Comparison of GBLUP and WGBLUP in genomic selection for beef cattle under different LD patterns and mixed multi-population scenarios: a simulation study Authors: Le Zhou, Fengying Ma, Lin Zhu, Chencheng Chang, Mingjuan Gu, Risu Na, Wenguang Zhang Published: 2025 ABSTRACT AbstractBackgroundGenomic selection (GS) is a key tool to enhance genetic progress in livestock and poultry breeding, but the effects of differences in linkage disequilibrium (LD) patterns between populations and the way reference populations are constructed on the accuracy of genomic prediction (GP) still need to be thoroughly investigated. Therefore, in this study, three beef cattle populations with different genetic backgrounds and LD patterns were simulated to compare the performance of GBLUP, WGBLUP, and their mixed reference population derivation methods (GBLUPmix and WGBLUPmix) in predicting genomic breeding values, with a focus on analyzing the effects of different LD patterns and reference population sizes on prediction performance.ResultsThe results showed that increasing the reference population size significantly improved the accuracy of GP, especially when the reference population was increased from 4000 to 5000, and the WGBLUPmix model showed the highest prediction accuracy in multi-population integration scenarios. In addition, the performance of the WGBLUPmix model was significantly different in different LD modes. The superiority may be due to the realization of capturing different LD patterns in the genome, adaptation to different genetic contexts of population structure, and better capture of large effect SNP contributions through weight adjustment or assignment.ConclusionThese results highlight the importance of using the WGBLUPmix model to flexibly adjust weights to accommodate different reference population sizes and LD patterns in multi-population genome prediction, demonstrate the superiority of WGBLUP in low-LD populations and its synergistic effect with mixed reference populations, and provide an effective strategy to improve prediction accuracy.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12864-025-12224-2.Keywords: Genomic selection, GBLUP, WGBLUP, Linkage disequilibrium, Mixed reference population, Beef cattle Background With the continuous development of single nucleotide polymorphism (SNP) microarray and resequencing technologies, GS has rapidly emerged as a superior research methodology in the field of plant and animal breeding, and GP has been widely used in breeding programs for major commercial livestock breeds, significantly improving annual genetic gain [1, 2]. Traditional breeding methods may be limited by lower accuracy and slower genetic progress in some traits, and these limitations are overcome by the GS approach, which allows candidate selection of superior individuals by relying on prediction equations derived from reference populations with available phenotypes and genotypes [3]. Based on experimental evidence and theoretical foundations, GS has been shown to be more accurate and faster than traditional breeding methods, resulting in better breeding results [4–6]. However, genomic evaluations have been primarily implemented in large breeds. However, small breeds face unique challenges. These breeds often have a limited number of progeny test bulls available. Additionally, they struggle to establish large reference populations consisting of animals with precise phenotypes. Achieving accurate GP typically requires large populations of individuals with both phenotypes and genotypes. This requirement is not always feasible for all traits or all varietiess. To overcome these limitations, there are several strategies that can be employed. One remedy is multi-breed prediction. This approach involves adding data from other breeds to improve the accuracy of SNP effect estimation. Another strategy is a cross-breed prediction. This method utilizes SNP effects from another breed that has a large reference population [7]. It has been shown that combining multiple closely related populations into a common reference population improves the accuracy of GP in dairy cattle [8]. Integrating multiple populations into a shared reference population has the potential to enhance the accuracy of GP, but many challenges remain in merging and evaluating genomic breeding values. In small breeding populations, it is often impractical to construct a sufficiently large reference population for GP, and in cross-breed GP, the value of the information provided by other breeds is often limited, especially when these breeds are genetically distant [9]. The size of the reference population is an important factor affecting the reliability of GP, and for breeds with small population sizes, there are many challenges in carrying out GP, among which the lack of a sufficiently large reference population has become a key obstacle to the accuracy of their GP [10]. In multi-population or cross-population scenarios, the reference population of a single population may not be able to adequately capture the genetic variation of the target population, leading to a reduction in prediction accuracy. In addition, the presence of large effect SNPs may also significantly affect prediction accuracy, especially when multiple trait prediction is involved and when populations are closely related. However, it has also been shown that for multi-breed prediction, a multi-breed reference population dominated by one breed reduces the prediction accuracy of smaller breeds. In Australian Red cattle, prediction accuracy increased significantly when a small number of Holstein bulls were included in the reference population, and while increasing the size of the reference population usually improved prediction accuracy, further increasing the number of Holstein bulls reduced accuracy, sometimes below that achieved within the breed [11, 12]. Thus, for breeds with small population sizes, it may be more effective to select a reference population that is more closely related to the target breed than to use a large reference population, so care needs to be taken in designing multi-breed reference populations rather than using all available data. In addition, accuracy is dependent on the proportion of genetic variance explained by genome-wide single nucleotide polymorphisms (SNPs) and the accuracy of the estimation of the effects of these SNPs, both of which are closely related to LD between SNPs and quantitative trait loci (QTLs) [13]. When multiple breeds are combined to form a reference population, prediction accuracy depends on SNP-QTL LD across breeds; however, LD may differ between breeds due to selection and genetic drift, or the phases of SNP and QTL alleles may be reversed between breeds [14]. Integrating QTL or SNPs closely linked with QTL into GP marker panels can be beneficial. This practice is applicable both in multi-breed reference populations and for cross-breed prediction. It may help to mitigate the issue of varying SNP-QTL LD between breeds. This assumption may not hold true if there are interactions between QTL and genetic background. Therefore, it may be more appropriate to assume that QTL (and therefore SNP effects) are different but correlated, rather than identical, between breeds. Thus, differences in LD between certain beef cattle breeds or populations and the limited size of their reference populations hinder the accuracy of genomic estimated breeding values (GEBV). Genomic Best Linear Unbiased Prediction (GBLUP), as a genome-based prediction method, is widely used in practical genetic assessment work. It assesses the genetic similarity between individuals by calculating the genetic relationship matrix (G) between individuals. This genetic relationship is constructed based on the deviation of allele doses of marker genotypes from the population mean [15]. However, in structured populations such as multi-species or mixed populations, there may be significant differences in marker allele frequencies between subpopulations. In this case, if the average allele frequency of the whole population is still used to construct the genetic relationship matrix, it may reduce the accuracy of the estimation of genomic breeding values, especially in the presence of large effect SNPs or more complex patterns of LD. Although GBLUP typically assumes that all genetic variants contribute relatively little to traits, by using a weighted genomic relationship matrix [16, 17], which could allow certain preselected sequence variants to have a greater effect on traits. In recent years, the weighted GBLUP (WGBLUP) method has been proposed, which reflects the degree of contribution of each SNP to the trait more accurately by assigning weights to different SNPs, and thus improves the accuracy of GP [18, 19]. In this study, the prediction models are mainly divided into two categories: one is the GBLUP model based on inter-individual genomic relationships, which can reflect the genetic similarity between individuals and is calculated by the deviation of allele doses of marker genotypes from the population mean; the other is the WGBLUP model that introduces a weighted genomic relationship matrix, which can more accurately reflect the contribution of each SNP to the trait. Therefore, this study focused on assessing the effects of different LD patterns, reference population size on prediction performance by simulating the performance differences in genomic breeding value prediction of three beef cattle populations with different genetic backgrounds and LD patterns, GBLUP, WGBLUP, and their hybrid reference population derived methods (GBLUPmix and WGBLUPmix). The study, emphasize the importance of integrating reference population size, LD differences and genetic background complexity in multi-population GP. In addition, we pay special attention to the performance of the WGBLUP model in the case of multi-population merging when the reference population size is increased from 4000 to 5000, as well as its adaptability to different LD modes. Methods Data simulation Firstly, the pedigree and genomic data were simulated using QMSim V1.10 [20]. To investigate the impact of different LD patterns on the accuracy of genomic estimated breeding values (GEBVs) in beef cattle populations, the initial historical population (HP) was simulated into three distinct beef cattle populations (PopA, PopB, and PopC) based on three different LD patterns. Secondly, the final breeding animals from these populations were used as candidates for alternative selection schemes. QMSim was used to simulate five replicates, i.e., five sets of final breeding animals. Within each replicate, the breeding schemes studied selected the first batch of breeding animals from the same population, although different schemes could use different groups of breeding animals. To save data storage space and without compromising computational speed, only the genotypic and phenotypic data of animals in the 9th and 10th generations were simulated. The phenotypes were sampled from a normal distribution with a mean of 0 and a variance of 1, with the overall mean being the only fixed effect. Table 1 summarizes the population structure and parameters of the simulation process. Table 1Parameters of the simulation processPopulation structurePopAPopBPopCStep 1: Historical Generations (HG) Number of generations phase 1 (size)0(200)0(500)0(1,000) Number of generations phase 2 (size)1,000(200)1,000(200)1,000(1,000) Number of generations phase 3 (size)1,095(1,000)1,095(1,000)1,095(200)Step 2: Recent Generations (RG) Number of founder males from HG505550 Number of founder females from HG2,0002,1002,200Selection and matingebv/hSire/Dam replacement0.6/0.30.5/0.30.5/0.2Mating systemRandomCulling designebv/lGenome Number of chromosomes29 (no X Chr) Genome length2715.85cM Number of markers58990 (50k) Marker/QTL positionsRandom Number of marker/QTL alleles2/2 3 4 Marker of allele frequenciesEqual QTL allele effectsEqual Mutation rate2.5e-05 Parameters of the simulation process Population structure In QMSim, the HP was generated based on equal sex ratios, non-overlapping generations, random mating, no selection, and no migration. These conditions were used to simulate initial values of LD, mutations, and genetic drift, thereby mimicking the structure and extent of LD present in beef cattle populations [21]. The population structures of the three distinct beef cattle populations (including historical, base, and current population structures) were simulated in detail as described in Fig. 1. Visual presentation of the simulated data. PopA, PopB, and PopC experienced different phases of bottlenecks and expansions to generate differences in LD. Founder animals were selected for the three populations and randomly mated for 10 generations, resulting in different population sizes and the number of genotyped animals selected PopA simulated a LD pattern that began with a stable population and then experienced continuous expansion. First, the HP of PopA was simulated, consisting of an initial 200 individuals that remained at a constant size of 200 individuals over 1,000 generations. In the subsequent 95 generations, the population size gradually expanded to 1,000 individuals (100 males and 900 females) to establish initial LD and mutation-drift equilibrium. Next, 250 males and 3,000 females were randomly selected from the last generation of the HP to expand the population size. All males and females were randomly mated, with each female producing one offspring per generation for 10 generations. Finally, 50 males and 2,000 females were randomly selected from the expanded base population. These selected individuals were simulated over the most recent 100 generations, with each female producing one offspring per generation. The selection design was based on phenotypic performance and BLUP (EBV) methods. The replacement rates were 60% for bulls and 30% for cows. The missing record rate for bulls and cows was 0.05. PopA simulated a LD pattern that began with a stable population and then experienced continuous expansion. First, the HP of PopA was simulated, consisting of an initial 200 individuals that remained at a constant size of 200 individuals over 1,000 generations. In the subsequent 95 generations, the population size gradually expanded to 1,000 individuals (100 males and 900 females) to establish initial LD and mutation-drift equilibrium. Next, 250 males and 3,000 females were randomly selected from the last generation of the HP to expand the population size. All males and females were randomly mated, with each female producing one offspring per generation for 10 generations. Finally, 50 males and 2,000 females were randomly selected from the expanded base population. These selected individuals were simulated over the most recent 100 generations, with each female producing one offspring per generation. The selection design was based on phenotypic performance and BLUP (EBV) methods. The replacement rates were 60% for bulls and 30% for cows. The missing record rate for bulls and cows was 0.05. PopB simulated a LD pattern that began with an initial stable population and then experienced contraction followed by expansion. First, the HP of PopB was simulated, consisting of an initial 500 individuals that decreased to a stable population size of 200 individuals over 1,000 generations. In the subsequent 95 generations, the population size gradually expanded to 1,000 individuals (100 males and 900 females) to establish initial LD and mutation-drift equilibrium. Next, 250 males and 2,550 females were randomly selected from the last generation of the HP to expand the population size. All males and females were randomly mated, with each female producing one offspring per generation for 10 generations. Finally, 55 males and 2,100 females were randomly selected from the expanded base population. These selected individuals were simulated over the most recent 100 generations, with each female producing one offspring per generation. The selection design was based on phenotypic performance and BLUP (EBV) methods. The replacement rates were 50% for bulls and 30% for cows. The missing record rate for bulls and cows was 0.05. PopB simulated a LD pattern that began with an initial stable population and then experienced contraction followed by expansion. First, the HP of PopB was simulated, consisting of an initial 500 individuals that decreased to a stable population size of 200 individuals over 1,000 generations. In the subsequent 95 generations, the population size gradually expanded to 1,000 individuals (100 males and 900 females) to establish initial LD and mutation-drift equilibrium. Next, 250 males and 2,550 females were randomly selected from the last generation of the HP to expand the population size. All males and females were randomly mated, with each female producing one offspring per generation for 10 generations. Finally, 55 males and 2,100 females were randomly selected from the expanded base population. These selected individuals were simulated over the most recent 100 generations, with each female producing one offspring per generation. The selection design was based on phenotypic performance and BLUP (EBV) methods. The replacement rates were 50% for bulls and 30% for cows. The missing record rate for bulls and cows was 0.05. PopC simulated a LD pattern that began with an initial stable population and then experienced expansion followed by contraction. First, the HP of PopC was simulated, consisting of an initial 1,000 individuals that remained at a stable population size of 1,000 individuals over 1,000 generations. In the subsequent 95 generations, the population size gradually decreased to 200 individuals (100 males and 100 females) to establish initial LD and mutation-drift equilibrium. Next, 250 males and 3,000 females were randomly selected from the last generation of the HP to expand the population size. All males and females were randomly mated, with each female producing one offspring per generation for 10 generations. Finally, 50 males and 2,200 females were randomly selected from the expanded base population. These selected individuals were simulated over the most recent 100 generations, with each female producing one offspring per generation. The selection design was based on phenotypic performance and BLUP (EBV) methods. The replacement rates were 50% for bulls and 20% for cows. The missing record rate for bulls and cows was 0.05. PopC simulated a LD pattern that began with an initial stable population and then experienced expansion followed by contraction. First, the HP of PopC was simulated, consisting of an initial 1,000 individuals that remained at a stable population size of 1,000 individuals over 1,000 generations. In the subsequent 95 generations, the population size gradually decreased to 200 individuals (100 males and 100 females) to establish initial LD and mutation-drift equilibrium. Next, 250 males and 3,000 females were randomly selected from the last generation of the HP to expand the population size. All males and females were randomly mated, with each female producing one offspring per generation for 10 generations. Finally, 50 males and 2,200 females were randomly selected from the expanded base population. These selected individuals were simulated over the most recent 100 generations, with each female producing one offspring per generation. The selection design was based on phenotypic performance and BLUP (EBV) methods. The replacement rates were 50% for bulls and 20% for cows. The missing record rate for bulls and cows was 0.05. The animals genotyped include all individuals from the 95th to the 100th generation. Individuals from the 95th to the 98th generations were randomly selected to form the reference population. Due to variations in the number of animals in the 99-100th generation across breeds, the number of genotyped individuals varied slightly for each breed. Therefore, 1,000, 2,000, 3,000, 4,000, and 5,000 individuals were randomly selected from the 99-100th generation to serve as the validation population. Genome The bovine genome assembly version ARS-UCD1.2 (< https://jul2023.archive.ensembl.org/Bos_taurus/Info/Index>) was used to simulate the 29 pairs of autosomes in beef cattle, with a total length of 2715.85 centimorgans. The aim was to create a more realistic scenario by considering the true distances between markers and quantitative trait loci (QTL) sites. SNP markers were uniformly distributed and randomly generated, with a density of 50k markers, representing diallelic loci with segregation. The number of SNPs on each chromosome was proportional to its size. The effects of markers on traits were assumed to be neutral. The whole genome consisted of 725 quantitative trait loci (QTL). The number of recombinations per Morgan was sampled from a Poisson distribution with a mean of 1, and crossovers were randomly placed along the chromosomes. In the final generation of the HP, a TLR effect for a production trait was introduced. This trait had a heritability of 0.42, similar to the heritability of birth weight in beef cattle [22]. In QMSim, the genetic effects for this trait were sampled from a gamma distribution with a shape parameter of 0.4. Over the 1095 generations of the HP, the mutation rate for both QTL and SNPs was 2.5 × 10⁻⁵, with the SNP mutation pattern set to “recurrent,” meaning mutations only switched between alleles without generating new mutation types. The detailed simulation parameters are summarized in Table 1. Genetic evaluations The simulated heritability was 0.42, and the phenotypic variance for the single trait was set to 1.0. The true breeding value (TBV) for each animal was calculated as the sum of the additive effects of the QTL, as follows: Where βj​ is the additive effect of QTLj, Qkj​​ is the QTL genotype at locus j, coded as 0, 1, or 2, as the number of copies of a specified QTL allele is carried by an individual (k). The phenotypes (yi​) were simulated by adding a residual term sampled as εi​∼N(0, ), where ​ is the residual variance. The EBVs were estimated for all individuals in the current population from generations 9 to 10 based on phenotypic values and pedigree data. The best linear unbiased prediction (BLUP) of breeding values are obtained by Henderson’s [23] mixed linear model. The BLUP predictor has the smallest prediction error variance among all possible linear unbiased predictors. The numerator relationship matrix (A) is used in the following mixed model equations to derive BLUP of random additive effects (including polygenes and QTL): where y is the vector of phenotypic records, Z is the incidence matrix relating the records to the random additive effects (a), is the residual variance, andis the additive genetic variance.The mixed model equations are solved by the conjugate gradient method. Genotype quality control The PLINK v1.9 software [24], a widely used toolset for whole-genome association and population-based linkage analyses, was employed to further filter SNPs based on the following criteria: minor allele frequency (MAF) < 0.05, genotype call rate < 0.10, individual call rate < 0.10, and Hardy-Weinberg equilibrium < 1e-5. After genotype quality control, PopA, PopB, and PopC retained 52,676, 53,340, and 52,081 segregating SNPs with MAF >0.05, respectively, for subsequent analyses. The simulation assumed that QTL allele effects were identical across all breeds, but the frequencies of QTL varied among breeds, resulting in differences in the variance explained by each breed. Model and analysis In this study, statistical methods used to estimate breeding values include the Genomic Best Linear Unbiased Prediction (GBLUP) based on the genomic relationship matrix and the Weighted Best Linear Unbiased Prediction (WGBLUP) based on genomic information. When conducting Genomic Best Linear Unbiased Prediction (GBLUP) using genomic information, we applied a general linear mixed model. We opted for the HIBLUP software to construct the model, which is specifically optimized for handling large-scale genomic data. It has efficient computational capabilities, enabling rapid data analysis and providing accurate predictions. Although GBLUP and PBLUP use the same fixed effects, GBLUP employs a genomic relationship matrix (GRM) based on SNP markers to estimate genomic estimated breeding values (GEBVs). The expression for GRM is as follows: Here, M is the matrix of individual genes (where homozygotes, heterozygotes, and alternative homozygotes are converted to 0, 1, and 2, respectively), m is the total number of SNP markers, and is the frequency at the i-th position in the SNP. Then, since only additive genetic effects are modeled, only these effects are considered here: The general linear mixed model equation for GBLUP is as follows: Where y is the vector of phenotypes, X is the matrix associating fixed effects with each animal individual, b is the vector of fixed effects, Z is the design matrix allocating records to genetic values, g is the vector of additive genetic effects for individuals, G is the genomic relationship matrix, e is the vector of residual error effects with a normal distribution ~ N(0, ), ​ is the residual variance. ​​. In the weighted GBLUP, the model and inference are the same, except for the use of a different SNP effect vector when constructing the G matrix. In the SLEMM software developed by Jiang et al. (2023), two schemes for optimizing GP by SNP weighting are provided, namely: (1) based on the minor allele frequency (MAF) dependence of SNP effect sizes; (2) based on the SNP effect estimates with weight W equal to the identity matrix. SLEMM fits the following linear mixed model: where y is a vector of phenotypes for a quantitative trait, β is a vector of fixed effects including the mean, X is the design matrix for β, α is a vector of SNP effects with a diagonal covariance matrix , Z is a matrix of standardized genotypes, and e is a vector of residuals with a diagonal covariance matrix . R is usually equal to an identity matrix, and the diagonal elements of W are weights, with a mean representing the relative contribution of the SNP effect to the genetic variance; that is, Wj represents the contribution of SNPj to the genetic variance. Due to LD, the effect of a quantitative trait locus (QTL) can be captured by nearby SNP loci, and the effects of adjacent SNPs tend to be similar in model fitting. Therefore, the second SNP weighting scheme is used in this study, with the formula as follows: where C is a scaling constant to control the mean weight to be 1, S is the number of SNPs on each side of SNPj, and is the estimate of the effect of the kth SNP in an existing BLUP with W equal to the identity matrix. This specification of the jth SNP’s weight borrows information from a window of 2S+1 SNPs. SLEMM first fits the model with training data where W is equal to the identity matrix, and then fits with W computed by Eq. (8). Admixed-reference population implementation (GBLUPmix/WGBLUPmix) In this study, the GBLUPmix and WGBLUPmix methods extend the standard GBLUP and WGBLUP approaches by constructing strategically mixed reference populations. For the purposes of this research, we created mixed reference populations by combining individuals from PopA, PopB, and PopC. Specifically, we randomly selected 1,000, 2,000, 3,000, 4,000, and 5,000 individuals from each population. This sampling scheme generated balanced mixed reference populations with total sizes of 3,000, 6,000, 9,000, 12,000, and 15,000 individuals, respectively, while maintaining consistent population proportions across different sizes. Subsequently, these mixed populations were utilized for the evaluation of genomic prediction models, which were constructed in the same manner as the standard forms: the model applications of GBLUP and WGBLUP methods have been described above. Accuracy of genomic prediction In this study, model performance was evaluated using prediction accuracy and unbiasedness. Since the true breeding values (TBV) were directly provided in the simulation, prediction accuracy is defined as the Pearson correlation coefficient between the genomic estimated breeding values (GEBV) and TBV, calculated as: where Corr ranges from 0 to 1, indicating the strength of the linear relationship between GEBV and TBV. Additionally, prediction unbiasedness is assessed by the regression coefficient of GEBV on TBV, calculated as: When b ≈ 1 and the intercept ≈ 0, the genomic predictions are considered unbiased. Two-way analysis of variance (2-way ANOVA) To statistically assess the differences in predictive accuracy across various populations and conditions, we utilized a two-way analysis of variance (2-way ANOVA) with Prism software. This analytical method enables us to examine the main effects of two categorical independent variables, as well as their potential interactive effects, on a continuous dependent variable—in our case, genomic prediction accuracy.The two-way ANOVA model incorporated two factors: (1) the impact of different populations (e.g., PopA, PopB, PopC) on genomic prediction accuracy; (2) the influence of condition factors (e.g., reference population size, prediction model type) on genomic prediction accuracy. The two-way ANOVA provided estimates of the main effects for each factor, as well as the interaction effect between the two factors. The significance of these effects was determined by the calculated F-statistics and the corresponding p-values. P-value less than 0.05 was considered to indicate a statistically significant difference. Results Population genetic characterization To explore genetic population structure, we performed PCA using genotyping datasets from three different populations, popA, popB, and popC (Fig. 2). The first two principal components (PC1 and PC2) were chosen in this study because together they explained most of the variation in the data. As shown in Fig. 1, the distributions of popA, popB, and popC on PC1 and PC2 were significantly different, suggesting that these three groups are genetically distinct at the genomic level. the differences between popA and popB may reflect their differences in some key genetic traits, whereas the independent distributions of popC may be indicative of a unique genetic background. Principal component analysis (PCA) results for three different groups (popA, popB, and popC). The horizontal axis represents principal component 1 (PC1) and the vertical axis represents principal component 2 (PC2). Dots of different colors and shapes represent different groups, where red dots represent popA, blue triangles represent popB, and green diamonds represent popC Figure 3 illustrates the pattern of LD decay in three different populations (PopA, PopB, PopC). LD is a non-random association between allele frequencies at two or more loci in a genome. In this study the strength of LD was quantified using the r2 statistic, where r2 values close to 1 indicate strong LD, while close to 0 indicates weak LD. As shown in Fig. 3, the LD values of all the three populations declined with the increase in intergenic distances, which is in line with the trend of LD decay reported in the existing literature. Specifically, PopA showed the highest r2 values at all distances examined, indicating the slowest rate of LD decay. In contrast, PopB and PopC had lower r2 values and PopC had the fastest rate of LD decay. The difference in r2 values between the three populations was more significant at smaller genetic distances (0 to 500 kb). As the distance increases, this difference gradually decreases, and by 1000 kb, the r2 values of the three populations tend to be close to each other. LD between markers in the last generation (generation 10) of the most recent populations of three different populations (PopA, PopB, PopC). x-axis shows the range of distances between markers in the genome, expressed in kilobase pairs (kb); y-axis shows the average LD, expressed as r2. The curves show that r2 values decreased in all populations as the distance between genes increased, but the rate of decrease varied between populations Additionally, we thoroughly assessed the impact of LD structure on the accuracy of heritability estimation by simulating data under different LD patterns. As shown in Table 2, under high LD patterns, the estimated heritability values closely approached the theoretical values, with a standard error of 0.0594. This indicates that under high LD patterns, our estimation method can very accurately capture the true values of heritability. However, under low LD patterns, there was a larger deviation between the estimated values and the theoretical values, with a standard error of 0.0805. This suggests that the accuracy of heritability estimation is somewhat affected under low LD patterns. Such differences may be related to the influence of LD patterns on gene association and variance estimation. Table 2Comparison of estimated heritability values with theoretical values under different LD patternsLD PatternTheoretical HeritabilityEstimated HeritabilityStandard ErrorPopA(high)0.420.40570.0594PopB(medium)0.420.39720.0603PopC(low)0.420.37280.0627 Comparison of estimated heritability values with theoretical values under different LD patterns The impact of different LD patterns on the accuracy of GEBV prediction Firstly this study compared the accuracy of genetic evaluation using GBLUP and WGBLUP methods in three beef cattle populations under different LD patterns and different reference population sizes, and the results are shown in Table 3. The results showed that PopC consistently achieved the highest GP accuracies when the reference set comprised 1,000, 2,000, 3,000 or 5,000 individuals, outperforming both PopA and PopB. Further inspection of the LD-decay plot (Fig. 3) revealed that PopA displayed the slowest LD decay, retaining elevated r² over longer genomic distances. This extended LD enables a larger proportion of SNPs to remain in effective linkage with causal QTL, thereby allowing GBLUP/WGBLUP to capture genetic effects more precisely and yielding high accuracy. In contrast, PopB and PopC exhibited rapid LD decay, causing marker–QTL associations to break down quickly and reducing the models’ ability to stably link genotypic information to phenotypes, which translated into lower accuracies. PopB showed an intermediate LD-decay profile, and accordingly its prediction accuracies ranked between those of PopA and PopC across all reference-population sizes. Table 3The accuracy of GP for three different populations (PopA, PopB, PopC) under various LD patterns using GBLUP and WGBLUP methodsReference population sizePopGBLUPWGBLUP1000A0.2170.36B0.2110.4C0.1480.2922000A0.2220.357B0.2370.339C0.150.2963000A0.2470.356B0.2380.354C0.1640.3084000A0.2510.367B0.2460.388C0.1850.3165000A0.2780.384B0.2580.418C0.1970.317 The accuracy of GP for three different populations (PopA, PopB, PopC) under various LD patterns using GBLUP and WGBLUP methods Notably, the prediction accuracy of the WGBLUP method was significantly higher than that of the GBLUP method regardless of the reference population size and LD pattern. In particular, the prediction accuracy of WGBLUP in PopC was better than GBLUP at reference population sizes of 1,000, 2,000, 3,000, and 5,000, and exhibited stronger LD decay between the 50–200-kb region, suggesting that it was able to capture associations between markers and QTL more effectively. Similarly, WGBLUP had significantly higher prediction accuracy in PopA than GBLUP at a reference population size of 4,000, further confirming the superiority of WGBLUP between the 50–200-kb region.The LD attenuation pattern of PopB was intermediate between that of PopA and PopC, and its prediction accuracy was lower than that of PopA at all reference population sizes, although the WGBLUP’s prediction accuracy in PopB was still significantly higher than that of GBLUP, suggesting that WGBLUP has higher robustness in all different LD modes. In order to further optimize the accuracy of genomic genetic assessment, hybrid versions of GBLUP and WGBLUP (GBLUPmix and WGBLUPmix) were also introduced, that is, samples with different reference population sizes were extracted from three populations (PopA, PopB and PopC) in equal proportions, and a hybrid reference population was constructed for evaluation, and the results are shown in Table 4. The results show that the prediction accuracy of GBLUPmix and WGBLUPmix is not significantly higher than that of their corresponding single-population versions (GBLUP and WGBLUP). Among them, GBLUPmix showed better prediction accuracy than GBLUP in different reference group sizes and LD mode, especially in PopB, and the prediction accuracy of GBLUPmix showed an improvement trend. For example, when the reference population size is 1,000 and 2,000, the prediction accuracy of GBLUPmix is significantly higher than that of GBLUP in PopB, indicating that the mixed reference population can better capture genetic diversity between populations, thereby improving prediction performance. Table 4The accuracy of GP for three different populations (PopA, PopB, PopC) under various LD patterns using GBLUPmix and WGBLUPmix methodsReference population sizePopGBLUPmixWGBLUPmix1000A0.1790.26B0.2250.23C0.1210.1852000A0.1840.264B0.2290.24C0.1230.1893000A0.1840.254B0.240.247C0.1330.1914000A0.2010.266B0.250.257C0.1460.1885000A0.2150.284B0.2590.291C0.1580.194 The accuracy of GP for three different populations (PopA, PopB, PopC) under various LD patterns using GBLUPmix and WGBLUPmix methods The impact of reference population size on the accuracy of GEBV prediction Secondly this study also compared the accuracy of genomic genetic evaluation (GEBV) using GBLUP and WGBLUP methods for three beef cattle populations under different reference population sizes, as shown in Fig. 4. The results showed that the prediction accuracy of both methods showed an increasing trend as the reference population size increased, but the prediction accuracy of WGBLUP was significantly higher than that of GBLUP at all reference population sizes. among them, the prediction accuracy of GBLUP showed a relatively gentle increasing trend with the increase of reference population size, especially when the reference population size increased from 1000 to 3000, its The prediction accuracy increased from 0.148 to 0.217 to 0.164–0.247, which is a small increase. However, when the reference group size increases from 4000 to 5000, GBLUP’s prediction accuracy improves more dramatically, suggesting that GBLUP’s marginal gains in large reference groups gradually diminish. Compared with the results of GBLUP, the prediction accuracy of WGBLUP consistently maintains a higher growth rate when the reference group size increases from 1000 to 5000, especially when the reference group size increases from 3000 to 5000, its prediction accuracy improves from 0.308 to 0.356 to 0.0.317–0.418, which is a significantly higher increase than that of GBLUP.From the overall results, under different reference group sizes, the WGBLUP method has significantly higher prediction accuracy than GBLUP method (****, p < 0.00001). Comparison of the Accuracy of GP for PopA (a), PopB (b) and PopC (c) Under Various Reference Population Sizes Using GBLUP and WGBLUP Methods Similarly to further optimize the accuracy of genomic genetic evaluation, this study also compared the accuracy of GEBV using GBLUPmix and WGBLUPmix methods under different reference population sizes, as shown in Fig. 5. The results show that the prediction accuracies of both hybrid methods significantly increase with the increase of reference population size, and the prediction accuracies of WGBLUPmix are significantly higher than that of GBLUPmix under all reference population sizes. among them, the prediction accuracy of 0.185–0.26 of WGBLUPmix has been significantly higher than that of 0.121–0.225 of GBLUPmix at a reference population size of 1,000, which indicating that WGBLUPmix has higher efficiency in small reference populations. As the reference population size increases, the prediction accuracy of WGBLUPmix continues to improve, reaching a maximum value of 0.188–0.266 at a reference population size of 5000, whereas the prediction accuracy of GBLUPmix is 0.146–0.25 at the same reference population size, which further confirms the advantage of WGBLUPmix in the genetic evaluation using mixed reference populations. In addition to this, the prediction accuracy of the WGBLUPmix method was significantly higher than that of the GBLUPmix method at smaller reference population sizes (e.g. 1000 and 2000) (****,p < 0.00001). However, when the reference population size increased to 3000, the difference in prediction accuracy between the two methods was no longer significant (ns, p > 0.05). At reference population sizes of 4000 and 5000, the prediction accuracy of the WGBLUPmix method was again significantly higher than that of the GBLUPmix method (****,p < 0.00001). All these results suggest that the WGBLUPmix method is able to utilize genomic information more efficiently for genetic assessment at different reference population sizes. Comparison of the accuracy of GP for PopA (a), PopB (b) and PopC (c) under various reference population sizes using GBLUPmix and WGBLUPmix methods The impact of mixed reference populations on the accuracy of GEBV prediction Finally, in order to more intuitively compare the prediction accuracy differences between the traditional GBLUP and GBLUPmix methods and the WGBLUP and WGBLUPmix methods under different reference group sizes, the comparison of their results is visualized, as shown in Figs. 6 and 7. Firstly, according to the comparison results of traditional GBLUP and GBLUPmix in Fig. 6 and Table S1, the prediction accuracy of both methods showed an upward trend, and with the increase of the size of the reference group, the GBLUPmix method did not show higher prediction accuracy in the three groups. Among them, for PopA and PopB, the prediction accuracy of GBLUPmix in all reference group sizes is almost significantly lower than that of traditional GBLUP, and the analysis of variance () is p < 0.0001 (***) and p, respectively < 0.001 (**), but the difference in prediction accuracy between the traditional GBLUP and GBLUPmix methods for PopB was not significant, and the results after ANOVA showed that it was not significant (*). In addition, the prediction accuracy of traditional GBLUP shows a relatively stable increase with the increase of reference group size, especially PopA and PopB, when the reference group size increases from 1000 to 3000, the prediction accuracy increases relatively slightly. However, when the reference group size increases from 3000 to 5000, the prediction accuracy of WGBLUP improves more significantly, indicating that the prediction accuracy of traditional GBLUP is higher when the reference population size is larger. In contrast, when the reference group size increases from 1000 to 5000, the prediction accuracy of GBLUPmix continues to increase at a higher rate, but the increase is not as large as that of traditional GBLUP. Comparison of the accuracy of GP for PopA (a), PopB (b) and PopC (c) under various reference population sizes using GBLUP and GBLUPmix methods Comparison of the accuracy of GP for PopA (a), PopB (b) and PopC (c) under various reference population sizes using WGBLUP and WGBLUPmix methods Then, according to the comparison results of WGBLUP and WGBLUPmix in Fig. 7 and Table S2, the prediction accuracy of both methods increases with the increase of the size of the reference group, and the WGBLUPmix method shows high prediction accuracy in all three groups. However, for PopA, PopB, and PopC WGBLUPmix showed significantly lower prediction accuracy than WGBLUP in all reference group sizes, and the ANOVA result was p < 0.0001(***)。 In addition, the prediction accuracy of WGBLUP showed a relatively modest improvement with the increase of the reference group size, especially when the reference group size increased from 1000 to 3000, the prediction accuracy of PopB and PopC groups increased slightly. However, when the reference group size increases from 3000 to 5000, the prediction accuracy of WGBLUP increases more significantly, indicating that the prediction accuracy of WGBLUP is higher when the reference population size is larger. In contrast, when the reference population size increases from 1000 to 5000, WGBLUP’s prediction accuracy continues to increase at a higher rate, but not as much as WGBLUPmix. Table 5 presents the unbiased estimates obtained using four methods—GBLUP, WGBLUP, GBLUPmix, and WGBLUPmix—across three populations (A, B, and C). These estimates were derived from reference populations of varying sizes, ranging from 1,000 to 5,000 individuals. In population A, the heritability estimates (b) obtained through GBLUP and WGBLUP were relatively stable across different population sizes, with consistently low standard errors (SE). Although GBLUPmix and WGBLUPmix provided comparable estimates, the SE was slightly higher, indicating that the inclusion of mixed populations had a minimal impact on the accuracy of heritability estimation. In population B, the heritability estimates from GBLUP and WGBLUP showed a slight increase with the increase in population size, demonstrating an improvement in estimation accuracy with larger reference populations. The GBLUPmix and WGBLUPmix methods again provided similar results, with a notable increase in SE as the population size increased, likely due to the increased complexity of the genetic background in the mixed reference panel. In population C, the estimates from GBLUP and WGBLUP were lower than those in populations A and B, possibly reflecting genuine differences in heritability or methodological limitations. The GBLUPmix and WGBLUPmix methods showed slightly higher estimates, suggesting that the mixed reference panel might capture more genetic variation, although the SE values were also higher. Overall, while all four methods are capable of providing unbiased estimates, the choice of method and the composition of the reference population can influence the accuracy and precision of heritability estimates. Table 5The unbiased estimates obtained using the GBLUP、WGBLUP、GBLUPmix and WGBLUPmix method in populations A, B, and CPopReference population sizeGBLUPWGBLUPGBLUPmixWGBLUPmixbSEbSEbSEbSEA10001.220.0161.240.0121.460.0261.230.01120001.40.0711.010.0511.650.0331.010.05430001.410.0611.020.0521.770.0261.020.05240001.590.061.120.0511.460.0221.120.0550000.980.0511.310.0530.970.0181.310.053B10001.240.0151.010.0341.140.0571.240.01520000.90.0111.040.0811.430.0231.280.01730001.510.0461.150.0930.880.0181.860.01740001.050.0351.080.0110.810.0171.160.02250001.350.0141.280.0721.350.0141.160.012C10001.340.0141.240.010.820.0131.110.04220001.050.0121.030.0771.110.0150.940.01930001.480.0121.340.0791.480.0121.020.01540001.550.0121.010.0651.550.0121.350.01250001.460.011.290.0621.460.0111.020.012 The unbiased estimates obtained using the GBLUP、WGBLUP、GBLUPmix and WGBLUPmix method in populations A, B, and C Discussion Accuracy of genomic genetic evaluation reveals high and low heritability [25–27], the size of reference population [Predicting the effect of reference population on the accuracy of within, across population on the accuracy of within, across, and multibreed GP; Increased accuracy of artificial selection by using the realized relationship matrix. Accounting for population structure in selective cow genotyping], marker densities [28, 29], magnitude and number of QTL effect [30], and the linkage between marker and QTL disequilibrium [31, 32] and other factors on GP performance. In GP of multi-breed beef cattle, the reasons for the limited improvement of prediction accuracy are multiple, and one of the key factors is the differences in LD between SNPs and QTLs among different beef cattle breeds [33]. Most previous studies have focused on GP of single breed populations or hybrid breed populations against purebred breed populations, and have neglected the potential impact of differences between different interlocking disequilibrium patterns on prediction accuracy when constructing mixed reference populations in multibreed populations. The objective of this study was to investigate multi-breed reference populations of beef cattle under different LD patterns by comparing four evaluation methods, GBLUP, WGBLUP and their hybrid versions (GBLUPmix and WGBLUPmix), and attempting to construct reference populations from multi-breed populations by extracting different reference population sizes and mixing them equitably, in order to determine whether multi-breed mixed reference populations to determine whether multi-species mixed reference populations could improve the likelihood of prediction accuracy. In studies of GP in multi-breed beef cattle, the impact of LD differences between SNPs and QTL among different breeds on prediction accuracy cannot be ignored. As pointed out by Schmid et al. [34], differences in LD between SNPs and QTLs among different breeds may lead to multi-breed GP with limitations, such as decreased prediction accuracy, genetic background differences among breeds, underrepresentation of reference populations, differences in QTL-SNP associations, and inconsistent LD phases. In small breed populations, these challenges are particularly evident due to the limited size of reference populations. Consequently, these populations require joint genetic evaluation across breeds to overcome difficulties in genetic assessment. In this study, we found that when different breeds of beef cattle were simply combined into a single reference population and GPs were made, prediction accuracy was not significantly improved in populations where the LD patterns were all different, which is consistent with previous findings [14, 35–37]. This suggests that prediction accuracy may not be effectively improved by increasing the size of the reference population alone without taking into account the LD differences between SNPs and QTLs across varieties. Even when there are partially overlapping QTL between breeds, these differences may lead to a decrease in prediction accuracy. To overcome the problem that prediction accuracy was not improved after simply combining different breeds of beef cattle into a single reference population, in this study, we adjusted the weights based on the weighted GBLUP method by introducing SNP weights to reflect the differences in the contribution of different SNPs to the phenotype, using the effect estimates of neighboring SNPs, and at the same time, we took into account the linkage imbalance among SNPs in the mixed reference population of multiple breeds (LD), an approach that assumes that effect sizes of neighboring SNPs tend to be similar in model fitting while temporarily setting aside other factors to improve prediction accuracy. The results show that the significant improvement in the effect of our method in terms of prediction accuracy improvement compared to training alone and simply merging the reference populations can be explained by the fact that when merging the reference populations of two study subjects, if a common substitution effect is estimated for a SNP in an additive model, then certain SNPs with a chained imbalance with the QTL in one of the study subjects, the If the phase of the chain disequilibrium with the QTL is different in the other study subject, this opposite phase of chain disequilibrium may generate false signals in estimating the effects of these SNPs. However, in this study, we used the weighted GBLUP method, in which we not only considered the effect of allele frequencies of SNPs on the weights, but also adjusted the weights of the current SNPs by a window-weighting scheme using the effect estimation of the neighboring SNPs, to efficiently filter out SNPs with similar chain disequilibrium phases in multiple study subjects and estimate their effects individually. By the above method, we are able to effectively screen out SNPs with similar chained imbalance phases in multiple research objects and estimate their effects individually. The subset of SNPs selected based on the weighted GBLUP was regarded as more informative SNPs, thus significantly improving the prediction accuracy by this scheme. This approach not only preserves population-specific estimates of QTL effects but also leverages cross-population information, resulting in a significant overall gain in prediction accuracy compared with either single-population or simply merged reference sets. In this study, it was found that an increase in the size of the reference population significantly improved the accuracy of GP, which is in line with the findings of Brito, Nwogwugwu, and Perez et al. [21, 38, 39] are consistent with the findings that the size of the reference population is an important factor in GP. However, we also note that simply increasing the reference population size does not completely solve the problem of prediction accuracy. As mentioned in the study of Yin et al. [40], LD reflects the degree of non-random association among genomic loci, which is affected by population structure, genetic background and historical selection. In multi-population GP, the closer the kinship between populations, the more similar the LD patterns are, and the higher the applicability of the prediction model; on the contrary, if the kinship is far away, the LD differences are large, and the applicability is reduced. The limited sample size of small populations and the high uncertainty of LD estimation also affect the prediction accuracy. Therefore, an increase in reference population size can only maximize its potential if LD differences between SNPs and QTLs across breeds are taken into account and acted upon accordingly. Some studies have also focused on the effect of LD differences among different breeds of beef cattle on the combined reference population, both from the perspective of selecting more genetically diverse animals as GWAS discovery sets [41], or from the perspective of Bayesian assigning higher weights to SNPs close to causal polymorphisms [8, 42], the former focuses more on improving the resolution and accuracy of GWAS, whereas the latter focuses more on standardizing genotypes through the use of varietal-specific allele frequencies to better reflect genetic differences among varieties. Our comprehensive simulation study indicates that the relative performance of GBLUP and WGBLUP methods is significantly influenced by the LD patterns and marker density in cattle populations. Under low LD conditions, GBLUP demonstrates stable predictive accuracy; however, when LD is high or when marker weights align with genetic structure, such as regions carrying growth-related QTLs, WGBLUP shows greater advantage. This difference may stem from WGBLUP’s ability to prioritize markers associated with functional variation. Nevertheless, the biological mechanisms underlying these differences, such as pleiotropy or epistasis interactions, still require further investigation through integrated transcriptomic or GWAS data analysis.Thus, this study emphasizes that in multi-population GP, the increase in reference population size, although important, must be considered in conjunction with differences in LD between populations and the complexity of the genetic background. Although some positive results were obtained in this study, we also recognize that there are some limitations of the model. For example, this study was based on simulated data for analysis, and while simulated data allow for flexibility in controlling parameters such as population and genetic background, there are still differences between real data. Despite the precision of QMSim, its default parameter settings have several limitations: (1) environmental effect modeling only considers additive variance components and does not incorporate genotype-environment interactions (G×E); (2) simplified assumptions regarding population structure (e.g., fixed migration rates, dispersed generations); (3) limited capability to simulate epigenetic effects and non-additive genetic effects.In addition, only GBLUP, WGBLUP and their hybrid versions of the assessment methods were examined in this study, while other GP models such as Bayesian methods were not used, which may not fully reflect the performance of all methods in different populations and reference population sizes. Finally, the accuracy of GP is affected by a variety of genetic parameters, such as the heritability of a trait and the number of QTL [9, 43, 44]. The present study did not explore in depth the variation of these genetic parameters across populations and evaluation methods and their impact on prediction accuracy, which somewhat limits in-depth interpretation of the results. These limitations suggest that in subsequent studies, we need to combine real data and a wider variety of assessment methods to explore in depth the mechanism of the role of genetic parameters and to expand the range of reference population sizes in order to further optimize the strategy of multi-population GP. In our study, we observed that the increase in reference population size significantly improved the accuracy of GP, especially when the reference population size was increased from 4000 to 5000, and the WGBLUPmix model showed the highest prediction accuracy when multiple populations were combined. In addition, there were significant differences in the performance of the WGBLUPmix model under different LD patterns.The superiority of the WGBLUPmix model in these several scenarios could be attributed to several potential impacts by adjusting or assigning different weights, including capturing different LD patterns in the genomes, adapting to the genetic backgrounds in the different population structures, and being able to better capture large effect SNP contributions [19, 45, 46]. These results emphasize the importance of flexibly adjusting or assigning different weights to accommodate different reference population sizes and chain imbalance patterns in multi-population GP through WGBLUPmix models, providing an effective strategy for improving prediction accuracy. Conclusions In conclusion, this study delved into the effects of reference population size and LD on the accuracy of GP by modeling beef cattle populations with different genetic backgrounds. Based on the findings, the importance of integrating reference population size, chain imbalance differences and genetic background complexity in multi-population GP was emphasized.The WGBLUPmix model showed significant advantages in handling multi-population data, especially when there were differences in chain imbalance patterns. These findings provide an important theoretical basis for beef cattle breeding and provide directions for future research, especially in optimizing reference population construction and evaluation methods. Supplementary Information Supplementary Material 1. Acknowledgements None. Abbreviations Genomic selection True breeding values Estimation of breeding values Best linear unbiased prediction Genomic estimated breeding value Single-nucleotide polymorphisms Minor allele frequency Linkage disequilibrium Quantitative trait loci Genomic best linear unbiased prediction Weighted best linear unbiased prediction Genomic relationship matrix Authors’ contributions ZL1conceived and performed statistical analysis and was a major contributor in writing the manuscript. MFY and CCC improved the design of the methodologies for the experiment and the manuscript. ZL, GMJ and NRS Mainly responsible for providing manuscript resources and project management and improved manuscript. ZWG(corresponding author) mainly reviewed and edited the manuscript and supervised its writing. All authors read and approved the final manuscript. Funding This work was supported by the Double First-Class Construction Funds for IMAU (BZX202201), the Fundamental Research Funds for Inner Mongolia Autonomous Region Direct Affiliated Universities(BR22-11-13 and BR221024), the Inner Mongolia Natural Science Foundation project (2021ZD05). Data availability The datasets generated and analyzed during the current study are not publicly available due to the nature of the data being generated through simulation software. However, the parameters used for the simulation are fully documented in the Methods section of this article. The simulation software used is publicly available and can be accessed at https://animalbiosciences.uoguelph.ca/~msargol/qmsim/. Any relevant data can be provided by the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate The experimental data in this study are simulated, therefore approval from the Animal Care and Use Committee is not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The datasets generated and analyzed during the current study are not publicly available due to the nature of the data being generated through simulation software. However, the parameters used for the simulation are fully documented in the Methods section of this article. The simulation software used is publicly available and can be accessed at https://animalbiosciences.uoguelph.ca/~msargol/qmsim/. Any relevant data can be provided by the corresponding author upon reasonable request.