Predicting microbe-disease association based on graph autoencoder and inductive matrix completion with...

Area of Science:

• Microbiology
• Computational Biology
• Genomics

Background:

• Microbes significantly impact human health and disease.
• Identifying microbe-disease interactions is key for pathogenesis insights, diagnosis, and treatment.
• Current computational methods for microbe-disease association screening lack accuracy and efficiency due to data inconsistencies and underutilized prior information.

Purpose of the Study:

• To develop an improved deep learning framework, GIMMDA, for identifying latent microbe-disease associations.
• To enhance the accuracy and efficiency of predicting microbe-disease relationships.
• To leverage graph autoencoder and inductive matrix completion for robust association prediction.

Main Methods:

• Proposed GIMMDA, a deep learning framework utilizing graph autoencoder and inductive matrix completion.
• Employed a co-training strategy across microbe and disease spaces to generate new representations.
• Implemented a similarity fusion strategy to boost prediction performance within an end-to-end framework.

Main Results:

• GIMMDA achieved competitive performance against state-of-the-art methods on three datasets (HMDAD, Disbiome, multiMDA).
• Achieved high Area Under the Receiver Operating Characteristic Curve (AUC) scores: 0.9735, 0.9156, and 0.9396.
• Case studies on asthma and obesity validated the model's effectiveness and reliability, confirming the benefits of similarity fusion.

Conclusions:

• The GIMMDA model exhibits strong capability in predicting microbe-disease associations.
• The framework is expected to aid in identifying potential microbe-related diseases.
• Further development may enhance the prediction of complex microbe-disease interactions.