The researchers, led by Yongxia Yang and colleagues, developed a framework named MAGEP, which stands for meta-learning with adversarial domain adaptation. This approach addresses two significant challenges in cross-region phenotypic prediction: the scarcity of labeled data in new environments and the distribution shift between source and target domains. By integrating meta-learning with adversarial domain adaptation, MAGEP allows the model to learn generalizable strategies that can be rapidly adapted to new tasks, while simultaneously aligning feature distributions to mitigate the impact of environmental heterogeneity. This dual strategy is particularly relevant for precision breeding, where high-throughput phenotyping is costly and data from new locations is often sparse.

The study utilized a comprehensive dataset from the Genomes to Fields (G2F) initiative, specifically the 2018 data for the PHW65 reference design population. The maize samples were genotyped using the Practical Haplotype Graph approach, resulting in a high-dimensional matrix of single nucleotide polymorphisms. Field experiments were conducted across ten locations in North America, including sites in Iowa, Arkansas, Georgia, and Texas. Environmental data, including air temperature, relative humidity, solar radiation, and rainfall, were recorded at 30-minute intervals throughout the growing season. This detailed environmental context allowed the researchers to model the complex interactions between genotype and environment, known as G x E interactions, which are crucial for understanding phenotypic plasticity.

At the core of the MAGEP framework is the GEfused module, a dual-channel structure designed to extract and fuse genotypic and environmental features. The genotypic branch employs multi-scale dilated convolutions to capture SNP interaction effects at different scales, while the environmental branch uses a Transformer encoder to model temporal dependencies in meteorological data. By projecting these features into a unified hidden space and computing their interaction via the Hadamard product, the model effectively captures the non-linear relationships between genetic background and environmental conditions. This fused representation serves as the input for the subsequent meta-learning and domain adaptation stages, ensuring that the model learns domain-invariant features that are robust to regional variations.

To handle the few-shot learning requirement, the framework utilizes the Reptile algorithm, a first-order meta-learning method that offers a computationally efficient alternative to Model-Agnostic Meta-Learning (MAML). Reptile learns a shared initialization that generalizes across domains, allowing the model to adapt to a new target environment with only a small number of labeled samples. In the study, the source domain data were partitioned into multiple meta-tasks, each corresponding to a specific location. The model was trained to minimize the prediction error on support sets and update its meta-parameters based on performance on query sets. This process enables the model to quickly fine-tune its parameters when applied to a new region, significantly reducing the data requirements for accurate prediction.

Addressing the domain shift problem, MAGEP incorporates adversarial domain adaptation through a gradient reversal layer (GRL) and a domain classifier. The GRL reverses the gradient flow during backpropagation, creating an adversarial relationship between the feature extractor and the domain discriminator. This forces the model to learn features that are indistinguishable between source and target domains, thereby aligning their distributions. Additionally, the framework employs statistical alignment methods such as Maximum Mean Discrepancy (MMD) and CORAL to further reduce distribution discrepancies. The combination of these adversarial and statistical alignment strategies ensures that the model remains robust even when the environmental conditions of the target domain differ significantly from those of the source domain.

Experimental results demonstrated that MAGEP outperformed a wide range of baseline models, including statistical methods like rrBLUP and BayesB, machine learning algorithms such as Random Forest and XGBoost, and deep learning models like DNNGP and GEFormer. Across four cross-domain scenarios, MAGEP achieved lower Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) for both plant height and grain yield prediction. Notably, the MAGEP_CORAL variant, which uses CORAL for distribution alignment, showed superior performance in terms of Pearson Correlation Coefficient (PCC), indicating a stronger linear association between predicted and observed phenotypes. The model also maintained competitive performance in non-cross-domain settings, suggesting that the meta-learning component provides a strong foundation for generalization regardless of the presence of domain shift.

Further analysis revealed that the model’s performance was influenced by the environmental composition of the target domain. Prediction errors varied depending on which specific trials were used as the target, with some environments presenting greater challenges due to more extreme or distinct climatic conditions. However, MAGEP consistently maintained lower prediction errors compared to baseline models, with fewer outliers in the error distributions. This stability is a critical advantage for breeding programs, where reliable predictions are necessary for making selection decisions. The model’s ability to handle heterogeneous environmental data without significant performance degradation highlights the effectiveness of the adversarial domain adaptation strategy in mitigating the impact of regional variability.

Interpretability analysis using gradient-based feature attribution identified key SNP loci and environmental factors associated with plant height and grain yield. Several highly ranked SNPs colocalized with previously reported quantitative trait loci (QTLs) for maize height and yield, validating the biological relevance of the model’s predictions. For example, SNPs associated with tassel length and branch number were identified as important for grain yield, consistent with known physiological mechanisms where excessive tassel growth can reduce grain production. Environmental factors such as rainfall, humidity, and solar radiation were also identified as dominant drivers of phenotypic variation, with their influence varying across different scenarios. These insights provide a quantitative basis for understanding the environmental regulation of crop phenotypes and offer valuable guidance for precision breeding strategies.

The study also evaluated the generalization ability of MAGEP across years and species. Using the 2019 G2F dataset as a target domain, the model demonstrated robust performance in cross-year prediction tasks, reducing RMSE and MAE compared to baseline models. Furthermore, the framework was tested on a multi-year rice dataset, where it achieved competitive predictive performance for plant height and grain yield. This cross-species validation suggests that the learned representations are transferable beyond maize, indicating the potential of MAGEP for broader application in crop breeding. By providing an effective and interpretable framework for cross-domain few-shot genomic prediction, this research offers a promising technical solution for accelerating the development of climate-resilient crop varieties in the face of global food security challenges.

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

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

Article References: Yang, Y., Peng, J., Wu, J., Lin, S., Gu, B., Yuan, A., & Hu, J. (2026). A meta-learning with adversarial domain adaptation method for few-shot cross-region maize height and yield joint prediction. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.09.010

Image Credits: AI Generated

DOI: 10.1016/j.aiia.2026.09.010

Keywords: meta-learning, adversarial, domain, adaptation, method, few-shot, cross-region, maize, height, yield, joint, prediction