Artificial Intelligence Meets Genomic Selection: Comparing Deep Learning and GBLUP Across Diverse Plant Datasets

Abelardo Montesinos-López¹OSVAL A. MONTESINOS-LOPEZ^2*Sofia Ramos-Pulido¹Alejandro Brandon³Edgar Alejandro Guerrero-Arroyo¹Jose Crossa^4,5Rodomiro Ortiz^6*

• ¹Departamento de Matemáticas, Centro Universitario de Ciencias Exactas e Ingenierías (CUCEI), Universidad de Guadalajara, Guadalajara, Mexico
• ²Facultad de Telemática, Universidad de Colima, Colima, Colima, Mexico
• ³Institut National des Sciences Appliquées de Lyon, Lyon, France
• ⁴International Maize and Wheat Improvement Center (Mexico), Texcoco, Tabasco, Mexico
• ⁵Colegio de Postgraduados (COLPOS), Montecillo, Mexico
• ⁶Swedish University of Agricultural Sciences, Uppsala, Sweden

The practical implementation of genomic selection (GS) in plant breeding experiments is challenging due to numerous factors that affect its reliability. To address these challenges, deep learning (DL) models have been adopted as a powerful framework to improve the accuracy of GS. To bolster the empirical evidence supporting DL models, we conducted a comparative study using 14 datasets from real plant breeding programs, comparing the genomic prediction performance of DL models with the best genomic best linear unbiased predictor (GBLUP) model. Our findings indicate that DL models are highly competitive and excel at capturing complex, non-linear patterns, even when working with small datasets like those used in this study. Nevertheless, the successful implementation of DL models requires meticulous parameter tuning to achieve optimal results. This comprehensive comparison highlights the complementary strengths of both DL and GBLUP models. It also underscores the importance of selecting models based on specific traits and prioritized metrics in predictive analyses, as neither DL nor GBLUP universally outperformed the other across all scenarios.