Introduction: Rice blast, caused by Magnaporthe oryzae, poses a severe threat to global rice production. Chemical control raises increasing environmental and economic concerns, while traditional resistance gene identification remains laborious and inefficient. Thus, understanding the molecular interplay between rice and the pathogen is essential for developing sustainable resistance strategies. Methods: We proposed a novel deep learning framework, the Biological Attentional Graph Transformer (BAGT), to integrate genomics, transcriptomics, and proteomics into a multi-omics heterogeneous graph (the source code for BAGT is publicly available at https://github.com/yuxinjin26- lab/BAGT). BAGT employed attention supervision guided by biological priors to identify functionally important sRNAs in the rice-M. oryzae interaction network. Results: BAGT achieved 85.71% accuracy and an AUC of 0.9094, significantly outperforming baseline graph neural network models (GCN, GAT, HAN, and RGCN). The model identified 20 key rice sRNAs with high regulatory potential. GO and KEGG enrichment analyses of their target genes revealed significant associations with plant immune signaling, phosphorylation, and metabolic pathways. Discussion: The attention mechanism in BAGT enhanced both prediction accuracy and biological interpretability. Key sRNAs showed central roles in the regulatory network and were enriched in defense- related pathways, suggesting that they may regulate rice’s immune response against fungal invasion. Conclusion: Compared to traditional machine learning methods, the proposed BAGT model demonstrated superior performance, achieving 85.71% accuracy and an AUC of 0.9094. These results underscore its broad applicability to other plant–pathogen systems and its potential to inform the development of crop varieties resistant to pathogens.
Jin et al. (Fri,) studied this question.