Controlling population structure in the genomic prediction of tropical maize hybrids

In tropical maize breeding programs where more than two heterotic groups are crossed, factors such as population structure (PS) can influence the achievement of reliable estimates of genomic breeding values (GEBVs) for complex traits. Hence, our objectives were (i) to investigate PS in a set of tropical maize inbreds and their derived hybrids, and (ii) to control PS in genomic predictions of single-crosses considering two scenarios: applying (1) the traditional GBLUP (GB) and four adjustment methods of PS in the whole group, and (2) homogeneous- (A-GB), within- (W-GB), multi- (MG-GB), and across-group (AC-GB) analysis in stratified groups. Three subpopulations were identified in the inbred lines and hybrids based on fineSTRUCTURE results. Adding four different sets of PS as covariates to the prediction model did not improve the predictive ability (r). However, using non-metric multidimensional scaling and fineSTRUCTURE group clustering increased the reliability of GEBV estimation for grain yield and plant height, respectively. The W-GB analysis in the stratified groups resulted in low r, mostly due to the reduction of training size. On the other hand, A-GB and MG-GB showed similar r for both traits. However, MG-GB presented higher broad sense genomic heritabilities compared to A-GB, efficiently controlling heterogeneity of marker effects between subpopulations. The r of the AC-GB method was low when predicting groups genetically distant. We conclude that predicting hybrid phenotypes by using PS covariates and multi-group analysis in stratified clusters may be an efficient method, increasing reliability and predictive ability, respectively.