Phage matching stops relying on one embedding: routing by query novelty lifts both-unseen AUPRC from 0.036 to 0.671
Synopsis
Phage therapy needs a matching phage for each clinical bacterial isolate, and conventional experimental screening is slow and labor-intensive. PhageHost-GNN encodes phage receptor-binding proteins and bacterial K-locus proteins as sequence, structure and genomic-context embeddings, feeds them to a graph neural network, and automatically selects the embedding model best suited to how novel a query strain is relative to the training set. On the Spain-derived Beamud2023 data, with both hosts and phages unseen, the routed model reached an AUPRC of 0.671, while fixed-embedding models stayed at or below 0.036.
Interpretation
Phage-host pairing no longer depends on a single fixed protein embedding; the framework switches embedding models according to how novel a query strain is relative to the training set. Earlier work either embedded phage or host proteins with one specific algorithm or concatenated and fused several embeddings, and rarely evaluated model performance separately by the novelty of query entities before choosing a model. On the Spain-derived Beamud2023 data, with both hosts and phages unseen, the routed model reached an AUPRC of 0.671 and an AUROC of 0.935, while all fixed-embedding models stayed at or below 0.036 AUPRC and 0.548 AUROC.
The framework builds phages and host bacteria as two node types in a heterogeneous graph, uses experimentally verified infection relations as supervision edges, and lets new nodes only receive messages from the training graph. GraphSAGE handles neighborhood sampling and aggregation, while GATv2 applies two attention heads to weight neighbors, so newly isolated strains can be admitted while context learned from pretraining is retained. Training and evaluation used the Wu2026 dataset (121 hosts, 74 phages, 8,954 interactions) and the Boeckaerts2024 dataset (200 hosts, 105 phages, 10,006 interactions), with leave-one-group-out cross-validation under host-unseen, phage-unseen and both-unseen modes.
A local laboratory can fine-tune an existing model onto its own small dataset instead of retraining from random initialization. Full fine-tuning updates the graph encoder and the MLP decoder, while decoder-only fine-tuning freezes the graph encoder and updates only the decoder, cutting trainable parameters to 19.1%. On an NVIDIA H100 80 GB, de novo end-to-end training took slightly over 8 minutes, full fine-tuning slightly over 2 minutes, and decoder-only fine-tuning slightly over 1 minute; on Beamud2023 the overall AUPRC was 0.338 for de novo training, 0.181 for full fine-tuning and 0.085 for decoder-only fine-tuning.
Perspective
The framework targets computational prescreening for Klebsiella phage therapy, with inputs limited to phage receptor-binding proteins and bacterial K-locus proteins, and outputs a ranking of candidate pairs for subsequent wet-lab validation. Evaluation covers host-unseen, phage-unseen and both-unseen inference modes, with retrospective benchmarks on the experimentally validated Boeckaerts2024-in-vitro and Beamud2023 datasets. The authors state that prospective multisite evaluation is required to determine whether it reduces screening time or cost.
The authors note that interactions not using the capsule or K antigen as the primary receptor may be missed, because model inputs are restricted to receptor-binding proteins and K-locus proteins. Receptor-mediated adsorption is only the first step of productive infection; post-adsorption host defenses and intracellular compatibility can block genome entry, replication, assembly or lysis, and these determinants are not represented by the current mechanism. kNN auxiliary edges may introduce similarity bias for divergent entities, and quantifying these effects requires denser interaction matrices and independent controls. These conclusions come from evaluation on the Beamud2023 dataset, and prospective validation on independent datasets is required before broader application.