Improving grain yield prediction in Southern US oat germplasm using genomics information and environmental covariates

Abstract

Genetic gains of oat (Avena sativa L.) grain yield have been historically low compared to other major cereal crops. The use of machine learning models to capture complex interactions and leveraging data types other than genomic information in prediction models has great potential for improving complex traits in oat breeding programs. This study assessed the performance of deep learning model for genomic prediction compared to other statistical models, examined the optimal training set size for grain yield prediction, and investigated the potential of incorporating environmental covariates for enhancing oat grain yield prediction. A total of 463 oat lines were evaluated in five environments in Southern United States, and genotyping of the lines gave 12,657 single‐nucleotide polymorphism markers. Our results showed that training set sizes 200–350 could be the optimal size for our panel, indicating the possibility of reducing phenotyping costs by reducing the size of the oat panel tested. The deep learning model was less superior to genomic best linear unbiased prediction and other models for grain yield, test weight, and heading days in the different environments. Incorporating interaction effects (G × E or G × W) into the multikernel prediction models across environments improved predictive abilities for grain yield by up to 0.21 compared to the baseline models. This reveals the potential of incorporating weather data to enhance predictive abilities in genomic prediction models. Our findings provide important information for improving genetic gains in oat breeding programs by integrating genomics and environmental information.

Core Ideas

• Incorporating environmental covariates can improve the predictive abilities of oat grain yield and other traits.

• Optimizing training set size for oat grain yield prediction could reduce phenotyping cost and improve genetic gain in breeding programs.

• Three training set optimization methods (coefficient of determination, D‐optimal, and maximin) performed similarly to the random method for prediction.

• Integrating deep learning models to capture complex interactions has the potential for improving complex trait prediction in oat breeding programs.

Plain Language Summary

Oat is an important small grain cereal crop cultivated globally, and consumption is increasing in popularity due to its health benefits. Development of superior oat genotypes that meet farmers' and consumers’ needs requires the use of different modern crop improvement tools and strategies. An artificial intelligence‐based method was used to predict grain yield and other traits. Furthermore, environmental information collected from different environments was combined with genotyping data to predict performance of the genotypes. The predictive ability of grain yield and other traits based on the artificial intelligence‐based model was lower compared to other less complex models. A combination of genotyping information with environmental data resulted in improved predictive ability of grain yield and other traits. This implies the possibility of obtaining higher grain yield and faster rate of releasing improved oat varieties for farmers.

Abbreviations

CDmean

coefficient of determination mean

GBLUP

genomic best linear unbiased prediction

PEVmean

mean prediction error variance

SNP

single‐nucleotide polymorphism

SOAP

Southern Oat Association Panel

1. INTRODUCTION

Oats (Avena sativa L.) are nutrient‐rich and low‐glycemic‐index whole grains comprising an abundant amount of a unique composition of protein, suitable levels of essential amino acids, and high levels of lysine and threonine (Yu et al., 2025). The health benefits of oat grains include lowering blood cholesterol levels, regulating blood glucose levels, and reducing the risk of colorectal cancer (Omondi et al., 2022). In the United States, the public breeding programs have primarily focused on enhancing grain yield and disease resistance in oat varieties. For grain yield improvement, breeders have mostly relied on pedigree or selected bulk‐pedigree breeding methods, as well as field‐based screening. This conventional breeding approach is time‐intensive, requiring 11–12 years to produce genetically and phenotypically stable lines suitable for commercialization. In addition, the quantitative nature and complex interactions with the environment make genetic improvement of grain yield challenging, thereby limiting genetic gains. Genetic gain for grain yield has been historically low in oats, unlike other major crops (Oliveira et al., 2025). Hence, there is a great need for more efficient breeding strategies based on high‐throughput genotyping or phenotyping to expedite the identification and development of elite lines.

Genomic selection has revolutionized plant breeding by enabling the prediction of complex traits through the integration of genome‐wide molecular marker data with phenotypic information to predict breeding values of untested germplasm. It has been demonstrated to increase genetic gain and efficiency of complex traits in plant breeding programs beyond marker‐assisted and phenotypic selection (Beche et al., 2021; Crossa et al., 2017; Haikka et al., 2020; Mellers et al., 2020; Rio et al., 2021; Sørensen et al., 2023; X. Wang et al., 2023). Thus, implementing genomic selection for grain yield improvement in breeding programs has the potential to shorten the breeding cycle and result in greater genetic gain (Sandhu, Lozada, et al., 2021). A few studies have implemented genomic selection for oat grain yield (Gui, 2021; Haikka et al., 2020; Oliveira et al., 2025; Sandro et al., 2024). However, the application of genomic selection to the genetic improvement of grain yield is still in its infancy in the Southern US oat breeding programs.

Genomic selection models are calibrated using a training set of lines genotyped and phenotyped for desired traits. These models are thereafter used to estimate the genetic values of the genotyped test set of individuals. Training set optimization is particularly important when phenotyping is costly in traits with low heritability and in situations where there is significant genotype × environment interaction (Sarinelli et al., 2019). However, estimating sample size is still unclear in genomic prediction studies. The most widely used methods for optimizing training population size are the coefficient of determination (CDmean) and the mean prediction error variance (PEVmean) (Akdemir et al., 2015; O. A. Montesinos‐López et al., 2023; Rincent et al., 2012; Rio et al., 2021). These two methods aim to select the maximum number of individuals in the calibration set for predicting as accurately as possible the target set of individuals. Contrarily, Fernández‐González et al. (2023) recommended Avg_GRM_self (average genomic relationship self) strategy (minimizes the average relationship within the training set to maximize variability) and Avg_GRM_MinMax strategy (minimizes average relationship within the training set but also maximizes the average relationship between training and test sets) for untargeted and targeted optimizations, respectively, to obtain a more precise estimation of the optimal training set size.

Artificial intelligence technology has been used to speed up the development of superior plant varieties, utilizing high‐throughput genotyping and phenotyping to advance plant breeding research (Khan et al., 2022). This has led to the increasing application of machine learning models in genomic selection. Since there is no perfect prediction model for the implementation of genomic selection, it is imperative to assess several models for a specific data set and then select the most appropriate for each specific situation (O. A. Montesinos‐López et al., 2021). Therefore, various statistical (ridge regression, mixed models, Bayesian regression, generalized regression, etc.) and machine learning (support vector machine, random forest, deep learning, etc.) models are employed for prediction in genomic selection. A major demerit of the linear methods in genomic selection is their inability to capture the high dimensionality of marker data versus the number of individuals and the presence of complex relationships that are difficult to elucidate (Danilevicz et al., 2022).

An important merit of machine learning methods compared to the linear genomic prediction techniques is the capability of integrating interaction effects other than the additive effects (such as epistasis) without giving any prior assumption, such that it includes all the variances, interactions, and environmental effects (González‐Camacho et al., 2018; Ray et al., 2023; Sandhu, Lozada, et al., 2021). Deep learning algorithms mimic how biological neural networks function. The main structure of a neural network consists of the input, hidden, and output layers (Tong & Nikoloski, 2021). The most employed deep learning approaches in genomic selection are the multilayer perceptron and the convolutional neural networks. Though thousands of single‐nucleotide polymorphism (SNP) markers are utilized in these two approaches, the square root matrix of the genomic relationship matrix is used to prevent model training using a huge number of parameters (A. Montesinos‐López et al., 2023). Thus, deep learning can examine a large volume of data, establish optimal breeding decisions, and speed up the development of superior genotypes, efficiently reducing breeding cycle time, improving breeding efficiency, reducing cost, and boosting crop productivity (X. Wang et al., 2023). There is presently no report on the implementation of deep learning models for determining the prediction accuracy of quantitative traits in oats. Sandhu, Lozada, et al. (2021) reported 0%–5% higher prediction accuracies using deep learning models compared to ridge regression best linear unbiased predictor model for the studied traits in wheat. Ray et al. (2023) found that the deep learning model had slightly lower performance than genomic best linear unbiased prediction (GBLUP), Gaussian kernel, and arc‐cosine kernel models for grain yield and other traits in soybean.

Obtaining high prediction accuracy in genomic selection is challenging due to trait complexity, environmental variability, and data limitations (Juliana et al., 2018; O. A. Montesinos‐López et al., 2024). This is because quantitative traits are influenced by several genes, and environmental variations can modify trait expression. Collection of environmental data across time and space to determine the relationship between the crops’ envirome and phenotypic variation of important factors driving genotype by environment interaction (enviromics) is crucial for improving predictive ability in genomic prediction (Costa‐Neto & Fritsche‐Neto, 2021). Based on such environmental data, an environmental relationship matrix may be obtained, which is an indication of environmental variability. Obtaining this data further allows for understanding of genomics by enviromics interactions and its influence on improving predictive ability in breeding programs. The approach of building an environmental relationship matrix from W covariates, just as the genomic relationship matrix was first introduced by Jarquín et al. (2014). This method considers variability from non‐genetic sources to build up information from the envirotyping outputs. Such information could characterize training and testing environments and improve the predictive ability of models by integrating weather data into the genotype by environment interaction component (Jarquin et al., 2021). Sandro et al. (2024) found that modeling genotype by environment interaction was useful in small but not in large mega‐environments in long‐term multi‐environment trials for oat grain yield genomic prediction.

Core Ideas

• Incorporating environmental covariates can improve the predictive abilities of oat grain yield and other traits.

• Optimizing training set size for oat grain yield prediction could reduce phenotyping cost and improve genetic gain in breeding programs.

• Three training set optimization methods (coefficient of determination, D‐optimal, and maximin) performed similarly to the random method for prediction.

• Integrating deep learning models to capture complex interactions has the potential for improving complex trait prediction in oat breeding programs.

The development of high‐yielding oat varieties with resistance to diseases is a crucial goal that must be met to keep up with the increasing demand for livestock feed and optimize animal performance. The SunGrains (Southeastern University GRAINS) cooperative research program provides extensive regional testing of different germplasm and offers great potential for applying the genomic selection technique to develop oat varieties in a relatively short time with high yield for Southern US growers. Despite being used as a major winter animal forage and feed crop, no effort has been made to successfully integrate modern molecular tools to increase the efficiency in oat breeding programs and accelerate the variety development process. In addition, the rising demand for nutritious animal feeds and increasing consumer preference for healthy whole grain foods necessitates the need to boost oat grain yield on farmers’ fields by integrating advanced technologies and approaches. The objectives of this study were to (1) optimize training set sizes for complex trait prediction in the Southern US oat germplasm panel using different strategies, (2) compare the accuracy of conventional and artificial intelligence‐based deep learning prediction methods for genomic prediction, and (3) investigate the potential of incorporating environmental covariates in enhancing the prediction of oat grain yield.

2. MATERIALS AND METHODS

2.1. Plant materials and experimental design

The grain trial evaluated in this study comprised 463 lines from the Southern Oat Association Panel (SOAP) developed by the public small grain breeding programs in the Southern United States. The grain trial was planted in five environments in the Southern United States. The environments include Citra, Florida, in the 2022/23 growing season; Winnsboro, Louisiana; Quincy, Florida; and Florence, South Carolina, in the 2023/24 growing season; and McGregor, Texas, in the 2024–2025 growing season. The trial was planted using an augmented block design with five repeated checks (Juggernaut, Horizon720, RAM99016, TAMO 412, and Savage) in each block, and the remaining 458 lines were unreplicated. There were 20 blocks in total, each consisting of 28 entries. The repeated checks used are broadly adapted and cultivated in the Southern United States. The plot area used for the trial at the different locations ranged from 5.57 to 6.97 m².

2.2. Management practices and trait phenotyping

Application of fertilizers, chemicals, and other field management practices was followed based on the recommendations for Southern United States to achieve optimal growth and yield potential (Mask et al., 2017). Grain yield (kg/ha) was measured per plot using the Wintersteiger Master combine plot harvester from the grain weight of individual plots. This was estimated by dividing the total grain weight per plot by the plot area using 13% moisture content. The test weight (kg/m³) was also determined post‐harvest. Days to heading were recorded in Julian days as the number of days from the first day of the year when approximately 50% of the plants in a plot have emerged panicle nodes from the flag leaf sheath. Heading days' data were only collected from Citra, Quincy, and McGregor locations.

2.3. Envirotyping

Weather covariate data were obtained from the Clemson University weather station at Pee Dee Research and Education Center for the Florence, South Carolina location; Macon Ridge Research Station R&D weather data for Winnsboro location, TX; the mesonet for the McGregor location in Texas; and Florida Automated Weather Network database for Citra and Quincy locations in Florida. The weather covariate dataset consisted of daily summaries of weather variables, which were used to estimate weekly averages. The weather data for Winnsboro, Florence, and McGregor were all from October 15 to May 31 of each year; December 4–May 31 for the Citra location; and January 7–June 30 for the Quincy location. These periods cover the growth period in each location in the year the trial was evaluated at the different locations. The weather variables used include total rainfall (mm), average temperature (°C), relative humidity (%), solar radiation (Wm⁻²), minimum temperature (°C), and maximum temperature (°C) (Tables S1–S5).

2.4. Genotyping

A genotyping‐by‐sequencing SNP genotyping approach was used for genotyping the 463 SOAP lines. The genotyping and SNP calling of the lines were carried out at the USDA‐ARS Small Grains Regional Genotyping Laboratory at Fargo, ND, as previously described by Bazzer et al. (2025). The raw SNP dataset consisted of 51,533 markers. Quality filtering of the dataset was done by eliminating markers with greater than 30% missing data, heterozygosity >10%, and <5% minor allele frequency, resulting in 12,657 SNP markers. Only 444 lines that had marker information were used for subsequent analysis. Missing marker data were imputed using the linkage disequilibrium k‐nearest neighbor imputation method in Trait Analysis by Association, Evolution and Linkage software (Bradbury et al., 2007).

2.5. Phenotypic data analysis

Adjusted means were estimated for the genotypes in each environment, adjusting for block effects using the “augmentedRCBD” package in R. The augmented block design model, which considers the effect of genotypes, was used as follows:

$$Y_{ij}=\propto +B_i+C_j+G_{j\left(i\right)}+e_{ij}$$ (1)

where Y_ij is the trait of interest, $\propto$ is the overall mean, B_i is the effect of the ith block, $G_{j(i)}$ is the random effect of unreplicated lines j nested within ith block and distributed as independent and identically distributed, G_j ∼ N(0, σ ² _g), C_j is the effect of replicated check cultivar, and e_ij is the standard normal error distribution e_ij ∼ N(0, σ² _e) (Sandhu, Mihalyov, et al., 2021). Broad‐sense heritability for each trait in the individual environment was estimated using the formula:

$$H^2=\frac{{\mathrm{\sigma }}_g^2}{{\mathrm{\sigma }}_g^2+\frac{{\mathrm{\sigma }}_{g\times \mathrm{e}}^2}{n}+\frac{{\mathrm{\sigma }}^2}{nr}}$$ (2)

where ${\mathrm{\sigma }}_g^2$, ${\mathrm{\sigma }}_{g\times \mathrm{e}}^2$, and ${\mathrm{\sigma }}^2$ denote the genotypic variance, genotype × environment variance and the error variance respectively. n is the number of environments and r is the number of replications. Correlation analyses among the traits across environments and among environments for each trait were performed using metan package in R (Olivoto & Lúcio, 2020).

2.6. Training population optimization

The oat panel was separated into a candidate set and a test set of individuals by randomly selecting 20% of the 444 lines as the test set, while the remaining lines were used as the candidate set (Rio et al., 2021). The training set of individuals (50, 100, 150, 200, 250, 300, and 350) was then selected from the candidate set using CDmean, maximin, and D‐optimality criteria implemented in the TrainSel v 3.0 R package (Akdemir et al., 2023) using 10,000 iterations. These optimization strategies were also compared with the random method. The CDmean optimization method is derived from the G‐BLUP linear mixed model. It selects a sample of genotypes for a homogenous design in an environment by relating genotypic data to phenotypic observations of the training data. It can be represented as follows:

$$y=1\mathrm{\mu }+Zu+e$$ (3)

where y represents an n vector of phenotypic observations in the training data, $\mathrm{\mu }$ is the scalar parameter for the mean of the phenotypic observations, $Z$ is the $n\times N$ design matrix of N genotypes in the candidate set, and $e$ is the residual error, $e\sim N_n(0,{\mathrm{\sigma }}_e^2)$ independent of $u\sim N_q(0,{\mathrm{\sigma }}_g^{2}G)$. The coefficient of determination matrix of $\widehat{u}$ is used for predicting u (Akdemir et al., 2023). The diagonals of this matrix are the coefficients of determination of the predictions for each genotype. Finding the best training set is done by selecting larger CDmean values.

The D‐optimal approach is also based on a linear model relating genotypic data to phenotypic observations in the training set. It can be applied for selecting a set of genotypes for a homogenous design in a single environment and represented as follows:

$$y=1\mathrm{\mu }+f\left(M\right)\mathrm{\beta }+e$$ (4)

where y, $\mathrm{\mu },$ and $e$ are the same as in the previous equation. M represents the $n\times m$ marker matrix for the n training individuals, f( M ) represents $n\times q$ genomic matrix, $\mathrm{\beta }$ is the q vector of effects of genomic features. The D‐optimal method of selecting $n$ training individuals from N genotypes in the candidate set involves minimizing the determinant of the f( M ) matrix, which is the first principal component matrix for the marker matrix M . The maximin strategy involves selecting optimal training set sizes n from N candidates by maximizing the minimum square genetic distance of the training individuals.

For the three traits (grain yield, test weight, and days to heading), predictive abilities of the different optimization methods and training set sizes were obtained by using the baseline GBLUP model trained on the training sets to predict the test set of individuals. Environments with the highest overall predictive abilities and those with the lowest overall predictive abilities for the three traits based on the GBLUP model were assessed separately using all optimization methods and training set sizes. In order to capture sufficient variability in our panel for trait prediction, 80% of all the lines in our panel were used as training and the remaining 20% as testing in subsequent genomic prediction analyses.

2.7. Genomic prediction models

The model equation for GBLUP is expressed as follows:

$$\mathrm{y}=\mathrm{\mu }1+\mathrm{Zg}+\mathrm{e}$$ (5)

where $\mathrm{y}$ is the vector of phenotypes for n individuals, $\mathrm{\mu }$ is the grand mean, $1$ is a vector of ones, and $\mathrm{e}$ is the random vector of residual effects with $\mathrm{e}\sim N(0,{\mathrm{\sigma }}_e^2)$. $\mathrm{Z}$ is the incidence matrix for genotype effects (linking phenotypes to breeding values), $\mathrm{g}$ is the vector of genomic effects $g_i$ such that with $\mathrm{g}=\{{\mathrm{g}}_{\mathrm{i}}\}\sim N(0,\mathrm{G}{\mathrm{\sigma }}_g^2),$ where G is the genomic relationship matrix (VanRaden, 2008) and ${\mathrm{\sigma }}_g^2$ is the additive genetic variance. The model equation for Bayesian Lasso is as follows:

$$\mathrm{y}=\mathrm{\mu }1+{\mathrm{X}}_{\mathrm{g}}+\mathrm{e}$$ (6)

with $\mathrm{g}|\mathrm{\lambda }-{\prod }_j\frac{\mathrm{\lambda }}{2}\exp (-\mathrm{\lambda }|{\mathrm{g}}_i|)$, where y, $\mathrm{\mu }$ and $\mathrm{e}$ are the same as in Equation (5) above. X is an (n × p) design matrix consisting of the genotypes of p SNPs for each of the n individuals, $\mathrm{g}=\{{\mathrm{g}}_i\}$ is the random vector of SNP effect, and λ is the “sharpness” parameter (Colombani et al., 2013). The Bayesian Lasso algorithm assumes that the markers follow the Laplace or double exponential distribution (O. A. Montesinos‐López et al., 2022). The BGLR R‐package was used for the GBLUP and Bayesian Lasso models (Pérez & de los Campos, 2014; Rio et al., 2021). The number of iterations was 12,000 with 2000 burn‐ins.

The random forest model was implemented using the “randomForest” R package (Liaw & Wiener, 2002). This model uses an ensemble of decision trees that are individually trained utilizing bootstrapping (random sampling with replacement of samples) and a random subset of SNPs. The test sample predictions are averaged to produce a final estimate to reduce the possibility of overfitting. A total of 500 trees were used in this study. The extreme gradient boosting method was also implemented using “XGBoost” package in R (Zhou et al., 2023). The extreme gradient boosting method is a more powerful prediction model than the gradient boosting decision tree. It uses a large number of decision trees, each new decision tree is trained based on the residual of the prediction from the previous trees, and the response variable involves a combination of a substantial number of trees. The prediction error of the model is therefore further minimized in each iteration. The number of trees sampled was 2000, with early stopping rounds of 50, a learning rate of 0.05, and a depth of 6.

The deep learning model typically has an input layer, an outer layer, and hidden layers. Hyperparameters, including the number of hidden layers, dropout rate, number of units, and the type of activation function, influence the ability of a deep learning model to learn and therefore determine the model's performance (Ray et al., 2023). The random search approach was used for hyperparameter tuning. Due to the continuous nature of traits examined in this study (grain yield test weight and days to heading), the rectified linear unit activation function on the input and hidden layers was used. Dropout rate of 0.2, number of hidden layers (1, 3, and 4), epochs (50, 200, 500, and 1000), and batch size (32 and 56) were assessed. There were no significant differences in the results obtained from the hyperparameter tuning. For this analysis, a 0.2 dropout rate, 200 epochs, a batch size of 56, and 3 hidden layers were used. The dropout and L2 regularization were included in each hidden layer to avoid overfitting, whereas only the L2 regularization was applied to the output layer. Deep learning model analysis was performed using the Keras and TensorFlow packages in R software. All the prediction model analyses were carried out in R software (R Core Team, 2024).

2.8. Multi‐omics prediction models

2.8.1. Incorporating environmental effects and weather covariates

As described by Jarquín et al. (2014), the simplest model, GBLUP, was further extended to include environmental effects (E), weather covariates (W), and their interaction effects. Extending the GBLUP model to include random environmental effects to give the G + E model is expressed as follows:

$$y_{ij}=\mathrm{\mu }+E_i+{\mathrm{g}}_j+{\mathrm{e}}_{ij}$$ (7)

where $y_{ij}$ is the phenotypic adjusted means of jth genotype in ith environment, $E_i\sim N(0,{\mathrm{\sigma }}_E^2)$, g and e are the same as described in Equation (5).

Replacing the environmental effects with weather covariates, the model can be expressed as:

$${\mathrm{y}}_{ij}=\mathrm{\mu }+w_i+g_j+{\mathrm{e}}_{ij}$$ (8)

where ${\mathrm{y}}_{ij}$, g and e are the same as in Equation (7), $w_i$ denotes the environmental conditions of the ith environment based on K environmental covariates, $W_{ik}$(k = 1, 2, …, K). ${\mathrm{w}}_i\sim N(0,\mathrm{\Omega }{\mathrm{\sigma }}_w^2)$, $\mathrm{\Omega }$ is the covariance matrix computed as $\frac{{\mathit{WW}}^{\prime }}{K}$ and ${\mathrm{\sigma }}_w^2$ is the variance component. The covariance structure explains the similarity between the environmental conditions of the genotypes in the same way as obtaining genetic similarity among the genotypes using the genomic data. Integrating the genomic data, weather covariates, and environmental effects components from Equations (7) and (8) above is presented below:

$${\mathrm{y}}_{ijk}=\mathrm{\mu }+w_i+g_j+{\mathrm{E}}_k+{\mathrm{e}}_{ijk}$$ (9)

All the components in this model are similar to those mentioned in Equations (7) and (8).

In addition, principal coordinate analysis (PCoA) was performed to assess the relationship among environments based on the weather data relationship matrix using the R software package (R Core Team, 2024).

2.8.2. Extending the main effects to include interaction effects

The above models in Equations (7) and (8) accounted for the genomic and weather covariates main effects, respectively. Combining the two models using their covariance structures as described by Jarquín et al. (2014) is as follows:

$${\mathrm{y}}_{ij}=\mathrm{\mu }+w_i+g_j+g{w}_{ij}+{\mathrm{e}}_{ij}$$ (10)

$\mathrm{where}g{w}_{ij}$ denotes the interaction between jth genotypes and ith environments, $gw$ = $g{w}_{ij}$ which is the Hadamard product of $Z_g\mathrm{G}{Z^{\prime }}_g$ and $\mathrm{\Omega }$. It follows a multivariate normal distribution, and it is expressed as $gw\sim N(0,Z_g\mathrm{G}{Z^{\prime }}_g\circ Z_E\mathrm{\Omega }{Z^{\prime }}_E{\mathrm{\sigma }}_{gw}^2)$, $Z_g$ and $Z_E$ are the incidence matrices connecting phenotypes with genotypes as well as the environments, respectively.

Furthermore, considering the interaction effect between markers and environments is crucial because the weather covariates may not completely capture the differences across environments, and regression on genetic markers may not fully capture the genetic differences among the genotypes. Thus, some proportion of the G × E may not be well captured by the interaction effect, $g{w}_{ij}$. The inclusion of environment and G × E effects in Equation (9) is expressed as follows:

$$y_{ijk}=\mathrm{\mu }+w_i+g_j+{\mathrm{E}}_k+\mathrm{g}{\mathrm{E}}_{jk}+e_{ijk}$$ (11)

where $gE\sim N(0,Z_g\mathrm{G}{Z^{\prime }}_g\circ Z_E{Z^{\prime }}_E{\mathrm{\sigma }}_{gE}^2)$ and all other terms are the same as in the previous models.

2.9. Cross‐validation and assessment of model performance

Two different cross‐validation approaches were used to assess the precision of the prediction models representing different predictive breeding scenarios. All the genomic (except for training set optimization) and multiomics prediction models were first evaluated using the CV2 fivefold cross‐validation strategy with a training population consisting of 80% of the lines for predicting 20% of the test set comprising unobserved lines (O. A. Montesinos‐López et al., 2023). In addition, the CV1 fivefold cross‐validation scheme was used to assess the performance of the multiomics models’ ability to predict the performance of new lines that have not been previously evaluated in field trials. Predictive abilities of each model were assessed based on the Pearson correlation between the predicted values and observed phenotypes of individuals in the validation set.

3. RESULTS

3.1. Phenotypic variation, heritability, and correlations

Broad‐sense heritability (H ²) and descriptive statistics of grain yield, test weight, and days to heading are presented in Table 1. The highest H ² for grain yield (51.56%) was observed for Texas 2025 (TX25) environment, and the lowest (25.10%) for Citra 2023 (Citra23) environment. Phenotypic performance of genotypes varied among environments, revealing the strong influence of environmental factors. Mean grain yield ranged from 1444.03 kg/ha for Citra23 environment to 3959.36 kg/ha for Clemson 2024 (Clem24) environment. The H ² for test weight ranged from 20.03% for Citra23 to 78.52% for Winnsboro 2024 (Winns24) environment. Similarly, Clem24 gave the highest mean test weight (478.39 kg/m³) and Citra23 gave the lowest (130.77 kg/m³).

TABLE 1.

Descriptive statistics of grain yield, test weight, and days to heading in five different environments.

Env	Mean	Range	SE	H 2 (%)
		Grain yield, kg/ha
Winns24	2680.64	663.06–4721.47	28.99	41.20
Clem24	3959.36	624.90–7566.50	57.76	47.33
Quin24	2361.75	266.04–6635.77	56.69	45.21
TX25	2356.94	413.34–3808.21	20.42	51.56
Citra23	1444.03	216.40–4238.96	38.60	25.10
		Test weight, kg/m3
Winns24	383.12	271.37–473.88	1.79	78.52
Clem24	478.39	367.83–554.99	1.51	43.51
Quin24	248.26	50.04–434.04	4.35	33.54
TX25	393.86	287.19–485.70	1.30	75.52
Citra23	130.77	41.09–387.86	3.78	20.03
		Days to heading
Citra23	84.54	70.11–106.11	0.46	84.64
Quin24	96.77	90.93–105.53	0.13	48.45
TX25	98.87	89.34–115.34	0.28	69.02

Abbreviations: Citra23, Citra 2023; Clem24, Clemson 2024; Env, environment; H ², heritability; Quin24, Quincy 2024; SE, standard error; TX25, Texas 2025; Winns24, Winnsboro 2024.

The H ² for days to heading ranged from 48.45% for Quincy 2024 (Quin24) environment to 84.64% for Citra23, and the mean days to heading varied from 84.54 for Citra23 to 98.87 for TX25. Pearson correlation showed low association among the five environments for each of the three traits (Figure S1). The highest correlation obtained for grain yield was between Quin24 and Winns24 (0.19**) as well as between Winns24 and Citra23 (0.15**). For test weight, Clem24 and TX25 (0.26**) as well as Clem24 and Winns24 (0.22**) had the highest correlations, whereas for days to heading, Citra23 and Quin24 (0.19**) as well as Quin24 and TX25 (0.14**) showed the highest correlation. Across environments, grain yield showed a strong positive association with test weight (0.56**) and a low negative association with days to heading (−0.03), whereas test weight showed a low but significant negative correlation with days to heading (−0.12*) (Figure S2).

3.2. Optimization of training set sizes for genomic prediction

The effects of different training population sizes on mean predictive abilities using CDmean, D‐optimal, and maximin, as well as random selection training population selection approaches, are presented in Figure 1. For environments with the overall highest predictive abilities for grain yield and test weight (Winns24) and days to heading (Citra23) based on GBLUP model (Figure S3), there was an increase in predictive abilities as the training set sizes increased. However, some dips/fluctuations were observed for D‐optimal and maximin methods for grain yield and CDmean and D‐optimal for test weight as well as CDmean for days to heading. Nevertheless, there was no significant increase in predictive abilities after training set size 200–300 for the three traits for most optimization strategies (Figure 1A; Table S6). None of the optimization strategies clearly performed better than the random method, except CDmean and maximin, which gave higher predictive abilities for days to heading.

FIGURE 1.

Effect of training set size on mean genomic best linear unbiased prediction (GBLUP) predictive abilities using environments with the highest predictive ability for the three traits (A). and those with the lowest predictive abilities for the three traits (B).

For grain yield, maximin and D‐optimal strategies gave higher predictive abilities than the random method at training population size 200. However, considering the environment, Quin24, with the overall lowest predictive abilities for the three traits (Figure S3), there were a lot of inconsistencies in the predictive abilities obtained as the training population sizes increased (Figure 1B; Table S7). This may be attributed to the low heritabilities obtained for traits in this environment, particularly for test weight and days to heading. Even though the predictive abilities obtained using population sizes 200–350 as training sets were comparable with using higher population sizes in our panel, we used 80% of the lines in our panel as a training set and 20% as a testing set in subsequent analyses to capture sufficient variability in the population.

3.3. Predictive abilities of statistical and machine learning models

Comparison of the performance of deep learning model with the most often used model for genomic prediction, GBLUP, as well as Bayesian Lasso, random forest, and extreme gradient boosting models for each environment and the three traits is presented in Figure 2. The GBLUP and Bayesian Lasso models gave comparable and the highest predictive abilities for the three traits, except days to heading in Citra23, where the random forest gave the highest predictive ability. The predictive abilities obtained using the extreme gradient boosting method were slightly lower than those obtained using GBLUP, Bayesian Lasso, and random forest methods. The deep learning model gave the lowest predictive abilities for all the traits in the different environments. The environment, Winns24, had the highest predictive abilities for grain yield (0.36) and test weight (0.60), whereas Citra23 had the highest predictive ability for heading date (0.59).

FIGURE 2.

Prediction abilities of statistical and machine learning models for grain yield, test weight, and days to heading. BL, Bayesian Lasso; Citra 23, Citra 2023; Clem 24, Clemson 2024; DNN, deep neural network; GBLUP, genomic best linear unbiased prediction; Quin 24, Quincy 2024; RF, random forest; TX 25, Texas 2025; Winns 24, Winnsboro 2024; XGBoost, extreme gradient boosting.

3.4. Incorporating weather covariates and environmental effects in multikernel prediction models

The variance components decomposition for the five models, which accounted for the main effects of markers (G), environmental effects (E), weather covariates (W), and their possible interaction effects (E + G, W + G, W + G + GW, W + G + E, and E + G + W + GE) for grain yield, test weight, and days to heading, are presented in Figure 3 and Table S8. For grain yield, the model M1 (E + G) had the highest residual percentage of 64%, which reduced to 53.4% and 53.3% for models M3 (W + G + GW) and M5 (E + G + W + GE), respectively (Figure 3). For test weight, the model M1 (G + E) gave a residual percentage of 18.2%, which reduced to 14.8% for model M3 (W + G + GW), while model M5 (E + G + W + GE) had a residual percentage of 18.3%. The simplest model, M1 (G + E), had the highest residual percentage (54.5%) for days to heading, which reduced significantly to 26.4% and 32.2% for models M3 (W + G + GW) and M5 (E + G + W + GE), respectively. The integration of weather covariates in model M4 (W + G + E) drastically reduced variability due to the environment for all the traits.

FIGURE 3.

Variance components decomposition (%) for the five models tested for grain yield (A), test weight (B), and days to heading (C). M1 − E + G, M2 − W + G, M3 − W + G + GW, M4 − W + G + E, M5 − E + G + W + GE. E, environmental effect; G, genomic data‐derived matrix; GE, G‐matrix × environment interaction; GW, G‐matrix × W‐matrix interaction; W, weather data‐derived matrix.

The percentage variance decomposition from model 5 (E + G + W + GE) for grain yield showed that 35.8%, 2.0%, and 0.1% of the total variability were explained by the main effects of environment, markers, and weather covariates, respectively, and 8.8% by genotype × environment interaction. For test weight, 75.2%, 1.7%, 0.1%, and 4.7% of the variability was explained by the environment, markers, weather covariate, and genotype × environment interaction, respectively. For days to heading, 0.1%, 4.2%, 44.9%, and 18.6% variability explained were attributed to the environment, markers, weather covariate, and genotype × environment interaction effects, respectively.

The five models were assessed based on CV2 and CV1 cross‐validation approaches (Figure 4). For the CV2 cross‐validation scheme, the predictive abilities of the models ranged from 0.03 for model E + G to 0.23 for models W + G + GW and E + G + W + GE for grain yield. For test weight, the models’ predictive abilities varied from 0.18 for model E + G to 0.32 for model E + G + W + GE, while for days to heading, model W + G showed the lowest predictive ability of 0.17, while model E + G + W + GE had the highest predictive ability of 0.39. In the CV1 cross‐validation approach, predictive abilities of the models were generally slightly lower than those obtained based on the CV2 approach. The predictive abilities obtained for grain yield in the CV1 scheme varied from 0.01 for models W + G and W + G + E to 0.22 for model E + G + W + GE. For test weight, the predictive abilities varied from 0.12 for model E + G to 0.30 for model E + G + W + GE. For days to heading, the predictive abilities ranged from 0.17 for the model W + G + E to 0.37 for W + G + GW. In both cross‐validation schemes, model E + G + W + GE showed the highest prediction ability for the three traits, except for heading date in the CV1 approach.

FIGURE 4.

Bar plots showing predictive abilities of different kernel models using CV2 (A) and CV1 (B) cross‐validation schemes for different traits. E, environmental effect; G, genomic data‐derived matrix; GE, G‐matrix × environment interaction; GW, G‐matrix × W‐matrix interaction; W, weather data‐derived matrix.

Similarity percentage of the top 30% genotypes based on the observed traits mean values versus their corresponding predicted genotypes was assessed using the best multiomics model (E + G + W + GE) (Figure 5). Coincidence percentages of 46%, 44%, and 41% were observed for grain yield, test weight, and days to heading, respectively, based on the CV2 prediction results. Comparably lower coincidence percentages were obtained based on the CV1 prediction results. The coincidence percentages obtained from the CV1 prediction results were 35%, 31%, and 36% for grain yield, test weight, and days to heading, respectively. The top 50 genotypes based on observed trait mean values were compared with their corresponding predicted genotypes ranking using the model E + G + W + GE based on the CV2 and CV1 prediction results.

FIGURE 5.

Coincidence percentage between the top 30% observed genotypes using the BLUEs and their corresponding predicted genotypes using predicted values from the E + G + W + GE model from the CV2 (A) and CV1 (B) as well as for the G + W + GW model based on CV2 (C) and CV1 (D) cross‐validation scenarios.

Comparable rankings of the top 50 observed versus the predicted genotypes were observed for some of the genotypes for grain yield, test weight, and days to heading using predicted and observed values from the two cross‐validation scenarios (Figures 6 and 7). However, less similarity in ranking was observed in the top 50 genotypes' ranking using the CV1 prediction results (Figure 7).

FIGURE 6.

Rankings of the top 50 observed genotypes based on observed trait means versus the corresponding predicted genotypes ranking based on the best CV2 model (E + G + W + GE).

FIGURE 7.

Rankings of the top 50 observed genotypes based on observed trait means versus the corresponding predicted genotypes ranking based on the best CV1 model (E + G + W + GE).

3.5. PCoA of environments

The PCoA plot showing clustering of the five environments based on the weather data matrix is presented in Figure 8. Principal component (PC) 1 and PC2 explained 41.62% and 26.37% phenotypic variation, respectively, indicating that the PCoA plot reasonably explained the variability in the weather data dissimilarity. The environment, Winns24, is closer to Quin24 and Citra23 based on the weather covariables, while TX25 is far distant from all other environments. Clem24 is located close to Quin24 and distinct from all other environments. This indicates the extent of similarity in the weather profiles of the five environments, implying that Winns24, Quin24, and Citra23 have somewhat close weather patterns; TX25 and Clem24 have completely distinct weather profiles compared to all other environments, except for the closeness of the weather profile of Clem24 to Quin24.

FIGURE 8.

Principal coordinate analysis showing the clustering of five environments based on the weather data matrix. Citra23, Citra 2023; Clem24, Clemson 2024; Quin24, Quincy 2024; TX25, Texas 2025; Winns24, Winnsboro 2024.

4. DISCUSSION

Enhancing genetic gains in oat breeding programs, particularly for complex traits, requires the implementation of modern breeding tools, including genomic selection, as well as integration of multikernel prediction approaches by incorporating different data sources. To date, the use of traditional breeding approaches for major traits is still prevalent in oat breeding programs, which are labor‐intensive, time‐consuming, and costly. Implementation of genomic selection in earlier generations in breeding programs significantly improves breeding efficiency by shortening the breeding cycle and ultimately increasing genetic gain (Vitale et al., 2025). This is achievable through the possibility of selecting or discarding untested genotypes in generations where yield trials were unfeasible.

Heritability estimates differ based on the population, environment, trait complexity, etc. The broad‐sense heritability range (25%–52%) observed for grain yield in our study is lower compared to that reported by Gui (2021), who found heritability ranging from 44% to 76% for grain yield in 305 Canadian elite oat breeding lines. Comparable to our findings, Mathias‐Ramwell et al. (2023) reported heritability of 27% for grain yield in 132 oat cultivars and pure lines of diverse origin. The variability in the broad‐sense heritability among environments in our study could be attributed to the differences in the environmental conditions in the different environments. This was also reflected in the low correlations observed among environments.

4.1. Optimal training set size for grain yield prediction

The choice of training set sizes and composition affects prediction accuracies, and it is crucial for the successful implementation of genomic selection, particularly when genotypes to be included in the training set are from large and genetically diverse germplasm groups (Akdemir & Isidro‐Sánchez, 2019; Sarinelli et al., 2019; Sørensen et al., 2023). Optimizing the training set involves choosing a training population as an ideal subgroup of the candidate set (Fernández‐González et al., 2023) to increase the precision of the predictions made on a test set as the size of the training set is minimized, thus reducing phenotyping cost. The findings from this study showed an overall increase in predictive abilities as the training population size increased, considering environments with the highest predictive abilities. The training set sizes of 200–350 could be the optimal size for this oat panel. Though there was no significant increase in predictive abilities for most optimization strategies and traits beyond training set sizes of 200–300. Previous studies on the assessment of the effects of training set sizes on prediction accuracy have mostly shown that larger training population sizes are preferable (Berro et al., 2019; Isidro et al., 2015; Lopez‐Cruz et al., 2022; Sun et al., 2022). Thus, population size should be sufficiently large to accurately estimate the effects of each of the effective chromosome segments segregating in the target population, particularly for complex traits (Rio et al., 2022).

The predictive abilities of the different optimization strategies were not distinctly better than the random method. This may be due to the lack of distinct genetic structure in the panel used in this study, as CDMean and D‐optimal optimization strategies perform better for a candidate population with strong genetic structure (Chen et al., 2024). The lack of strong genetic structure in the panel used in this study is most likely due to the exchange of germplasm over the years by the different small grains breeding programs in the Southern United States, from which lines in this panel were obtained.

4.2. Performance of conventional and artificial intelligence‐based deep learning models

An important objective of this study was to compare the predictive abilities of conventional and artificial intelligence‐based deep learning prediction methods for grain yield improvement. Our results showed differences in the mean predictive abilities of models assessed in individual environments for grain yield, test weight, and days to heading. The conventional GBLUP and Bayesian Lasso models generally gave the highest predictive abilities for the three traits. However, for days to heading, random forest had comparable performance with these two models. The deep learning model had the lowest predictive abilities for all traits and in different environments. Previous studies have reported superior performance of non‐parametric models for traits largely influenced by non‐linear and epistatic effects, whereas parametric genomic prediction strategies usually perform better for traits largely controlled by additive genetic effects, particularly with small datasets (Budhlakoti et al., 2022; Howard et al., 2014; Kunwar et al., 2025). Similarly, increasing the dimensionality and size of the dataset used in this study could help to improve the performance of the deep learning models. In consonance with our findings in this study, Ray et al. (2023) reported slightly lower performance of the deep learning model compared to the conventional GBLUP model. The poor performance of the AI model in their study was attributed to the genotype‐by‐environment interaction effects. A. Montesinos‐López et al. (2025) performed a detailed comparison of deep learning and GBLUP models using 14 plant datasets obtained from different breeding programs. The authors found that the deep learning model efficiently captured complex, non‐linear genetic patterns in wheat grain yield, giving superior predictive ability to GBLUP. They underscored the complementary nature of deep learning and GBLUP models, as neither of the two models generally performed better than the other, considering all traits and scenarios evaluated.

4.3. Incorporating environmental covariates to enhance genomic prediction

Though different genomic prediction models have been extensively used in breeding programs, these models mainly focus on genetic effects while ignoring environmental effects, limiting the possibility of improving the predictive ability of models (J. Wang et al., 2025). For the practical implementation of genomic selection in breeding programs, different approaches have been utilized for improving the performance of prediction models. Incorporation of environmental covariates has been reported to occasionally improve the predictive abilities of genomic prediction models (Fernandes et al., 2024; O. A. Montesinos‐López et al., 2024).

Great variation was observed in predictive abilities across the different models within each cross‐validation scheme in this study. The models that involved genomic data alone or in combination with environmental covariate information gave the lowest predictive abilities. However, integration of G × W or G × E interaction effects into the multiomics prediction models across environments significantly increased the predictive abilities of the models. Two different cross‐validation scenarios (CV2 and CV1) were used in this study. The CV2 scheme gave slightly higher predictive abilities than CV1 for the three traits due to the availability of both genotype and environment information in the CV2 scheme. The predictive abilities of the models for grain yield in both scenarios were lower than what was observed for test weight and days to heading, indicating the genetic complexity of the primary trait.

Compared to the main effects of markers or combinations of environmental covariates, marker, or environments, introducing the interaction effects increased the predictive abilities for grain yield from 0.02 for E + G to 0.22 for E + G + W + GE and from 0.01 for both W + G and W + G + E to 0.21 for W + G + GW based on the CV1 cross‐validation scheme. Similarly, incorporating the interaction effects reduced the residual variance up to 14% for grain yield, 16% for test weight, and 37% for days to heading. Our findings corroborate that of Vitale et al. (2025), who reported better prediction ability when multiple years were modeled together, integrating the G × Y interaction effect into the genomic prediction model. Similarly, Jarquin et al. (2021) reported improved predictive ability across all prediction schemes incorporating G × E interaction effect in a study to assess the importance of naively integrating environmental covariates in genomic prediction models across 30 environments using 1,481 maize hybrids. Nevertheless, the authors observed no improvement in predictive ability by integrating environmental covariates in G × E models. Gui (2021) reported no improvement in predictive abilities in reaction norm models, which included main effect information plus the interaction terms, compared to using the main effect information alone in the CV1 scheme for oat grain yield prediction. However, leveraging interaction terms gave better predictive abilities in the CV2 scheme. Nguyen et al. (2023) found that only the main effects, G + E or G + W, were enough to give a high level of accuracy compared to incorporating environment interaction effects, G × E or G × W, which did not result in any improvement in the prediction accuracy. Semagn et al. (2022) reported inconsistencies in predictive abilities of models when the G × E interaction effect was incorporated, depending on the populations used, and concluded that this should be considered on a population basis. The model W + G + GW in this study gave comparable predictive abilities as E + G + W + G × E for all traits and in all scenarios, indicating the usefulness of incorporating weather data in genomic prediction models. Integrating high‐throughput phenotyping data in the multikernel prediction model could further improve grain yield prediction in oat breeding programs.

Accurate selection of the top‐performing individuals is more relevant than the mean prediction in breeding programs (Fernandes et al., 2024). Coincidence percentages between the top 30% observed genotypes versus their corresponding predicted genotypes using the best model (E + G + W + G × E) and model G + W + G × W were similar for the three traits. This result emphasizes the importance of integrating environmental data into prediction models, particularly for complex traits such as grain yield. In the case of the CV1 scenario, where we are forecasting the performance of new genotypes in tested environments, this is a more challenging situation that breeders face in a breeding program. Our results showed a 36% chance of identifying the top 30% observed genotypes using the best model (E + G + W + G × E) for grain yield. Incorporating historical weather data could further help in improving the impact of integrating weather data on prediction models in future studies. This study showed that combining environmental information with genomic data, particularly with the inclusion of interaction effects, could significantly enhance the predictive ability of models compared to the use of genomic information alone. Our finding shows great promise, particularly with the growing interest in combining enviromics and genomics data for complex trait prediction. Testing higher number of oat genotypes in more locations and years could improve the performance of the deep learning model and the effect of incorporating environmental covariates in the multikernel model for oat grain yield prediction.

5. CONCLUSIONS

Our study showed a corresponding increase in predictive abilities as the training population size increased, and the training set sizes of 200–350 could be the optimal size for the Southern US oat panel used. This indicates great potential for reducing phenotyping cost by reducing the size of the oat panel tested in this study to about half. The deep learning model gave the lowest predictive abilities for all traits and in different environments compared to GBLUP and other models tested. Compared to the main effects of markers or combinations of environmental covariates, marker, or environments, integrating the interaction effects into the multikernel prediction models across environments resulted in a significant increase in the predictive abilities for grain yield based on CV1 and CV2 scenarios. The model W + G + GW gave similar predictive abilities as E + G + W + GE for all traits and in all scenarios, indicating the potential of incorporating weather data in genomic prediction models. These findings provide crucial information for improving genetic gains in oat breeding by integrating genomics and environmental data.

AUTHOR CONTRIBUTIONS

Samuel A. Adewale: Conceptualization; data curation; formal analysis; investigation; methodology; writing—original draft; writing—review and editing. Md Ali Babar: Conceptualization; funding acquisition; investigation; project administration; resources; supervision; writing—review and editing. Diego Jarquin: Formal analysis; software; writing—review and editing. Naeem Khan: Data curation; investigation; writing—review and editing. Stephen Harrison: Data curation; investigation; resources; writing—review and editing. Noah DeWitt: Data curation; investigation; resources; writing—review and editing. Rick Boyles: Data curation; investigation; resources; writing—review and editing. Shuyu Liu: Data curation; resources; writing—review and editing. Ellen Melson: Data curation; investigation; resources; writing—review and editing. Daniel Hathcoat: Data curation; investigation; writing—review and editing. Jason D. Fiedler: Data curation; investigation; resources; software; writing—review and editing. Raja Sekhar Nandety: Data curation; investigation; resources; software; writing—review and editing.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

Supporting information

Supplementary Figure S1: Pearson correlation among environments for grain yield, test weight, and heading date.

Supplementary Figure S2: Pearson correlation among grain yield, test weight, and heading date across environments.

Supplementary Figure S3: Predictive abilities of grain yield, test weight and days to heading based on GBLUP model.

Supplementary Table S1: Clemson 2023/2024 weekly weather data.

Supplementary Table S2: Winnsboro 2023/2024 weekly weather data.

Supplementary Table S3: Quincy 2024 weekly weather data.

Supplementary Table S4: Texas 2024/2025 weekly weather data.

Supplementary Table S5: Citra 2022/2023 weekly weather data.

Supplementary Table S6: Mean and standard error of predictive abilities based on different training optimization methods using environments with the highest predictive ability for the three traits.

Supplementary Table S7: Mean and standard error of predictive abilities based on different training optimization methods using environments with the lowest highest predictive ability for the three traits.

Supplementary Table S8: Variance estimates decomposition for the five models tested for three traits.

Adewale, S. A. , Babar, Md. A. , Jarquin, D. , Khan, N. , Harrison, S. , DeWitt, N. , Boyles, R. , Liu, S. , Melson, E. , Hathcoat, D. , Fiedler, J. D. , & Nandety, R. S. (2026). Improving grain yield prediction in Southern US oat germplasm using genomics information and environmental covariates. The Plant Genome, 19, e70307. 10.1002/tpg2.70307

DATA AVAILABILITY STATEMENT

All data generated or analyzed during this study are available at: https://github.com/SAdewale75/GY