PBIP: a deep learning framework for predicting phage-bacterium interactions at the strain level.

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Phage therapy is a promising antimicrobial strategy, with phage-bacterium interaction (PBI) prediction being key.
• Current computational methods for PBI prediction lack strain-level specificity and deep sequence feature analysis.
• This limits the clinical application of phage therapy and discovery of novel phage-bacterium relationships.

Purpose of the Study:

• To develop a novel deep learning framework, PBIP, for accurate strain-level phage-bacterium interaction prediction.
• To address limitations in existing methods by incorporating deep embedding representations and strain-specific data.
• To enhance the potential of phage therapy through improved PBI prediction.

Main Methods:

• Strain-level PBI data was generated from *Klebsiella pneumoniae* clinical isolates.
• A pretrained unified representation model generated deep embeddings from phage and bacterium protein sequences.
• The synthetic minority oversampling technique balanced the dataset.
• A deep neural network (CNN, Bi-GRU, attention) was designed for feature extraction and PBI prediction.

Main Results:

• PBIP demonstrated superior performance in strain-level PBI prediction compared to state-of-the-art methods.
• The framework effectively captures complex biological patterns from sequence data.
• Deep embeddings and advanced neural network architecture significantly improved prediction accuracy.

Conclusions:

• PBIP offers a powerful new tool for strain-level PBI prediction, advancing phage therapy research.
• The framework's ability to analyze deep sequence features opens new avenues for discovering specific phage-bacterium interactions.
• This work facilitates the optimization of therapeutic strategies in phage therapy.