Development and validation of embedded multilayer attention graph convolution neural network models for predicting...

Area of Science:

• Microbiome research
• Computational biology
• Precision medicine

Background:

• Identifying gut microbe-disease associations is vital for understanding disease mechanisms and advancing precision medicine.
• Experimental validation of these associations is costly and time-consuming, with existing computational methods facing challenges like data imbalance and difficulty modeling complex interactions.
• There is a need for efficient computational tools to predict and prioritize potential microbe-disease links.

Purpose of the Study:

• To develop a robust computational framework, MAGMDA, for predicting gut microbe-disease associations.
• To address limitations of existing methods, including class imbalance, nonlinear relationships, and sparse network connectivity.
• To improve the accuracy and efficiency of identifying potential microbial biomarkers for various diseases.

Main Methods:

• MAGMDA utilizes a graph convolutional neural network with a multi-layer attention mechanism to integrate microbe-disease associations, microbial functional similarity, and disease semantic similarity into a heterogeneous network.
• A graph convolutional encoder learns low-dimensional embeddings, while a multi-head additive attention mechanism preserves feature contributions and mitigates information decay.
• The model is optimized using a weighted binary cross-entropy loss to enhance sensitivity to positive associations.

Main Results:

• MAGMDA demonstrated superior performance on the HMDAD and Disbiome datasets, achieving significant AUC and AUPR improvements over existing methods.
• Case studies on asthma and type 2 diabetes highlighted literature-supported predictions, validating the model's ability to identify relevant microbes.
• Preliminary external validation using clinical data and animal experiments provided supporting evidence for MAGMDA's potential in identifying associations with osteoarthritis and diabetic cardiomyopathy.

Conclusions:

• MAGMDA provides a powerful computational approach for prioritizing gut microbe-disease associations.
• The framework has the potential to guide hypothesis-driven research and reduce the experimental burden in microbiome studies.
• This method can accelerate the discovery of microbial targets for therapeutic interventions and diagnostic purposes.