Enhancing Wheat Genomic Prediction by a Hybrid Kernel Approach

Jaime Cuevas¹Jose Crossa^2,3Abelardo Montesinos-López⁴Johannes W R Martini⁵Guillermo Sebastián Gerard²Jaime Ortegón¹Susanne Dreisigacker²Velu Govindan²Paulino Pérez-Rodríguez³Carolina Saint Pierre²Leonardo Abdiel Crespo Herrera²OSVAL A. MONTESINOS-LOPEZ^6*Paolo Vitale^2*

• ¹División de Ciencias, Ingeniería y Tecnologías (DCIT), Universidad Autónoma del Estado de Quintana Roo, Chetumal, Quintana Roo, México, Chetumal, Mexico
• ²International Maize and Wheat Improvement Center (Mexico), Texcoco, Mexico
• ³Colegio de Postgraduados (COLPOS), Montecillo, Mexico
• ⁴Departamento de Farmacobiología, Centro Universitario de Ciencias Exactas e Ingenierias, Universidad de Guadalajara, Guadalajara, Jalisco, Mexico
• ⁵Aardevo B.V., Nagele, The Netherlands, Aardevo, Netherlands
• ⁶Facultad de Telemática, Universidad de Colima, Colima, Colima, Mexico

This study integrates genomic and pedigree data by leveraging advanced modeling techniques, aiming to enhance the predictive performance of genomic selection models by capturing complex genetic relationships through the interaction of both matrices and exploring the utility of non-linear methods, such as kernel matrices. Our goal was to improve genomic prediction accuracy by combining the pedigree-based or genetic similarity matrix (A) with the genomic similarity matrix (G). Using various wheat datasets, we performed five single-environment models and five multi-environment models that incorporated genotype-by-environment (G × E) interactions. The proposed models S5 and M5 significantly enhanced prediction accuracy by incorporating two novel symmetric kernels, C and P, derived from the interaction of genomic and pedigree matrices. These hybrid kernels captured additional, independent genetic variation not explained by conventional matrices. The proposed prediction model outperformed the standard conventional models in most single-environment and multi-environment models. The genomic models with non-linear kernels were better predictors than the linear prediction models.