Genomic Prediction of Sweet Sorghum Agronomic Performance under Drought and Irrigated Environments in Haiti

Résumé

High-throughput phenotyping remains costly and inaccessible to most plant breeding programs. Over the last decade, genomic selection (GS) has gained momentum as a tool for predicting genetic gain in plant breeding populations, while lowering costs associated with phenotyping. Different statistical models and approaches have been developed to implement GS in plant breeding, and strategies that promote accurate and resource-efficient prediction are of increasing interest. Since its establishment in 2010, the sweet sorghum breeding program at CHIBAS, Haiti, has led efforts to develop and release cultivars resilient to abiotic and biotic stress. Among abiotic constraints, drought stress is the most limiting since growers depend on erratic rainfall for sorghum production in Haiti. The goal of the present study was to predict the genomic estimated breeding values of a sweet sorghum breeding population (n=250) under contrasting environments in Haiti using four statistical models (Bayes A, B, C and Bayesian Ridge Regression (BRR)). We evaluated twelve sorghum traits and performed within and across irrigated, pre-flowering and vegetative water stress prediction scenarios. Overall, the four methods showed similar results, however Bayes B and BRR were superior in prediction accuracy and computation time, respectively. Generally, prediction accuracy was higher for within-environment (0.31 to 0.7) than across-environment (0.06 to 0.7) involving vegetative water stress scenarios. Prediction accuracy varied substantially for all traits, with total green leaf showing the highest mean value (0.70), and grain yield showing the least (0.49). Overall, there was no improvement in the prediction accuracy of grain yield with multi-traits genomic selection.