Heterotic pools and realistic cross-validation reveal the usefulness of genomic selection in tropical maize haploid inducer populations

Abstract

The development of doubled haploid (DH) lines through in vivo haploid induction is a powerful tool to accelerate maize breeding. However, the process is limited by the low heritability and phenotyping complexity of haploid induction traits. This study evaluated the potential of genomic selection (GS) for improving putative (HIRp) and real (HIRr) haploid induction rates in a tropical maize population subjected to three recurrent selection cycles (C0, C1, C2), using 190 inducer progenies and testers from Flint and Dent heterotic groups. Phenotypic data were analyzed with linear mixed models, and genomic data were obtained through genotyping-by-sequencing (GBS). Heritability estimates were low (0.16 for HIRp, 0.47 for HIRr), with tester correlations of 0.11 and 0.56, highlighting the challenges of phenotypic selection. Genomic predictions with GBLUP were assessed under k-fold and leave-one-cycle-out (LOCO) cross-validation. LOCO consistently outperformed k-fold, reaching predictive ability (PA) of up to 0.79 and relative efficiency (RE) of 426 for HIRr. A key finding was the strong heterotic pool effect: although Flint sometimes presented higher PA and RE numerically, R1-nj marker suppression was substantially higher (46.8%) compared with Dent (0.7%). This indicates that Dent germplasm provides more reliable phenotypic inputs and therefore constitutes a more consistent foundation for GS. These results confirm that GS can significantly improve haploid induction prediction in tropical maize, particularly when validation mimics inter-cycle prediction and heterotic group differences are accounted for. The study highlights the importance of precise phenotyping, group-specific modeling, and temporal validation frameworks, reinforcing GS as a strategic tool to enhance efficiency and sustainability in DH breeding programs.