Multivariate Genomic Best Linear Unbiased Prediction (GBLUP) Can Improve Genomic Selection in Mungbean

The application of genomic selection (GS) in mungbean has been limited. This study evaluated GS for mungbean yield using genomic best linear unbiased prediction (GBLUP), including multi-trait GBLUP to maximize prediction accuracy. With yield (YLD), days-to-flowering (DTF), and plant height (PHT) data collected from 313 diverse mungbean germplasm panel tested across four environments and genotyped for 3,310 SNPs, we compared multivariate GBLUP models including single-trait−multi-environment (ST-ME), multi-trait−single-environment (MT-SE), and multi-trait−multi-environment (MT-ME) against univariate single-trait−single-environment (ST-SE) models. For MT models, YLD was the primary trait, while DTF and PHT were the indicator traits. Indicator traits are correlated traits which are easy and relatively inexpensive to measure early in crop development. Cross-validation schemes varied depending on whether the validation set had data for indicator traits (IND) or not (NOIND ). Next, we evaluated two prediction scenarios. In Scenario 1, validation sets were completely untested, so performance predictions relied solely on training data. In Scenario 2, validation sets were tested in non-target environments, allowing predictions to incorporate both training data and cross-environment performance of the validation set. ST-SE models prediction accuracies ranged from 0.33±0.02 to 0.61±0.01 for YLD, 0.39±0.02 to 0.49±0.02 for DTF, and 0.38±0.01 to 0.59±0.01 for PHT. Multi-trait models showed no improvement under NOIND-Scenario 1 but provided up to +0.26 improvement in predicting YLD under IND schemes. The moderate to high prediction accuracies suggest promising GS application in mungbean. Benefits are maximized using multi-trait approaches when strong genetic correlations exist between primary and indicator traits or across environments in multi-environment trials.