Genomic Selection in Rice Using Targeted Sequencing: Model Evaluation and Training Population Optimization for Breeding Applications

Background/Objectives: Genomic selection (GS) can accelerate crop improvement, but its application in breeding programs is often constrained by the cost of high-density genotyping. This study evaluated genomic prediction using 4647 targeted sequencing-derived SNPs in 829 rice accessions and investigated the effects of prediction model selection, population structure, training population optimization, and GWAS-assisted marker selection on predictive performance. Methods: Seventeen GS models were evaluated for eight agronomic traits using repeated five-fold cross-validation, cross-population prediction, STPGA-based training population optimization, and GWAS-derived marker subsets. Resaults: Predictive performance varied substantially among traits and models. For grain length (GL), EnsembleGS achieved the highest predictive performance among the 17 models (r = 0.838), while RKHS showed the highest performance for grain number per panicle (r = 0.698), RFR for heading date (r = 0.854), and EnsembleGS for length-to-width ratio (r = 0.844). Cross-population prediction generally resulted in lower predictive performance than within-population cross-validation, with substantial variation among transfer directions. STPGA retained substantial predictive performance with reduced training population sizes; for GL, the highest performance reached r = 0.800 with a 400-individual STPGA-selected training population. GWAS-guided marker prioritization also retained substantial predictive information after marker reduction. For GL, the highest performance reached r = 0.789 and 0.802 using the GWAS-Top500 and GWAS-Top1000 subsets, respectively, compared with 0.751 and 0.775 for the corresponding Random500 and Random1000 subsets. For length-to-width ratio, the corresponding values were r = 0.764 and 0.800 for GWAS-Top500 and GWAS-Top1000, compared with 0.727 and 0.783 for the random subsets. Conclusions: Integrating targeted sequencing with trait- and model-aware prediction, optimized training population design, and GWAS-guided marker prioritization can retain substantial genomic prediction performance while reducing marker and training population requirements, providing a potentially cost-effective framework for genomic selection in rice breeding.