AgriNetFusion: Guided hierarchical attention fuses multimodal data to unlock drivers of crop performance

Accurate prediction of crop phenotypes is critical for breeding and agricul-tural sustainability but challenged by complex interactions between genetics, environment, and management (G×E×M). Existing models often struggle to integrate diverse data modalities effectively or incorporate domain knowledge, limiting predictive power and interpretability. Here, we introduce AgriNetFu- sion, a novel deep learning framework employing Hierarchical Multimodal At- tention (HMMA) explicitly guided by phenotype-specific domain knowledge. AgriNetFusion integrates high-dimensional genotype data (using PCA or a novel attentive autoencoder) with dynamic environmental and static field data from large-scale corn trials. Rigorous benchmarking across 11 corn phenotypes shows AgriNetFusion significantly outperforms standard machine learning (Random Forest, XGBoost, LASSO) and simpler deep learning baselines. Ab- lation studies confirm the crucial contributions of hierarchy, guided attention, and genotype representation. Interpretability analyses reveal learned modality importances consistent with agronomic priors while also highlighting novel interactions. AgriNetFusion provides a powerful and interpretable approach for modeling complex biological systems, advancing predictive capabilities in agriculture and offering a framework adaptable to other multimodal scientific domains.