Comparative Evaluation of SNP Selection Methods for Stable and Informative Genomic Prediction in Barley Breeding

Genomic selection (GS) enhances plant breeding by predicting complex traits using genome-wide markers, yet optimal single nucleotide polymorphism (SNP) selection strategies remain unclear. In barley, where polygenic traits arise from numerous small-effect loci, selecting informative SNPs is critical for achieving accurate, stable, and biological relevance. Biological relevance refers to selecting markers that are strongly associated with traits and overlap with causal variants, which simulate functionally important loci contributing to trait variation. We compared four SNP selection methods: standard GS with ridge regression best linear unbiased prediction (rrBLUP), GWAS-Assisted GS, stability-informed selection via the GWAS Stability Index (GSI-GS), and minor allele frequency-based selection (MAF-GS). The study used 318 barley accessions genotyped with 3,974 SNPs and simulated a polygenic trait with a heritability of 0.8. Performance was evaluated across 15 independent 80:20 train-test splits, assessing prediction accuracy (R²), stability (coefficient of variation of R², CV_R²), and biological informativeness. GWAS-Assisted GS achieved the highest prediction accuracy (R² = 0.526), closely followed by GSI-GS (R² = 0.514, p = 0.54). However, GSI-GS showed the greatest prediction stability (CV_R² = 0.165) compared to GS (0.185), GWAS-Assisted GS (0.203), and MAF-GS (0.279), and achieved the lowest root mean square error (16.847). GSI-GS also selected SNPs with stronger trait associations and greater overlap with functional loci than MAF-GS (R² = 0.341). Its SNP selection was largely independent of minor allele frequency (r = 0.008, p = 0.27), supporting robust marker performance across different allele classes. These findings establish GSI-GS as an effective strategy for balancing predictive performance, model stability, and biological relevance in GS pipelines for barley.