Predicting potential microbe-disease associations with graph attention autoencoder, positive-unlabeled learning, and...

Area of Science:

• Computational biology
• Microbiome research
• Genomics and bioinformatics

Background:

• Microbial imbalances are linked to various human diseases.
• Accurate identification of microbe-disease associations (MDAs) is crucial for diagnosis and therapy.
• Experimental MDA prediction is costly and time-consuming.

Purpose of the Study:

• To develop an efficient computational method for predicting microbe-disease associations (MDAs).
• To leverage advanced machine learning techniques for improved MDA prediction accuracy.

Main Methods:

• Developed GPUDMDA, integrating graph attention autoencoder, positive-unlabeled learning, and deep neural networks.
• Computed microbe and disease similarity matrices using functional and Gaussian association profile kernel similarities.
• Employed positive-unlabeled learning to identify reliable negative MDA pairs.
• Utilized deep neural networks for final MDA prediction based on learned features.

Main Results:

• GPUDMDA outperformed four state-of-the-art methods in cross-validation tests on HMDAD and Disbiome databases.
• Achieved high AUC values, including 0.9501 on HMDAD and 0.8948 on Disbiome.
• Identified potential associations between *Enterobacter hormaechei* and asthma/inflammatory bowel disease.

Conclusions:

• GPUDMDA demonstrates significant potential for predicting microbe-disease associations.
• The method can accelerate the screening of therapeutic clues for microbe-related diseases.