A Stacked Ensemble Learning Framework to Improve Predictive Performance in Genomic Selection for Quantitative Traits

Genomic Selection (GS) utilizes genome-wide markers to estimate Genetic Estimated Breeding Values (GEBV), optimizing germplasm selection for costly or late-expressed traits. Traditional models, like GBLUP, assume a polygenic architecture, yet the underlying genetic complexity of many traits includes non-additive effects. Since no single model is universally optimal for diverse genetic architecture, stacking ensemble learning, which trains a meta-model using predictions from diverse base models, presents an interesting approach to capture this complexity. This study evaluated various stacking configurations for genomic prediction across 8 simulated traits, encompassing additive, dominance, and epistatic architectures. Using a 5-fold cross-validation scheme, the stacking ensemble demonstrated superior predictive ability in all scenarios. Gains were highly significant, achieving an 83% increment over the best individual model (BayesA with dominance) in complex architectures (100 QTLs, h²=0.3), and a 27.59% gain in oligogenic scenarios with epistasis (10 QTLs, h²=0.6). The success of the stacking strategy can be attributed to careful base learner selection and the use of robust meta-learners (such as penalized regression) to deal input multicollinearity effectively.