Accelerating Sugarcane Breeding with Machine Learning: A Concurrent Multi-trait Predictive Approach

A large and highly polyploid genome, long breeding cycles, low narrow-sense heritability of yield and its components, and persisting yield stagnation continue to challenge sugarcane variety improvement through both conventional and molecular breedings. To address this, this study developed an efficient phenotype-based concurrent multi-trait prediction framework using artificial intelligence, leveraging 2700 parent combinations and their offspring field performance data from the Sugarcane Research Institute of Guangxi Academy of Agricultural Sciences. Five machine learning paradigms were selected based on data characteristics: tree-based models (GBDT, XGBoost) for structured tabular data and nonlinear trait interactions; attention-based tabular models (TabTransformer, TabNet) for subtle cross-trait dependencies; and an MLP as a baseline. All models were optimized via 80/20 stratified train–test split, class-weighted loss, and cross-validated grid search. GBDT achieved the best overall performance with a macro F1 score of 0.7316, Hamming loss of 0.2052, and sample accuracy of 0.5630, excelling in predicting sucrose content, cane yield, and smut resistance. XGBoost performed comparably, while TabTransformer offered complementary advantages with a higher F1 score (0.6829) on overall agronomic performance. These results indicate the potential of using artificial intelligence (AI), particularly machine learning (ML), to increase the predictive power and robustness through integrated AI approaches in the future. This phenotype-based framework shows the value of ML as an efficient decision support tool that can help breeders prioritize cross combinations and optimize resource use, thereby reducing reliance on extensive field trials. While the potential of AI in accelerating variety development is evident, its routine application in real-world commercial breeding requires validation.