Predictive tools in corn breeding: from automated phenotyping to genomic selection

Abstract

The use of predictive tools has become a common practice in maize breeding programs, particularly due to the large number of hybrids to be evaluated and the need to support decisionmaking throughout the breeding process. In this context, approaches based on phenomics and genomic selection have expanded the possibilities for phenotypic and genotypic prediction at different stages of breeding. In the first study, a methodology for automated phenotyping was developed and validated to count the number of kernels per ear from digital images, using convolutional neural networks based on the YOLOv8 algorithm. The method showed high agreement with manual counting, with an intraclass correlation coefficient (ICC) of 0.89 and a strong linear association, with a Pearson correlation coefficient of 0.89, in addition to providing an approximate 63% reduction in evaluation time. In the second study, the efficiency of genomic selection was evaluated for predicting the performance of single-cross hybrids not evaluated in the field for kernel number per ear, considering scenarios of partial parental genotyping. Models based exclusively on pedigree information (A matrix) and models integrating pedigree and genomic information (H matrix) were compared under different crossvalidation schemes. Models based on the H matrix showed higher mean predictive ability (0.51) and greater consistency in hybrid ranking compared with pedigree-based models (0.46), leading to changes in the selected set and increases in expected genetic gain. Overall, the results demonstrate that the integrated application of predictive tools, including automated phenotyping and genomic prediction, contributes to a more efficient and robust hybrid selection process in maize, reinforcing their potential as decision-support tools in breeding programs.