Uso de covariáveis ambientais para predição da interação de clones de cana-de-açúcar com ambientes

Abstract

The evaluation of genotypes in multi-environment trials that represent the diversity of cultivation conditions is one of the main challenges of plant breeding programs. In perennial or semi-perennial species, such as sugarcane, this challenge is amplified by prolonged experimentation over several harvest seasons. The use of environmental covariables has proven to be a promising approach to increase the efficiency of genotype selection. This study aimed to evaluate different models for predicting the performance of sugarcane clones in distinct environments using environmental covariables. Phenotypic data for yield (tons of stalks per hectare) were obtained from 53 trials conducted by the Sugarcane Breeding Program (PMGCA/UFG/Ridesa) between 2012 and 2022. Envirotyping data for 32 environmental covariates were collected from NASA's POWER platform. Six Bayesian ridge regression models were developed, integrating different combinations of genotypic effects (G), trial effects (E), environmental covariates (V), and their interactions. To evaluate the models, two cross-validation schemes were used, 5-fold and leave-one-out, representing the following scenarios, respectively: (1) the ability of the models to predict genotype performance in trials where they were not evaluated; (2) the ability of models with covariates to predict genotype performance in untested environments. Predictive ability was assessed globally and by individual trial. In scenario 1, global predictive abilities ranged from 0.93 to 0.13, while predictive ability by trial ranged from 0.50 to 0.16. In scenario 2, global predictive abilities ranged from 0.47 to 0.01, and by trial from 0.58 to 0.21. Models without environmental covariates, G + E and G + E + G × E, proved suitable for filling information gaps in breeding programs. The G + V model, although it did not consider genotype × environmental covariables interactions, stood out as the most promising option for cultivar recommendations in locations without prior data. The results highlight differences in the performance of the tested models and the importance of considering environmental variability in the interpretation of predictions. Identifying trials with low predictive ability is essential to guide data collection strategies and improve models capable of effectively handling environmental complexity.