Predicting phage–host interactions is fundamental to microbial ecology and phage therapy, and is essential for elucidating phage–host networks and identifying optimal therapeutic phages. Traditional approaches primarily rely on sequence alignment tools such as NCBI BLAST (19) to infer host relationships through homology searches. However, the high diversity and rapid evolution of phage genomes often lead to weak homology signals, making these methods insufficiently sensitive for metagenomic data, especially when dealing with short fragments or distant evolutionary relationships. To address these limitations, a series of alignment-free methods has been proposed. For example, oligonucleotide frequency–based approaches predict hosts by comparing differences in k-mer distributions (20), achieving high prediction accuracy for metagenomic viral sequences. Similarly, WIsH (21) applies a Markov model to predict hosts for phage contigs, while PHP (22) employs a Gaussian model to analyze prokaryotic virus sequences. By eliminating dependency on sequence alignment (23, 24), these methods enhance both robustness and computational efficiency. With the advancement of machine learning, protein feature–based methods have further improved prediction performance. For instance, RaFAH (25) accurately identifies bacterial and archaeal hosts by analyzing viral protein composition combined with a classifier model. HoPhage (26), a de novo prediction tool designed for metavirome fragments, integrates both sequence and protein features to effectively handle short viral contigs. In addition, network-based methods such as (27) integrate multiple data sources—including sequence similarity and co-occurrence networks—into a unified framework for systematically predicting virus–prokaryote interactions.

In addition to phage–host interaction prediction, AI is increasingly being applied to other key aspects of phage therapy, accelerating its transition from laboratory research to clinical practice. In the prediction of phage protein function, the tool PhageDPO employs a support vector machine model to identify depolymerases—encoded within phage genomes—that are capable of degrading bacterial polysaccharide structures, thereby offering novel targets for disrupting biofilm barriers (28). In the realm of phage genome design, the Prophage Activation platform utilizes a random forest model to predict regulatory elements in prophages and enables the activation of dormant prophages by simulating edits to transcription factor binding sites (13). This approach provides a new strategy for precise antimicrobial intervention from within drug-resistant bacteria. At the clinical level, AI is further harnessed to optimize phage cocktail formulations and monitor the emergence of bacterial resistance (29). Together, these applications underscore the considerable potential of AI technologies in advancing phage therapy from empirical screening toward a rational, design-based framework.

This study adopted the dataset construction scheme consistent with VirHostMatcher-Net (27) to ensure comparability. We collected 2,288 phage genomes with known host information from the NCBI RefSeq database (accessed November 11, 2019). From these, 826 phages with specifically identified hosts at the strain level were selected to construct the positive training set, including 817 bacteria-infecting and 9 archaea-infecting phages. The remaining 1,462 phage-host pairs were held out as the final independent test set for all performance evaluations. Host information was extracted from the “isolate_host” or “host” fields in the GenBank records (27). For negative sample construction, we implemented a taxonomic distance-based sampling strategy by randomly selecting 826 phage-host pairs with the requirement that each selected host belongs to a different phylum from the phage's true host. While we acknowledge potential false negatives in these non-interacting pairs, the impact is minimized by the specificity of phage-host interactions and the phylum-level exclusion criterion (27).

We propose a deep Siamese network framework to accurately predict phage-host interactions, thereby selecting a precise phage for the treatment of pulmonary infections. The model formulates the task as a binary classification problem (match/non-match), with the core idea being to learn high-level features from the complete genome sequences of both phages and hosts, and then evaluate the likelihood of their interaction.

The overall architecture, as shown in Figure 1, consists of four core modules: a Sequence Preprocessing and Feature Representation module responsible for segmenting long genome sequences and converting them into numerical feature matrices; a Local Feature Learning module that uses CNN to extract local patterns from sequence segments; a Global Dependency Learning module that employs Transformer encoders to integrate all local features, forming a whole-genome representation; and a Fully Connected Classifier module for match prediction. Model training aims to minimize the cross-entropy loss function, optimizing all parameters via the backpropagation algorithm.

The model was trained for 200 epochs using the Adam optimizer with a learning rate of 1 × 10⁻⁴. The exponential decay rates for the moment estimates (β₁ and β₂) were set to 0.9 and 0.999, respectively. We employed a mini-batch training strategy with a batch size of 32. To prevent overfitting, an L2 weight decay regularization term with a coefficient of 1 × 10⁻³ was incorporated into the loss function. Furthermore, gradient clipping was applied to mitigate the exploding gradient problem, with the threshold set to 5.0. During testing, the population statistics (mean and variance) used as fixed parameters for normalization were estimated from the training data. These statistics were computed as the moving average of the mini-batch statistics during training, with a decay rate of 0.999.

All experiments were conducted on a workstation equipped with an NVIDIA GeForce RTX 4090 GPU. The deep learning framework was implemented using PyTorch 1.9.0 with CUDA 11.1 support.

To systematically evaluate prediction performance, we followed the standard evaluation protocol in phage-host prediction literature and quantitatively assessed the results across five taxonomic levels: Genus, Family, Order, Class, and Phylum. For any given taxonomic level L, the prediction accuracy is defined as:

where N_{correct, L} represents the number of phages whose predicted host matches the true host at taxonomic level L, and N_total denotes the total number of phage-host pairs in the test set (1,462 in this study). This hierarchical evaluation framework comprehensively measures model performance at different taxonomic resolutions. We placed particular emphasis on Genus-level and Phylum-level accuracy, which reflect the model's capability for fine-grained host prediction and broad taxonomic assignment, respectively. All evaluations were performed on an independent test set that was completely separate from the training data, ensuring unbiased assessment of model generalization capability.

To comprehensively evaluate the predictive performance of our model, we conducted a systematic comparative analysis against current mainstream phage-host prediction methods on a standardized test set (containing 1,462 phages and 62,493 candidate hosts). The compared methods included representative algorithms such as WIsH (based on Markov models), PHP (which uses k-mer similarity), and VirHostMatcher (VHM) (27). All methods were rigorously evaluated at five major taxonomic levels (Genus, Family, Order, Class, and Phylum) to ensure comprehensive and fair comparison. The experimental results shown in Table 1 fully demonstrate the superior performance of our model. In the most challenging genus-level classification task, our model achieved an accuracy of 36.1%, representing an improvement of 2.1% over PHP's 34.0% and 2.7% over VHM's 33.4%, with this enhancement being statistically significant. As the taxonomic level increases, the prediction accuracy of all methods shows an upward trend, which aligns with the general principle that higher taxonomic levels exhibit stronger sequence conservation. Particularly, our model reached 82.8% accuracy at the phylum level, surpassing PHP's 80.0% by 2.8%, demonstrating the model's stable advantage in high-level classification tasks.

The performance differences among various methods profoundly reflect the technical characteristics of their underlying algorithmic principles: The WIsH method, based on Markov models, performs well in short sequence prediction tasks but has limited feature extraction capability when processing complete genome sequences; The VirHostMatcher method utilizes k-mer frequency features and enhances prediction performance through multi-feature fusion; The PHP method employs Gaussian models to process k-mer features and demonstrates good adaptability at higher taxonomic levels; In contrast, our model innovatively introduces deep learning technology into the field of phage-host prediction, effectively capturing local sequence patterns through CNN networks, combined with Transformer architecture to model long-range dependencies, achieving end-to-end learning from raw sequences to prediction outcomes, thereby avoiding the limitations of manual feature design in traditional methods and consequently demonstrating significant advantages across multiple taxonomic levels.

To validate the effectiveness of each key component in our model, we conducted a series of ablation experiments on the testset (1,462 phages, 62,493 candidate hosts). All experiments used the identical training set and hyperparameter settings. The results across five taxonomic levels (Genus, Family, Order, Class, and Phylum) are shown in Table 2.

The ablation experiments reveal distinct contributions from each architectural component: The Skip-gram + CNN combination achieved 31.5% accuracy at the genus level, demonstrating that local pattern extraction through convolutional networks combined with semantic embeddings provides a solid foundation for sequence representation. However, the absence of Transformer encoding resulted in a 4.6% drop compared to the complete model, highlighting the importance of global dependency modeling for capturing long-range interactions in genomic sequences. The performance of CNN + Transformer (without Skip-gram) (32.0% at genus level) indicates that while the CNN-Transformer pipeline can learn meaningful representations from raw sequences, the pre-trained semantic embeddings from Skip-gram provide additional discriminative power. The 4.1% improvement when incorporating Skip-gram embeddings demonstrates their value in capturing biological context and k-mer semantics.