Avaliação e seleção de porta-enxertos de videira (Vitis spp.) tolerantes ao déficit hídrico através de aprendizagem de máquina

Abstract

Plant breeding programs seek to select superior genotypes due to human needs, like productivity increase, stability, and quality of economically important species, besides reducing environmental impacts and production costs. Thus, the grapevine (Vitis spp.) is considered an economic, social, and food fruit tree important. However, climate change and the scarcity of water resources have provoked a growing investment in the development and use of drought-tolerant rootstocks. However, new drought-tolerant cultivar obtention is a long time-consuming, and complicated process as it's a polygenic characteristic. Because of this, use new tools like machine learning algorithms can leading to the identification and selection of new drought-tolerant cultivars due to the ability to manage large amounts of data and identify relevant patterns. In this sense, this work aimed to assess the genetic divergence of 45 grapevine rootstocks cultivars and predict the drought tolerance classes of the three Brazilian cultivars IAC313, IAC572, and IAC766, whose information is still unknown in the literature, through machine learning algorithms. For genetic divergence analysis, K-means and Principal Components Analysis algorithms were applied. The results obtained indicated five heterotic groups and 37 crossover options viable in the genetic divergence analysis for the evaluated traits, indicating divergence between the cultivars. The performance of six different algorithms was compared, such as Decision Tree, Random Forest, K-Nearest Neighbors, XGBoost, Support Vector Machines, and Linear Discriminant Analysis to predict drought tolerance classes. And the best performing algorithm was used to predict the degree of drought tolerance of the three Brazilian cultivars. The results indicated Random Forest as the best model, which predicted high drought tolerance for IAC 313 and IAC 766 and a low tolerance for IAC 572. In this sense, it was possible to achieve easy-to-understand results with the machine learning algorithms in our study, showing itself as another option helpful and accessible tool to breeders for identifying better crosses for a specific characteristic and predicting classes.