Deep Learning Transforms Phage-Host Interaction Discovery from Metagenomic Data.

Area of Science:

• Microbiome research
• Computational biology
• Virology

Background:

• Microbial communities are vital for ecosystem functions, with phages playing a key role in shaping them.
• Existing methods for inferring phage-host interactions (PHIs) from metagenomic data suffer from low sensitivity and limited ecological accuracy.

Purpose of the Study:

• To introduce PHILM, a novel deep learning framework for predicting PHIs directly from metagenomic taxonomic profiles.
• To overcome the limitations of current computational approaches for PHI inference.

Main Methods:

• Developed PHILM, a deep learning framework utilizing metagenomic taxonomic profiles.
• Validated PHILM using synthetic datasets from ecological models and real-world metagenomic data.
• Compared PHILM against co-abundance-based and assembly-based methods for PHI inference.

Main Results:

• PHILM consistently outperformed co-abundance-based methods in inferring PHIs.
• Applied to a large dataset, PHILM identified 90% more genus-level PHIs than assembly-based approaches.
• PHILM's latent representations effectively captured microbial succession and improved disease classification accuracy.

Conclusions:

• PHILM provides a powerful new computational framework for predicting phage-host interactions from metagenomic data.
• The framework offers significant advancements for microbiome science and translational medicine by revealing complex phage-host dynamics.