A meta-learning with adversarial domain adaptation method for few-shot cross-region maize height and yield joint prediction

High accuracy phenotypic prediction driven by genotype (G), environment (E) and their interactions is the foundation for precision crop breeding. However, under cross-region few-shot scenarios with domain shift, prediction accuracy and model generalization are often declined. To address this challenge, a novel phenotype prediction framework, namely Meta-learning with Adversarial domain adaptation for G × E Prediction (MAGEP), is established. The framework first introduces a GEfused module with a dual-branch architecture to extract fused genotypic and environmental features. It then employs Reptile-based meta-learning to learn generalizable features across environments. A gradient reversal layer is incorporated to perform adversarial training for learning domain-invariant features. Meanwhile, Maximum Mean Discrepancy (MMD) or Correlation Alignment (CORAL) domain alignment strategies are applied to quantify and reduce the feature distribution discrepancy between source and target domains. Using a publicly available maize G × E dataset, the MAGEP was compared with eleven baseline models in four cross-domain prediction scenarios. The results show that MAGEP outperforms all baselines in both accuracy and robustness. Notably, MAGEP_CORAL achieves the best performance, reaching average Pearson correlation coefficients (PCC) of 0.75 for plant height and 0.45 for grain yield, improvements of 44.2% and 55.2% over MAGEP_MMD, respectively. Experiments with varying target-domain sample sizes further confirm the superior adaptability of the CORAL strategy. Furthermore, cross-year and cross-species experiments confirm that MAGEP achieved lower RMSE and MAE than baseline models. Finally, post-hoc interpretability analysis was performed on MAGEP using gradient-based attribution methods. The findings reveal that plant height and grain yield are regulated by multiple genetic loci, and the model can identify key environmental regulators for each trial locations. This supports the biological plausibility of MAGEP model and provide a basis for analyzing the genetic and environmental co-regulation of crop phenotypes.