Genomic Selection for Milk Yield and Milk Composition Traits in Dairy Goats Using Machine Learning and Prior-Information Models

Genomic selection (GS) provides an effective approach to accelerating genetic gain in dairy goats, but the prediction performance is strongly influenced by the statistical model, marker density, phenotype adjustment strategy, and biological architecture of the target trait. In this study, dairy goat populations comprising Xinong Saanen and Saanen dairy goats from major production regions in China were used to evaluate genomic prediction for milk yield (MY), milk fat percentage (MFP), and milk protein percentage (MPP). Genotypes from 1034 dairy goats were generated using low-coverage whole-genome sequencing (lcWGS), imputed to improve genotype completeness and accuracy; a high-quality chip-based dataset was also constructed from previously developed 25K single-nucleotide polymorphism (SNP) chip loci. Conventional genomic best linear unbiased prediction (GBLUP) models, Bayesian regression models, and machine learning algorithms were compared using 10-fold cross-validation. Bayesian models showed clear trait-specific advantages, with BayesB improving MFP prediction by approximately 12.9% relative to GBLUP under the 25K chip-based strategy. Among machine learning methods, gradient boosting models performed strongly; extreme gradient boosting (XGBoost) improved the prediction accuracy for MY, MFP, and MPP by 14.3%, 17.9%, and 18.5%, respectively, relative to GBLUP under the chip-based strategy. Incorporating genome-wide association study (GWAS)-derived prior information and selection signature priors further improved the prediction accuracy, particularly for milk composition traits. Overall, the results indicate that genomic prediction in dairy goats can be optimized by matching models, genotyping platforms, and prior biological information to the genetic characteristics of the target trait.