MVPHI: a multi-view learning framework for predicting complex microbial interactions Authors: Yun Xie, Jie Pan, Dan Li, Qi Wang, Yanmei Sun, Shiwei Wang Published: 2025 ABSTRACT Bacteriophages (phages) are viruses that infect bacteria. As the natural regulators of microbial communities, it plays a crucial role in microbiome turnover. Predicting phage–bacteria interactions (PBIs), as well as bacteria–bacteria interactions (BBIs) is essential for advancing microbiome research. Given the high cost and risk of wet-lab techniques, computational biology and bioinformatics methods are reasonable alternatives. However, existing approaches suffer from the low predictive accuracy and poor efficiency. In this study, we proposed a multi learning-based model named MVPHI for predicting complex microbial interactions. More specifically, we first construct a heterogeneous multi-attributed microbial network (MAMN) based on pathogenic bacteria and associated phages. Next, MVPHI introduces three different view microbial characteristics for training the model, including the statistical-view, textual-view and topology-view features. Experimental results on seven benchmark datasets indicated that MVPHI achieves superior performance compared with the six variant models and eight baseline algorithms. Moreover, case study and protein docking experiments further demonstrated the robustness and generalization of our model. In conclusion, the proposed MVPHI model has potential ability to predict novel PBIs and BBIs, and can also provide valuable insights for phages screening and bacterial community research.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-32359-2. Introduction Bacteriophages (Phages) are natural predators that invade bacteria and have co-evolved with bacteria for nearly 4 billion years1–3. As one of the most abundant organisms on earth (about 1031 particles in the biosphere), they can affect the function, structure, and evolution of natural microbial communities4,5. Recently, phages have been found with the potential for food preservation, environmental governance, and nanotechnology, while all of which depended on the bactericidal action of phages6. Additionally, antibiotic-resistant bacteria are rapidly emerging owing to the misuse of antibiotics, and phage therapy is considered to be one of the most promising alternatives for treating bacterial infections7–9. Host ranges are the key to explain the role of phages in complex molecular mechanisms, and to be designed as effective tools for manipulating microbial communities10. Therefore, prediction of phage–bacteria interactions (PBIs) is essential in modifying phages and exploring complex relationships between microorganisms11. In addition, microbial interactions are inherently governed by heterogeneous biological information spanning multiple levels, including genomic structure, sequence composition, protein semantic attributes, and complex network topology12. Each level reflects a distinct aspect of microbial behaviour and interaction mechanisms. Therefore, relying on any single type of information is insufficient to fully characterize the inherent complexity of PBIs and bacteria–bacteriainteractions (BBIs)13. In the field of microbiology, most techniques for identifying PBIs still rely on traditional wet-lab experiments, including plaque assays, liquid culture, microfluidic PCR, and flow cytometry14,15. However, such biological approaches are laborious, time-consuming and costly. With the rapid development of biotechnology, biomedical data of phages and bacteria have emerged and drawn extensive attention16. Therefore, it is reasonable to develop computational models for predicting BBIs and PBIs17. To date, many computational approaches have been proposed and can be broadly categorized into three groups: sequence similarity−based methods, machine learning–based methods, and network-based methods. Sequence similarity-based methods require the evolutionary information of microbes to construct the similarity network. After phages infect bacterial, their genome will be integrated into the bacterial genome. Thus, sequence similarity-based methods are preliminary used to predict BBIs and PBIs. For example, Andrzej et al.18 designed an extension for BLAST, named Phirbo. It related the sequence and taxonomic similarity of phages and target bacterial sequences through intermediate public database, that have potential homologous for both phages and bacteria hosts. Wang et al.19 proposed a new model named GERMAN-PHI, which collected the host−host and phage-phage similarity information to construct a heterogeneous network, and combined multi-head attention mechanism with graph embedding representation learning technique to predict PBIs. Pan et al.20 proposed a deep learning-based prediction framework named GSPHI to identify PBIs by integrating DNA and protein sequence data with graph embedding techniques. The model employs a natural language processing approach for initial node representation, followed by structural deep network embedding (SDNE) to extract topological features, and finally uses a deep neural network for accurate prediction. Gonzales et al.21 proposed PHIStruct, a structure-aware phage–host interaction prediction model that leverages embeddings derived from SaProt, a structure-informed protein language model. By inputting receptor-binding protein features into a multilayer perceptron, PHIStruct can effectively predict phage hosts across ESKAPEE genera. Experimental results demonstrate that PHIStruct achieves superior and more stable F1 scores, particularly under low sequence similarity scenarios, outperforming traditional machine learning tools by a notable margin. Despite their interpretability and biological relevance, sequence similarity-based methods suffer from several limitations22,23. First, these approaches rely on the availability and completeness of genomic sequences, which can be challenging due to incomplete or fragmented data in public databases. Second, sequence similarity may not fully capture the functional relationships or phenotypic effects of microbial interactions, especially in cases where sequence conservation is low but functional similarity remains high. Third, such methods may fail to identify novel interactions, as they are typically constrained by known homologies and cannot generalize well to unseen patterns. These drawbacks highlight the need for more advanced computational models that integrate diverse biological information beyond sequence similarity. Machine learning-based models are always combined with network analysis-based methods. After combining, computational models can analyze the topological structures of microbial network to regrade their complex relationships. For example, Li et al.24 developed a novel deep learning framework named PHIAF, which integrates GAN-based data augmentation to alleviate data scarcity in PHI prediction. It fuses DNA and protein sequence features to enhance representation learning. Furthermore, an attention-enhanced CNN module is employed to improve interpretability and classification performance. Wang et al.25 developed a convolutional neural network (CNN)-based model called DeepHost that leverages a novel encoding strategy. DeepHost transforms phage genomes into three-dimensional matrices and applies CNN to infer host taxonomy. It is alignment-free, eliminating the need for sequence or k-mer comparisons with existing datasets, thereby improving computational efficiency on large-scale data. It also incorporates spaced k-mers, which enhance tolerance to SNPs and InDels. In addition, it employs a learning-based model that directly outputs classification results, avoiding manual selection of distance measures or thresholds. LMFH-VH26 and ILMF-VH27 extracted microbial information from three networks, including phage–bacteria network, bacteria–bacterianetwork, and phage−phage network. Both of them adopted neighborhood regularization logistic matrix factorization to predict PBIs. The difference of LMFH-VH and ILMF-VH is that ILMF-VH adds the similar network fusion to integrate multiple-source features. Wei et al.28 proposed a novel computational model named DeepPBI-KG that predicts PBIs by focusing on key proteins and genes involved in phage–bacteria interactions. A validation module based on essential genes is introduced to highlight biological relevance. Unlike traditional genome-wide approaches, DeepPBI-KG selectively targets interaction-related genes. In addition, K-means clustering is used to generate high-quality negative samples, improving data balance and predictive performance. Benchmark results confirm the model’s robustness and effectiveness. However, machine learning-based approaches still encounter several inherent limitations. Many models rely on predefined features or simplified representations, which may not sufficiently capture the complex biological mechanisms underlying PHI29,30. In addition, the predictive performance of these models can be affected by the imbalance between positive and negative samples, especially in datasets with sparse validated PHIs. Deep learning models, while powerful, often function as black boxes, limiting their interpretability and hindering biological insight extraction. Furthermore, most existing methods fail to incorporate the heterogeneous biological contexts, such as host immunity, horizontal gene transfer, or environmental conditions that are critical for accurately modeling microbial interactions. These limitations highlight the need for more interpretable and biologically informed learning strategies. Network-based methods model microbial entities (e.g., phages and hosts) as nodes and their interactions as edges, enabling the use of graph theory and network representation learning to capture topological and relational patterns. These approaches are particularly effective for revealing complex interaction structures that are difficult to identify through sequence information alone. For example, Xiao et al.31 developed PHISGAE, a host specificity-aware graph autoencoder framework for phage−host interaction prediction. This method integrates both phage genome and protein sequence data to construct initial phage-phage relationships, which are further refined using a K-nearest neighbor strategy to preserve meaningful local semantics in a heterogeneous network. By explicitly incorporating host specificity into the representation learning process, PHISGAE achieves superior predictive performance at both species and genus levels, and demonstrates robust generalization in identifying potential hosts for novel phages discovered in metagenomic data. Zhou et al.32 proposed PHISDetector, a comprehensive computational framework that predicts phage–host interactions by detecting and integrating diverse genomic signals, such as CRISPR matches, prophage regions, and protein–protein interactions. The model employs machine learning techniques to assess the likelihood of interaction based on these features. PHISDetector is suitable for both individual predictions and large-scale metagenomic analyses, offering a flexible solution for exploring phage-microbe relationships. While network-based methods effectively capture topological patterns in phage−host relationships, they also present several limitations33,34. These approaches often depend on the availability of high-quality interaction data, making them susceptible to noise and sparsity. In addition, they may struggle to incorporate dynamic or heterogeneous biological information, which can reduce their interpretability and generalizability. Compared to previous studies, the proposed MVPHI model has several distinct advantages. MVPHI extracted multi-view features from four different level (structural-level, constituent-level, textual level and topology-level). Additionally, we utilized a graph attention mechanism to combined the structural-level and constituent-level information as the statistical-view features to reduce the noise and redundancy, and explores the robustness and superiority of our model. Owing to these inherent limitations of single-view models, multi-view learning has recently emerged as an effective paradigm for integrating complementary biological information from different perspectives35. By jointly modeling structural, compositional, semantic, and topological features, multi-view frameworks enable a more comprehensive and robust characterization of complex microbial interactions. In this study, we propose a novel multi-view learning model named MVPHI, which integrated statistical-view features, protein textual-view features and graph topology-view features. More specifically, we first collect the ESKAPE (Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, and Enterobacter spp.) based bacteria and their target phages to build a heterogeneous multi-attributed microbial network (MAMN), from which we can capture multi-view information to characterize microbial nodes from different perspectives. We first used the composition/transition/distribution (CTD) and k-mer algorithm to extract the structural-level and constituent-level information of microbial genomes. Next, we introduce a graph attention mechanism to integrate them as statistical-view features. Then, considering the attribute information of protein sequences, we adopted a pre-trained natural language processing method, doc2vec, to aggregate neighborhood protein textual-level information. To further enhance the predictive power, we use a graph transformer network to capture the topology pattern features of MAMN. Following the preceding steps, we can obtain informative attribute representation, which incorporate the statistical-view, textual-view and topology-view information. Finally, a two-layer multilayer perceptron (MLP) is adopted to training and predicting the potential interactions between phages and target bacterial. Extensive experiments have performed and demonstrated that MVPHI can fully leverage multi-source biological information to improve the prediction accuracy. Additionally, we also conducted the case study of K. pneumoniae to further demonstrated the effectiveness of our model. Materials and methods In this section, we collected the ESKAPE pathogenic associated bacteria and phages to construct a heterogeneous multi-attributed microbial network (MAMN), which consisted of bacteria–bacteria interactions (BBIs) and phage–bacteria interactions (PBIs). Then, we introduce the different extraction methods to capture biological information from different perspectives. The workflow of the proposed MVPHI model is shown in Fig. 1. Fig. 1Over architecture of MVPHI model. a Constructing the heterogeneous network. b Capturing multi-view features from MAMN. c Model training and prediction. Over architecture of MVPHI model. a Constructing the heterogeneous network. b Capturing multi-view features from MAMN. c Model training and prediction. Dataset In this work, the performance of MVPHI model was comprehensively evaluated on six bacteria–bacteriainteractions (BBIs) and a phage–bacteria interactions (PBIs) dataset. The basic information about these datasets can be described as follows.Bacteria–bacteriainteractions datasets Bacteria–bacteriainteractions datasets Bacteria–bacteriainteractions datasets We constructed a benchmark BBIs datasets to explore the bacterial infections. The specific of these datasets is that it covers ESKAPE organisms, which constructed by E. faecium, S. aureus, K. pneumoniae, A. baumannii, P. aeruginosa, and Enterobacter spp36. In the hospital settings, these bacteria often cause serious infections, while exhibit degrees of resistance to various antibiotic drugs. These BBIs datasets are collected from STRING (https://cn.string-db.org/)37database, and the details of BBIs datasets are listed in Table 1. (b)Phage–bacteria interactions dataset (b)Phage–bacteria interactions dataset Phage–bacteria interactions dataset The phage–bacteria associated databases are downloaded from three public databases, including UniProt38 (https://www.uniprot.org/) and Millard Lab39 (https://millardlab.org) database. Finally, we collected 1232 PBIs pairs that consisted of 959 phages and 793 bacteria. The prediction of BBIs and PBIs datasets is essentially a binary classification task. To construct balanced datasets, the down-sampling technique was used to select the same number of pairs from unlabeled associations as negative samples. In this way, the negative samples may have uncharacterized interaction pairs, but considering the down-sampling ratio, we can ignore this statistical and machine learning perspectives. Table 1ESKAPE pathogenic corresponding BBIs datasets.BacteriaNodeSamplesDown-sampling E. faecium 175722,5340.73% S. aureus 183221,6690.65% K. pneumoniae 370767,9640.49% A. baumannii 198121,5360.55% P. aeruginosa 360848,7820.38%Enterobacter spp.271424,8700.34% Table 1ESKAPE pathogenic corresponding BBIs datasets.BacteriaNodeSamplesDown-sampling E. faecium 175722,5340.73% S. aureus 183221,6690.65% K. pneumoniae 370767,9640.49% A. baumannii 198121,5360.55% P. aeruginosa 360848,7820.38%Enterobacter spp.271424,8700.34% ESKAPE pathogenic corresponding BBIs datasets. Structural-level information The composition/transition/distribution (CTD) algorithm can describe the protein sequences globally, and was primarily developed for predicting protein fold class40. Here, composition information is the ratio of the composition of different amino acids in target sequences.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{i}^{c}=\frac{{{N_i}}}{{{N_{length}}}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{i}^{c}$$\end{document} characterizes the composition of the i-th amino acid. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{length}}$$\end{document} defined the count of i-th amino acid and the length of whole protein, respectively. Transition information describes the frequency of transitions between adjacent amino acid pairs.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{{i,j}}^{t}\frac{{{N_{i,j}}}}{{{N_{length}} - 1}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{{i,j}}^{t}$$\end{document} denotes the transition features between i-th and j-th amino acid. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{i,j}}$$\end{document} is the number of adjacent occurrences of the i-th and j-th amino acids in the protein sequences. Distribution information represents the distribution of distances between different amino acids.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{k}^{d}={N_k}/{N_{length}} - k$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{k}^{d}$$\end{document} is the distribution features of amino acid pairs with a distance of k and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_k}$$\end{document} represents the number of these amino acid pairs. Then we used these CTD information to convert each phages/bacteria sequence as 64-dimensional structural features, which can provide multifaceted information to characterize the structure of sequences more comprehensively. Constituent-level information In the field of bioinformatics, k-mer is a simple but powerful method for analyzing sequence, which is used to express the occurrence frequency of k neighbouring nucleic acids41. This model can capture the pattern information of local sequences. In this work, we applied 3-kmer to encode the protein sequence as a 64-dimensional vector to capture the constituent information from MAMN. The details calculation is shown as follows:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_i}=\frac{{{n_i}}}{{{N_{length}} - k+1}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_i}$$\end{document} represents the frequency of subsequence i occurs in length k, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n_i}$$\end{document} denotes the number of i, and k represents their length. Statistical-view feature fusion After extracting the statistically based information (CTD and k-mer), we obtain multi-view microbial features of phages and bacteria. In general, protein statistical information offers many useful insights. However, previous graph neural network-based strategies failed to utilize their potential power. Mean operation or concatenation method is the most direct way to integrate them. The achieved features may contain noise and redundancy. In addition, inappropriate method may lead to poor quality of the fusion features, and have negative influence on the models’ prediction performance. Thus, we used a graph attention (GA) mechanism to integrate these statistical features42.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( {\alpha _{i}^{1},\alpha _{i}^{2}, \cdots \alpha _{i}^{B}} \right){\mathrm{=}}Attention\cdot \left[ {\left( {x_{c}^{1},x_{c}^{2}, \cdots x_{c}^{B}} \right),\left( {x_{k}^{1},x_{k}^{2}, \cdots x_{k}^{B}} \right)} \right]$$\end{document} Here, A represents the attention, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_c}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_k}$$\end{document} represents the multi-source features, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _{i}^{{(t)}}=\left( {\alpha _{i}^{1},\alpha _{i}^{2} \cdots \alpha _{i}^{B}} \right)$$\end{document} denotes the fusion statistical features. The attention value of each statistical feature can be calculated as:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{i}^{{(t)}}=\delta \left( {{b^{(t)}}+{W^{(t)}}{{\left( {\alpha _{i}^{{(t)}}} \right)}^T}} \right){\left( {{q^{(t)}}} \right)^T}$$\end{document} where T, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b^{(t)}}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W^{(t)}}$$\end{document}represents the transposition, bias vector and trainable weight matrix. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta (\cdot )$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${q^{(t)}}$$\end{document} denotes the tanh activation function and attention vector respectively. After normalizing the attention value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{i}^{{(t)}}$$\end{document}, we can yield the weight of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _{i}^{{(t)}}$$\end{document} through a softmax layer:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{i}^{{(t)}}=\frac{{{e^{\left( {S_{i}^{{(t)}}} \right)}}}}{{\sum\nolimits_{{t=1}}^{B} {{e^{\left( {S_{i}^{{(t)}}} \right)}}} }}$$\end{document} The larger value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{i}^{{(t)}}$$\end{document}, the more importance of the association features. Next, we can define the statistical information of node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v^{(i)}}$$\end{document} based on the weight values:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L_i}=\sum\limits_{{t=1}}^{B} {S_{i}^{{(t)}}\cdot \alpha _{i}^{{(t)}}}$$\end{document} After the above calculation process, the structural-level and constituent-level microbial information can be integrated as the statistical-view features. Protein textual-view feature To fully mine the biological information of phages and bacteria, we used doc2vec to capture the textual features based on the context semantic features from protein sequences. In particular, each microbial sequence can be seen as a sentence and transformed biology functions that can be semantically interpreted by doc2vec. First, we performed a sliding window strategy with the sliding set of 5 and window size of 1 to split the microbial genome sequences. Subsequently, the segmented sentence fragments are used to train the classification model. Lastly, we can embed any protein sequence into an association fixed-size feature.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P_i}=c+\sum\nolimits_{{a \in se{q_i}}} {({v_a}+{u_i})}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P_i}$$\end{document} is the vector representation of protein i, c is a specific label vector which can denotes the semantics of the whole protein sequences, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v_a}$$\end{document} is the word vectors representation of amino acid a, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u_i}$$\end{document} represents the document vector. Graph topology-view feature Graph transformer network (GTN) is capable for characterizing novel graph topological information43. Unlike the above statistical (CTD and k-mer) or textual (doc2vec)-based methods, GTN is adopted to capture the topology information of MAMN. Taking the PBIs network topology-view feature extraction process as an example. Firstly, we employed an attention layer to calculate the attention weights.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Attention(Q,K,V)=soft{\mathrm{max}}\left( {\frac{{Q{K^T}}}{{\sqrt {{\operatorname{d} _k}} }}} \right)V$$\end{document} where Q, K and V represent the matrix of query, key and value, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d_k}$$\end{document} represents the dimensions of K and V. The softmax function is used to normalize the attention weights, and the weighted summation results are the characteristic representation of the microbial nodes. In addition, we introduced the multi-head attention mechanism.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Attention(Q,K,V)=concat\left( {hea{d_1},hea{d_2} \cdots hea{d_h}} \right){W^m}$$\end{document}12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$hea{d_i}=Attention\left( {QW_{i}^{Q},KW_{i}^{K},VW_{i}^{V}} \right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$hea{d_i}$$\end{document} represents the calculation result of i-th attention head. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{i}^{Q}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{i}^{K}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{i}^{V}$$\end{document} is the projecting weighting matrix of Q, K, and V, respectively. The output projecting matrix is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W^m}$$\end{document} and we utilized Concat to link results from multiple attention heads. After the linear transformation by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W^m}$$\end{document}, the final graph topology features of MAMN can be obtained. Model training and prediction In this work, we utilized a two-layer multilayer perceptron (MLP) to train the multi-view features and predicting the likelihood of various interactions. More specifically, given a phage (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${p_i}$$\end{document}) and a bacterium (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u_j}$$\end{document}), and their latent representations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z_j}$$\end{document} from the whole biological embedding Z. The prediction score of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${p_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u_j}$$\end{document} can be calculated by:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathop G\limits^{ \wedge } _{i,j}}={\delta _2}\left( {W_{2}^{T}\left( {{\delta _1}\left( {W_{1}^{T}\left( {\left[ {{z_i}\left\| {{z_j}} \right.} \right]} \right)+{b_1}} \right)} \right)+{b_2}} \right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\delta _1}( \cdot )$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\delta _2}( \cdot )$$\end{document} represent the ReLU activate function and softmax function, respectively. The higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathop G\limits^{ \wedge } _{i,j}}$$\end{document} represents that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${p_i}$$\end{document} is more likely to interact with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u_j}$$\end{document}. In addition, we adopted the binary cross-entropy loss function to optimize the proposed MVPHI model.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Loss= - \sum\limits_{{({p_i},{a_j}) \in {H^+} \cup {H^ - }}} {{G_{i,j}}\log {{\mathop G\limits^{ \wedge } }_{i,j}}+\left( {1 - {G_{i,j}}} \right)\log \left( {1 - {{\mathop G\limits^{ \wedge } }_{i,j}}} \right)}$$\end{document} where the positive and negative samples of training samples is represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H^+}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H^ - }$$\end{document}. When \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$({p_i},{a_j}) \in {H^+}$$\end{document}, the ground truth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${G_{i,j}}$$\end{document} is equal to 1, otherwise it will be set to 0. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathop G\limits^{ \wedge } _{i,j}}$$\end{document} is the prediction score of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${G_{i,j}}$$\end{document}. To obtain the better performance, we also used the Adam optimizer to minimize the loss function. Moreover, the dropout regularization technique is also used to reduce overfitting. The entire MVPHI can be described as Algorithm 1. Algorithm 1The overall process of MVPHI model The overall process of MVPHI model Results and discussion Evaluation metrics In this study, the PBIs and BBIs identification is actually a binary classification task. In order to assess the predictive ability of MVPHI, we used several evaluation indicators, including accuracy (ACC), recall (REC), precision (PRE), and F1-scoure (F1), which can be expressed as:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Accuracy=\frac{{TN+TP}}{{TN+FP+FN+TP}}$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{Precison}}=\frac{{TP}}{{FN+TP}}$$\end{document}17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Recall=\frac{{TP}}{{FP+TP}}$$\end{document}18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F1 - scoure=\frac{{precision \times recall}}{{recall+precision}} \times 2$$\end{document} where TN, TP, FP, and FN represent the number of true negatives, true positives, false positives, and false negatives, respectively. To translate the model performance more intuitively, we also used plot two common curves to intuitively indicate the prediction performance of MVPHI, including the receiver operating characteristic (ROC) curve and precision-recall (PR) curve. In addition, the area under ROC curves and PR curves were also calculated to display the performance more intuitive. Performance of MVPHI in PBIs and BBIs We validate the performance of MVPHI through 5-fold cross-validation (CV) framework. Taking the E. faecium BBIs task as an example, the whole E. faecium datasets were divided into 5 equal subsets. One subset is chosen for testing and the four remaining were selected as training set. All experiments were conducted on a workstation equipped with an Intel i7-0700 CPU, 32 GB RAM, and an NVIDIA RTX 5060 GPU. Under this setting, the average training time of MVPHI for one fold is approximately 5–8 min, depending on the dataset size. Table 2 summarizes the experimental results of MVPHI obtained on seven benchmark datasets. As can be seen that MVPHI achieves the average values of ACC with 0.8243, 0.8332, 0.8109, and 0.8401 on the E. faecium, A. baumannii, P. aeruginosa, and PBIs dataset, respectively. They show a relatively lower level on the S. aureus, K. pneumoniae, and Enterobacter spp. datasets with the ACC value of 0.8088, 0.8033, and 0.7955, respectively. The main reason may be due to the presence of some specific biological features or genetic variation that may not have been fully captured. However, it also achieves high AUC values with 0.8835, 0.8822, and 0.8718, respectively. The corresponding ROC and PR curves are presented in Fig. 2. Table 2Experimental results of MVPHI model performed on the ESKAPE−related datasets.DatasetACCPRERECF1AUCAUPR E. faecium 0.82430.81680.83670.82640.90090.9024 S. aureus 0.80880.80600.81400.80960.88350.8779 K. pneumoniae 0.80330.80070.80810.80410.88220.8770 A. baumannii 0.83320.83640.89900.83250.90270.9103 P. aeruginosa 0.81090.79880.83110.81450.88250.8771Enterobacter spp.0.79550.78210.81950.80030.87180.8774phage–bacteria0.84010.81810.87710.84570.91190.8890 Table 2Experimental results of MVPHI model performed on the ESKAPE−related datasets.DatasetACCPRERECF1AUCAUPR E. faecium 0.82430.81680.83670.82640.90090.9024 S. aureus 0.80880.80600.81400.80960.88350.8779 K. pneumoniae 0.80330.80070.80810.80410.88220.8770 A. baumannii 0.83320.83640.89900.83250.90270.9103 P. aeruginosa 0.81090.79880.83110.81450.88250.8771Enterobacter spp.0.79550.78210.81950.80030.87180.8774phage–bacteria0.84010.81810.87710.84570.91190.8890 Experimental results of MVPHI model performed on the ESKAPE−related datasets. Fig. 2ROC and PR curves obtained from MVPHI model on seven datasets. ROC and PR curves obtained from MVPHI model on seven datasets. Parameter optimization analysis In MVPHI model, parameter selection plays essential role to model fitting and performance, with learning rates and hidden units is even critical. In order to obtain the optimum model, we conducted optimization experiments for these two hyper parameters. For the parameter of learning rate, we increased it from 0.00001 to 0.01, and the experimental results of PBIs prediction task is shown in Fig. 3a. While these prediction values of ACC, PRE, REC, F1, AUC and AUPR indicators vary, they showed a pattern of first rising and then decline tendency. According to the experimental results, we selected the 0.001 as the optimal learning rate, at which MVPHI can obtain the best results. Similarly, we also conduct the experiments for searching the best hidden unites from the unites 16 to 256, and the comparison results are shown in Fig. 3b. Fig. 3Hyper-parametric optimization of MVPHI model. Hyper-parametric optimization of MVPHI model. Impact of iteration number In order to find the best combination of predictive performance and operating efficiency, we explore the impact of the parameter about iteration number on the PBIs prediction task. In the MVPHI model, the number of iterations is the key parameter of the local smoothing strategy. The optimal iterations number is able to fully learn the pattern in the data and avoid overfitting. Specifically, we set different iteration numbers including 5, 15, 50, 150, and 500 to train the model, and the rest parameters were unchanged. The impact of diverse iteration numbers is shown in Fig. 4. It can be observed that with the increase in the number of iterations, the AUC and AUPR values initially rise rapidly and then levelled off after 50 iterations. It can be concluded that when the iteration number is less than 50, MVPHI cannot achieve the optimal performance, whereas beyond 50 iterations, only minimal improvement is obtained. Additionally, exceeding 50 iterations increases the requirement of computational resources. Therefore, considering these factors, we choose the 50 iterations as the optimal training parameter. Fig. 4Line chart of the performance under diverse iteration numbers on the PBIs task. Line chart of the performance under diverse iteration numbers on the PBIs task. Model difference analysis As mentioned above, to obtained the better performance of MVPHI, we used multi-view feature formation. In this part, we conduct an ablation study to analyze the influence of different views. In particular, we designed six variant models of MVPHI as follows: MVPHI-G: A variant model of MVPHI that only used topology-view features to predict PBIs.MVPHI-S: A variant model of MVPHI that only used statistical-view features to predict PBIs.MVPHI-D: A variant model of MVPHI that only used textual-view features to predict PBIs.MVPHI-GS: A variant model of MVPHI that combined topology-view and statistical-view features.MVPHI-DG: A variant model of MVPHI that combined topology-view and textual-view features.MVPHI-DS: A variant model of MVPHI that combined textual-view and statistical-view features. MVPHI-G: A variant model of MVPHI that only used topology-view features to predict PBIs. MVPHI-G: A variant model of MVPHI that only used topology-view features to predict PBIs. MVPHI-S: A variant model of MVPHI that only used statistical-view features to predict PBIs. MVPHI-S: A variant model of MVPHI that only used statistical-view features to predict PBIs. MVPHI-D: A variant model of MVPHI that only used textual-view features to predict PBIs. MVPHI-D: A variant model of MVPHI that only used textual-view features to predict PBIs. MVPHI-GS: A variant model of MVPHI that combined topology-view and statistical-view features. MVPHI-GS: A variant model of MVPHI that combined topology-view and statistical-view features. MVPHI-DG: A variant model of MVPHI that combined topology-view and textual-view features. MVPHI-DG: A variant model of MVPHI that combined topology-view and textual-view features. MVPHI-DS: A variant model of MVPHI that combined textual-view and statistical-view features. MVPHI-DS: A variant model of MVPHI that combined textual-view and statistical-view features. To ensure the fairness, we adopt the same parameters for each variant model and performed them on the same A. baumannii BBIs dataset. As shown in Table 3, we can see that the ACC and AUC values of the variant models all decreased to some extent, which further demonstrated the effectiveness of our strategies. Obviously, the performance of MVPHI-S significantly decreased when only statistical-level feature used. However, when the statistical-level features combined with Textual-level (MVPHI-DS) or topology-level (MVPHI-GS) features, the performance has improved a lot. We attribute this phenomenon to that the statistical-level features have limited expressive power for complex PBIs networks. In addition, textual-level features can capture the evolutionary information of bacteria and phages, while topology-level features reflect the comprehensive relationships and connectivity patterns in microbial network structures. Table 3Ablation studies of MVPHI on A. baumannii dataset.ModelACCPRERECF1AUCAUPRMVPHI-G0.79840.83950.73960.78550.86690.8804MVPHI-S0.59570.85520.23140.35750.60210.5960MVPHI-D0.80210.83600.75210.79160.86870.8838MVPHI-GS0.80270.79710.81510.80520.87590.8819MVPHI-DG0.81220.84070.77230.80440.87940.8918MVPHI-DS0.81140.79710.83550.81590.87880.8715MVPHI0.84010.81810.87710.84570.91190.8890 Table 3Ablation studies of MVPHI on A. baumannii dataset.ModelACCPRERECF1AUCAUPRMVPHI-G0.79840.83950.73960.78550.86690.8804MVPHI-S0.59570.85520.23140.35750.60210.5960MVPHI-D0.80210.83600.75210.79160.86870.8838MVPHI-GS0.80270.79710.81510.80520.87590.8819MVPHI-DG0.81220.84070.77230.80440.87940.8918MVPHI-DS0.81140.79710.83550.81590.87880.8715MVPHI0.84010.81810.87710.84570.91190.8890 Ablation studies of MVPHI on A. baumannii dataset. Comparison with state-of-the-art algorithms To further verify the necessity of each view, we make a compared MVPHI with some baselines algorithms which had been widely used in bioinformatic tasks: i.Laplacian eigenmaps (LAP): Lap extract the manifold topology information of biological data, facilitating improved the performance in various bioinformatics tasks.ii.Large-scale information network embedding (Line): Line captures both first-order and second-order proximities in large-scale microbial network, and preserving the local and global structure information.iii.Structural Deep Network Embedding (SDNE): SDNE uses the deep autoencoders to extract the topology information, leveraging the non-linear capabilities for more accurate embeddings.iv.High-Order Proximity preserved Embedding (Hope): Hope preserves higher-order proximity by capturing global network structure information.v.Support Vector Machine (SVM): SVM is a powerful binary classification model that find a hyperplane in a separating feature space.vi.Random Forest (RF): RF is a bagging algorithm that aggregating multiple decision trees to reduce overfitting and improve predictive performance.vii.K-nearest neighbor (KNN): KNN is a majority-voting model that calculates the distance between the known samples and unknown samples to find the possible samples.viii.Gradient Boosting Decision Tree (GBDT): GBDT employ an ensemble learning model based on decision trees, which is an additive model that estimated by gradient descent. i.Laplacian eigenmaps (LAP): Lap extract the manifold topology information of biological data, facilitating improved the performance in various bioinformatics tasks. Laplacian eigenmaps (LAP): Lap extract the manifold topology information of biological data, facilitating improved the performance in various bioinformatics tasks. ii.Large-scale information network embedding (Line): Line captures both first-order and second-order proximities in large-scale microbial network, and preserving the local and global structure information. Large-scale information network embedding (Line): Line captures both first-order and second-order proximities in large-scale microbial network, and preserving the local and global structure information. iii.Structural Deep Network Embedding (SDNE): SDNE uses the deep autoencoders to extract the topology information, leveraging the non-linear capabilities for more accurate embeddings. Structural Deep Network Embedding (SDNE): SDNE uses the deep autoencoders to extract the topology information, leveraging the non-linear capabilities for more accurate embeddings. iv.High-Order Proximity preserved Embedding (Hope): Hope preserves higher-order proximity by capturing global network structure information. High-Order Proximity preserved Embedding (Hope): Hope preserves higher-order proximity by capturing global network structure information. v.Support Vector Machine (SVM): SVM is a powerful binary classification model that find a hyperplane in a separating feature space. Support Vector Machine (SVM): SVM is a powerful binary classification model that find a hyperplane in a separating feature space. vi.Random Forest (RF): RF is a bagging algorithm that aggregating multiple decision trees to reduce overfitting and improve predictive performance. Random Forest (RF): RF is a bagging algorithm that aggregating multiple decision trees to reduce overfitting and improve predictive performance. vii.K-nearest neighbor (KNN): KNN is a majority-voting model that calculates the distance between the known samples and unknown samples to find the possible samples. K-nearest neighbor (KNN): KNN is a majority-voting model that calculates the distance between the known samples and unknown samples to find the possible samples. viii.Gradient Boosting Decision Tree (GBDT): GBDT employ an ensemble learning model based on decision trees, which is an additive model that estimated by gradient descent. Gradient Boosting Decision Tree (GBDT): GBDT employ an ensemble learning model based on decision trees, which is an additive model that estimated by gradient descent. To provide a fair comparison result, these baselines algorithms adopt the default parameter and performed on the same PBIs dataset. The details prediction results of 5-fold CV are shown in supplementary Table S1–S9. We can see that the proposed MVPHI model achieved the best predictive results with the average AUC of 0.9119 and AUPR of 0.8890, which are 1.41% and 1.83% better than that of the second-best LAP-based model. In terms of ACC and F1 value, it is remarkably better (at least 1.2% and 1.12%) than these eight comparison algorithms (Fig. 5). This phenomenon can be attributed to the introduce of graph attention mechanism that can better aggregate the representations of neighbors. These comparison results further confirmed that MVPHI is an efficient and robust model for exploring the complex associations between phages and target bacteria. Fig. 5Box plots of different comparison methods under the PBIs datasets. Box plots of different comparison methods under the PBIs datasets. Case study To further evaluate the effectiveness and biological relevance of MVPHI, we conducted case studies on Klebsiella pneumoniae and Staphylococcus aureus, two clinically important pathogens belonging to the ESKAPE group and major contributors to multidrug-resistant infections worldwide. These species are representative targets for phage therapy and therefore provide biologically meaningful benchmarks for validating PBIs. In the K. pneumoniae case study, all known phages that associated with K. pneumoniae were excluded from the training data and used solely for testing, while the remaining data were used to train the model. After prediction, all candidate phages were ranked by their predicted interaction scores with K. pneumoniae, and the top 20 phages were selected for further validation. As shown in Table 4, among these top-ranked candidates, 17 phages were verified through publicly available databases, while 3 phages (ranked 3rd, 9th, and 15th) were not previously reported, indicating potential novel phage-host associations. Similarly, in the S. aureus case study, we applied the same leave-out strategy by removing all phages interacting with S. aureus from the training data. The top 20 high-scoring phages were then selected based on their predicted interaction probabilities and compared with known database records. As presented in Table 5, 15 out of the 20 predicted phages were successfully confirmed, while the remaining 5 lacked existing experimental validation, suggesting their potential as novel candidates for future biological investigation. Table 4The top 20 predict K. pneumoniae-interacted phages ranked by predicted scores.RankPhage IDScoreEvidenceRankPhage IDScoreEvidence1A0A0A8J9 × 40.996038Confirmed11A0A0A8JA060.869535Confirmed2A0A0A8JA020.994911Confirmed12D1L2 × 00.851551Confirmed3BBI55720.10.990348N/A13K9L8R50.850821Confirmed4A0A0A8JBQ80.982874Confirmed14D1L2Z40.838986Confirmed5A0A0A8J9A50.977797Confirmed15A0A1W5LJ610.836008N/A6A0A0A8J9V70.967909Confirmed16V9QKP60.835699Confirmed7A0A0A8J8T20.953412Confirmed17A0A068Q6B20.832012Confirmed8A0A0A8JBR20.948926Confirmed18D5JFT20.809896Confirmed9A0A0M7QAU40.926474N/A19A0A0A8J9W20.804612Confirmed10A0A0A8J9B00.922198Confirmed20A0A0A8J8S80.798332Confirmed Table 4The top 20 predict K. pneumoniae-interacted phages ranked by predicted scores.RankPhage IDScoreEvidenceRankPhage IDScoreEvidence1A0A0A8J9 × 40.996038Confirmed11A0A0A8JA060.869535Confirmed2A0A0A8JA020.994911Confirmed12D1L2 × 00.851551Confirmed3BBI55720.10.990348N/A13K9L8R50.850821Confirmed4A0A0A8JBQ80.982874Confirmed14D1L2Z40.838986Confirmed5A0A0A8J9A50.977797Confirmed15A0A1W5LJ610.836008N/A6A0A0A8J9V70.967909Confirmed16V9QKP60.835699Confirmed7A0A0A8J8T20.953412Confirmed17A0A068Q6B20.832012Confirmed8A0A0A8JBR20.948926Confirmed18D5JFT20.809896Confirmed9A0A0M7QAU40.926474N/A19A0A0A8J9W20.804612Confirmed10A0A0A8J9B00.922198Confirmed20A0A0A8J8S80.798332Confirmed The top 20 predict K. pneumoniae-interacted phages ranked by predicted scores. To further investigate the potential relationships between confirmed and unconfirmed phages, we computed the cosine similarity between their learned feature representations derived from the model44. Cosine similarity is widely used as a measure of vector orientation and has been shown to reflect potential functional similarity when applied to protein representations; in general, a higher cosine similarity between two protein sequences implies a greater likelihood of functional or structural resemblance. The resulting similarity matrices were visualized as heatmaps, as shown in Fig. 6. Specifically, Fig. 6a presents the similarity scores between K. pneumoniae-associated confirmed phages and three unconfirmed candidates, while Fig. 6b illustrates the corresponding results for S. aureus. Notably, several unconfirmed phages exhibited high similarity scores with known phages, suggesting potential functional convergence. These observations not only support the biological plausibility of our model’s predictions but also highlight the value of similarity-based evidence in identifying novel phage-host interactions that warrant further experimental validation. Fig. 6Heatmaps of cosine similarity between confirmed and unconfirmed phages. Heatmaps of cosine similarity between confirmed and unconfirmed phages. Table 5The top 20 predicted S. aureus-interacted phages ranked by predicted scores.RankPhage IDScoreEvidenceRankPhage IDScoreEvidence1A0A0K1LKC90.98939Confirmed11YP_002332534.10.98203Confirmed2E9LT630.98896Confirmed12AXY82734.10.98112N/A3A0A0N9BAZ10.98762Confirmed13A0A0E4BZU60.97958Confirmed4AUS03394.10.98706Confirmed14YP_009006774.10.97758Confirmed5AUS03379.10.98676Confirmed15A0A126DJD50.97694N/A6ATN93936.10.98651Confirmed16YP_001004393.10.97603Confirmed7A0A023ZWA70.98554N/A17A0A220BYL50.97582Confirmed8A0A097J2320.98551N/A18AUS03378.10.97244Confirmed9A0A1D6Z2830.98534Confirmed19A0A160CBC20.97119N/A10BAO47150.10.98455Confirmed20S4SUQ80.968686Confirmed Table 5The top 20 predicted S. aureus-interacted phages ranked by predicted scores.RankPhage IDScoreEvidenceRankPhage IDScoreEvidence1A0A0K1LKC90.98939Confirmed11YP_002332534.10.98203Confirmed2E9LT630.98896Confirmed12AXY82734.10.98112N/A3A0A0N9BAZ10.98762Confirmed13A0A0E4BZU60.97958Confirmed4AUS03394.10.98706Confirmed14YP_009006774.10.97758Confirmed5AUS03379.10.98676Confirmed15A0A126DJD50.97694N/A6ATN93936.10.98651Confirmed16YP_001004393.10.97603Confirmed7A0A023ZWA70.98554N/A17A0A220BYL50.97582Confirmed8A0A097J2320.98551N/A18AUS03378.10.97244Confirmed9A0A1D6Z2830.98534Confirmed19A0A160CBC20.97119N/A10BAO47150.10.98455Confirmed20S4SUQ80.968686Confirmed The top 20 predicted S. aureus-interacted phages ranked by predicted scores. Molecular docking To further verify the biological plausibility of the experimental results in case studies, we conducted molecular docking experiments on phages that have not been confirmed by public databases. For more specific, we first used the AlphaFold 345 sever to generate the three-dimensional structures of phages and bacteria hosts. Subsequently, we performed rigid-body docking simulations through the GRAMM server, and visualized the resulting binding interfaces using PyMOL46. Binding energy is a key indicator which represents the change in free energy when two molecules form a stable complex via non-covalent interactions47. In protein-protein docking, it reflects the energetically favourable nature of the predicted binding pose. The more negative the binding energy, the stronger and more stable the interaction. Based on experience, binding energies below -5 kcal/mol are generally considered to indicate biologically meaningful interactions. As shown in Fig. 7a-c, the computed binding energies ranged from −6.6 to −25.0 kcal/mol. Among them, A0A1W5LJ61 displayed the most favourable binding profile with a binding energy of −25.0 kcal/mol, which suggested a highly stable and potentially functional interaction with the bacterial host. Similarly, Fig. 7d-h illustrates the docking outcomes for five unconfirmed phages predicted to target Staphylococcus aureus, with binding energies ranging from − 7.0 to −25.1 kcal/mol. The lowest binding energy was observed for AXY82734.1 (−25.1 kcal/mol), indicating a strong interaction with the host receptor. These results offer important structural and energetic support for the computational predictions, suggesting that even phage–host pairs lacking prior experimental evidence may represent true biological interactions worthy of future validation. Fig. 7Molecular docking results of unconfirmed phage-host pairs. (a-c) Unconfirmed phages predicted by MVPHI to interact with K. pneumoniae. (d-h) Unconfirmed phages predicted by MVPHI to interacted with S. aureus. Molecular docking results of unconfirmed phage-host pairs. (a-c) Unconfirmed phages predicted by MVPHI to interact with K. pneumoniae. (d-h) Unconfirmed phages predicted by MVPHI to interacted with S. aureus. Conclusion In this study, we proposed a novel computational model named MVPHI to predict PBIs and BBIs through multi-view features. Specifically, we constructed a heterogeneous multi-attributed microbial network (MAMN) based on ESKAPE pathogenic bacteria and phages, which can help researchers to elucidate the intricate relationship between pathogens and target viruses. Next, MVPHI fully leverages multi-view features to release the best performance of MAMN. To this end, we used CTD and k-mer algorithm to extract the structure and constituent information from the evolutionary information of microbial nodes. Then a graph attention mechanism was adopted to combine structure and constituent information as statistical features. The doc2vec algorithm is performed to extract protein textual features from amino acid sequences, and a graph transformer network is also used to capture the topological features of MAMN. MVPHI designed a two-layer MLP to train these multi-type features and predict potential associated pairs. Extensive experimental results demonstrate that the proposed MVPHI model significantly outperforms the state-of-the-art methods. Additionally, to verify the efficacy of MVPHI in real-life problems, the experiment of case study on K. pneumoniae and molecular docking experiments further indicated that our method is effective and reliable. MVPHI is expected to facilitate in-depth understanding of microbial genomics and advance research of phages discovery. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1