Uso de aprendizado de máquina para otimização da seleção recorrente recíproca em milho

Abstract

Reciprocal Recurrent Selection (RRS) has been widely used in maize breeding to maximize genetic potential and heterosis between populations. However, its effectiveness is intrinsically linked to the precise analysis of maize traits and progenies. In this context, machine learning-based methodologies and advanced statistical models can be jointly employed to optimize selection and predict performance in RRS programs. The main objective of this study was to optimize the selection process of maize progenies in the Federal University of Lavras RRS program using machine learning techniques. The experiments were conducted at the Federal University of Lavras between 2021 and 2024 using an Alpha Lattice design. The study was structured into four main stages: i) Analysis of the genetic potential of progenies through strategies based on the integration of the Best Subset Regression methodology to optimize the simultaneous selection of traits using selection indices. ii) Recombination of the top 15% of selected progenies followed by the generation of four intrapopulation and interpopulation crosses using the Bulk Pollen Pollination method. These populations were evaluated and subjected to genetic and factorial parameter analyses based on phenotypic and genetic correlations, aiming for the simultaneous selection of progenies with optimal phenotypic and genetic expression of their traits, optimized through K-means clustering. iii) Prediction of progeny yield using multilevel models in the context of nested breeding progenies within populations and populations within crosses. iv) Integrating high-throughput phenotyping with interpretable tree-based machine learning models and simultaneous selection indices. In conclusion, the integration of Best Subset Regression and simultaneous selection indices proved to be an effective approach for identifying promising lineages and enhancing the genetic progress of maize populations in RRS. Factorial selection strategies in populations demonstrated efficacy in identifying progenies with simultaneous phenotypic and genetic performance, contributing to the development of cultivars with greater agronomic and genetic stability. The multilevel model stood out as a tool with high predictive power, supporting the accurate selection of superior progenies. The integration of machine learning models with simultaneous selection indices maximized selection response by choosing maize progenies with optimal genetic gains. This study validates machine learning and its integration with statistical models as an innovative and effective strategy for optimizing RRS in maize, accelerating genetic breeding progress, and contributing to the development of high-performance hybrids.