Predicting human microbe-disease associations via graph attention networks with inductive matrix completion

Area of Science:

• Microbiology
• Computational Biology
• Genomics

Background:

• Human microbes are crucial in complex diseases and represent a key area for precision medicine.
• Current in silico methods often fail to capture nonlinear microbe-disease associations and struggle with novel predictions.

Purpose of the Study:

• To develop an advanced computational framework for predicting human microbe-disease associations.
• To overcome limitations of existing linear models and label propagation techniques.

Main Methods:

• Proposed GATMDA, a deep learning framework utilizing graph attention networks (GAT) with inductive matrix completion.
• Engineered an optimized GAT with talking-heads and a bi-interaction aggregator for robust node representation and neighbor aggregation.
• Integrated multiply sources of biomedical data for feature construction.

Main Results:

• GATMDA demonstrated superior performance over baseline methods on HMDAD and Disbiome datasets.
• Case studies on asthma and inflammatory bowel disease validated the model's predictive accuracy and clinical relevance.
• The framework effectively captures complex, nonlinear microbe-disease associations.

Conclusions:

• GATMDA offers a powerful new tool for identifying microbe-disease links, advancing our understanding of disease pathogenesis.
• The model's ability to handle novel microbes and diseases opens new avenues for drug target screening and personalized medicine.