Adversarial regularized autoencoder graph neural network for microbe-disease associations prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in healthcare

Background:

• Microorganisms are linked to numerous human diseases, necessitating accurate association prediction.
• Experimental methods for identifying microbe-disease links are costly and time-consuming.
• Deep learning and biological networks offer a promising approach for large-scale association prediction.

Purpose of the Study:

• To develop an efficient computational method for predicting microbe-disease associations.
• To leverage graph neural networks and autoencoders for enhanced prediction accuracy.
• To identify novel microbial contributions to human diseases.

Main Methods:

• Proposed Stacked Adversarial Regularization for Microbe-Disease Associations Prediction (SARMDA), an adversarial regularized autoencoder graph neural network algorithm.
• Constructed a heterogeneous network integrating microbial and disease topological and functional similarity.
• Employed a GraphSAGE-based autoencoder and an adversarial regularized autoencoder graph neural network embedding model.

Main Results:

• SARMDA achieved high performance on the Human Microbe-Disease Association Database (HMDAD) with AUC 0.9891 and AUPR 0.9902.
• On the Disbiome dataset, SARMDA yielded AUC 0.9328 and AUPR 0.9233, outperforming eight other methods.
• Model effectiveness validated through case studies on asthma and inflammatory bowel disease.

Conclusions:

• SARMDA demonstrates superior performance in predicting microbe-disease associations compared to existing methods.
• The model's ability to integrate network topology and attributes enhances prediction.
• This approach facilitates a deeper understanding of microbe-disease relationships and potential therapeutic targets.