Predicting phage-host interactions via feature augmentation and regional graph convolution.

Area of Science:

• Bioinformatics
• Computational Biology
• Microbiology

Background:

• Antibiotic resistance in superbugs necessitates novel therapeutic strategies like phage therapy.
• Accurate identification of phage-host interactions (PHIs) is critical for developing effective phage therapies.
• Current prediction methods, often sequence-based, struggle with data sparsity and fail to capture complex PHI relationships.

Purpose of the Study:

• To develop an advanced deep learning approach for more accurate phage-host interaction prediction.
• To address limitations of existing methods, including sequence-only reliance and susceptibility to overfitting.
• To enhance the modeling of intricate relationships within phage-host systems.

Main Methods:

• Proposed MI-RGC, a novel approach integrating mutual information for feature augmentation and regional graph convolution.
• Constructed a mutual information-based heterogeneous network to capture phage dependency relationships.
• Employed a regional graph convolutional model with attention mechanisms to learn node embeddings from heterogeneous networks.

Main Results:

• MI-RGC demonstrated superior performance in predicting phage-host interactions across three benchmark datasets.
• The method effectively addressed data sparsity and improved the comprehensive modeling of PHIs.
• Feature augmentation via mutual information and regional graph convolution enhanced predictive accuracy.

Conclusions:

• MI-RGC offers a significant advancement in predicting phage-host interactions, crucial for phage therapy development.
• The integration of mutual information and regional graph convolution provides a robust framework for complex biological network analysis.
• This approach holds promise for accelerating the discovery and application of bacteriophages in combating antibiotic resistance.