Prediction-based breeding: Modern tools to optimize and reshape programs

Prediction-based breeding reshapes plant genetic improvement by prioritizing the predictive ability of models over causal interpretation. This review examines recent advances in the use of tools such as genomic selection, high-throughput phenotyping, multi-omics integration, and enviromics to enhance genetic gain and improve the efficiency of breeding programs. Predictive models, while powerful, rely on validation within the genetic and environmental domains represented in the training set, with evident risks when extrapolated to unrelated scenarios. Traditional approaches such as marker-assisted selection and genome-wide association study remain limited for quantitative traits, reinforcing the need for prediction-oriented models. Moreover, the expansion of omics data sources, although capturing greater biological complexity, must be accompanied by rigorous validation practices to avoid fragile interpretations. Stochastic simulations are a strategic tool for testing selection schemes, optimizing training populations, anticipating overfitting risks, reducing costs, and guiding decisions based on prospective scenarios. This review also highlights the importance of ensuring independence between calibration and prediction, focusing on practical accuracy evaluation, and prioritizing operational utility over mechanistic explanation. In summary, prediction-based breeding is a core strategy for modernizing breeding programs, connecting computational tools, high-dimensional data, and pragmatic decision-making to deliver consistent results.