Genomic Prediction of Kernel Water Content in a Hybrid Population for Mechanized Harvesting in Maize in Northern China

Abstract

Genomic prediction enables rapid selection of maize varieties with low kernel water content (KWC), facilitating the development of mechanized maize harvesting and reducing costs. This study evaluated and characterized the KWC and grain yield (GY) of hybrid maize in northern China and used genomic prediction to identify superior hybrid combinations with low kernel water content at maturity (MKWC) and high GY adapted to northern China. A total of 285 hybrids obtained from single crosses of 34 inbred lines from Stiff Stalk and Non-Stiff Stalk heterotic groups were used for genomic prediction of KWC and GY. We tested 20 different statistical prediction models considering additive effects and evaluating the impact of dominance and epistasis on prediction accuracy. Employing 10-fold cross-validation, it showed that the average prediction accuracy ranged drastically from 0.386 to 0.874 across traits and models. Eight linear statistical methods displayed a very similar prediction accuracy for each trait. The average prediction accuracy of machine learning methods was lower than that of linear statistical methods for KWC-related traits, but the random forest model had a high prediction accuracy of 0.510 for GY. When genetic effects were incorporated into the prediction model, the prediction accuracy for each trait was improved. Overall, the model with dominant and epistatic effects (G:AD(AA)) performed best. For the same number of markers, predictions using trait-specific markers resulted in higher prediction accuracy than randomly selected markers. When the number of trait-specific SNPs was set to 100, the prediction accuracy of GY increased by 33.27%, from 0.406 to 0.541. Out of all the 561 potential hybrids, the TOP 30 hybrids selected by genomic prediction would lead to a 1.44% decrease in MKWC compared with Xianyu335, a hybrid with a fast kernel water dry-down, and these hybrids also had higher GY simultaneously. Our results confirm the value of genomic prediction for hybrid breeding low MKWC suitable for maize mechanized harvesting in northern China. In conclusion, this study highlights the potential of genomic prediction to optimize maize hybrid breeding, enhancing efficiency and providing insights into genotype-accuracy relationships. The findings offer new strategies for hybrid design and advancing mechanized harvesting in northern China.

1. Introduction

Maize is in great demand worldwide as food and feed, and with global warming, it has the potential to be an excellent cereal crop in the future as a C4 plant. As human labor costs continue to rise, mechanized maize harvesting is an inevitable trend in the modernization of agriculture. Low kernel water content at harvest (HKWC) is a key factor for efficient mechanical harvesting, as it significantly reduces the need for post-harvest drying. When the HKWC was low, the grain required less drying time and energy, thereby reducing drying costs and processing time. This was particularly crucial in large-scale farming operations, where rapid processing and cost efficiency are essential [1]. The HKWC, in turn, depends on MKWC (kernel water content at maturity) at the physiological maturity stage and kernel water dry-down rate (DR). Therefore, accelerating the selection of maize varieties suitable for mechanization is an important goal for breeders and is also of great significance for the development of agricultural modernization.

HKWC is determined by two stages: the first stage is pre-physiological maturity, which is the rapid filling stage of the maize kernel. During this stage, dry matter increases rapidly and reaches a maximum value. Grain filling continues with a gradual increase in kernel water content (KWC), which peaks and then begins to decline gradually to a state of net water loss. This stage is primarily a developmental dehydration stage, where water dynamics vary from variety to variety and are regulated mainly by genetic factors. The second stage is the dehydration process after physiological maturity, marked by the formation of a black layer at the base of the kernel, which cuts off the material connection from the plant; the dry weight of the KWC remains stable, and the water content gradually decreases. At this time, the KWC is not solely determined by genetics but is also affected by external environmental variables [2,3]. The dehydration rate after the physiological maturity of maize kernel is closely related to HKWC. Fast kernel filling and kernel water dry-down in the field are distinctive characteristics of maize varieties with low KWC while maintaining high yield potential [4,5].

KWC is strongly associated with kernel composition, such as sugars and certain hydrophobic compounds [6,7], as well as a range of agronomic characteristics. These traits encompass kernel row number, ear length and diameter, and husk tightness and length [8,9,10,11,12,13]. KWC, as a quantitative trait, can be affected by multiple factors at varying points in time, thereby complicating the process of trait estimation. Previous studies have revealed that KWC was controlled by numerous quantitative trait loci (QTL), and hundreds of QTL for maize KWC have been identified [14,15,16,17,18,19,20,21,22,23,24]. The use of marker-assisted selection (MAS) for KWC in maize breeding programs was not ideal due to the lack of large effect QTL, and the incorporation of additional markers was necessary. Unlike MAS, genomic prediction (GP) used the collective contribution of all genetic markers [25], offering a more comprehensive and accurate method for selecting hybrids with low KWC, regardless of the small effect sizes of individual QTL. Therefore, GP can overcome the challenges posed by QTL-based approaches, providing a more efficient and precise solution for improving KWC in maize breeding programs.

GP is a highly effective method of utilizing markers across the entire genome for forecasting crop trait outcomes, even in the absence of information regarding genes associated with the traits [26]. To date, a number of statistical methods have been developed for GP, including genomic best linear unbiased prediction (GBLUP) [27,28], Bayesian approach (Bayes A, B, C) [29,30], ridge regression (rrBLUP) [31,32], and least absolute shrinkage and selector operator (LASSO) [33,34]. In addition, nonparametric methods for genomic prediction based on machine learning (ML) techniques such as random forest (RF) [35,36,37], support vector machine (SVM) [38,39], and feed-forward artificial neural networks (ANN) were proposed [40]. Some progress has been made in GP related to maize breeding. By optimizing parameters including single-nucleotide polymorphism (SNP) marker density and training population size, a high-accuracy GP model was developed for evaluating traits such as plant height [41], husk length [42], and GY [43,44]. Modeling using the training population is a key step in GP. Prediction models, such as LASSO, rrBLUP, and ML methods, can be used for GP [45]. In maize breeding, GP is used to predict and select target traits in breeding lines [26,46]. It has also been utilized to predict the performance of hybrids and accurately assess the general and specific combining abilities (GCA and SCA) of the parental lines [47].

Maize mechanical harvesting places high demands on controlling grain water content [48]. Excessive grain water affects threshing efficiency, increases drying and storage costs, and reduces grain quality and market value. Globally, controlling KWC is a critical factor in enhancing maize production efficiency. In regions facing similar challenges, such as the US corn belt, genetic improvements and agronomic practices have been implemented to accelerate kernel dry-down [49]. In Europe, efforts have also focused on breeding maize hybrids with lower KWC to meet the demands of mechanized systems [50]. In northern China, due to rapid temperature drops in autumn, the natural drying process of maize kernels during maturity is relatively slow, making high water content a major challenge for mechanical harvesting [51]. Recent studies have explored the genetic basis of kernel dry-down traits and their application in hybrid breeding [24,52]. However, most research has relied on traditional phenotypic selection, which is time-consuming and less precise. Northern China’s unique climate further amplifies the challenge, necessitating innovative breeding strategies to address the slow natural drying process. Ensuring that maize is harvested at an optimal KWC is essential for achieving efficient, low-cost production, boosting farmer profits, and advancing agricultural mechanization.

GP has been performed on maize, and all have yielded promising results; however, few studies have used GP for KWC-related traits in maize. This study evaluated 285 maize hybrids, with phenotypic data collected across four environments. These data were used to (1) understand the distribution of KWC in maize hybrids in northern China, (2) estimate the prediction accuracy of different prediction models for KWC, (3) appraise the influence of genetic effects and specific markers on prediction accuracy, and (4) explore the potential of GP for KWC in maize breeding.

2. Materials and Methods

2.1. Plant Material and Field Trials

In this study, 34 diverse inbred lines were chosen, of which 18 inbred lines belonged to the Stiff Stalk (SS) heterotic group, and the remaining 16 lines belonged to the Non-Stiff Stalk (NSS, Sipingtou/Iodent) heterotic group. Two hundred and eighty-five single-cross hybrids were produced from a partial diallel cross design between the two groups of lines [53].

The field experiments were performed in four geographically distinct locations in 2022: Harbin, Heilongjiang (HLJ, 130° E, 46° N), Gongzhuling, Jilin (JL, 125° E, 44° N), Tieling, Liaoning (LN, 118° E, 38° N), and Urumqi, Xinjiang (XJ, 86° E, 45° N), respectively. In every location, hybrids were planted in double-row plots utilizing an incomplete block design comprising two replicates. The length of each row was 4.0 m with a row spacing of 0.6 m. The number of plants sown in each double-row plot for each hybrid was fourteen. Field management followed the best local practices.

2.2. Phenotypic Data Collection and Analysis

The two hundred and eighty-five maize hybrids were sown on four dates based on the growing season of maize hybrids to achieve similar physiological maturity in each environment. All hybrids were under open pollination. The time of tassel anthesis and ear silking was recorded for each hybrid in each environment, and kernel water content at grain physiological maturity (MKWC) collection was performed based on the time of maturity (55–65 days after anthesis). The MKWC for 5 uniformly growing plants in each row was measured using a hand-held moisture meter (TimberMaster). MKWC was calibrated according to the temperature and humidity at the time of measurement. In general, the effect of temperature on the MKWC calibration can be approximately compensated as follows: for every 5 °C above 20 °C, subtract 0.5 from the value displayed. For every 5 °C below 20 °C, add 0.5mc from the value displayed. The effect of humidity on the MKWC calibration can be approximately compensated as follows: for every 10% above 65%, subtract 0.5 from the value displayed. For every 10% below 65%, add 0.5 from the value displayed. Kernel water content at harvest (HKWC), kernel water dry-down rate (DR), and grain yield (GY) were also estimated. GY was adjusted to a 14% standard water content. HKWC was measured using a grain moisture meter after harvest at each location. DR was calculated using the following formula: DR = MKWC–HKWC.

Phenotypic data were evaluated with R version 4.3.1 on Windows (https://www.r-project.org/, accessed on 3 March 2024). To reduce environmental impacts, maize phenotype BLUP (best linear unbiased prediction) was computed across four locations through the lme4 package’s lmer function [54]. The ggplot2 and patchwork package were used to analyze the correlation coefficients and make a box plot. Generalized heritability was determined from variance components:

$$H^2=\left({\delta }_{\mathrm{g}}^2\right)/\left({\delta }_{\mathrm{g}}^2+{\delta }_{\mathrm{g}}{\mathrm{e}}^2/y+{\delta }_e^2/yr\right)$$

where $y$ and $r$ represent the number of growing locations and the number of field replicates per location, respectively; ${\delta }_{\mathrm{g}}^2$ represents genetic variance; ${\delta }_{\mathrm{g}}{\mathrm{e}}^2$ represents genotype-by-location interaction variance; and ${\delta }_{\mathrm{e}}^2$ is residual variance [55].

2.3. Genotyping

The 34 diverse parental inbred lines, including 18 from the SS and 16 from the NSS heterotic group, were genotyped using the GenoBaits Maize 45K Panel [56]. Briefly, the final 23,692 SNPs were used for the subsequent analysis after quality control [53]. Finally, both genotypic and phenotypic data were available for 285 crosses, including crosses between the 18 SS lines and the 16 NSS lines. Based on the mating design, genomic profiles of the 285 single-cross hybrids were derived using genotypes of the parental lines.

2.4. Genomic Prediction with Different Models

We compared twenty different statistical prediction models considering the genetic effects of the four traits. Based on the characterization of the Reproducing Kernel Hilbert Space (RKHS) with a user-arbitrarily defined covariance structure [30], we performed genomic prediction using the RKHS, considering additive, dominance, and epistatic effects. The difference with previous studies was that we considered the effect of gamete origin and used a genomic relationship matrix completely developed from theory [57] (rather than transplants based on the concept of pedigree).

2.4.1. G Model

Studies have indicated that this model was defined as grounded in hybrid genotypes and overlooked the origin of gametes (Stiff Stalk and Non-Stiff Stalk parental origins) [58]. The G-model can be written as follows:

$$y=1\mu +g_{A^{\left(H\right)}}+g_{D^{\left(H\right)}}+g_{A{A}^{\left(H\right)}}+E$$

where $g_{{\mathrm{A}}^{\left(\mathrm{H}\right)}}$ are the additive genetic effects of hybrids distributed as $g_{{\mathrm{A}}^{\left(\mathrm{H}\right)}}\sim \mathrm{MVN}\left(0,{\mathrm{G}}_{{\mathrm{A}}^{\left(\mathrm{H}\right)}}{\sigma }_{{\mathrm{A}}^{\left(\mathrm{H}\right)}}^2\right)$, the dominance genetic effects are $g_{{\mathrm{D}}^{\left(\mathrm{H}\right)}}\sim \mathrm{MVN}\left(O,D_{\mathrm{H}}{\sigma }_{{\mathrm{D}}^{\left(\mathrm{H}\right)}}^2\right)$, and the additive-by-additive epistatic. For hybrids, the details of computing these various genomic relationship matrices were described in the study by González-Diéguez and colleagues [57].

We also tested the $g_{A^{\left(H\right)}}$ under different statistical models, such as linear models and ML models. Linear models used include GBLUP, RR, LASSO, Bayes A, B, C, partial least squares (PLS), generalized linear models (GLM); ML models include RF, SVM, gradient boosted machine (GBM), and ANN. GBLUP was implemented using the R package “sommer” [59]. RR and LASSO models were implemented using the R package “glmnet”, and the lambda parameters were optimized by 10-fold cross-validation [33]. RKHS employs a Gaussian process with an arbitrary user-defined covariance structure to realize various genomic relationships in regression. Bayes (A, B, C) and RKHS were implemented using the R package “BGLR” [31]. PLS, GLM, RF, SVM, GBM, and ANN were implemented using the Sparse Kernel Methods (SKM) R library [60].

2.4.2. GCA Model

The GCA model defines the genetic effect based on gamete origin [57]. Thus, the GCA model for the hybrid (S; N) resulting from crossing the parental lines S (from the Stiff Stalk population) and N (from the Non-Stiff Stalk population) can be formulated as follows:

$$y=1\mu +Z_d{g}_{A_S}+Z_f{g}_{A_N}+Z{g}_D+Z_d{g}_{A{A}_S}+Z_f{g}_{A{A}_N}+Z{g}_{A{A}_{SN}}+e$$

where $y$ is the vector of the original phenotype of the $n$ phenotyped hybrids, and $\mu$ is the intercept. $g_{A_S}$ and $g_{A_N}$ are the vectors of the random additive effect (A) from the stiff stalk and Non-stiff stalk parental lines with $g_{A_S}\sim \mathcal{N}\left(0,K_{A_S}{\sigma }_{A_S}^2\right)$ and $g_{A_N}\sim \mathcal{N}\left(0,K_{A_N}{\sigma }_{A_N}^2\right)$, respectively. $g_D$ is the vector of random dominance effect (D) with $g_D\sim \mathcal{N}\left(0,K_D{\sigma }_D^2\right)$. $g_{A{A}_S}$ is the vector of random A-by-A epistatic effect (E) within the stiff stalk (resp. Non-stiff stalk) A-by-A E across the stiff stalk and Non-stiff stalk populations. $g_{A{A}_{SN}}\sim \mathcal{N}\left(0,K_{A{A}_{SN}}{\sigma }_{A{A}_{SN}}^2\right)\cdot K_{A_S},K_{A_N}$, epistasis within the stiff stalk population, A-by-A E within the Non-Stiff Stalk population, and the A-by-A E across populations genomic kinship matrices calculated according to González-Diéguez et al. [57]. $e$ is the vector of error terms, with $e\sim \mathcal{N}\left(0,I{\sigma }_e^2\right)$. The various random effects are considered to be independent. Refer to

Table 1. Definition of tested genomic prediction models.

Models	Model Code	Random Genetic Effects	Reference
G	G: A (GBLUP)	gA	VanRaden (2008) [28]
	G: A (rrBLUP)		Endelman (2011) [32]
	G: A (LASSO)		Friedman et al. (2010) [33]
	G: A (Bayes A)		Karkkainen et al. (2012) [29]
	G: A (Bayes B)
	G: A (Bayes C)
	G: A (PLS)		Montesinos-López et al. (2023) [60]
	G: A (GLM)
	G: A (RF)
	G: A (SVM)
	G: A (GBM)
	G: A (ANN)
	G: AD	gA + gD	Vitezica et al. (2017) [58]
	G: A(AAH)	gA + gAA
	G: AD(AAH)	gA + gD + gAA
GCA	GCA: A	gAS + gAN + r	González-Diéguez et al. (2021) [57]
	GCA: AD	gAS + gAN + gD
	GCA: A(AASN)	gAS + gAN + gAASN
	GCA: AD(AASN)	gAS + gAN + gD + gAASN
	GCA: AD(AAS)(AAN)(AAH)	gAS + gAN + gD + gAAS + gAAN + gAASN

G and GCA models are models that successively added additive effects (A), dominance effects (AD), and additive-by-additive genetic effects (AD(AA)). The list of the random genetic effects considered in the GCA models corresponds to Stiff Stalk (gA_S), Non-Stiff Stalk (gA_N), intragroup additive-by-additive epistasis for the Stiff Stalk (gAA_S) and Non-Stiff Stalk group (gAA_N), and intergroup additive-by-additive epistasis (gAA_SN) effects. In the G models, random genetic effects within hybrids(_H) correspond to additive (gA), dominance (gD), and additive-by-additive epistasis (gAA) effects.

for a summary of the models that were tested.

2.5. The Design of SNP Marker Dataset and Prediction of Specific Markers as Fixed Effects

The study chose specific markers through genome-wide association analysis (GWAS) within the hybrid population and then utilized them for genomic prediction through two methods. Two strategies, GWAS-assistant selection and random selection, were used to build the SNP marker dataset. In the GWAS-assistant selection strategy, the MLM (mixed linear model) algorithm was used for GWAS. Furthermore, we investigated the effect of the number of ts_SNPs on the prediction accuracy. Six random-SNP marker datasets with 50, 100, 500, 1000, 5000, and 10,000 SNPs were established for GP. Then, GP analysis was performed with two SNP marker datasets. The markers specific to the trait were regarded as fixed effects in the analysis [61]. The trait-specific SNP sets (ts_SNPs) were selected based on the score of SNPs (−log¹⁰, P). For the ts_SNPs dataset, the top 50, 100, 500, 1000, 5000, and 10,000 SNPs with the highest −log¹⁰(P) value were selected as six ts_SNPs marker datasets, respectively. For the random-SNP dataset, the same number of SNPs were retrieved using the biomaRt package and the script retrieve_SNP_info.R randomly. In addition, we tested the effect of changes in the top SNP dataset on prediction accuracy when the top SNPs were modeled as a fixed effect (fixed).

We used 10-fold cross-validation to evaluate the accuracy of GP in this study. In each cross-validation, 90% of the hybrids were randomly selected as the training set and the remaining 10% as a test set. The cross-validation was repeated 100 times, and the prediction accuracy was assessed using Pearson’s correlation coefficient between the observed and the predicted phenotypic values.

2.6. Superior Hybrids Selected by Genomic Prediction

We used all 285 hybrids as a training set and predicted the hybrid performance of all possible 561 single-cross combinations between the 34 parent inbred lines based on the GBLUP (A + D) model. We selected the 30 hybrids with the lowest kernel water content at maturity (5%, 30/561, MKWC_TOP 30) and compared the predicted phenotypic values of these hybrids with all other single-cross combinations (Others). Based on the predicted breeding values, we calculated the selection difference, that is, the difference between the mean performance of the selected hybrids and the performance of the control “Xianyu 335” (a commercial hybrid with fast dehydration in China). Based on the 30 hybrids with the lowest MKWC of predicted value, predicted values of HKWC, DR, and GY based on these hybrids were used to compare gains with all remaining hybrids.

3. Results

3.1. Extensive Phenotypic Variation About KWC-Related Traits Observed in Field Trials

The mean values of MKWC in the four locations, HLJ, JL, LN, and XJ, were 37.3%, 37.7%, 39.3%, and 38.7%, respectively (Figure 1A). The MKWC for all hybrids in the four environments ranged from 31.2% to 43.0%, reflecting considerable phenotypic variation across environments. This variation was critical because it underscored the influence of environmental factors, such as temperature and humidity, on MKWC. The MKWC at these four environments showed a good correlation with an average Pearson correlation of 0.328 (p < 0.001) (Figure 1B). MKWC was influenced by the genotypic effect (G), environmental effect (E), and genotype-environment interaction (G × E). Significant genotypic effect (G) was observed (p < 0.001), and the broad-sense heritability (H²) reached 0.75 (

Table 2. Variance analysis of maize kernel water content-related traits and yield.

Source	Traits	Df	Sum Sq	Mean Sq	F Value	p-Value	H2
G	MKWC	284	0.674	0.002	6.221	<2 × 10−16 ***	0.75
	HKWC	284	0.573	0.002	8.189	<2 × 10−16 ***	0.85
	DR	284	0.561	0.002	3.439	<2 × 10−16 ***	0.76
	GY	284	1110.060	3.909	3.476	<2 × 10−16 ***	0.53
E	MKWC	3	0.145	0.048	127.068	<2 × 10−16 ***
	HKWC	2	0.358	0.179	725.118	<2 × 10−16 ***
	DR	2	0.068	0.034	58.908	<2 × 10−16 ***
	GY	2	309.340	154.669	137.557	<2 × 10−16 ***
G × E	MKWC	847	0.512	0.001	1.587	2.463 × 10−13 ***
	HKWC	563	0.178	0.000	1.280	6.233 × 10−4 ***
	DR	562	0.462	0.001	1.430	1.519 × 10−6 ***
	GY	563	800.420	1.422	1.264	1.091 × 10−3 ***

*** Significant at p = 0.001; G, genotype; E, environment; G × E, genotype-by-environment interaction.

), indicating that genetic factors play a significant role in determining MKWC traits.

Given the high correlation of MKWC in multiple environments, the BLUP method was used to integrate MKWC data from the four environments to obtain stable individual genetic MKWC_BLUP values. The distribution of MKWC_BLUP in the maize hybrid population was in line with the normal distribution (Figure 1C). Analysis of variance (ANOVA) showed that different maize hybrids differed significantly (p < 0.01) in MKWC_BLUP, with a mean coefficient of variation (CV) over 4.91% (Figure 1C).

We also collected three traits, HKWC, DR, and GY, but the phenotypic data of these three traits were not collected in XJ due to weather reasons. The distribution of these three traits in the population conformed to a normal distribution (Figure S1). Heritabilities were 0.85, 0.76, and 0.53, respectively. For all three traits, the genotypic effect and genotype-environment interaction effect were highly significant (p < 0.001). However, the correlations between the traits were low and even negative for some of the traits (Table S1).

3.2. Performance of Prediction Models Based on Hybrid Genotypes

The accuracies of 12 different genomic prediction models relying on additive effects were compared by using 10-fold cross-validation (

Table 3. Prediction accuracy of maize kernel water content-related traits and yield.

Genetic Model	Prediction Method	MKWC	HKWC	DR	GY
G: A	GBLUP	0.788 ± 0.065	0.846 ± 0.073	0.693 ± 0.115	0.463 ± 0.123
	rrBLUP	0.797 ± 0.088	0.861 ± 0.052	0.700 ± 0.098	0.445 ± 0.121
	LASSO	0.791 ± 0.075	0.848 ± 0.062	0.702 ± 0.123	0.433 ± 0.145
	Bayes A	0.816 ± 0.065	0.850 ± 0.045	0.714 ± 0.109	0.424 ± 0.143
	Bayes B	0.793 ± 0.067	0.853 ± 0.071	0.703 ± 0.107	0.455 ± 0.158
	Bayes C	0.799 ± 0.072	0.850 ± 0.065	0.682 ± 0.103	0.491 ± 0.134
	PLS	0.767 ± 0.041	0.839 ± 0.031	0.673 ± 0.094	0.423 ± 0.140
	GLM	0.787 ± 0.075	0.844 ± 0.058	0.716 ± 0.081	0.435 ± 0.095
	RF	0.739 ± 0.064	0.801 ± 0.067	0.708 ± 0.071	0.510 ± 0.098
	SVM	0.781 ± 0.069	0.863 ± 0.065	0.709 ± 0.061	0.473 ± 0.099
	GBM	0.758 ± 0.088	0.811 ± 0.067	0.676 ± 0.093	0.431 ± 0.096
	ANN	0.684 ± 0.121	0.811 ± 0.048	0.685 ± 0.105	0.386 ± 0.098
G: AD	RHKS	0.793 ± 0.074	0.843 ± 0.061	0.730 ± 0.103	0.432 ± 0.138
G: A(AA)	RHKS	0.811 ± 0.084	0.874 ± 0.051	0.751 ± 0.075	0.486 ± 0.151
G: AD(AA)	RHKS	0.810 ± 0.055	0.873 ± 0.049	0.747 ± 0.111	0.520 ± 0.116
GCA: A		0.801 ± 0.062	0.851 ± 0.052	0.721 ± 0.086	0.481 ± 0.114
GCA: AD		0.754 ± 0.074	0.864 ± 0.047	0.664 ± 0.120	0.409 ± 0.113
GCA: A(AASN)		0.794 ± 0.072	0.867 ± 0.048	0.728 ± 0.086	0.421 ± 0.165
GCA: AD(AASN)		0.773 ± 0.072	0.851 ± 0.068	0.695 ± 0.111	0.415 ± 0.127
GCA: AD(AAS)(AAN)(AAH)		0.778 ± 0.084	0.849 ± 0.059	0.701 ± 0.098	0.465 ± 0.113

). The average prediction accuracy ranged between 0.386 and 0.874 across traits and models. For each trait, we observed that the prediction accuracies of the eight linear methods were similar, while the prediction accuracies of these ML models RF and SVM were similar to that of the linear methods, with the RF model performing the best in predicting GY at 0.510. ML models GBM and ANN had lower prediction accuracies. Prediction accuracies of the three traits MKWC, HKWC, and DR were not significantly different among the twelve models. The prediction accuracy of GY was significantly different among the 12 models. For example, the linear model Bayes C had a higher prediction accuracy of 0.491 for GY. In contrast, the ML model ANN had the lowest prediction accuracy of 0.386 for GY, and the prediction accuracy of ANN for MKWC was also the lowest, 0.684.

The model was extended based on hybrid genotypes to include dominance and epistatic effects based on the RHKS model. Based on the results in Table 3, it can be found that the prediction accuracy of MKWC and DR was not improved in the prediction model, considering dominance and epistatic effects. In contrast, the prediction accuracy of the G: A(AA) and G:AD(AA) models for HKWC were improved to 0.874 and 0.873, respectively. The prediction accuracies of the two traits, MKWC and DR, in the models with the addition of dominance and epistatic effects were similar to those of the linear and ML models. The average prediction accuracy for GY improved to 0.479 after considering both dominance and epistatic effects compared to the model with only additive effects. The prediction accuracy of GY in the model with both dominance and epistatic effects (G:AD(AA)) reached 0.520, which was higher than that of the linear and ML models.

3.3. Prediction Accuracy of Prediction Models Based on Parental Genotypes

Several nested models were tested by successfully incorporating dominance and epistatic effects into the additive effects based on the genotypes of the two parents. According to the results in Table 3, it was found that the prediction models based on parental genotypes did not improve the prediction accuracy when incorporating dominance and epistatic effects. In the GCA model, the average prediction accuracy of MKWC was 0.781. The average prediction accuracies of the remaining three traits, HKWC, DR, and GY, were 0.858, 0.702, and 0.436, respectively. The prediction accuracies of the GCA model were all reduced compared to the G model.

3.4. The Potential of Trait-Specific SNPs in Genomic Prediction

Figure 2A,B shows the p-values of the markers obtained by the MLM GWAS method and the p-value comparison of SNPs (MKWC) in the ts_SNPs dataset and the random_SNPs dataset. For other traits, see Figures S2 and S3. In our study, for all four traits, prediction accuracies improved when using the top 100 SNPs with the largest effect on trait markers, with accuracies of 0.811, 0.867, 0.737, and 0.541, respectively (Figure 2C–F). When predicting with randomly selected SNPs (random_SNPs), prediction accuracies of MKWC, HKWC, DR, and GY were lower than that of ts_SNPs. When the ts_SNPs marker set was modeled as a fixed effect, prediction accuracy decreased slightly as the number of SNPs increased but was mostly better than random_SNPs.

In addition, where GWAS was performed for four traits in each environment, and using environmentally specific SNPs (ts_SNPs = 100) as genotypes improved prediction accuracy for all of them compared to the random_SNPs model (Figure S4).

3.5. Identification of 19 Single-Cross Combinations with Low MKWC and Higher GY

Based on the predicted performance of 561 single-cross combinations, we selected 30 promising combinations with low MKWC (Table S2, Figure 3). These data were used to compare the selection gain of MKWC_ TOP30 with the remaining combinations. The mean value of MKWC_TOP30 was 36.28%, while the mean value of MKWC in others was 38.34%, which was significantly lower than MKWC with all the remaining combinations. Interestingly, based on the yield performance of MKWC_TOP30, the yields of these combinations were also higher than the yields of all the remaining combinations (Figure 3D). MKWC was reduced by 1.44% in MKWC_TOP30 compared with Xianyu335 (Figure 3). Among these TOP 30 hybrids, the performance of 11 hybrids was evaluated in the field (Table S3, Figure S5), and the other 19 novel hybrids were predicted and selected that have not yet been tested. Therefore, these hybrids are promising candidates for the next evaluation stage in field trials.

4. Discussion

4.1. Accurate Identification of KWC in Northern China

The development and application of maize varieties with low HKWC is a prerequisite to ensure that mechanical harvesting can be applied. The use of whole genomic prediction technology holds practical importance for the rapid selection of low KWC maize varieties. Because KWC is controlled by multiple small-effect QTLs and is highly influenced by environmental conditions, one of the biggest challenges lies in obtaining accurate KWC values, particularly at the kernel’s physiological maturity stage. A number of methods have been developed to measure the KWC. For example, moisture determination indicators by detecting changes in capacitance [62], the SK-300 moisture-determination meter [63], the digital timber-moisture meter [64], the wooden moisture meter Voltcraft FM-200 Humidity [65], the grain moisture meter with microwave attenuation at 10.5 GHz [66], and the hand-held moisture meter [67]. Although the classical drying method is regarded as the most reliable for determining MKWC values, it is labor-intensive and time-consuming, making it unsuitable for large-scale phenotyping in genomic prediction studies. To balance precision and practicality, we chose the Protimeter Moisture Meter (BLD5605) for the phenotypic acquisition of MKWC. This method provided reliable readings after calibration with corresponding temperature and humidity data, ensuring accuracy without the excessive resource demands of classical methods.

Considering that MKWC was mainly influenced by varietal factors (genetics), we recorded MKWC during the maturity period of kernel development of each hybrid to minimize variability caused by differing maturation stages. While environmental factors significantly affect MKWC, we standardized the harvesting period to ensure a consistent dehydration process and reduce environmental errors. Our current results indicated that there were significant correlations among the four growing locations for KWC-related traits and also a high heritability of 0.75–0.85. These findings suggest that our methodological choices, including the Protimeter Moisture Meter and standardizing the harvest period, effectively addressed the trade-offs between precision, efficiency, and practicality. This allowed us to obtain accurate KWC values, which were instrumental in achieving the desired genomic prediction accuracy.

4.2. Influence of Different Statistical Models on the Prediction Accuracy of Traits

Prediction accuracy varied widely among agronomic traits in this study, with HWKC having the highest prediction accuracy of 0.876 and GY having the lowest prediction accuracy of 0.386. It was evident from the heritability that the prediction accuracy reaching large variations was significantly affected by different genetic structures, as shown in previous studies [68,69]. We used different model predictions for comparative analysis, and the results showed that for traits with higher heritability, statistical models had less effect on prediction accuracy, such as MKWC, HWKC, and DR, which reached similar prediction accuracy with all models (Table 3). For the lower heritability trait GY, the prediction accuracy of different statistical models varied considerably. For example, the lowest prediction accuracy of ANN was 0.386 for GY, and the highest prediction accuracy of the G:AD(AA) model was 0.520, which was an improvement of 0.134. Therefore, our results were consistent with previous studies that the prediction accuracy was mainly determined by the genetic structure [53]. For traits with low heritability, larger sample sizes or multi-environment experimental data can be used to improve the generalization ability of the model. However, by using several different models for prediction, it is possible to select the optimal model for different traits. For example, the highest prediction accuracy of Bayes A was 0.816 for MKWC.

According to earlier studies [70,71,72,73], trait-specific markers have demonstrated advantages for genomic prediction by capturing key genetic signals. In this study, we tested various subsets of specific SNPs using the G:A (RKHS) model, incorporating specific SNPs as fixed effects and the remaining SNPs as random effects. Consistent with previous findings [73,74], certain combinations significantly increased prediction accuracy. These results reinforced the importance of leveraging trait-specific markers to enhance accuracy, particularly in breeding programs targeting complex traits. By aligning with earlier studies while applying these insights to the unique genetic and environmental context of maize breeding, our findings provided further evidence for the efficacy of this approach and offered a practical framework for improving molecular breeding efficiency.

4.3. Prediction Accuracy and Practical Application of Genomic Prediction for Hybrid Performance

In this study, we evaluated the prediction accuracy of hybrid performance using two types of genetic models: one assuming a single genetic effect for the hybrid (G model) and another defining genetic effect based on allele origin (GCA model). Additive, dominance, and epistatic effects were considered. The results showed that the G model outperformed all GCA models across the four traits, with the G:AD(AA) model achieving the highest prediction accuracy for GY when non-additive effects were included. While previous studies have reported similar predictive power between G and GCA models in hybrid populations [75,76], our findings suggest a clear advantage of the G model for trait prediction. The lower performance of the GCA model in predicting hybrids may stem from its over-reliance on the general combining ability of the parents while neglecting the unique non-additive effects in hybrids. In this study, the G model more comprehensively captured the genetic potential of hybrid combinations by integrating additive, dominance, and epistatic effects, thereby demonstrating higher prediction accuracy. Nonetheless, the GCA model remains valuable for estimating parental line values and supporting the selection of next-cycle parental lines.

Results from many previous studies have demonstrated the potential of genomic prediction in crop breeding, particularly in maize hybrid breeding [53]. For example, GP has been successfully applied to traits such as GY and KWC in maize in America and Europe, highlighting its versatility in different environmental contexts and breeding programs [52,77,78]. Additionally, in wheat and rice, GP has been employed to improve traits such as grain quality, disease resistance, and abiotic stress tolerance in regions ranging from South Asia to South America [79,80,81,82]. These studies underline the global applicability and adaptability of GP models across species and regions. In this study, utilizing GP, we identified MKWC_TOP30 hybrids with the lowest MKWC, with an average MKWC of 36.28%. These hybrids also exhibited high GY, demonstrating the efficiency of this approach in selecting hybrids suited for mechanized maize harvesting. The GP model, developed using hybrid populations specifically from northern China, has significant practical implications. It not only reduces costs and time in the breeding process but also directly supports the adoption of mechanized harvesting by enabling the rapid selection of hybrids with optimal moisture content. Compared with previous studies conducted in regions like America, where kernel dry-down rates have been improved through both genomic and phenotypic selection, our findings emphasize the necessity of tailoring prediction models to specific regional climates. In northern China, where rapid temperature drops hinder natural kernel drying, the ability to predict and select hybrids with low MKWC was particularly impactful. These results advanced the regional breeding strategies for northern China and contributed to global efforts in optimizing crop performance.

5. Conclusions

Our study aimed to identify superior low MKWC hybrids and evaluate the potential of genomic prediction using maize hybrids from northern China. The results showed that prediction models had minimal impact on trait prediction accuracy, which largely depended on the genetic basis of the trait. Combining genetic effects and trait-specific markers significantly improved yield prediction accuracy. These findings highlighted the potential of GP to enhance breeding efficiency by enabling precise selection of hybrids with low MKWC and high yield, reducing the time and costs of traditional phenotypic selection. This approach was particularly valuable for supporting mechanized harvesting in northern China, boosting agricultural efficiency and farmer profitability. Moreover, the study contributed to understanding the genetic basis of key traits and offered a foundation for applying GP to other crops and regions with similar challenges.