PhageHost-GNN: novelty-aware and locally adaptable strain-level Klebsiella-phage prediction

Klebsiella infections are becoming increasingly resistant to antibiotics, prompting researchers to explore bacteriophages as alternative treatments. However, the narrow host range of these phages makes it challenging to identify the right phage for a specific bacterial isolate using traditional experimental methods. Existing computational models using biomolecular features have shown potential for phage-host matching, but their performance varies across datasets with different levels of novelty.

This study introduces PhageHost-GNN, a novel framework that utilizes routed models to predict strain-level interactions between Klebsiella and phages. This framework was evaluated under three clinical scenarios, including host-unseen, phage-unseen, and both-unseen conditions. By employing multiple embedding models to encode sequence, structural, and genomic-context features, PhageHost-GNN achieved Area Under the Precision-Recall Curve (AUPRC) scores of 0.784, 0.731, and 0.642 for the respective scenarios.

Benchmarks on experimentally validated datasets demonstrated the framework's clear advantage over fixed embeddings when both hosts and phages were novel to the trained model.

In addition to direct inference and de novo training, PhageHost-GNN also implements a fine-tuning pipeline to adapt existing models to local custom datasets. This reduces model-training time and provides flexibility in model selection, evaluation, and local adaptation for practical phage-therapy screening.