Protein embeddings improve phage-host interaction prediction.

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Phage therapy is a promising strategy against antimicrobial resistance.
• Accurate prediction of phage-host interactions is essential for selecting effective phages.
• Current computational methods often require manual feature engineering, limiting efficiency.

Purpose of the Study:

• To develop a more effective computational method for predicting phage-host interactions.
• To leverage protein language models for automated feature extraction from phage proteins.
• To assess the performance of receptor-binding protein embeddings in host prediction.

Main Methods:

• Phage-host interaction prediction framed as a multiclass classification task.
• Utilized embeddings from protein language models (e.g., ProtT5) for receptor-binding proteins.
• Compared performance against models using handcrafted genomic and protein sequence features.

Main Results:

• Protein language model embeddings outperformed handcrafted features for host prediction.
• The ProtT5 model achieved the highest accuracy.
• A 3% to 4% increase in weighted F1 and recall scores was observed with ProtT5 embeddings.

Conclusions:

• Receptor-binding protein embeddings provide a powerful input for phage-host interaction prediction models.
• Automated feature extraction using protein language models enhances prediction performance.
• This approach offers a more efficient and accurate method for identifying candidate phages.