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

Area of Science:

• Microbiology and Bioinformatics
• Computational Biology and Disease Association Studies

Background:

• Microbes significantly impact human health and disease, making microbe-disease association identification crucial for clinical applications.
• Current computational methods for predicting microbe-disease associations are limited by data scarcity, insufficient prior knowledge, and suboptimal performance.

Purpose of the Study:

• To develop an accurate and efficient deep learning framework, M²BA-MDA, for predicting potential microbe-disease associations.
• To address limitations of existing methods by integrating multi-modal, multi-view data and employing a bidirectional attention mechanism.

Main Methods:

• Proposed M²BA-MDA, a deep learning framework utilizing multi-modal, multi-view, and bidirectional attention.
• Implemented an Enhanced Graph Attention Network (EGAT) to extract robust microbial and disease features, mitigating gradient issues.
• Employed a bidirectional attention mechanism to capture complex microbe-disease interactions and generate enhanced embeddings for prediction via a deep neural network (DNN).

Main Results:

• M²BA-MDA demonstrated superior performance compared to five state-of-the-art methods on the HMDAD and DisBiome datasets.
• Parameter analysis and ablation studies confirmed the model's robustness and sensitivity.
• Case studies successfully validated the model's effectiveness in identifying disease-associated microbes.

Conclusions:

• M²BA-MDA offers a powerful and effective computational approach for predicting microbe-disease associations.
• The framework's advanced architecture enhances feature extraction and interaction modeling, paving the way for improved disease diagnosis and treatment strategies.