Predicting Maize Hybrid Performance with Machine Learning and a Locus-Specific Degree of Dominance Transformation

The genetic architecture of a trait plays a vital role in the predictive ability of genomic models. While classical models dominate plant breeding, ML methods are gaining traction for their superior handling of non-linear effects. This study assessed ML and classical models incorporating a novel locus-specific weighted dominance effect transformation matrix for genomic prediction in hybrid maize. Five models were compared: (1) XGBoost combined with locus-specific weighted dominance, (2) XGBoost only, (3) GBLUP model with additive, (4) GBLUP model including additive and dominance, (5) GBLUP combined with locus-specific weighted dominance. The models were evaluated using two simulated maize populations (polygenic scenario and oligogenic scenario) and a real maize hybrid population. Aiming to encompass a range of diverse genetic architectures, the assessment of the real maize population focused on their performance across several traits, including grain yield, test weight, ear height, plant height, Pollen DAP, Silk DAP, and grain moisture. The results show that the transformation with dominance effects did not improve ML methods, and while ML models performed better than the GBLUP model with additive effects only and the GBLUP model including additive and dominance effects, the GBLUP Transformed model outperformed all others, including ML models across the three levels of heritability in the polygenic scenarios, demonstrating the superiority of genomic prediction with transformed GBLUP for polygenic traits. This cannot be said for oligogenic scenarios, as ML models outperformed all other models in oligogenic scenarios. Also, we observed that the performance of genomic models generally decreases as dominance increases. Our findings contribute to optimizing breeding strategies and advancing genomic prediction methodologies.