Predicting potential microbe-disease associations based on auto-encoder and graph convolution network.

Area of Science:

• Microbiology
• Computational Biology
• Bioinformatics

Background:

• The human microbiome significantly influences health and disease, including drug efficacy and cancer progression.
• Identifying microbe-disease associations is crucial for clinical practice.
• Traditional experimental methods are costly and time-consuming, necessitating advanced computational approaches.

Purpose of the Study:

• To develop an accurate computational model for predicting potential microbe-disease associations.
• To overcome limitations of existing methods, such as low node feature utilization and suboptimal prediction accuracy.

Main Methods:

• Proposed the DAEGCNDF model, integrating graph convolutional networks (GCN) for low-rank feature extraction and a deep sparse auto-encoder (DAE) for high-rank feature extraction.
• Fused four similarity features for microbes and diseases into a comprehensive matrix.
• Employed Deep Forest for the final prediction of microbe-disease relationships.

Main Results:

• The DAEGCNDF model effectively integrates low-rank and high-rank features, improving prediction performance.
• Combining feature types enhanced the utilization of node information.
• Deep Forest demonstrated superior classification performance compared to baseline models.

Conclusions:

• The DAEGCNDF model offers a promising computational strategy for identifying microbe-disease associations.
• Integrating diverse feature representations and advanced machine learning techniques can significantly advance microbiome research.
• Accurate prediction of microbe-disease links can aid in clinical diagnostics and therapeutic strategies.