M²BA-MDA: A Multi-Modal Multi-View Bidirectional Attention Network for Microbe-Disease Association Prediction

Numerous studies have shown that microbes play vital roles in human health and disease. Identifying microbe-disease associations can aid in disease diagnosis, treatment, and prevention. However, traditional biological experiments are time-consuming and costly. Although various computational methods have been developed, accurate and efficient approaches remain limited due to single-source data, insufficient prior knowledge, and suboptimal model performance. This paper proposed M2BA-MDA, a deep learning framework based on multi-modal, multi-view, and bidirectional attention mechanism for predicting potential microbe-disease associations. Firstly, microbe and disease features are extracted using multiple similarity measures and fused for consistency. Secondly, to mitigate gradient vanishing and over-smoothing issues in deep graph attention networks, we propose a stable enhanced graph attention network (EGAT) module incorporating cross-layer connections to extract microbial and disease features from each perspective. Moreover, to more effectively capture the complex interactions between microbes and diseases, we introduce an interaction module based on a bidirectional attention mechanism. This module enhances the mutual dependencies between the two entities and generates their final embeddings. Finally, a deep neural network (DNN) classifier is employed to predict potential associations. Extensive experiments conducted on the HMDAD and DisBiome datasets demonstrate that M2BA-MDA consistently outperforms five state-of-the-art methods. Parameter analysis and ablation studies further confirm the robustness and sensitivity of the model. In addition, case studies validate its effectiveness in identifying disease-associated microbes.