Abstract
INTRODUCTION: Ridge regression BLUP (rrBLUP) is a widely used model for genomic selection. Different genomic prediction (GP) models have their own niches depending on the genetic architecture of traits and computational complexity. Haploid inducers have unique trait performances, relevant for doubled haploid (DH) technology in maize (Zea mays L.). METHODS: We evaluated the performance of single-trait (ST) and multi-trait (MT) GP models, which include rrBLUP, BayesB, Random Forest, and xGBoost, using data from multifamily DH inducers (DHIs). We integrated multi-trait and de novo genome-wide association studies (GWAS) within the rrBLUP framework to model four target traits: days to flowering (DTF), haploid induction rate (HIR), plant height (PHT), and primary branch length (PBL). Predictive ability (PA) was assessed through five-fold cross-validation and further validated in multi-parent advanced generation intercross (MAGIC) DHIs. RESULTS: The average PAs of different GP methods across traits were 0.51 to 0.69. ST/MT de novo GWAS rrBLUP methods increased PA of HIR. In addition, MT GP models improved PA by 12% on average across traits relative to ST GP models in MAGIC DHIs. DISCUSSION: These findings highlight the potential benefits of integrating multi-trait modeling or de novo GWAS into the rrBLUP framework. Such GP approaches in this study enhance PAs and provide empirical evidence for accelerating the genetic improvement of maize haploid inducers.