (Acc), as follows: where m is the number of harvest years. To estimate the optimal number of harvest years for genotype selection, we used the criterion of Resende & Alves49, which allows classification of accuracy according to the desired p-value. We defined minimum p- values of 0.10 and 0.01 for selection in advanced breeding cycles and for cultivar release, respectively, as used by Chaves et al.50. Genetic correlation and relative ranking of clones across harvest years The genetic correlation matrix () between pairs of harvest years was estimated as proposed by Cullis et al.51. where is a diagonal matrix containing the inverse square roots of the diagonal elements of the covariance matrix ; represents the matrix of rotated factor loadings; and is the diagonal matrix of specific variances for the harvest years. Selection of clones for productive performance (OP) and stability (RMSD) and genetic gain with selection We applied the Factor Analytic Selection Tools (FAST)34,52 to assess the performance and stability of the clones. Thus, performance was estimated based on the OP (Overall Performance) parameter, and stability was estimated using the RMSD (Root Mean Square Deviation) parameter, as described below: where m is the number of harvest years, are the rotated loadings of the first factor in the j-th environment, are the rotated scores of the first factor for the v-th clone, and After selecting the five most adapted and stable clones (16.7%), we estimated the genetic gain (SG%) as: where is the mean of the BLUPs of the selected clones and is the overall mean of the clones. All analyses were processed using the ASReml-R package, version 4.153, in the R software54. Ethical statement The plant species Theobroma cacao L. used in this study is a cultivated plant, and the evaluated clones are grown at the “Frederico Afonso” Experimental Station (ESTEX-OP), located in Ouro Preto do Oeste, Rondônia, Brazil. ESTEX-OP is managed by the Executive Committee of the Cocoa Crop Plan (CEPLAC), which granted us permission to conduct field research with the cultivated cacao clones. We confirm that all necessary standards for this type of research were fully observed. Results Individual analyses by harvest year The likelihood ratio test (LRT) showed significant effects (p < 0.05) for genotypes with respect to cacao yield. The mean dry bean yield (476 kg ha−1) varied over the seven years of harvests. The most productive year (837 kg ha−1) was 2020, while 2024 was the least productive (122 kg ha−1) (Table 1). Estimates of the experimental coefficient of variation (CV) ranged from 29.2% (2020) to 58.6% (2024), the latter resulting from the low productivity recorded in 2024. Estimates of the genetic coefficient of variation () ranged from 39.2% (2022) to 76.5% (2021), reflecting favorable conditions for selection. This favorable condition was reinforced by the estimates of the ratio between and CV with a mean of 1.65. The heritability () ranged from 0.874 (2024), the year of worst productive performance of the clones, to 0.960 (2021), with a mean of 0.932. Table 1Estimates of genetic parameters for dry cocoa bean yield (kg ha⁻1), by harvest years.YearsParameters12019202020212022202320242025MeanMean (kg ha−1)716837270815224122348476CV(%)30.829.236.432.438.458.636.637.5(%)54.052.976.539.270.670.161.360.71.751.812.101.211.841.201.681.650.9490.9510.9600.9040.9490.8740.9400.9321 CV: Experimental coefficient of variation; : Genetic coefficient of variation; : Ratio between e CV e; : Heritability51. Estimates of genetic parameters for dry cocoa bean yield (kg ha⁻1), by harvest years. 1 CV: Experimental coefficient of variation; : Genetic coefficient of variation; : Ratio between e CV e; : Heritability51. Joint multi-year analyses by harvests Of the ten covariance matrix models evaluated to model the matrix related to the genotype × harvest year interaction, the Factor Analytic model with three factors (FA3) provided the best fit (Table 2), according to the AIC criterion45. The AIC criterion establishes that the best-fitting model is the one that returns the lowest test value, in this case FA3 with 6742.70. Table 2Selection of covariance matrix models to model the matrix related to the genotype × harvest year interaction, according to the AIC criterion, for dry cocoa bean yield.NrModels of covariance matrixAIC criterionBIC criterion1CS—Compound symmetry6860.796878.162AR1—First-order autoregressive6847.196864.553DIAG—Heterogeneous diagonal6824.196863.274CORGH—Heterogeneous compound correlation6770.586900.855CORH—Homogeneous compound correlation6766.046809.466FA1—Factor analytic with 1 factor6761.036826.167AR1H—Heterogeneous autoregressive6755.116798.538AD—Ante-dependence6748.006808.389FA2—Factor analytic with 2 factors6744.076826.1610FA3—Factor analytic with 3 factors6742.706838.22 Selection of covariance matrix models to model the matrix related to the genotype × harvest year interaction, according to the AIC criterion, for dry cocoa bean yield. After selecting the FA3 covariance matrix model as the best fit for modeling the matrix related to the genotype × harvest year interaction, we evaluated two structures for the residual covariance matrix. As a result, the DIAG structure was considered, according to the AIC criterion (Table 3). Table 3Selection of residual covariance matrix models according to the AIC criterion, for dry cocoa bean yield.NrModel selected of residual covariance matrix1AIC criterionBIC criterion1FA3 + AR1 residual6735.026839.232FA3 + DIAG residual6591.846726.441 See model codes in Table 2. Selection of residual covariance matrix models according to the AIC criterion, for dry cocoa bean yield. 1 See model codes in Table 2. Repeatability analysis across harvest years Based on the estimates of accuracies (Acc) of repeatability coefficients (ρ), we estimated the optimal number of harvest years for the selection of cacao clones in advanced breeding cycles and for cultivar release. According to the evaluations of Resende and Alves49, p-values of 0.10 and 0.01 for the above selection conditions correspond to accuracy values of 0.79 and 0.92, respectively. In this context, the productivity of cacao clones reached an accuracy of 0.85 as early as the second year of consecutive harvests, and above 0.93 from the fifth year of consecutive harvests onward. It is worth noting that the accuracy of dry bean yield ranged from 0.75 (2019) to 0.95 (2025) (Suplementary Fig. 2). Genetic correlation and relative ranking of clones across harvest years The genetic correlation matrix between pairs of harvest years (Fig. 2) reveals that cacao yield remained moderately associated () throughout the experimental evaluation period. Intersting, the first two consecutive years (2019 and 2020) maintained a high correlation () between their yields. There were also strong correlations between yields recorded in 2022 and 2023 (), 2022 and 2025 (, and between 2024 and the first two years (2019 and 2020). The range of genetic correlations between pairs of harvest years results from the pronounced changes in the ranking of clones in terms of cacao yield (Fig. 3). Heatmap of genetic correlations between harvest years, related to dry cocoa bean yield. Relative ranking of clones across harvest years, with their BLUPs, related to dry cocoa bean yield. Selection of clones for productive performance (OP) and stability (RMSD) and genetic gain with selection With the factor analytic selection tools (FAST), we selected the five most productive and stable clones, namely EEOP 33, 63, 65, and 80, and Ipiranga 1 (Fig. 4). According to these selection tools, the genotype of interest should have the highest OP (Overall Performance) and the lowest RMSD (Root Mean Square Deviation). By selecting the five most productive and stable clones, we achieved a substantial genetic gain of 64.66%. Plot of the five clones with the highest performance (OP) and stability (RMSD) selected for dry cocoa bean yield. The most productive clones are positioned at the top of the Y-axis, and the most stable ones on the left side of the X-axis. Discussion Generally, cacao breeding programs in producing countries focus on two main lines of action: mass selection of clones and the selection of biclonal hybrid families (crosses between heterozygous clones in pairs)55,56. The primary objective of these programs has been breeding for increased yield, improve quality, and resistance to pests and diseases7,57. Although these programs are consistent and have clear goals, the progress achieved has been lower than expected. Dias7 lists several factors that hinder this progress, such as the existence of genetic self-incompatibility in many populations, the particularities of propagation, management, and crop physiology, and long-term field evaluations that affect the survival rate of cacao trees. Furthermore, successful genetic improvement programs presuppose rigid and robust field experimentation, as well as broad access to cacao genetic resources, as argued by Nieves-Orduña57. Finally, estimates of genetic parameters, such as heritability, coefficients of variation, and genetic correlations, are rarely obtained in cacao breeding, hindering the understanding of the genetic architecture of the main traits of interest. Our research revealed expressive genetic variability among cocoa clones (LRT test, p < 0.05), a necessary and favorable condition for selection practice. This result also confirmed the effectiveness of the selective process carried out in Rondônia state37–40, given that 15 of the 30 clones evaluated in the study were locally selected. It is important to note that the average yield of dry beans recorded in the agroforestry trial (476 kg ha⁻1) was in the same order of magnitude as the Brazilian average cocoa yield (455 kg ha⁻1)58. However, yields in the trial ranged from 122 kg ha⁻1 (2024) to 837 kg ha⁻1 (2020), revealing strong yield oscillation among clones over the years. This marked amplitude in productivity (nearly sevenfold) reinforces the need to select not only high-yielding clones but, above all, those that are productively stable over the years. Experiments in Rondônia involving 40 clones resulted in yields between 935 and 1350 kg ha⁻1 for the 13 best clones40. It is worth noting that an average yield of 258 kg ha⁻1 was recorded in experiments in Costa Rica59. The experimental precision, reflected by the mean experimental coefficient of variation (37.5%), can be considered acceptable and in the same order of magnitude as reported in other studies23,38–40,60. However, an experimental CV of 58.6% was recorded in 2024, which may be explained by the clones’ poorest performance over the seven harvest years. The genetic variation expressed as a percentage of the mean by the genetic coefficient of variation was significant (60.7%), consistent with the LRT test for clones. The which serves as an indicator of the clones’ potential for selection application, was high (1.65). In maize progeny trials, Vencovsky61 states that values above 1.0 reflect a favorable situation for selection. Heritability ()51 was developed for unbalanced experimental data conditions and complex correlation structures. It differs from conventional heritability, which is the ratio between genetic and phenotypic variance, as it measures the proportion of genetic variance relative to the variance of the differences between the genotype BLUP means. Thus, it is interpreted as a broad-sense heritability estimate yet corrected for the experimental design structure. Consequently, indicates how reliable the genotype ranking is based on their predicted values. If values are close to 1, it indicates that the genetic variability among clones is almost entirely of a genetic nature . In the present study, the average was 0.932, indicating that selection for yield will be highly efficient. For illustration purposes only, and not for comparison, Duval et al.29 and Mustiga et al.59 have estimated narrow-sense heritability values for dry bean yield ranging between 0.09 and 0.50, based on individual plant data. In the selection of the covariance matrix for the genotype × harvest year interaction (), we identified the factor analytic model with three factors (FA3) as the best fit, based on the Akaike Information Criterion (AIC). Compared to the BIC, the AIC tends to favor more complex models, with a minimax optimality for prediction62. The use of FA structures to model the G × A interaction covariance matrix is widely recommended in multi-environment trials, as it allows for the description of complex covariance patterns using a reduced number of parameters, providing parsimony in the estimation of variance components63,64. Subsequently, heterogeneous diagonal matrix model, in addition to FA3, was selected to model the residual covariance matrix. Among the residual structures evaluated, the heterogeneous diagonal matrix (DIAG) showed the best fit, in line with findings reported by Chaves et al.24,33, allowing for the modeling of specific residual variances for each harvest year without imposing residual correlation between observations. This structure is particularly suitable for data where differences in experimental quality between crops and environmental conditions result in residual heteroscedasticity. The adoption of specific residual variances per year increases model flexibility and reduces potential biases in the estimation of genetic parameters, when the assumption of residual homogeneity is not met33. The ante-dependence model has often been reported as the most suitable approach for analyzing longitudinal data of cocoa clones evaluated over 1319 and 18 years18. This model assumes that correlations between two successive years remain relatively high and stable, while correlations between years diminish as the temporal gap increases. However, the erratic yield behavior observed in our dataset did not align with such expectations. We identified the factor analytic model with three factors (FA3) as the best fit. The superior fit of the FA3 model likely results from its capacity to capture the latent structure of the G × A interaction without being confined to a strictly time-series framework. This flexibility enables the model to account not only for autoregressive correlations but also for non-autoregressive patterns. The number of successive harvest years is an important parameter for defining the minimum observation time necessary to practice selection in cocoa trees. For this purpose, the repeatability coefficient and its associated accuracy (Acc) are estimated49. It is important to emphasize that accuracies differ depending on the stage of the breeding program and the desired probability level (p-value). Thus, we adopted a p-value of 0.10 (equivalent to a repeatability accuracy of 0.79) for selecting clones in the advanced stage of the breeding program. Additionally, we adopted a p-value of 0.01 (equivalent to a repeatability accuracy of 0.92) for selecting clones in the cultivar release phase. For the first selective scheme, two consecutive harvest years are sufficient (Acc = 0.85), while five consecutive harvest years (Acc = 0.93) are required to select clones for release. Dias and Kageyama22 estimated a minimum period of two successive harvest years (with Acc = 0.90) necessary for selecting hybrid cocoa trees with 11 years in the field, in southern Bahia. Although genetic correlations between pairs of harvest years revealed that productivity remained moderately associated () throughout the seven harvest years, some years were more closely associated with each other. The first two harvest years (2019 and 2020), as well as 2022 and 2023, were found to be strongly associated (). Part of the fluctuation in the association between pairs of harvest years originated from changes in the performance ranking of the clones. Another important factor that may have determined this oscillation in the correlations between annual yields was rainfall, both in terms of annual total and distribution. The year 2024 presented the lowest volume of rainfall (1572.7 mm), and the most intense drought (Fig. 1), correlating moderately with 2022 (), 2023 (), and 2025 () (Fig. 2). Alvim et al.65 concluded that cocoa trees respond strongly to water availability, with accumulated rainfall over the preceding six months being one of the most reliable indicators for predicting future productivity. This cumulative precipitation influences key physiological processes, including leaf water potential, vegetative growth rhythms (flushing and leaf fall), flowering, and fruit development. Building on Alvim et al.’s hypothesis, we represented the rainfall distribution from 2018 to 2025 as occurring in two distinct seasons: a wet season (October to April) and a dry season (May to September) (Fig. 1). Within this framework, it is plausible that the dynamics of cocoa yields across years have been shaped by the occurrence of extreme wet and dry seasons. The dynamics of cocoa yield over time has been investigated18,19,66. Previous studies reported that correlations between yields in consecutive years are typically high, declining progressively as the interval between years increases18,19. This pattern of yield behavior was observed in datasets spanning 1319 and 18 consecutive years of harvests18. However, the erratic yield trends observed in our dataset did not conform to these findings. For instance, 2024 was both the driest year and the least productive among those evaluated. Its yield correlations with 2022 (C = 0.45), 2023 (C = 0.48), and 2025 (C = 0.41) were only moderate, whereas the correlation with 2019 was unexpectedly high (C = 0.92) (Fig. 2). These results highlight that extreme climatic conditions can disrupt the expected temporal correlation structure of cocoa yield, underscoring the importance of incorporating environmental variability into longitudinal analyses. The application of factor analytic selection tools resulted in the identification of the five most productive and stable clones. The locally selected clones EEOP 33, 63, 65, and 80, along with Ipiranga 1, selected in the state of Bahia, exhibited superior performance and stability. Temporal stability, evaluated over successive harvest years at the same site, is of paramount importance to the cocoa grower18–24,60. The implications of temporal stability are distinct from those of spatial stability, as annual fluctuations cannot be predicted. Breeding programs must select buffered cultivars capable of adjusting their vital processes to maintain productivity under unpredictable environments67. Traditional growers who cultivate cocoa under shade seek not only high-yielding genotypes but also those that combine high productivity with temporal yield stability. Dias et al.23 have reinforced the need to select clones locally with temporal stability. Productive and stable clones guarantee a relatively stable income over the years and ensure the grower’s long-term engagement in the activity. Clones coded as EEOP were selected by the CEPLAC Experimental Station, in Ouro Preto do Oeste, RO, from mass selection practiced in cacao populations composed of hybrid varieties established in the 1970s. This mass selection aimed to identify productive genotypes tolerant to pests and diseases. Specifically, ‘EEOP 33’ has a partially documented pedigree and originated from a controlled cross between clones POUND 7 and BE 10, indicating introgression from Upper Amazon genetic materials. In the literature, the clone EEOP 33 stood out at the time for its high yield (945 kg ha⁻1)40. In contrast, ‘EEOP 63’, ‘EEOP 65’, and ‘EEOP 80’ were selected from improved seed populations derived from Amazonian plantations during regional breeding initiatives conducted by CEPLAC in Rondônia. ‘EEOP 63’ and ‘EEOP 65’ show high productivity and lower incidence of witches’ broom, attributed to balanced nutritional status and greater nutrient-use efficiency68. ‘EEOP 80’ combines resistance to the cacao fruit borer (Conotrachelus humeropictus) with high productivity. These four selected EEOP clones resulted from crosses between contrasting parents and are all self-compatible69. The EEOP series 55 to 97 resulted from controlled crosses between contrasting parents from the EEOP series 23 to 54, based on the pedigrees of the latter, in order to avoid inbreeding. Therefore, the EEOP clones exhibit significant genetic variability. Finally, the Ipiranga 1 clone, also self-compatible, shows yield potential above 1,100 kg ha⁻1 and is among those recommended for cultivation in southern Bahia41. The selection of the five most productive and stable clones (EEOP 33, 63, 65, and 80, and Ipiranga 1) resulted in a expressive genetic gain of 64.66%, in line with gains reported by Dias & Resende31. The productivity of these five top-performing and most stable clones reached 800 kg ha⁻1, equivalent to 1.75 times the Brazilian average (corresponding to 475 kg ha-1)58. For cocoa growers in agroforestry systems, this productivity increase may be essential for sustaining their activity. Beyond increased financial returns, these growers will benefit from a more sustainable, ecologically balanced, and potentially longer-lived crop due to shading. It is worth noting that in trials in Bolivia, cocoa agroforestry systems have shown lower performance than monocultures, and that locally selected clones are two to five times more productive than some international Trinitario hybrid clone, regardless of the cultivation system70. It is important to emphasize that cocoa research offers growers willing to invest in a technological package (primarily productive and stable clones, fertigation, and integrated pest and disease management, among other technologies) the possibility of reaching productivities of up to 3,000 kg ha⁻1 of dry beans, as reported by Benjamin et al.9. However, we understand that agroforestry systems are the future of cocoa cropping in the Amazon, due to their environmental, economic, and social benefits. Extrapolating our results Finally, one question remains: What is the representativeness and applicability of these experimental results generated at ESTEX-OP for the cocoa-growing region of Rondônia? Although our data do not allow for a conclusive analysis, it is perfectly possible to draw a comparison between the soil types and climate of both to support the extrapolation of our results. Climatologically, both ESTEX-OP and the Rondônia’s cacao belt are dominated by a tropical humid climate (Köppen’s Am type). This climate is characterized by elevated mean annual temperatures ranging from 24 to 26 °C and substantial annual rainfall spanning 1,400 to 2,600 mm, marked by an intense rainy season from October to April and a distinct, drier window between May and September35,71,72. The soils at ESTEX-OP form a heterogeneous landscape composed of Ultisols, Alfisols, Red-Yellow Podzolic soils, and Latosols. These soils differ in texture, drainage, and nutrient-holding capacity, reflecting the pedological diversity characteristic of the Amazon Basin. Such soils are typically acidic, low in base saturation, and strongly weathered, requiring amendments and careful nutrient management for perennial crops like cacao36. ESTEX-OP acts as an exceptional structural micro-model of the region’s best agricultural soils. In the cacao-growing region of Rondônia, similar soil classes predominate, particularly Latosols and Podzolic soils, which present comparable chemical and physical limitations. These soils, although challenging, are suitable for cacao when managed under agroforestry systems that improve nutrient cycling and microclimate regulation36. This alignment reinforces the relevance of ESTEX-OP as a research environment that mirrors the conditions faced by cocoa growers across the state. Overall, ESTEX-OP combines a representative humid tropical climate with a broad spectrum of soil types that reflect the environmental conditions of Rondônia’s cacao belt. This makes the station a strategic site for evaluating cacao adaptability, productivity, and resilience, generating insights that can be reliably transferred to commercial agroforestry systems in the region. For future studies, it is recommended to conduct multi-environment trials for the selected clones. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Acknowledgements The authors express their gratitude to the National Council for Scientific and Technological Development (CNPq), Research Support Foundation of the State of Minas Gerais (FAPEMIG), and the Coordination of Superior Level Staff Improvement (CAPES), Brazil, for providing scholarships. Author contributions CMVCA: Acquisition of data and drafted the work; FLOC: Acquisition of data and drafted the work; FAN: Acquisition of data; GMB: Data analyses and paper review; LASD: Conception, design of the work, and paper review. Data availability The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.