Deep homo-heterogeneous association mining with hybrid scholars and multidimensional mixed moment networks:...

Drug repurposing accelerates microbial therapy development by bypassing the costly and time-consuming traditional drug discovery process. However, existing computational methods for predicting drug-microbe associations (MDAs) struggle to capture complex feature distributions and lack mechanistic interpretability. To overcome these challenges, we propose the Scholar-guided and Multi-dimensional Moment Neural Network (GSSMMT), a novel MDA prediction framework. First, biomedical data (e.g., chemical structures, microbial genomes, interaction databases) are integrated into drug and microbe homogeneous graphs. Second, multi-view random walk algorithms are applied to combine multimodal drug and microbial features from these homogeneous graphs, followed by constructing a heterogeneous network with drug-microbe interaction data. Third, GSSMMT employs a dual-path architecture: a scholar-guided network extracts domain-specific features from the homogeneous graphs, while a multi-dimensional moment neural network captures higher-order statistical patterns from the heterogeneous graph. Cross-graph fusion, driven by attention-based transpose matrices, dynamically adjusts interaction weights. Finally, a dual-tower support vector machine decodes the fused representations to predict MDA likelihoods, ensuring accuracy and traceability. Experiments on three benchmark datasets show GSSMMT outperforms five state-of-the-art baselines in prediction accuracy. Ablation studies further confirm the contributions of its key components, and case studies on three clinically relevant drugs reveal over 80 % alignment with PubMed-reported interactions. Overall, these results highlight GSSMMT's reliability in identifying latent MDAs with both computational rigor and biological interpretability.