Towards accurate artificial intelligence models for strain-level phage-host prediction.

Area of Science:

• Computational biology
• Genomics
• Microbial ecology

Background:

• Strain-level prediction of phage-host interactions (PHIs) is vital for precision phage therapy.
• Existing methods lack the resolution and scalability for clinical applications.
• Artificial intelligence (AI) models offer new approaches using genomic data.

Purpose of the Study:

• To review recent advances in AI-driven strain-level PHI prediction.
• To analyze challenges including data limitations and evaluation strategies.
• To assess the clinical translatability of current PHI prediction frameworks.

Main Methods:

• Review of biologically grounded feature-based models.
• Examination of hybrid representation-learning frameworks.
• Analysis of phylogeny-agnostic machine learning and deep learning architectures.
• Investigation of evaluation design, negative definition, and train-test splitting.

Main Results:

• AI models leverage genomic information for strain-level PHI inference.
• Challenges include sparse/imbalanced data, assay-dependent labels, and generalization.
• Inappropriate benchmarking can overestimate model performance and biological resolution.

Conclusions:

• Current AI frameworks face biological, experimental, and data constraints for clinical phage therapy.
• Robust, interpretable, and clinically translatable PHI prediction systems are needed.
• Pragmatic pathways are outlined for improving future PHI prediction models.