Trait-Informed Machine Learning Improves Precise Genomic Selection and Hybrid Design in Wheat.

Machine learning can unlock genomic selection (GS) for polygenic traits, but only if millions of SNPs are distilled into breeder-actionable loci. Here, we presented genome-wide association signal ranking with linkage disequilibrium pruning and Random Forest feature importance selection (GLRS), a trait-informed framework for deriving ultra-sparse marker sets for pure-line and hybrid GS breeding in wheat. Training on 768 diverse accessions phenotyped for nine yield-related traits across six environments and 570,441 SNPs, GLRS compressed markers for each trait to a few hundred yet raised cross-validation accuracy by 7.2% relative to conventional GS using random selected markers. Independent evaluation using a publicly available wheat dataset further supported the robustness and general applicability of the GLRS strategy. Validation in 1,163 advanced breeding lines from seven independent trials further demonstrated the robustness of GLRS, with prediction accuracy improvement of 14.4% and 5.8% over random markers and genome-wide association signal ranking with linkage disequilibrium pruning markers, respectively. GLRS also delivered accurate predictions of F1 performance (r = 0.790), general combining ability (r = 0.851) and heterosis (r = 0.541) in 292 hybrid combinations, enabling simultaneous hybrid wheat and elite inbred variety development. Finally, we proposed a combining-ability-based GS strategy for wheat pure-line and hybrid breeding. This study provided an efficient GLRS pipeline to facilitate the practical application of trait-informed GS in wheat high-yield breeding.