Genomic Prediction in Livestock: A Decision Framework for Breeders

Genomic prediction uses genome-wide single nucleotide polymorphism (SNP) markers to estimate the genetic merit of animals before their own performance or progeny information is available. This article provides a practical decision framework for breeders, researchers, and analysts who need to implement genomic prediction in livestock breeding programs. The framework covers data requirements, model selection, interpretation of results, and common implementation pitfalls, with emphasis on matching statistical methods to population structure and available resources.

At a Glance

Genomic prediction replaces traditional pedigree-based relationships with genomic relationships derived from dense SNP marker data. The core workflow involves assembling a reference population with both genotypes and phenotypes, selecting a statistical model, validating prediction accuracy, and applying the resulting equation to select candidates. The table below summarizes the main model families and their practical considerations.

No single algorithm performs best across all species and traits. A systematic comparison of ten genomic prediction algorithms across eight traits in five aquaculture species found that no algorithm achieved the best performance for every trait, although reproducing kernel Hilbert space and support vector machine methods showed relatively high accuracy in most tested traits, while BayesA and random forest better prevented noise interference in phenotypic data. Breeders should therefore evaluate multiple models on their own data instead of adopting a default method.

Core Principles of Genomic Prediction

From Pedigree Relationships to Genomic Relationships

Traditional estimated breeding values rely on pedigree records to model the expected proportion of genes shared between relatives. Genomic prediction instead uses SNP genotypes to estimate realized relationships, capturing the actual segregation of parental genomes within families. This distinction matters because animals within a breed are all related to some extent, and genomic relationships provide a finer resolution than pedigree expectations. The concepts underlying genomic prediction of estimated breeding values in livestock and polygenic risk scores in humans share the same statistical foundation, but large family sizes in livestock create smaller effective population sizes, increased linkage disequilibrium, and higher average genetic relationships within populations. These population features directly influence which models are appropriate and how accuracy should be interpreted.

The Reference Population Is the Foundation

Prediction accuracy depends more on the composition of the reference population than on the choice of statistical model. The reference population must contain animals with both reliable phenotypes and genotypes, and it must be genetically connected to the selection candidates. The importance of information on relatives for predicting genomic breeding values has direct implications for the makeup of reference data sets in livestock breeding schemes. Animals that are closely related to the prediction candidates contribute more information than distantly related animals, so reference populations should include recent relatives of the candidates whenever possible.

Reference population size is a practical constraint, particularly for small and medium scale breeding programs. A study of Yorkshire pigs with three populations of different genetic backgrounds and sizes demonstrated that combining reference populations of the same breed or related breeds can improve genomic prediction accuracy when individual reference data are limited. The study used a 50K SNP array and evaluated three scenarios: combining all three populations, combining two populations to predict the third, and predicting each population independently. Results showed that the optimal combination strategy depended on the genetic relationship between populations and the trait being predicted.

Heritability and Genetic Architecture Determine Achievable Accuracy

The upper limit of prediction accuracy is constrained by the heritability of the target trait and the proportion of genetic variance captured by the markers. Traits with higher heritability generally yield more accurate predictions because phenotypic records contain more genetic signal. A study of yak coat color found heritability of 42.8% for head coat color but only 18.0% for rump coat color, and genomic selection using gBLUP and Bayesian generalized linear regression showed moderate to high accuracy with higher accuracy under random validation settings than aligned settings. Breeders should establish realistic accuracy expectations based on trait heritability before investing in genotyping.

Data Requirements and Quality Control

Genotype Data

SNP arrays have become a cornerstone of modern livestock breeding because they facilitate identification of genetic markers linked to economically important traits and provide a tool for predicting breeding values. However, conventional breeding programs often overlook additional genomic features contained in SNP array data, including copy number variation, runs of homozygosity, inbreeding levels, and aneuploidy. The genomeprofile tool processes genotype data in finalreport or PLINK ped format and produces output containing copy number variations, runs of homozygosity, selection signatures, aneuploidy, and inbreeding per individual and across populations. This allows breeding companies and researchers to identify unique individuals or genomic regions of interest based on routinely collected data.

Quality control should include checks for call rate, minor allele frequency, Hardy-Weinberg equilibrium, and sample identity verification. Genotype errors can bias relationship estimates and reduce prediction accuracy. Breeders should document the SNP array version, genome assembly, and quality control thresholds used in each analysis.

Phenotype Data

Phenotypes must be recorded consistently across the reference population. For production traits such as milk yield, growth, or carcass composition, standardized recording protocols are essential. A study of Mehsana buffalo used test day records from over 10,000 animals for milk yield, fat yield, solid-not-fat yield, and protein yield, with 3,887 animals genotyped using a 54K SNP array after quality control. The study compared pedigree best linear unbiased prediction (PBLUP) with single-step genomic best linear unbiased prediction (ssGBLUP) and found that ssGBLUP provided more accurate and less biased breeding value predictions than PBLUP for production traits.

Phenotypes should be adjusted for known environmental effects before genomic prediction, or these effects should be included in the statistical model. Contemporary groups, herd-year-season effects, and management groups account for systematic environmental variation. Failure to model these effects inflates residual variance and reduces prediction accuracy.

Combining Reference Populations

When individual reference populations are small, combining populations of the same breed or related breeds can improve accuracy. The Yorkshire pig study demonstrated this approach for growth traits and lean meat percentage. However, combining populations with different genetic backgrounds requires careful modeling of breed or line effects. The benefit of combining populations depends on the genetic correlation between the populations and the traits being predicted. Breeders should test whether combined analyses improve validation accuracy before committing to a multi-population reference.

Model Selection Framework

GBLUP and Single-Step Methods

Genomic best linear unbiased prediction (GBLUP) replaces the pedigree relationship matrix with a genomic relationship matrix computed from SNP genotypes. The method assumes that all SNPs contribute equally to genetic variance, which is computationally efficient and robust for many traits. Single-step GBLUP (ssGBLUP) combines pedigree and genomic information in a single evaluation, allowing animals with genotypes and phenotypes, animals with only phenotypes, and animals with only genotypes to be included in one analysis. The Mehsana buffalo study showed that ssGBLUP improved prediction accuracy by 11.37% on average compared to PBLUP and reduced prediction bias.

GBLUP is the most widely used prediction model in animal breeding. A comparative evaluation in Holstein cattle analyzed 122,672 SNPs from 16,122 animals with estimated breeding values for nine traits and found that GBLUP maintained the best balance between accuracy and computational efficiency. Bayesian models achieved the highest average accuracy in that study, but all advanced methods required more than six times the computational time of GBLUP, limiting their practical scalability.

Bayesian Methods

Bayesian methods differ from GBLUP in their assumptions about the distribution of SNP effects. BayesA assumes each SNP has its own variance, BayesB assumes many SNPs have zero effect with a subset having nonzero effects, and BayesC and BayesR use mixture distributions to model the proportion of SNPs with effects. These methods can improve accuracy for traits with few large-effect loci or skewed genetic architectures. The Holstein study found that BayesR achieved the highest predictive performance among the methods tested, with BayesCπ performing nearly as well.

Bayesian methods require specification of prior distributions, which can influence results. The choice of prior should reflect knowledge of the genetic architecture of the trait. For traits known to be controlled by many genes of small effect, GBLUP and BayesC with a high proportion of SNPs in the zero-effect class may perform similarly. For traits with known major genes, BayesB or BayesR may capture more of the genetic variance.

Machine Learning Approaches

Machine learning models including random forest, support vector regression, kernel ridge regression, and deep learning frameworks have been applied to genomic prediction. A review of machine learning models in animal breeding found that these methods performed well in fitting large noisy datasets and modeling minor nonadditive effects in some studies, but conventional methods sometimes outperformed machine learning models, confirming that no universal method exists. The review noted that adoption of machine learning in animal breeding remains low due to data limitations, complex genetic interactions, lack of standardization and reproducibility, and limited interpretability of models trained with biological data.

A systematic evaluation of ten genomic prediction algorithms in aquaculture species found that reproducing kernel Hilbert space and support vector machine algorithms achieved relatively high prediction accuracies in most tested traits, while BayesA and random forest better prevented noise interference in phenotypic data. The study also found that prediction performance in one species was improved by using genome-wide association study results to select subsets of significant SNPs before prediction.

Deep learning frameworks have shown promise for specific traits. The Gsformer framework, which combines convolutional neural networks with self-attention mechanisms, was evaluated on six datasets spanning pig, cattle, chicken, mouse, wheat, and maize. It generally ranked among the top two models across datasets, with notable improvements for cattle fat percentage, chicken first egg weight, pig age at 100 kg body weight, and mouse anxiety. However, the computational demands of deep learning methods are substantial, and their advantage over simpler methods varies by trait and dataset.

Weighted and Multi-Omics Models

Incorporating prior biological information can improve prediction accuracy. SNP weights derived from genome-wide association studies or Bayesian analyses can be incorporated into GBLUP to construct weighted GBLUP (WGBLUP). The Holstein study found that WGBLUP using BayesBπ weights outperformed GBLUP across all traits, with an average accuracy gain of 1.1% and a notable 4.9% gain for one trait. However, WGBLUP using GWAS weights improved accuracy by 1.3% but showed a 9.1% loss in unbiasedness, highlighting the tradeoff between accuracy and bias.

Transcriptomic data can complement genomic information. A study in Japanese quail used transcript abundances from intestinal tissue to predict efficiency-related traits and found that transcriptomic information explained a larger portion of phenotypic variance than host genetics. Models incorporating both genetic and transcriptomic information outperformed those using only one type of information. The derived GTCBLUPi model addressed redundant information between genomic and transcriptomic data.

A self-attention-based framework called AbGP was developed to extract stable, major-effect SNPs from small multi-omics datasets and apply them to larger populations with genomic data alone. Using a discovery population of 218 Huaxi cattle with matched genotype and transcriptome data, AbGP identified a compact subset of SNPs in the top 1.25% and validated predictive power in an independent population of 1,496 cattle, where it outperformed GBLUP and machine learning baselines for economic traits.

Model Comparison and Selection Criteria

The choice of model should be based on validation accuracy, computational resources, and practical constraints. The following criteria should guide model selection:

A study simulating a selected livestock population with known QTL and effects compared GBLUP, weighted 2GBLUP, random forest, and support vector regression. The 2GBLUP model resulted in the highest accuracy, with accuracy increasing when included QTL explained up to 80% of genetic variance, after which accuracy dropped. With a weighted 2GBLUP model, accuracy always increased when more QTL were included. GBLUP accuracy was consistently higher than support vector regression, and random forest resulted in the lowest prediction accuracy without improvement from QTL information.

Practical Implementation Workflow

Step 1: Define Breeding Objective and Target Traits

Identify the traits that contribute to the breeding objective and determine their economic weights. Genomic prediction is most valuable for traits that are difficult or expensive to measure, expressed late in life, sex-limited, or measured after slaughter. For traits already recorded routinely, traditional pedigree-based evaluation may remain cost-effective.

Step 2: Assess Current Data Infrastructure

Review existing pedigree, phenotype, and genotype records. Determine the size and genetic connectedness of the current reference population. If the reference population is small, consider strategies for combining populations of the same breed or related breeds, as demonstrated in the Yorkshire pig study. Evaluate whether additional phenotyping or genotyping is needed before implementing genomic prediction.

Step 3: Select Genotyping Platform

Choose a SNP array with appropriate density for the species and breeding objective. The genomeprofile tool demonstrates that SNP array data can provide comprehensive profiles including copy number variants, runs of homozygosity, and aneuploidy screening. Higher density arrays capture more genetic variation but increase cost. Low-density arrays with imputation to higher density can reduce genotyping costs while maintaining accuracy.

Step 4: Establish Quality Control Protocols

Define thresholds for SNP call rate, sample call rate, minor allele frequency, and Hardy-Weinberg equilibrium. Document all quality control steps and retain records of excluded SNPs and samples. Verify sample identity using pedigree checks or genotype concordance with parents. The NCBI Data Resources provide reference genomes and variation databases that support genotype quality assessment.

Step 5: Build and Validate Prediction Models

Partition the reference population into training and validation sets. Use cross-validation to estimate prediction accuracy. Compare multiple models including GBLUP, at least one Bayesian method, and one machine learning approach if resources permit. The systematic evaluation of algorithms in aquaculture species provides a template for comparing methods across traits and species. An R package called ASGS integrates commonly used traditional and machine learning algorithms for efficiently finding the optimal algorithm.

Step 6: Implement Routine Evaluation

Once a model is selected, establish a routine evaluation pipeline. Genotype selection candidates at the earliest age possible to maximize the reduction in generation interval. Genomic selection enables prediction of genetic merit from genome-wide SNP markers and has been adopted by dairy industries worldwide, with expectations of doubling genetic gains for milk production and other traits. The reduction in generation interval is a major contributor to increased genetic gain.

Step 7: Monitor and Update

Genomic prediction equations require periodic updating as the population evolves and new phenotypes become available. Monitor prediction accuracy over time and rebuild models when accuracy declines or when the genetic composition of the population changes. Long-term response to genomic selection depends on managing inbreeding and maintaining genetic diversity.

Records and Measurements

Essential Records for Genomic Prediction

Breeders implementing genomic prediction should maintain the following records:

Measuring Prediction Accuracy

Prediction accuracy is typically measured as the correlation between predicted breeding values and corrected phenotypes in a validation set, divided by the square root of heritability. The regression coefficient of corrected phenotypes on predicted breeding values measures prediction bias, with values close to 1 indicating unbiased predictions. The Mehsana buffalo study reported regression coefficients ranging from 0.44 to 0.52 for PBLUP and from 0.45 to 0.61 for ssGBLUP, with average values of 0.48 and 0.53 respectively, indicating reduced prediction bias under the ssGBLUP model.

Reference Population Composition

The composition of the reference population affects prediction accuracy for different groups of selection candidates. A study of Hanwoo cattle evaluated the accuracy of genomic estimated breeding values for carcass traits using cow-based reference populations compared to steer-based populations. Steer-based reference populations achieved the highest accuracy, averaging 0.78, likely due to standardized management and higher trait heritability. Cow-based reference populations exhibited prediction accuracies averaging 0.64 but remained highly practical alternatives. Growth traits showed significant bias in cross-sex predictions, while fat deposition traits remained stable across sexes. The study concluded that incorporating cows into the reference population is strategically important even though steer-based populations provide optimal accuracy.

Common Failure Patterns

Inadequate Reference Population Size

The most common cause of disappointing genomic prediction accuracy is a reference population that is too small or poorly connected to selection candidates. Small reference populations produce imprecise marker effect estimates and unstable predictions. The challenge of sufficient reference population data is significant in breeding programs aimed at improving livestock on a small to medium scale. Combining reference populations of the same breed or related breeds can help overcome this limitation.

Ignoring Population Structure

Failure to account for population stratification or breed composition can inflate accuracy estimates in validation and produce biased predictions in practice. Cross-validation should be designed to reflect the intended use of the prediction equation. Predicting across breeds or lines requires models that account for breed-specific effects. A study of crossbred performance derived equations for predicting the reliability of genomic breeding values and showed that a model assuming breed-specific effects of SNP alleles outperformed a model assuming identical SNP effects across breeds for specific scenarios.

Overfitting in Machine Learning

Machine learning models can overfit training data, particularly when the number of markers greatly exceeds the number of animals. The review of machine learning models in animal breeding noted that there is no remarkable outperformance of machine learning methods compared to traditional methods in genomic prediction. Overfitting is detected when validation accuracy is much lower than training accuracy. Proper cross-validation with repeated folds and hyperparameter tuning is essential.

Ignoring Environmental Covariates

Genomic prediction models are often tested for their ability to predict new individuals under the same environmental conditions, but management and diet change over time. A study using milk infrared spectral data as environmental covariates in genomic prediction models for Canadian Holstein found that including spectral data improved predictive ability when predictions had to be extrapolated to new environments. Fourier-transformed infrared spectral data can be used to calculate environmental coordinates for a farm at a given time, improving predictions under changing conditions.

Poor Phenotype Quality

Genomic prediction cannot compensate for poor phenotype recording. Inconsistent trait definitions, unrecorded environmental effects, and measurement errors reduce heritability and prediction accuracy. The systematic evaluation of genomic prediction algorithms in aquaculture species found that BayesA and random forest better prevented noise interference in phenotypic data compared to other algorithms, but no method can recover information that was never recorded.

Limitations and Interpretation

Accuracy Ceilings

Prediction accuracy is bounded by the heritability of the trait and the proportion of genetic variance captured by markers. Even with perfect models, traits with low heritability will have modest prediction accuracy. Breeders should establish realistic expectations based on trait heritability and reference population size before implementing genomic selection.

Across-Population Transfer

Prediction equations developed in one population may not transfer to another population, particularly if the populations are distantly related or have different allele frequencies. The Yorkshire pig study showed that combining populations with different genetic backgrounds can improve accuracy for some traits but not others. Breeders should validate prediction equations in their own population before relying on external equations.

Non-Additive Effects

Standard genomic prediction models capture additive genetic effects. Non-additive effects such as dominance and epistasis are generally not modeled, which limits accuracy for traits where these effects are important. Machine learning methods can capture some non-additive patterns, but their advantage is inconsistent across studies.

Computational Constraints

Advanced methods including Bayesian models and deep learning require substantial computational resources. The Holstein study found that advanced methods required more than six times the computational time of GBLUP. Breeders with limited computing infrastructure should weigh the accuracy gains of advanced methods against their computational costs.

Data Sharing and Reproducibility

Genomic data sharing is governed by policies that protect animal owner privacy and data security. The NIH Genomic Data Sharing Policy provides a framework for responsible data sharing that balances scientific benefit with privacy protection. The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable. Breeders should establish data management plans that comply with applicable policies and support reproducibility.

Welfare and Safety Context

Inbreeding Management

Genomic selection can accelerate genetic gain but also increase the rate of inbreeding if not managed carefully. The genomeprofile tool identifies runs of homozygosity and inbreeding levels per individual and across populations, enabling breeders to monitor and manage inbreeding. Long-term response to genomic selection depends on maintaining genetic diversity. Breeders should use genomic relationships to manage mating decisions and limit the rate of inbreeding.

Genetic Defect Screening

Genomic data can be used to identify carriers of genetic defects. The review of genomic selection in livestock noted that whole-genome sequence data should accelerate the rate of gain and enable rapid discovery and elimination of genetic defects from livestock populations. SNP array data can also screen for aneuploidy, as demonstrated by the genomeprofile tool. Breeders should incorporate genetic defect screening into their genomic selection programs.

Environmental Adaptation

Genomic selection can contribute to adaptation to changing environmental conditions. A review of the future of livestock breeding noted that as climate change escalates, reduced methane emissions from cattle and sheep will be a critical goal, and genomic selection for efficiency and reduced emissions intensity will play an important role. Genotype-by-environment interaction can affect prediction accuracy, as demonstrated by the yak coat color study that incorporated genotype-by-environment interaction-associated signals. Breeders should consider genotype-by-environment interaction when developing prediction models for animals raised across diverse environments.

Professional Escalation Criteria

Breeders should seek professional assistance from quantitative geneticists or bioinformaticians when:

• Reference population size is below the minimum needed for reliable predictions and combining populations is being considered
• Validation accuracy is substantially lower than expected based on heritability and reference population size
• Machine learning models show large discrepancies between training and validation accuracy
• Cross-population prediction is required and the genetic relationship between populations is unknown
• Genotype quality control reveals unexpected patterns such as aneuploidy or excessive runs of homozygosity
• Integration of multi-omics data such as transcriptomics or spectral data is being considered
• Regulatory or data sharing requirements apply to the genomic data being collected

The EMBL-EBI Training portal provides educational resources for researchers who need to develop bioinformatics skills for genomic data analysis. The NCBI Data Resources provide access to reference genomes, variation databases, and tools that support genomic prediction workflows.

Frequently Asked Questions

What is the minimum reference population size for genomic prediction?

There is no universal minimum reference population size because the required size depends on trait heritability, effective population size, and the genetic relationship between reference and selection candidates. Small reference populations produce imprecise predictions, and combining populations of the same breed or related breeds can improve accuracy when individual reference data are limited. Breeders should use cross-validation to determine whether their reference population supports reliable predictions.

How do I choose between GBLUP and Bayesian methods?

GBLUP is the appropriate starting point for most breeding programs because it is computationally efficient, robust, and works well for traits controlled by many genes of small effect. Bayesian methods may improve accuracy for traits with known major genes or skewed genetic architectures, but they require more computational resources and prior specification. Compare multiple models using cross-validation on your own data before selecting a final model.

Can I use a prediction equation developed in another population?

Prediction equations may transfer between populations that are closely related, but accuracy typically declines as the genetic distance between populations increases. The Yorkshire pig study showed that combining populations with different genetic backgrounds can improve accuracy for some traits. Validate any external prediction equation in your own population before relying on it for selection decisions.

What causes prediction bias and how do I detect it?

Prediction bias occurs when predicted breeding values are systematically too high or too low relative to true breeding values. It is detected by regressing corrected phenotypes on predicted breeding values in a validation set, with regression coefficients close to 1 indicating unbiased predictions. The Mehsana buffalo study found that ssGBLUP reduced prediction bias compared to PBLUP. Bias can result from poor phenotype adjustment, incorrect model specification, or genotype quality problems.

How often should I update my genomic prediction model?

Prediction models should be updated when new phenotypes become available, when the genetic composition of the population changes, or when validation accuracy declines. The rate of model decay depends on the generation interval and selection intensity. Breeders should monitor prediction accuracy over time and rebuild models on a regular schedule, typically annually or when the reference population grows substantially.

Are machine learning methods better than traditional genomic prediction?

Machine learning methods can capture complex patterns and non-additive effects, but they do not consistently outperform traditional methods. The review of machine learning models in animal breeding found no remarkable outperformance of machine learning compared to traditional methods. Machine learning requires large datasets, careful validation, and substantial computational resources. Evaluate machine learning methods on your own data before adopting them.

How does trait heritability affect genomic prediction accuracy?

Higher heritability traits produce more accurate predictions because phenotypic records contain more genetic signal. The yak coat color study found heritability of 42.8% for head coat color compared to 18.0% for rump coat color, with corresponding differences in prediction accuracy. For low heritability traits, larger reference populations and more precise phenotyping are needed to achieve useful accuracy.

What additional information can I extract from SNP array data?

SNP array data contain information beyond marker genotypes for prediction. The genomeprofile tool identifies copy number variants, runs of homozygosity, selection signatures, aneuploidy, and inbreeding levels per individual and across populations. This information can support genetic diversity management, genetic defect screening, and identification of unique individuals or genomic regions of interest.

Related Bioinformatics Guides

• Genomic Selection in Animal Breeding
• African Swine Fever: Computational Models for Early Detection and Spread Prediction in Wild Boar Populations
• Predicting AMR from Genomic Data
• Data Sharing and Privacy in Genomic Research
• Computational Strategies in Structure Based Drug Design

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• genomeprofile: Unveiling the genomic profile for livestock breeding through comprehensive SNP array-based genotyping.. Animal genetics, 2024.
• Systematic Evaluation of Genomic Prediction Algorithms for Genomic Prediction and Breeding of Aquatic Animals.. Genes, 2022.
• Complex Trait Prediction from Genome Data: Contrasting EBV in Livestock to PRS in Humans: Genomic Prediction.. Genetics, 2019.
• The future of livestock breeding: genomic selection for efficiency, reduced emissions intensity, and adaptation.. Trends in genetics : TIG, 2013.
• Prediction of the reliability of genomic breeding values for crossbred performance.. Genetics, selection, evolution : GSE, 2017.
• Enhancing Genetic Gain through Genomic Selection: From Livestock to Plants.. Plant communications, 2020.
• Genomic Prediction of Growth Traits in Yorkshire Pigs of Different Reference Group Sizes Using Different Estimated Breeding Value Models.. Animals : an open access journal from MDPI, 2024.
• Genomic selection in livestock populations.. Genetics research, 2010.
• Incorporating GO/KEGG Functional Annotations Improves the Accuracy and Stability of Genomic Prediction Across Diverse Beef Cattle Populations. 2026.
• Genome-wide detection and genomic selection based on genotype-by-environment interaction-associated signals for yak coat color.. 2026.
• Accuracy of Genomic Prediction for Meat Quality Traits Using Cow Reference Populations in Hanwoo Cattle. 2026.
• Enhancing genomic prediction accuracy in Huaxi cattle through integration of transcriptomic data and a self-attention-based SNP selection strategy.. 2026.
• Accuracy of genomic prediction for milk production traits in Mehsana buffalo.. 2026.
• Gsformer: a dual-architecture deep learning framework with CNN-self-attention and sparse-attention for genomic selection.. 2026.
• A review of machine learning models applied to genomic prediction in animal breeding. Frontiers in Genetics, 2023.
• Incorporating transcriptomic data into genomic prediction models to improve the prediction accuracy of phenotypes of efficiency traits. Genetics Selection Evolution, 2025.
• Comparative evaluation of SNP-weighted, Bayesian, and machine learning models for genomic prediction in Holstein cattle. BMC Genomics, 2025.
• Incorporating information of causal variants in genomic prediction using GBLUP or machine learning models in a simulated livestock population. Journal of Animal Science and Biotechnology, 2025.
• Use of Milk Infrared Spectral Data as Environmental Covariates in Genomic Prediction Models for Production Traits in Canadian Holstein. Animals, 2022.
• Comparison of genomic prediction methods for early growth traits of Inner Mongolia cashmere goats based on multi trait models.. Yi chuan = Hereditas, 2024.
• Genomic best linear unbiased prediction (gBLUP) for the estimation of genomic breeding values. Methods in Molecular Biology, 2013.
• Genomic selection needs to be carefully assessed to meet specific requirements in livestock breeding programs. Frontiers in Genetics, 2015.
• The importance of information on relatives for the prediction of genomic breeding values and the implications for the makeup of reference data sets in livestock breeding schemes.. Genetics Selection Evolution GSE, 2012.
• Empowering precision livestock farming: Artificial intelligence applications in animal genomic breeding and multi-dimensional phenotypic measurement. Smart Agricultural Technology, 2025.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.