Impact of adding filtered whole-genome sequencing variants on genomic prediction accuracy in Chinese Holstein cattle

Genomic prediction based on medium-density SNP chips is widely used in dairy cattle breeding; however, causal variants may be missing or insufficiently tagged, potentially limiting prediction accuracy. The increasing availability of imputed whole-genome sequence (WGS) -level variants provide an opportunity to prioritize informative variants, yet the optimal strategy for integrating GWAS-selected variants into routine chip-based prediction remains unclear. Based on 100K SNP chip data, whole-genome sequence-level variants were obtained through imputation, and significant SNPs were filtered using the mixed linear model (MLM) and Fixed and random model Circulating Probability Unification (FarmCPU) at different feature selection thresholds. These SNPs were then merged with the chip data to construct models under different genetic architectures for comparison. The results suggested that genomic prediction accuracy was significantly superior to that of pedigree-based best linear unbiased prediction (PBLUP), and the Bayesian four-distribution mixture model generally outperformed genomic best linear unbiased prediction (GBLUP). Although FarmCPU detected a greater number of SNPs than MLM, the SNPs selected by MLM provided more stable improvements in prediction performance. Under feature selection thresholds (0.001 and 0.0001), incorporating MLM-selected SNPs significantly enhanced the prediction accuracy for milk fat rate (MFR) and milk fat yield (MFY) ( p < 0.05). Furthermore, the two-component GBLUP architecture better utilized the selected SNPs to improve prediction performance, while the single-component Bayesian model already showed strong adaptability. Stringent SNP selection via MLM, combined with either a two-component GBLUP or a single-component Bayesian model, can effectively enhance genomic prediction accuracy, offering a feasible strategy for optimizing genomic prediction in Chinese Holstein cattle. The selection of informative variants from imputed WGS-level data based on genome-wide association studies (GWAS) represents an efficient strategy to utilize genomic information not captured by conventional SNP arrays. The optimal methodology involves employing the MLM for SNP screening under feature selection thresholds (e.g., 0.001 and 0.0001), in combination with either a two-component GBLUP framework or a single-component Bayesian mixture model. This integrated approach consistently enables superior predictive performance.