Next Generation Mining of Genebanks: Harnessing Sorghum ( Sorghum bicolor (L.) Moench) Germplasm in the Uganda National Genebank Through Genomic Prediction

ABSTRACT Germplasm in genebanks is vital for crop improvement, yet it often lacks information on key traits of interest to users. Genomic prediction offers great potential to use genomic estimated breeding values as a proxy for phenotypic records. This minimizes the expensive and laborious large‐scale evaluation of large genebank collections. In this study, we assessed the use of genomic prediction to exploit the genetic diversity of 3333 sorghum accessions conserved by the Uganda National Genebank. A core collection was defined using GenoCore software and evaluated as a training population for genomic prediction for flag leaf length and width, plant height, peduncle length and days to flowering. We investigated how training population size (10%–90%), genetic composition and model choice (rrBLUP, BayesA, BayesB and Bayes LASSO) influenced prediction accuracy. All models performed similarly in the core collection, with cross‐validation accuracies ranging from 0.24 to 0.83. The core collection achieved higher accuracy (0.26–0.86) than a randomly assembled training set (0.27–0.79). Increasing training size improved prediction accuracy for most traits investigated, though the flag leaf traits and plant height plateaued at 80% accuracy. Days to flowering and plant height had the highest prediction accuracies in all cross‐validation strategies. These findings demonstrated the utility of genomic prediction and optimized composite core collection sets to guide allele mining and prebreeding applications, enhancing the value of large genebank collections.