Development and validation of embedded multilayer attention graph convolution neural network models for predicting gut microbe-disease associations.

Introduction: Identifying associations between gut microbes and diseases is crucial for uncovering disease mechanisms and advancing precision medicine. However, experimentally validated microbe-disease associations remain scarce due to time-consuming and costly validation. Existing computational methods face three measurable challenges: (1) severe class imbalance (positive associations constitute only 1.89-2.56% of all possible pairs); (2) difficulty in capturing nonlinear relationships inherent in microbial community interactions; and (3) limited capacity for propagating information across heterogeneous networks with sparse connectivity (graph density < 0.03). Methods: To address these limitations, we propose MAGMDA, a graph convolutional neural network embedded with a multi-layer attention mechanism. MAGMDA integrates known microbe-disease associations, microbial functional similarity, and disease semantic similarity into a heterogeneous network. A graph convolutional encoder learns low-dimensional embeddings of microbes and diseases, while a multi-head additive attention mechanism (8 attention heads) is incorporated into each convolutional layer. Attention weights are computed via softmax normalization over layer-wise embeddings and shared across all nodes to preserve input feature contributions and mitigate information decay. A linear decoder reconstructs the association matrix, and the model is optimized using a weighted binary cross-entropy loss to enhance sensitivity to positive associations. Results: Evaluated on the HMDAD dataset and independently validated on the Disbiome dataset (4,351 associations between 218 diseases and 1,052 microbes), MAGMDA achieves average AUC improvements of 1.87% (HMDAD) and 1.44% (Disbiome) over the second-best method, with an average AUPR improvement of 1.83% (HMDAD). Case studies on asthma and type 2 diabetes-both closely linked to gut microbiota-show that the top 30 predicted microbes are strongly supported by published literature. Preliminary external validation using real-world clinical data (290 osteoarthritis patients vs. 290 healthy controls) and animal experiments (6 rabbits per group for osteoarthritis; 6 mice per group for diabetic cardiomyopathy) provides supporting evidence for the model's ability to identify gut microbes associated with osteoarthritis and diabetic cardiomyopathy, though further large-scale validation is needed. Discussion: MAGMDA offers a robust computational framework for prioritizing gut microbe-disease associations, with strong potential to guide hypothesis-driven research and reduce the burden of experimental validation in microbiome studies.