Beyond Predictive Ability: Multidimensional Evaluation of Single and Multi-Trait Genomic Selection Models in Oat for Enhanced Breeding Efficiency

Genomic selection (GS) can accelerate genetic gain through early marker-based selection, yet model evaluation in many studies remains centered on predictive ability (PA), which may inadequately reflect operational breeding value. Here, we evaluated single-trait (STM) and multi-trait (MTM) genomic prediction in a public oat breeding program using a multidimensional framework integrating predictive performance, selection efficiency, stability, and feasibility. A total of 1,941 oat breeding lines were phenotyped for five key traits (plant height, test weight, grain yield, plump kernels, and thin kernels) over five years (2018–2022) and genotyped using 5,126 genome-wide SNPs. Spatially adjusted phenotypes were used to derive BLUPs for genomic prediction with GBLUP (STM) and Bayesian MTM models. Model performance was assessed across four validation strategies: forward prediction, backward temporal prediction, historical-to-future prediction, and within-year cross-validation (RandomCV and CDmean). PA was highly trait-dependent, with consistently higher performance for plant height and thin kernels than for yield and plump kernels. In cross-validation, CDmean provided marginal PA improvements but required substantially greater computation and showed significantly lower rank correlation with phenotypic rankings, establishing RandomCV as superior for operational implementation. STM and MTM models showed comparable, trait-dependent performance. Importantly, multi-trait selection frameworks achieved hit rates exceeding 50%, demonstrating that moderate PA can deliver high breeding value when aligned with operational selection criteria. These results demonstrate that GS evaluation should complement PA with selection-aligned metrics, and we provide an implementation-focused framework to guide routine GS deployment in oat and similarly structured breeding programs.