Early-life stage phenomic prediction of field agronomic traits across breeding cycles in intermediate wheatgrass

Perennial grains hold promise for sustainable agriculture, yet breeding advancements are hampered by limited genotyping and the complexity of evaluating intricate traits over multiple years in various environments. Phenomic selection could alleviate these issues by utilizing low-cost, scalable, high-dimensional phenotypes gathered at an early developmental stage.

However, the reliability of such predictions across successive breeding cycles remains unclear. In this study, researchers examined the performance of genomic selection and phenomic selection across two breeding cycles of Thinopyrum intermedium, commonly known as intermediate wheatgrass or Kernza. The breeding cycle encompassed roughly 2,280 individuals from half-sib families, assessed across numerous field locations and years.

Relationship matrices were built using genomic markers and early-life stage phenomic data, encompassing seed and leaf color (HSV), CropReporter multispectral reflectance and indices, and cycle-specific hyperspectral reflectance sensors. Genomic models consistently yielded the strongest predictions for the majority of field traits in both breeding cycles.

Among the phenomic predictors, leaf HSV emerged as the most informative, whereas CropReporter and hyperspectral data demonstrated lower and trait-dependent performance, with seed HSV offering minimal predictive value. Genomic, leaf HSV, and CropReporter models retained their predictive abilities across breeding cycles with minimal loss of accuracy compared to within-cycle validation, indicating that their predictive signals were not confined to a single breeding cycle.

Early-life stage leaf HSV was identified as a practical and accessible tool for germplasm thinning and early-stage selection in perennial breeding programs. Notably, despite the limited similarity among relationship matrices, multi-relationship-matrix models rarely surpassed the accuracy of the superior individual single-relationship-matrix model.

Overall, these findings suggest that early-life stage phenomic data offer dependable insights into agronomic performance manifested years later, but that increased predictor complexity and data integration do not necessarily translate to enhanced prediction.