Genomic Selection in Animal Breeding: Methodological Foundations, Statistical Frameworks, and Species-Specific Applications

Introduction

Genomic selection (GS) represents a paradigm shift in animal breeding, enabling the prediction of genetic merit for complex traits using genome-wide marker data without requiring phenotypic records on all selection candidates [1, 2]. Unlike traditional marker-assisted selection that relies on a few significant quantitative trait loci (QTL), GS simultaneously estimates the effects of thousands of single nucleotide polymorphisms (SNPs) spanning the entire genome, thereby capturing both major and minor additive genetic contributions [3, 4]. The fundamental premise is that linkage disequilibrium (LD) between markers and causal variants allows accurate prediction of breeding values from marker genotypes alone.

The discipline draws on principles from population genetics, biostatistics, and high-performance computing. The reference population (or training population) comprises individuals with both genotypic and phenotypic records; the statistical model is trained on these data to estimate marker effects. Subsequently, selection candidates (often young animals without phenotypes) are genotyped and their genomic estimated breeding values (GEBVs) are computed as the sum of marker effects weighted by their genotypes [5, 6]. This approach dramatically shortens generation intervals, increases selection intensity, and can improve accuracy for sex-limited, low-heritability, or late-expressing traits.

The present article provides a technical review of GS methodology, statistical prediction models, implementation strategies across major livestock and aquaculture species, and emerging computational innovations. The discussion is situated within a veterinary and computational biology context, emphasizing practical considerations for breeding program design.

Statistical Frameworks for Genomic Prediction

Linear Mixed Models and the GBLUP Approach

The classic genomic best linear unbiased prediction (GBLUP) replaces the pedigree-based numerator relationship matrix (A) with a genomic relationship matrix (G) constructed from SNP genotypes. The model is:

y = Xb + Zg + e

where y is the vector of phenotypes, b fixed effects, g the vector of additive genetic effects with variance Gsigma^2_g, and e the residual. The G matrix is computed as G = (M-P)(M-P)' / [2sum(p_j(1-p_j))], where M is the matrix of marker genotypes (coded 0,1,2) and P contains twice the allele frequencies [6, 7]. This framework efficiently combines information from all markers and yields predictions with high accuracy for many polygenic traits.

Bayesian Alphabet and Variable Selection

Bayesian methods (BayesA, BayesB, BayesC, BayesR) incorporate different prior distributions for marker effects, accommodating a mixture of large-effect and small-effect loci. BayesB assumes most markers have zero effect, using a mixture distribution with a probability pi of having a non-zero effect drawn from a t-distribution [7, 8]. These models are computationally intensive but can improve prediction accuracy when traits are controlled by a few QTL of moderate to large effect. A recent implementation, 2NPLGBM, merges classical linear mixed models with machine learning gradients, demonstrating utility in both plant and animal contexts [9].

Machine Learning Algorithms

Deep learning architectures have entered GS. Yang et al. [7] introduced Gsformer, a dual-architecture framework combining convolutional neural networks (CNNs) with self-attention mechanisms and sparse-attention layers. This model captures both local LD patterns and long-range genomic interactions, outperforming GBLUP in simulated and real datasets for prediction accuracy. Convolutional denoising autoencoders, such as ALGI, leverage local genomic information for genotype imputation prior to prediction, reducing missing data bias [4].

Table 1 summarizes key model categories and their characteristics.

Genomic Imputation and Data Preprocessing

Missing genotype data is a pervasive challenge, especially with low-coverage sequencing or sparse SNP panels. Tan et al. [4] developed ALGI, a sparse convolutional denoising autoencoder that uses local genomic windows to impute missing genotypes. This method preserves LD structure and significantly improves downstream prediction accuracy compared to traditional imputation methods such as Beagle or FImpute. For time-series phenotypes like egg production, random forest imputation has been validated for filling missing longitudinal records, enabling more complete training datasets [10].

Applications Across Species

Dairy Cattle: Global Implementation and Methane Emissions

GS is most advanced in dairy cattle, where it has been commercially applied for fertility, production, and health traits. Richardson et al. [1] provided a global census on the development of direct genetic selection for enteric methane emissions. This trait, measured via respiration chambers or proxy measurements like milk mid-infrared spectra, is now being incorporated into national selection indices. The study documented that over 20 countries have established reference populations for methane, with prediction accuracies ranging from 0.20 to 0.50 depending on training size and phenotyping methodology. Genomic selection for methane offers a sustainable route to reduce the carbon footprint of dairy production.

Semen quality traits have also been subjected to GS in Italian Brown Swiss cattle [2]. Genetic parameters and genome-wide scans revealed moderate heritabilities (0.10 to 0.30) for ejaculate volume, sperm concentration, and motility. Gene-set analyses identified enrichment for spermatogenesis and ion transport pathways, demonstrating how GS can simultaneously improve reproductive efficiency.

Beef Cattle: Crossbred Prediction Using Reference Populations

In beef production, GS often faces challenges of small reference populations and multi-breed structures. Haque et al. [11] evaluated prediction accuracy for carcass traits in Jeju Black cattle using a Hanwoo reference population. Accuracies for backfat thickness and marbling score ranged from 0.35 to 0.55 when genomic relationships between breeds were moderate. The study underscores the importance of genetic proximity between reference and target populations.

Poultry: Semen Traits and Egg Production

Native chicken breeds are reservoirs of genetic diversity. Juiputta et al. [12] applied GS for semen traits in Thai native roosters, using a 50K SNP panel. Prediction accuracies for sperm motility and viability exceeded 0.40, indicating that GS can be cost-effectively deployed in non-commercial populations. For broiler breeders, Lan et al. [10] developed random forest imputation for missing egg production data, then applied GBLUP and Bayesian methods to predict cumulative egg number. Bayesian models with variable selection outperformed GBLUP when the data contained irregular missing patterns.

Goats: Combining Genic and Genomic Selection

Palhiere et al. [8] tested frequency-based methods of genic selection alongside GS in dairy goats. Genic selection targets specific alleles with known favorable effects (e.g., on milk composition), while GS captures the remaining polygenic variance. The combined approach yielded higher accuracy than either method alone, particularly for low-heritability traits. This hybrid strategy is especially relevant for populations where major genes (e.g., alphaS1-casein) have been characterized.

Aquaculture: Shrimp Genomics

Genomic selection is expanding rapidly in aquaculture due to the availability of cost-effective SNP panels. Fu et al. [3] developed a 1K SNP panel for whiteleg shrimp (Litopenaeus vannamei) and demonstrated applications from pedigree reconstruction to genomic prediction for growth and survival. The panel provided sufficient power to estimate parentage and produced GEBV accuracies of 0.30 to 0.50 for body weight. The low marker density makes implementation feasible for commercial hatcheries.

Swine: Fertility Prediction with Low-Coverage Sequencing

Zhang et al. [13] compared machine learning (random forest, gradient boosting) and GBLUP for fertility traits in Holstein heifers using low-coverage whole-genome sequencing (1X coverage). Machine learning models marginally outperformed GBLUP for age at first service and conception rate, particularly when the training set included diverse parity records. The approach reduced genotyping cost by over 60% compared to standard SNP arrays.

Bovine Embryo Genomic Selection

Wang et al. [14] developed an integrated single-tube whole-genome amplification and library construction system for bovine preimplantation embryos. This method enables genotyping from as few as five cells, allowing selection of embryos for transfer based on predicted genetic merit. Implementation in commercial embryo transfer programs could increase genetic gain by reducing generation intervals to a single gestation period.

Technical Considerations and Optimization

Reference Population Size and Composition

Gowane et al. [6] examined genetic parameter predictivity in large populations under strong GS. They demonstrated that as selection advances, the variance of marker effects changes due to allele frequency shifts, and prediction accuracy declines unless the reference population is periodically updated. Maintaining a diverse reference set that includes recent selection candidates is critical for sustained accuracy.

Detection of Ongoing Selection

Jansen et al. [15] developed methods to detect selection history in populations under ongoing directional selection. Their approach uses haplotype homozygosity and FST-like statistics to identify genomic regions experiencing recent selection. This methodology can inform GS model updates by pinpointing loci where effect sizes may be changing due to selection pressure.

Long-Term Maintenance of Genetic Diversity

Schrauf et al. [5] assessed GS strategies for maintaining rare alleles and de novo mutations over 40 generations of simulated selection. Strategies that incorporated optimal contribution selection (OCS) or mate allocation based on genomic relationships retained 15-20% more rare alleles than truncation selection alone. The study emphasizes that GS must be coupled with genetic diversity management to avoid inbreeding depression and loss of adaptive potential.

Computational Advances

Deep Learning for Genomic Selection

Gsformer [7] represents a significant advance in integrating deep learning with GS. The dual-architecture includes a CNN module to extract local SNP window features and a self-attention module to model distant interactions. Sparse-attention reduces computational load by focusing on informative marker pairs. In benchmark comparisons, Gsformer improved prediction accuracy by 5-12% over GBLUP for milk yield and somatic cell score.

Imputation with Convolutional Autoencoders

The ALGI algorithm [4] employs an encoder-decoder structure where a sparse convolutional layer learns local genomic patterns. The decoder reconstructs missing genotypes. When applied to a simulated dataset with 30% missingness, ALGI achieved imputation accuracy of 0.97, compared to 0.92 for standard imputation. Downstream GS accuracy improved by 8% after imputation.

Ensemble Methods and Gradient Boosting

Osatohanmwen et al. [9] introduced 2NPLGBM, a two-stage model that first fits a linear genomic relationship model to capture additive variance, then applies light gradient boosting to capture residual non-additive effects. In wheat data, 2NPLGBM outperformed GBLUP by 10% for grain yield. The framework is species-agnostic and has potential for animal breeding. Munroe et al. [16] provided a multi-metric evaluation of stochastic gradient boosting for wheat, establishing parametric optimization guidelines.

Workflow of Genomic Selection

The following Mermaid diagram illustrates the canonical GS workflow, from reference population establishment to selection of elite candidates.

flowchart TD
    A[Establish Reference Population], > B[Genotype with SNP Array or Sequencing]
    B, > C[Collect High-Quality Phenotypes]
    C, > D[Train Statistical Model\n (GBLUP, Bayes, Deep Learning)]
    D, > E[Estimate Marker Effects]
    E, > F[Genotype Selection Candidates\n (Young Animals)]
    F, > G[Compute GEBVs]
    G, > H[Select Top Candidates\n (Truncation or OCS)]
    H, > I[Mate Selected Animals\n (Diversity Management)]
    I, > J[Produce Next Generation]
    J, > K[Update Reference Population\n with New Phenotypes?]
    K, Yes, > B
    K, No, > F

Challenges and Future Directions

Genotype-by-Environment Interaction

GS assumes marker effects are stable across environments, but for traits like heat tolerance or disease resistance, genotype-by-environment (GxE) interactions can reduce accuracy. Multi-trait reaction norm models that incorporate environmental covariates are being developed.

Integration with High-Throughput Phenomics

Automated phenotyping (e.g., mid-infrared spectroscopy, image analysis) can feed thousands of records into GS models. The coupling of such phenomics with low-cost genotyping will enable accelerated cycles in species with long generation intervals.

Ethical and Regulatory Considerations

As GS moves into embryo selection and gene editing, ethical frameworks must address welfare and biodiversity concerns. The maintenance of rare alleles [5] is particularly relevant for conservation breeding programs.

Cross-Species Prediction and Multi-Breed Models

Predicting across breeds remains challenging due to differences in LD patterns and allele frequencies. However, combining reference populations from multiple breeds with weighted relationship matrices has shown promise for traits with high genetic correlation.

Conclusion

Genomic selection has transformed animal breeding by enabling accurate prediction of genetic merit from molecular markers alone. Statistical models range from simple GBLUP to sophisticated deep learning architectures like Gsformer, each suited to different genetic architectures. Applications span dairy cattle (methane, semen traits), beef cattle, poultry (fertility, egg production), goats, swine, and aquaculture species such as shrimp [1-3, 12-14, 17]. Critical to success are adequate reference population size, periodic model updating, maintenance of genetic diversity, and robust imputation methods [4-6, 15]. As computational tools and genotyping costs continue to improve, GS will become standard practice in an expanding number of species, supporting both productivity and sustainability goals.

References

[1] Richardson CM, Amer PR, Post M, et al. Invited review: Global census on the development and implementation of direct genetic selection for enteric methane emissions in dairy cattle. J Dairy Sci. 2026. https://pubmed.ncbi.nlm.nih.gov/42285492/

[2] Sani N, Cavani L, Kawu M, et al. Genetic parameters, genomic scans, and gene-set analysis of semen traits in Italian Brown Swiss dairy cattle. J Dairy Sci. 2026. https://pubmed.ncbi.nlm.nih.gov/42285474/

[3] Fu Q, Qiang G, Wang P, et al. Development and Applications of a 1K SNP Panel for Whiteleg Shrimp: From Pedigree Reconstruction to Genomic Selection. Int J Mol Sci. 2026. https://pubmed.ncbi.nlm.nih.gov/42278199/

[4] Tan T, Gao B, Zhang R, et al. ALGI: Sparse Convolutional Denoising Autoencoder Utilizing Local Genomic Information for Genotype Imputation. Animals (Basel). 2026. https://pubmed.ncbi.nlm.nih.gov/42278023/

[5] Schrauf MF, Vandenplas J, Mulder HA, et al. Genomic selection strategies and their potential to maintain rare alleles and de-novo mutations: a long-term assessment. Genet Sel Evol. 2026. https://pubmed.ncbi.nlm.nih.gov/42277635/

[6] Gowane G, Hidalgo J, Hollifield MK, et al. Genetic parameters via predictivity in large populations under strong genomic selection. Genet Sel Evol. 2026. https://pubmed.ncbi.nlm.nih.gov/42252414/

[7] Yang T, Wu W, Xue Y, et al. Gsformer: a dual-architecture deep learning framework with CNN-self-attention and sparse-attention for genomic selection. Genet Sel Evol. 2026. https://pubmed.ncbi.nlm.nih.gov/42237097/

[8] Palhiere I, Gousseau V, Colleau JJ. Testing frequency-based methods of genic selection in addition to genomic selection in goats. Animal. 2026. https://pubmed.ncbi.nlm.nih.gov/42229035/

[9] Osatohanmwen BE, Vieira IC, Sharifi AR, et al. 2NPLGBM: a genomic model that merges the strengths of classical and machine learning methods in genomic prediction. Plant Methods. 2026. https://pubmed.ncbi.nlm.nih.gov/42210373/

[10] Lan J, Jia S, Ma J, et al. Random forest imputation and genomic prediction for missing egg production time-series data in yellow-feathered broiler breeders. BMC Genomics. 2026. https://pubmed.ncbi.nlm.nih.gov/42129641/

[11] Haque MA, Jang JH, Mou MA, et al. Genomic Breeding Value Prediction and Accuracy Assessment for Carcass Traits in Jeju Black Cattle Using Hanwoo Reference Populations. Anim Biosci. 2026. https://pubmed.ncbi.nlm.nih.gov/42167322/

[12] Juiputta J, Tuntiyasawasdikul V, Chankitisakul V, et al. Genomic selection and association analyses for improving semen traits in Thai native roosters. Poult Sci. 2026. https://pubmed.ncbi.nlm.nih.gov/42134141/

[13] Zhang ZB, Wang A, Wang QY, et al. Cost-effective genomic prediction for fertility traits: a comparison of machine learning and conventional models using low-coverage sequencing in Holstein heifers. Animal. 2026. https://pubmed.ncbi.nlm.nih.gov/42150295/

[14] Wang K, Yang X, Feng C, et al. Development and validation of an integrated single-tube whole-genome amplification and library construction system for genomic selection in bovine preimplantation embryos. Genomics. 2026. https://pubmed.ncbi.nlm.nih.gov/42119679/

[15] Jansen ACM, Calus MPL, Wientjes YCJ. Methods to Detect Selection History in a Population under Ongoing Directional Selection. Genetics. 2026. https://pubmed.ncbi.nlm.nih.gov/42198933/

[16] Munroe HN, Osatohanmwen BE, Reza Sharifi A. Multi-metric Evaluation and Parametric Optimization of Stochastic Gradient Boosting Machines for Genomic Prediction and Selection in Wheat (Triticum aestivum) Breeding. G3 (Bethesda). 2026. https://pubmed.ncbi.nlm.nih.gov/42175721/

[17] Annicchiarico P, Nazzicari N, Pecetti L, et al. White Lupin Genomic Selection for Adaptation to Drought or Moderately Calcareous Soil: A Proof-of-Concept Study. Int J Mol Sci. 2026. https://pubmed.ncbi.nlm.nih.gov/42123632/