PHIEmbed

PHIEmbed predicts phage-host interactions by using receptor-binding protein (RBP) sequence representations to identify likely host genera, supporting computational selection of phages for targeting bacterial hosts.

Key Features:

• RBP-centric input: Uses receptor-binding protein amino acid sequences as the primary biological signal for host recognition.
• Protein language models: Transforms RBP sequences into dense embeddings that capture complex biological information without requiring alignment or structural data.
• ProtT5 (transformer) embeddings: Employs the ProtT5 transformer-based protein language model to encode protein sequences automatically.
• Multiclass classification: Frames host prediction as a multiclass classification problem that predicts host genus from RBP embeddings.
• Elimination of manual feature engineering: Removes the need for handcrafted features and proteome-wide feature selection by relying on learned embeddings.
• Performance improvement: ProtT5-derived embeddings yielded a 3%–4% increase in weighted F1 and recall scores across prediction confidence thresholds compared to traditional handcrafted genomic and protein features.

Scientific Applications:

• Phage therapy candidate selection: Helps shortlist candidate bacteriophages for targeting specific bacterial hosts, aiding development of treatments for antibiotic-resistant infections.
• Phage-host interaction research: Supports studies of host range and recognition mechanisms to inform investigations into antimicrobial resistance and therapeutic phage application.

Methodology:

Transforms RBP amino acid sequences into dense embeddings using the ProtT5 protein language model and applies a multiclass classifier to predict host genus.