M-GAL: A Multi-head Graph Autoencoder Framework for Deciphering Rice-Magnaporthe oryzae Regulatory Networks

Understanding cross-kingdom regulatory interactions between rice and Magnaporthe oryzae (M. oryzae) is critical for developing sustainable strategies to combat rice blast disease. At the same time, elucidating these interactions through biological experiments alone is complex, costly, and time-consuming. We proposed M-GAL, a framework that integrates genomics, transcriptomics, proteomics, and metabolomics data from rice and M. oryzae. It combines a Multi-head Adversarial Variational Graph Autoencoder (MHAVGAE) to learn cross-layer interactions and an Adaptive-Rules Louvain algorithm (AR-Louvain) to detect biologically meaningful modules by dynamically adjusting resolution parameters based on biological rules. M-GAL successfully uncovered key cross-kingdom interactions and highlighted that M. oryzae disrupts rice cellular processes, including protein synthesis, metabolism, and nutrient uptake, facilitating pathogen infection. Furthermore, the framework predicted 33 small RNAs from M. oryzae targeting avirulence and resistance genes in rice, illustrating bidirectional regulatory interactions critical for pathogen-host dynamics. M-GAL demonstrates improved capability in identifying complex host-pathogen interactions by integrating multi-omics data and using adaptive community detection. Compared to conventional approaches, it provides more accurate and biologically relevant results. M-GAL represents an effective and robust alternative to conventional methods for identifying cross-kingdom regulatory factors and constructing reliable host-pathogen regulatory networks.