Dynamic Weighting of GWAS Signals Enhances Genomic Selection Accuracy: Validation Across Trait Architectures in Wheat

Genomic selection (GS) has advanced plant breeding by predicting complex traits using genome-wide markers. Integrating genome-wide association study (GWAS) results in GS can further improve accuracy, but most methods rely on static marker selection that does not adapt well to diverse trait architectures. In addition, reported gains are often inflated due to flawed validation, where data leakage, or the unintended sharing of information between training and testing sets, leads to overestimated prediction accuracy. This study addresses these gaps by introducing GWAS_GS2, a novel method that dynamically weights genome-wide markers based on GWAS-derived effect sizes and p-values, enhancing GS models while preserving genome-wide variation. We compared GWAS_GS2 with standard GS (genomic best linear unbiased prediction), marker-assisted selection, and GWAS_GS1 (fixed GWAS effects), using a wheat dataset of 136 lines and 32,412 single nucleotide polymorphisms. Under strict validation, which involved a three-way data split with 50 lines for GWAS, 60 for training, and 26 for testing, GWAS_GS1 performed best for Mendelian traits (R² = 0.1920), while GWAS_GS2 outperformed other methods for polygenic traits (R² = 0.1365). In addition, 5-fold cross-validation overestimated R² for GWAS-informed methods by up to 57.6%. These findings confirm the importance of rigorous validation to avoid bias and highlight GWAS_GS2 as a scalable, unbiased approach to improve selection for complex traits in wheat breeding.