PhageMind: Generalized Strain-level Phage Host Range Prediction via Meta-learning

TL;DR

PhageMind addresses the challenge of predicting strain-level phage–host interactions across diverse bacterial genera under data scarcity. It combines Model-Agnostic Meta-Learning (MAML) with a heterogeneous graph convolutional network that encodes molecular interfaces via tail fiber proteins and O-antigen biosynthesis clusters, enabling rapid adaptation to new genera with limited data. The framework uses genus-specific tasks, inner/outer optimization loops, and a dynamic edge-masking strategy to learn transferable priors and robust topology-informed predictions. Across Escherichia, Klebsiella, Vibrio, and Alteromonas, PhageMind achieves high predictive accuracy, faster convergence, and strong cross-genus transfer, offering a scalable tool for targeted phage therapy, diagnostics, and ecosystem studies.

Abstract

Bacteriophages (phages) are key regulators of bacterial populations and hold great promise for applications such as phage therapy, biocontrol, and industrial fermentation. The success of these applications depends on accurately determining phage host range, which is often specific at the strain level rather than the species level. However, existing computational approaches face major limitations: many rely on genus-specific features that do not generalize across taxa, while others require large amounts of training data that are unavailable for most bacterial lineages. These challenges create a critical need for methods that can accurately predict strain-level phage-host interactions across diverse bacterial genera, particularly under data-limited conditions. We present PhageMind, a learning framework designed to address this challenge by enabling efficient transfer of knowledge across bacterial genera. PhageMind is trained to identify shared principles of phage-bacterium interactions from well-studied systems and to rapidly adapt these principles to new genera using only a small number of known interactions. To reflect the biological basis of infection, we represent phage-host relationships using a knowledge graph that explicitly incorporates phage tail fiber proteins and bacterial O-antigen biosynthesis gene clusters, and we use this representation to guide interaction prediction. Across four bacterial genera (Escherichia, Klebsiella, Vibrio, and Alteromonas), PhageMind achieves high prediction accuracy and shows strong adaptability to new lineages. In particular, in leave-one-genus-out evaluations, the model maintains robust performance when only limited reference data are available, demonstrating its potential as a scalable and practical tool for studying phage-host interactions across the global phageome.

Figures (9)

• Figure 1: The Framework of PhageMind. It is designed to predict interactions between phages and bacterial strains within target genera, and is capable of adapting to new genus with limited data. Genomes are translated into proteins, screened, and encoded into node features using sequence‑based properties and statistical summaries, while known interactions provide edge priors for the graph convolutional network. To enhance generalization, we apply meta‑learning, which optimizes the starting parameter set $\theta$ across tasks. The meta‑learned $\theta'$ serves as a robust initialization, and can rapidly adapt to different genera, enabling improved performance across diverse genera.
• Figure 2: Heatmap of Vibrio phage-host interactions (black = interaction; white = no interaction). Rows correspond to individual phages and columns to bacterial strains. Right color bar represents bacterial species while bottom color bar represents phage morphology. The map shows wide variation in phage host breadth and pronounced intra‑species differences in bacterial susceptibility, with some strains targeted by many phages while others are rarely or never infected.
• Figure 3: Structure of our meta-learning framework. Each independent task corresponds to a mutually exclusive dataset representing a single genus. Within each task, training data is split into a support set for inner-loop adaptation and a disjoint query set for evaluation. The outer loop performs a meta-update by aggregating losses across tasks to optimize meta-parameters.
• Figure 4: Structure of our GCN Model. Bacteria and phages nodes form mutually exclusive subsets, connected only through experimentally verified infection relationships.
• Figure 5: Strategy we used for dataset partition. (a) Double‑isolation is used for data partition in each bacterial genus. Bacteria and phages are independently partitioned into training and test set with ratio 7:3. The training set contains only interactions among training organisms, while the test set includes all remaining pairs (training-test, test-training, and test-test). (b) Each training set is randomly split into support set (30%) and query set (70%), which are needed to implement meta-learning framework.