2NPLGBM: a genomic model that merges the strengths of classical and machine learning methods in genomic prediction

Genomic prediction (GP) is a central component of modern plant breeding, enabling the early selection of superior genotypes based on genomic marker data. Classical GP models, such as genomic best linear unbiased prediction (GBLUP), operate within the data modeling culture and typically assume additive genetic effects, with extensions required to model non-additive effects such as dominance and epistasis. In contrast, machine learning (ML) models from the algorithmic modeling culture can flexibly model complex, non-additive genetic relationship but often lack direct grounding in quantitative genetic theory and interpretability. To bridge these gaps, we propose 2NPLGBM, a hybrid genomic prediction approach that integrates quantitative genetics with ML. This method introduces a two-matrix (2NP) genotype representation by concatenating additive (Z) and dominance (W) matrices, which are then used as input to a Light Gradient Boosting Machine (LGBM), enabling the simultaneous modeling of additive, dominance, and higher-order genetic interactions (AA, AD, DD).

The 2NPLGBM model was evaluated using six years of test-cross hybrid maize trial data across four agronomic traits (grain yield, plant height, days to silking, and days to anthesis) under five cross-validation schemes simulating temporal: Leave-One-Year-Out (LOYO), Rolling Window (RW), and genetic generalization: Five-Fold, and tester-based schemes (Tester CV0 and Tester CV00). Compared to GBLUP, 2NPLGBM achieved an average of 5% improvement in predictive accuracy under temporal validations and over 15% gains under tester-based schemes, particularly for flowering traits (days to silking and days to anthesis). Performance was generally comparable to LGBM, with both ML models outperforming GBLUP for most traits. Under Tester CV0, 2NPLGBM showed its strongest relative advantage over LGBM for flowering traits, suggesting improved capture of interaction-related genetic signals, whereas LGBM generally performed best for plant height and grain yield. In five-fold CV and Tester CV00, GBLUP remained competitive for some traits, while both machine learning models showed reduced gains, with LGBM slightly outperforming 2NPLGBM. In addition, 2NPLGBM generally improved selection efficiency over GBLUP and, in most cases, LGBM, indicating enhanced ability to capture complex genetic signals relevant for hybrid ranking, particularly for flowering traits, whereas LGBM tended to achieve the highest selection efficiency for plant height and grain yield. Feature interpretation using Shapley Additive exPlanations (SHAP) confirmed that non-additive interactions contributed substantially to prediction accuracy for highly heritable traits. It also revealed trait-specific architectures, additive effects dominated flowering traits, while dominance effects contributed more to plant height and yield. Classical variance component analysis supported these findings, indicating high dominance contributions of 17.3% for yield and 8.2% for plant height.

The 2NPLGBM model integrates quantitative genetic theory with machine-learning, bridging classical statistical (data-model) and algorithmic modeling cultures. bridging classical statistical (data model) and algorithmic modeling cultures. By jointly modeling additive and non-additive effects it can improve predictive accuracy, interpretability, and selection efficiency in test-cross hybrids. Future work should explore multi-trait and multi-environment extensions, integration of environmental covariates, and the inclusion of multi-omic data to further strengthen predictive power and interpretability.