High-resolution phage-host assignment through key proteins using large language models

Abstract

Viral sequences in diverse environments remain largely uncharacterized, impeding our comprehension of their genetic makeup, biological interactions, and potential applications. This underscores an urgent need for innovative analytical methods. Here, we present the VirHost Hunter framework, which employs phage tails and lysins, bypassing the requirement for full genomes, for efficient and high-resolution host assignment. By harnessing Protein Language Models and Vision Transformers, VirHost Hunter captures protein functional homology despite sequence dissimilarity, significantly boosting prediction accuracy. In the scenario of disease-associated gut bacteria, the calibrated VirHost Hunter surpasses existing methods, doubling phage host assignments, expanding taxonomic reach, and revealing previously uncharacterized phages targeting gut bacteria, including Akkermansia and Prevotella. Therefore, we establish a gut phage lysin database, enabling the synthesis of a lysin that effectively and specifically targets an obesity-promoting bacterium. VirHost Hunter’s precision and scalability mark a significant leap forward in virome research and present a promising avenue for microbiome therapies.

Here, the authors present VirHost Hunter, an AI-based approach to phage–host assignment using tail and lysin proteins, showing it improves host resolution, expands functional discovery of gut phage, and enables targeted lysin identification for microbiome research.

Introduction

Virome is a significant component of Earth’s ecosystems and has a profound impact on ecological and human health. In various environments, uncharacterized viral genomes and sequences widely exist due to limitations in current analytical techniques and are referred to as viral dark matter. This concept highlights the need for innovative approaches to uncover and understand these hidden viral entities¹. The intricate interplay between bacteria and their viruses - bacteriophages (phages)—has garnered significant attention in recent years, fueled by advances in predictive modeling and therapeutic applications. Identifying the host range of phages is essential in studying phage resistance of bacteria, coevolution of phage-bacteria^2,3, the influence of community context on phage-bacteria systems⁴, and the role of phages in human health and diseases^5,6. In clinical settings, phages have already been adopted to treat infections caused by drug-resistant bacteria, offering advantages in precision medicine due to their host specificity and minimal disturbance of normal gut flora^7,8. Whereas phages also hold great promise in modulating gut microbiota, their efficacy hinges on the availability of phages targeting gut bacteria, particularly those associated with chronic diseases. Only a handful of phages have been reported targeting gut anaerobes, and it has been implicated that isolating gut phages is arduous^9,10.

Phage lysins have demonstrated effective antimicrobial effects in animal models, food industry, and clinical therapies^11–15, presenting broad industrial and medical applications. Lysins often exhibit a broader range of specificity than phages, which typically target at the genus or species level, and are easy to synthesize, making them especially suitable under scenarios where phages are unavailable or where available phages are too host-specific to apply¹⁶. However, existing lysin databases mainly focus on clinical pathogens that cause infectious diseases and have not been developed to target the broader gut microbial community¹⁷, limiting their application in gut microbiota modulation. Predicting phage hosts and establishing a phage lysin database by leveraging gut phage databases to specifically target gut bacteria serves as an alternative solution.

Various computational approaches have been developed for predicting phage hosts, falling into two categories: alignment-dependent and alignment-free methods (Table S1). Alignment-dependent methods rely on phage marker genes^18–20, phage-host relatedness²¹, and CRISPR spacers^22,23, but they have limitations such as database size, strong data source dependency, alignment parameters (e.g., sequence similarity thresholds, allowed mismatches/gaps), and applicability only to phages with specific marker genes or CRISPR signals^24,25. Alignment-free approaches utilize phage-bacterium interaction matrices^26–37, phage whole genomes^29,38 or sequences of receptor-binding proteins (RBPs)^39,40 to predict phage hosts. Notably, Gonzales et al.⁴⁰. utilized protein language models (PLMs) for feature extraction from RBPs, highlighting the potential of computational techniques in this area. Protein language models (PLMs) are computational models inspired by techniques developed in the field of natural language processing (NLP). This approach significantly improves contextual understanding and enables the identification of complex patterns that were previously difficult to discern.

Current viral databases predominantly use alignment-dependent methods and CRISPR spacers for host assignment, resulting in incomplete coverage and limited recall. For instance, Paez-Espino et al.⁴¹ identified 9992 putative virus-host associations covering only 7.7% of metagenomic viral contigs (mVCs) in their study of Earth’s virome⁴¹. In the past three years, several human gut virome databases have also been released. The Metagenomic Gut Virus (MGV, 2021) database assigned hosts to 81% (n = 153,892) of the phages, followed by 69% (n = 31,259) in the Cenote Human Virome Database (CHVD, 2021), 42% (n = 13,954) in the Gut Virome Database (GVD, 2020), and 29% (n = 40,932) in the Gut Phage Database (GPD, 2021)^42–45. The GPD had the most stringent criteria, resulting in the lowest recall of the four databases, i.e., it only utilized CRISPR spacers from 2898 high-quality genomes of cultured human gut bacteria and tolerated zero mismatches across the whole length of the spacers. Therefore, a high- quality alignment-free method, with improved machine learning models and input features, can be complementary to the CRISPR spacers method to increase the sensitivity of host prediction without compromising precision.

Indeed, a recent tool, iPHoP⁴⁶, integrates alignment-dependent and alignment-free methods for host prediction, including Blast⁴⁷, CRISPR^22,23, WIsH²⁶, VHM^31,48, and PHP³². Although iPHoP is a comprehensive framework integrating six approaches for phage host prediction, the authors discussed its limitations, including slow running time and the fact that it only achieves genus-level resolution, which may impact its practical applications. Alignment-free computational methods based on host-specific proteins such as tails and lysins, instead of the whole genomes of phages, offer a complementary strategy. In some contexts, these methods can be as effective or even more practical than whole-genome-based methods. There are several advantages: (1) they require minimal data input, avoiding vast redundant information and overuse of computing resources; (2) they can make predictions based on incomplete genome assemblies resulting from metagenome or metavirome sequencing, as long as key host-interacting proteins, such as tails or lysins, are present; (3) they can achieve high-resolution host prediction, potentially at the species or strain level, for phage therapy applications^49–54; and (4) they facilitate applications in synthetic biology, including host range modulation by swapping or engineering phage RBPs^50,55, delivery vehicles based on proteins recognizing and attach to host surfaces⁵⁶, and therapeutic agents based on lytic proteins breaking down bacterial cell walls⁵⁷.

In this study, we develop VirHost Hunter, a framework for phage host prediction that combines protein and DNA feature representations, constructs datasets from specific proteins, and leverages Protein Language Models together with a Vision Transformer to enhance prediction performance. We verify the roles of each key component of our design by conducting control analyses, followed by a comprehensive comparison to other methods across family to species levels. We calibrate the model to facilitate its application towards disease-associated gut bacteria and validate its robustness under targeted scenarios. We apply the calibrated model to the GPD and identify a great number of phages targeting disease-associated gut bacteria, including previously uncharacterized phages targeting renowned bacterial species whose phages have hardly been characterized before. To further promote the application of the resource, we extract lysins from the GPD with expanded host assignments to establish a repository. As a proof of concept, we select a lysin from the repository and synthesize it to verify its function against an obesity-promoting bacterium. This work elucidates the design of a predictive framework for phage host prediction and provides insights into how to utilize machine learning to advance genomics data mining and protein function prediction. Deciphering gut phages using this tool not only enhances our understanding of phage diversity and phage-bacteria interactions but also facilitates the downstream application of gut phage resources into disease intervention.

Results

Designing a phage host prediction framework

To predict the host of phages, full genome sequences of phages and bacteria are usually used. However, whole-genome-based methods introduce a significant amount of nonessential data, including proteins unrelated to host recognition or infection, which can create noise and interfere with the prediction accuracy, resolution, and efficiency. Concentrating on phage proteins conferring specificity—those directly involved in the infection cycle—offers a more targeted approach. In our previous study⁵⁸, we showed that models using only tail protein sequences achieved performance comparable to those using full viral genomes. Some methods have utilized receptor-binding proteins (RBPs) to predict hosts^39,40, but it can be challenging to annotate RBPs for many phages. We initially counted the number of RBPs and tail proteins in 7598 phage genomes from NCBI (December 29, 2021), revealing an average of 1.33 RBPs and 15.24 tail proteins per phage (Fig. S1). Additionally, other tail components also mediate specific phage-host interactions and are structurally associated with RBPs, collectively influencing host recognition specificity⁵⁹. Therefore, we expanded the dataset to include specific proteins beyond tail fibers and tail spikes: non-RBPs of phage tails, such as tail sheath, tail tube, baseplate, and tail collar proteins; and lysins, which are enzymes highly active against bacterial cell wall⁶⁰. These proteins are key to the infection cycle, while more widely annotated and included for prediction as well.

Proteins may share low sequence similarity while still performing similar functions across diverse species, rendering traditional sequence alignment methods less effective in capturing these functional similarities. To overcome the challenge of predicting host specificity using these proteins, particularly when sequence similarity is low, we employed protein language models (PLMs). ProtT5 was chosen because of its superior performance in functional prediction tasks and exceptional ability in handling low-homology sequences, as demonstrated by Gonzales et al., who compared multiple PLMs⁴⁰. PLMs provide a powerful solution by learning deep contextual and functional patterns within protein sequences, enabling them to capture viral protein function and viral biology even in cases of minimal sequence homology⁶¹.

Because the same protein sequence can be encoded by different DNA sequences, we incorporate DNA sequence features of tail proteins and lysins into the framework. DNA sequences provide additional insights into phages’ genomic context, such as codon usage bias, GC content, and nucleotide frequency, which can further refine predictions by accounting for genomic stability and evolutionary constraints³⁹. To uncover long-range dependencies and global patterns in DNA sequence data, we utilized a Vision Transformer (ViT)⁶², a language model capable of handling variable-length sequences. As a language model, ViT⁶², through its self-attention mechanism, captures intricate patterns and contextual relationships within DNA sequences, such as nucleotide interactions that influence genetic structures and functions. These patterns can reveal insights into genetic structures, functions, or relationships that are not easily discerned by examining individual sequences alone.

As a result, we present the VirHost Hunter framework with the above characteristics (Fig. 1A). It consists of two primary components: a feature extractor and a classifier. The feature extractor integrates protein sequence embeddings from the ProtT5 model⁶³, along with physicochemical and k-mer features derived from DNA sequences via a deep neural network (DNN)⁶⁴ comprising a three-path convolutional neural network (CNN; Fig. S2) and ViT for multi-scale feature extraction. By jointly leveraging protein and DNA representations, the VirHost Hunter implements a multimodal approach for robust host prediction. The final classification step uses a multi-layer perceptron (MLP) and a Random Forest (RF) classifier⁶⁵, with RF refining high-confidence predictions from the MLP to improve accuracy. We next investigate whether combining protein and DNA features, constructing datasets from specific proteins, and using language models such as PLMs and ViT enhances host prediction as expected, respectively.

Fig. 1. The design, validation, and application of the VirHost Hunter framework.

A The model framework, including Feature Extractor and Classifiers. B Model validation with multi-taxonomic datasets, disease-associated datasets and previous experimental evidence. C Enhancement of the host assignment ratio of the Gut Phage Database. D Genetic analyses of the refined Gut Phage Database. E Construction of a gut phage lysin database with host assignment. F Synthesis and verification of lysin targeting disease-associated gut bacteria. Phage icons used in panels (A, C) were adapted from Servier Medical Art (https://smart.servier.com/smart_image/bacteriophage) and are licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).

Using both protein and DNA features improves learning over either alone

To evaluate whether integrating protein and DNA features offers superior performance compared to using either individually, we conducted ablation experiments using two benchmark datasets: the Bacteriophage RBP (Drug-Resistant receptor-binding proteins, DRRBP) dataset (n = 4845)³⁹ and the Bacteriophage Tail Proteins (Drug-Resistant tail, DRTail) dataset (n = 12,509). We measured performance using accuracy (ACC), precision, and F1 scores under three experimental conditions: using only protein features, using only DNA features, and using a combination of both. It is shown that relying on a single type of feature led to inconsistent model performance across different datasets (Fig. 2A). Particularly, in the DRRBP dataset, models that used only protein features outperformed those that used only DNA features. In contrast, for the DRTail dataset, DNA features alone provided better performance than protein features. This inconsistency reveals the limitations of using only one feature type, as neither approach fully captures the complexity of phage-host interactions.

Fig. 2. Contribution of feature integration, protein-specific datasets, and language models to VirHost Hunter performance.

A Accuracy (ACC), precision, and F1 of VirHost Hunter and two individual base modules on DRRBP and DRTail datasets. VirHost Hunter-both features: combining protein and DNA features. VirHost Hunter-only protein: utilizing protein features only, VirHost Hunter-only DNA: utilizing DNA features only. Data are presented as box plots derived from independent cross-validation experiments (n = 10 folds), conducted on the DRRBP dataset (n = 4845 proteins) and the DRTail dataset (n = 12,509 proteins), where the center line indicates the median, the box boundaries represent the 25th and 75th percentiles, and whiskers extend to 1.5 times the interquartile range (IQR)(Source Data). B ACC of VirHost Hunter across a series of sequence similarities for specific (tail, lysin) and non-specific (head, terminase) proteins at genus levels. C ACC of different methods across a series of sequence similarities of phage tail proteins at the genus level.

On the other hand, integrating both protein and DNA features consistently improved model performance across all datasets and metrics. For instance, using both feature sets together resulted in the highest performance, with an accuracy of 0.9081 and 0.8927, precision of 0.9090 and 0.8930, and F1 scores of 0.9077 and 0.8925 on the DRRBP and DRTail datasets, respectively, significantly outperforming models that used either feature set alone. This demonstrates that integrating protein and DNA features not only enhances predictive accuracy but also provides greater consistency and stability across datasets, particularly in the context of bacteriophage host prediction.

Phage tail components and lysins drive host prediction without full-genome data

To confirm that using all tail components - RBPs, tail sheath, tail tube, baseplate, tail collar, etc - for host prediction is feasible, we conducted a 10-fold cross-validation on the DRRBP and DRTail datasets using our method, DeepHost (CNN-based DNA feature processing)²⁹, Boeckaerts et al.’s method (RF with handcrafted features)³⁹, and M. Gonzales et al.’s method (ProtT5 protein embeddings with RF)⁴⁰ (Table S2). The results suggest that host prediction accuracy through all tail proteins is comparable to that via RBPs, emphasizing the utility of incorporating all tail proteins since they exist 10 times more than RBPs. Additionally, VirHost Hunter outperformed the other methods, achieving an accuracy of 0.9081 and 0.8927, precision of 0.9090 and 0.8930, and F1 scores of 0.9077 and 0.8925 on the DRRBP and DRTail datasets, respectively (Table S2).

To compare the efficacy of phage tails and lysins with that of non-specific proteins for host prediction, we conducted an evaluation using head proteins and terminases from the same phage datasets. Phage tails and lysins consistently outperformed head proteins and terminases across family, genus, and species levels (Figs. 2B, S3). Across all tested sequence similarity thresholds, phage tails achieved the highest accuracy, followed closely by lysins. In contrast, head proteins and terminases reached significantly lower accuracy, with a notable decline in performance at lower sequence similarity thresholds, especially at the species level (Fig. S3). This highlights that phage tails and lysins maintain their predictive power, even at reduced sequence similarity, unlike the non-specific control proteins. We further demonstrated that VirHost Hunter can reach species-level resolution and outperformed existing methods in accuracy, precision, and F1 score on a multi-taxonomic dataset of 7598 phage genomes (Fig. 1B, Supplementary Results, Fig. S4A and Tables S3, S4). As shown, the training loss decreases steadily and converges to near zero, while the validation loss stabilizes without divergence, indicating that the model achieves good generalization and does not suffer from overfitting (Fig. S5).

Functional homology is captured even in low-similarity sequences

To demonstrate the ability of VirHost Hunter to capture functional homology using protein language model (PLM) embeddings and DNA features extracted via a Vision Transformer (ViT), we evaluated its performance across datasets with varying sequence similarities. Using CD-HIT⁶⁶, we partitioned the multi-taxonomic dataset into subsets with sequence similarity thresholds of 50%, 60%, 70%, 80%, and 90%, enabling us to assess VirHost Hunter’s capability of predicting phage host based on functional relationships rather than strict sequence homology.

We benchmarked VirHost Hunter against existing models, including Boeckaerts et al.³⁹, DeepHost²⁹, and M. Gonzales et al.⁴⁰, across different similarity thresholds. VirHost Hunter consistently outperformed the other methods across taxonomic ranks—family, genus, and species—highlighting its superior ability to leverage functional homology in low-similarity datasets (Figs. 2C, S6).

Crucially, as sequence similarity decreased, the performance gap between VirHost Hunter and the other methods widened, particularly at the family and genus levels. This underscores the increasing importance of capturing functional homology in low-similarity regions, where conventional sequence similarity-based methods typically fail. VirHost Hunter’s integration of protein language models (PLMs), such as ProtT5, and DNA sequence features enables it to move beyond reliance on sequence similarity alone. Instead, it identifies deeper functional relationships, resulting in robust and accurate predictions, even under low similarity conditions.

Robust phage host prediction for application under targeted scenarios

Given the substantial impact of gut bacteria on human health, such as inflammatory bowel disease (IBD)^67,68, colorectal cancer^69–71, and metabolic diseases^72–76, obtaining more information on phages targeting these bacteria is advantageous. We can expand our knowledge of gut phage-bacteria interactions, gut phage diversity, and facilitate their therapeutic application. Phage information can be obtained either through co-culturing with a bacterial host or mining data from high-throughput sequencing. However, gut phages, especially those targeting obligate anaerobes, are hard to culture and isolate. Investigating gut phages by analyzing sequencing data is, therefore, usually considered more efficient. Having validated the superior performance of VirHost Hunter, including accuracy, precision, and resolution, we next evaluate its effectiveness in identifying phages targeting disease-associated gut bacteria. We compiled a dataset consisting of 60 gut bacterial species associated with various diseases, including carotid atherosclerosis, inflammatory bowel disease (IBD), and obesity (Fig. 1B, Supplementary Data 1). We annotated prophage tails and lysins from the dataset, yielding a total of 328,701 unique tail proteins and 312,565 lysins. We calibrated the VirHost Hunter model using these sequences across 29 families, 40 genera, and 60 species.

Consistent with previous evaluations, VirHost Hunter outperformed the other three tested methods when applied to this dataset. At the family, genus, and species levels, VirHost Hunter-tail (based on gut phage tails) yielded ACC scores of 0.9516, 0.935, and 0.9132, Precision scores of 0.9513, 0.9341, and 0.9112, and F1 scores of 0.9512, 0.9342, and 0.9115, respectively (Fig. S4B, Table S5). VirHost Hunter-lysin (based on gut phage lysins) exhibited ACC scores of 0.9817, 0.9756, and 0.9590, Precision scores of 0.9817, 0.9755, and 0.958, and F1 scores of 0.9817, 0.9755, and 0.9582, respectively (Fig. S4B, Table S6). We also examined how sample sequence similarity would affect model performance. VirHost Hunter consistently outperformed other methods across all similarity thresholds and taxonomic levels (Fig. S7), further highlighting its robustness in predicting gut phage hosts associated with chronic diseases.

To further validate VirHost Hunter’s performance on isolated gut phages, we used a previously reported collection of cultivated gut phages⁷⁷ targeting Bacteroides, Bifidobacterium, Clostridium, Dorea, Eggerthella, Enterococcus, Phocaeicola, Streptococcus, and Lactococcus (Fig. 1B). 702 tail proteins and 373 lysins were extracted from 156 gut phages, all with experimentally verified host data. To ensure comparability with existing CRISPR-based host prediction benchmarks, we adopted the same precision thresholds reported by Dion et al.²². In their study, allowing different numbers of mismatches led to trade-offs between precision and recall: E-value of 10⁻⁹ led to the highest precision of 95% and the lowest recall of 2.5%, zero mismatches resulted in 84% precision and 31% recall, while allowing two mismatches yielded 69% precision and 49% recall. These thresholds were then chosen for evaluation and comparison. At a 95% precision cutoff, VirHost Hunter correctly identified hosts for 73/156 phages at the family level and 58/156 at the genus level, while the CRISPR-based method yielded no assignments due to its low recall rate (Table 1). At 84% and 69% cutoffs, VirHost Hunter performed comparably with the CRISPR-based method, and combining both methods further improved the accuracy to 101/156 (84% cutoff), 113/156 (69% cutoff) at the family level and 107/156 (84% cutoff), 117/156 (69% cutoff) at the genus level (Table 1). At the species level, VirHost Hunter correctly identified 9/63, outperforming CRISPR, which yielded no valid assignments (95% cutoff). At a 84% cutoff, VirHost Hunter and CRISPR achieved 20/63 and 28/63 correct predictions, respectively, showing similar performance. At a 69% cutoff, CRISPR outperformed VirHost Hunter with 32/63 versus 23/63 correct predictions. Combining both methods improved prediction across cutoffs, with 36/63 correct predictions at a 84% cutoff and 43/63 at a 69% cutoff compared to either method alone (Tables 1, S11), highlighting their complementary strengths.

Table 1.

Host prediction for cultivated gut phages by VirHost Hunter and the CRISPR-based method

	95% precision			84% precision			69% precision
	VirHost Hunter	CRISPR-based	combined	VirHost Hunter	CRISPR-based	combined	VirHost Hunter	CRISPR-based	combined
Family	73	0	73	82	95	101	96	105	113
Genus	58	0	58	94	95	107	105	105	117
Species	9	N.D.	9	20	28	36	23	32	43

ND not determined.

To sum up, VirHost Hunter demonstrated superior performance in comparison to the other three alignment-free methods tested. Furthermore, it significantly outperformed the CRISPR-based method in an independent isolated gut phages dataset at a 95% precision cutoff and achieved comparable performance at 84% and 69% precision cutoffs. Additionally, our experiment revealed that the combination of VirHost Hunter and the CRISPR-based method significantly enhances the proportion of true positive predictions, particularly for high-resolution phage-host predictions in the gut microbiota at the species level. Overall, these results highlight the scalability of VirHost Hunter across different environments.

The repository of phages targeting disease-associated gut bacteria is vastly expanded

The four most recently published gut virus databases typically adopted commensal bacteria as their CRISPR libraries. Among them, the GVD⁴², the MGV⁴⁴, and the CHVD⁴⁵ employed loose cutoffs compared to the GPD⁴³, which allowed zero mismatch and resulted in the lowest host assignment rate. Although these databases provide an unbiased representation of gut viral diversity, a tailored approach is necessary for specific application scenarios, such as for intestinal pathobionts. Considering that the GPD had the lowest host assignment recall of 28.66% among the four databases, and as evaluated by Dion et al.²². the precision was 84% at the genus level, we used VirHost Hunter to assign hosts for GPD at both 95% and 84% precision thresholds, aiming to explore the dark matter in the human gut associated with chronic diseases (Fig. 1C).

Using our customized annotation pipeline, we identified 163,590 lysins and 388,894 tail proteins from 142,809 assembled gut phages in the GPD. We applied precision filters of 84% and 95% to predict hosts at different taxonomic levels. Based on phylogenetic composition analysis, the prediction results covered 8 phyla, 13 classes, 21 orders, 29 families, 40 genera, and 58 species, including 42 species of obligate anaerobes (Fig. 3, Supplementary Data 4). The host assignment for each phage integrated the predictions from both tail and lysin proteins. Notably, 7 families could only be assigned by VirHost Hunter-lysin but not VirHost Hunter-tail, including Eubacteriaceae, Atopobiaceae, Leuconostocaceae, Prevotellaceae, Peptoniphilaceae, Gemellaceae, and Aerococcaceae (Fig. 3). We evaluated the host assignment results of VirHost Hunter using 95% and 84% precision and compared that with the previous results of the GPD. We found that both VirHost Hunter-tail or VirHost Hunter-lysin can enhance the host assignment of gut phages. At 95% precision, VirHost Hunter newly assigned host to 15.91% (22,724/142,809) of the GPD phages, with 10.98% (15,677/142,809) by VirHost Hunter-tail and 9.41% (13,432/142,809) by VirHost Hunter-lysin (Fig. 4A). At 84% precision, VirHost Hunter newly assigned host to 33.99% (48,545/142,809) of the GPD phages, with VirHost Hunter-tail contributing 20.16% (28,790/142,809) and VirHost Hunter-lysin contributing 25.37% (36,236/142,809), boosting the final host assignment ratio to 62.66% (89,478/142,809) (Fig. 4A). These data illustrate that excelling VirHost Hunter on either tails or lysins can enhance the host assignment of gut phages, while combing the results of VirHost Hunter based on different key proteins and that of the CRISPR method could optimize the outcome.

Fig. 3. Tens of thousands of host assignments newly uncovered.

Taxonomic tree of host assignments to the GPD phages by VirHost Hunter under precision cutoffs of 84%. From inside to outside: taxonomic-driven tree (phylum-class-order-family-genus-species); the circular heatmap presentations for successfully assigned phage genome, tail, and lysin proteins under a precision cutoff of 84%; Light green triangles mark host families newly assigned by VirHost Hunter.

Fig. 4. Expansion of phage-host associations by VirHost Hunter across taxonomic levels and genome features of newly identified hosts.

A Massive expansion of phages assigned hosts by VirHost Hunter, VirHost Hunter-lysin, and VirHost Hunter-tail under a precision cutoff of 84% and 95%. Bright indigo: number of phages previously assigned by the CRISPR-based method within the GPD. Green: overlap assignments by both methods. Salmon: number of phages newly assigned by VirHost Hunter. VHH-tail is the abbreviation for VirHost Hunter-tail. VHH-lysin is the abbreviation for VirHost Hunter-lysin. B–D The top 10 and newly assigned taxa at the C family, D genus, and E species levels. The names of the newly assigned taxa are highlighted in red. VHH-tail is the abbreviation for VirHost Hunter-tail. VHH-lysin is the abbreviation for VirHost Hunter-lysin. E Genome map of phage representatives targeting Akkermansia muciniphila and Prevotella copri.

By integrating VirHost Hunter and the CRISPR-based method, we assessed the improvement and refinement of host assignment results in the GPD. Both VirHost Hunter-tail and VirHost Hunter-lysin significantly enhanced host taxonomic classification compared to the previous results. The host assignment results of VirHost Hunter-tail newly covered 3 families, 8 genera, and 20 species, and those of VirHost Hunter-lysin covered an additional 5 families, 12 genera, and 25 species (Fig. 4B–D). Overall, at the family level, VirHost Hunter identified phages targeting 5 previously uncharacterized families, accounting for 1.38% of total assignments under 84% precision, while phages targeting Aerococcaceae were not detected at the 95% cutoff. Lachnospiraceae and Bacteroidaceae, recognized as the two most prevalent host families by both VirHost Hunter and the CRISPR-based method, collectively accounted for over 50% of total assignments at both the 84% and 95% cutoffs (Fig. 4B). At the genus level, VirHost Hunter identified phages targeting 12 previously uncharacterized genera accounting for 21.58% of total assignments at the 84% cutoff, while three of the previously uncharacterized genera were not detected at the 95% cutoff. Bacteroides is the most abundant host genus identified by both VirHost Hunter and the CRISPR-based method (Fig. 4C). At the species level, VirHost Hunter identified phages targeting 25 previously uncharacterized species, accounting for 0.14% of total assignments at the 84% cutoff, while four of the previously uncharacterized species were not predicted at the 95% cutoff. Notably, VirHost Hunter identified phages targeting Cronobacter sakazakii as predominant, which was not detected by the CRISPR-based method, likely due to differences in training datasets (Fig. 4D).

In the refined database, there are five newly annotated host families, including Aerococcaceae, Akkermansiaceae, Gemellaceae, Prevotellaceae, and Xanthomonadaceae (Figs. 1D, 3, 4B). Phages have potential in diverse applications, including lysing specific bacteria to inhibit pathogens and integrating into host genomes for in situ microbiota manipulation. Akkermansia muciniphila is commonly regarded as a beneficial gut bacterium, while it has also shown potential associations with certain autoimmune diseases and neurological disorders, such as schizophrenia⁷⁸. Identifying phages that target this bacterium may be utilized for developing undescribed gene-editing and phage display tools, and for modulating the gut microbiota to intervene in disease development. However, phages targeting Akkermansia muciniphila have never been characterized by any previous publications. We successfully identified 36 phages targeting Akkermansia muciniphila at 95% precision cutoff and 95 phages at 84% precision cutoff, and we examined the 36 phages at the more stringent cutoff. The genome sizes of the Akkermansia muciniphila phages range from 11,830 to 92,135 bp, and the GC content ranged from 49.11% to 60% (Fig. S8). The number of CDS is between 21 and 127, and the annotation rate is between 23.81% and 52.17% (Fig. S8). Prevotella copri within the Prevotellaceae family, is another renowned species associated with rheumatoid arthritis^79–81 and type 2 and type 1 diabetes mellitus^82–84. Megaphages were the only phages reported to target the Prevotella copri, but previous attempts to isolate them failed⁸⁵. We successfully identified 15 phages targeting Prevotella copri at a 95% cutoff and 22 phages at an 84% cutoff, and we examined the 15 phages at the most stringent cutoff. It was shown that the genome sizes of the Prevotella copri phages range from 12,114 to 127,100 bp, and the GC content ranges from 39.17% to 48% (Fig. S8). The number of CDS is between 16 and 166, and the annotation rate is between 23.17% and 56.25% (Fig. S8). We selected representative phages targeting Akkermansia muciniphila and Prevotella copri using CD-hit with thresholds of coverage of 0.6 and identity of 0.6 and annotated their genomes using our refined pipeline. It was shown that the functional elements of phages mainly include lysis, lysogenic-related, structure, DNA maintenance, packaging and assembly, replication and transcription, transport, and regulation (Fig. 4E).

Diversity and geographic distribution of gut phages

To gain further insights from the expanded host assignments, we first analyzed the phylogenetic lineages of phages in the refined database (Fig. 1D). Out of the 89,478 phages, 11.42% were classified under six viral families, including Siphoviridae, Myoviridae, Podoviridae, Herelleviridae, Tectiviridae, and Microviridae, covering all taxonomic classifications identified in the GPD (Fig. 5A). The remaining 88.57% of assigned phages were unclassified (Fig. 5A). Compared to the previous results of the GPD, VirHost Hunter increased host assignments by nearly one-fold to Siphoviridae and Herelleviridae phages, approximately two-fold to Podoviridae and Myoviridae phages, and 33.3% to Microviridae phages, significantly enhancing host assignments across multiple taxonomic levels. As a result, we assigned hosts to 79,332 unclassified phages, 4566 Siphoviridae phages, 2902 Myoviridae phages, 2598 Podoviridae phages, 75 Herelleviridae phages, 4 Microviridae phages and 1 Tectiviridae phage (Fig. 5A). It is noteworthy that Microviridae, a class of tailless phages, were assigned hosts by VirHost Hunter-lysin instead of VirHost Hunter-tail as expected. Therefore, it is important to combine the results of VirHost Hunter-tail and VirHost Hunter-lysin for downstream analyses. These findings demonstrate the broad applicability of VirHost Hunter for host prediction across diverse phage lineages, regardless that the phages are with or without tails.

Fig. 5. Diversity and global distribution of gut phages.

A The number of gut phages assigned with hosts across different viral families. Bars are colored according to phage taxonomy. B Viral diversity is defined by the ratio of viral cluster number to phage number. Bars are colored according to host phylum. C Viral cluster percentage and phage number targeting the most abundant host families across different continents globally. The bars are colored according to host families consistently with that in Fig. 4B. D Principal Coordinates Analysis (PCoA) of the most abundant host families across different continents globally.

Given the large number of phages predicted to target identical hosts, we evaluated phage diversity within bacterial families across diverse phyla by calculating the ratio of viral cluster numbers to phage counts sharing the same host family, with a higher value indicating greater viral diversity. We observed a wide distribution of phage diversity across bacterial families, especially within Bacteroidetes. Notably, 23 bacterial families exhibited the highest viral diversity (with a value of 1), indicating that each viral cluster within these host families contained only one viral sequence. Among these, 15 families belong to the phylum Firmicutes (Fig. 5B), which is consistent with the GPD database showing that Firmicutes harbor significantly higher viral diversity. Additionally, eight other families, including Pseudomonadaceae, Neisseriaceae, Muribaculaceae, Moraxellaceae, Dermabacteriaceae, Corynebacteriaceae, Coprobacteriaceae, and Cellulomonadaceae, which belong to the phyla Actinobacteria, Bacteroidetes, and Proteobacteria, also demonstrated the highest viral diversity (Fig. 5B). In contrast, the top 5 families with the lowest viral diversity all belong to the phylum Bacteroidetes, including Bacteroidaceae, DTU089, Marinifilaceae, Rikenellaceae, and Tannerellaceae (Fig. 5B).

To gain insights into the relationship between the viral cluster number of host families and their geographic distribution, we analyzed the dominant families and performed principal coordinate analysis (PCoA) (Figs. 1D, 5C–D). The results showed that Asia and Europe have a higher total phage count compared to other regions, which may be attributed to the greater number of human metagenomic sequencing studies conducted in Asia and Europe (Fig. 5C). Host families show similar geographic distribution patterns across continents, with Lachnospiraceae and Bacteroidaceae dominating across all continents, indicating their role as hosts for globally prevalent gut phages (Fig. 5C). Similarly, Asia, Europe and Africa have more overlapping regions, suggesting similar phage compositions, while Oceania, North America and South America are more distinct, indicating different phage communities (Fig. 5C). The PCoA result reveals a similar pattern, showing that while some continents do not completely overlap with others in host bacterial community composition, the slight differences observed are not statistically significant (Pr(>F) = 0.065), indicating similar phage compositions (Fig. 5D).

An expansive lysin repository countering a broad array of gut bacteria

Considering VirHost Hunter’s precision in predicting gut phage and lysin hosts and its complementarity with the CRISPR-based method (Fig. 4A, Supplementary Results), we established the Gut Phage Lysin Database (GPLD), which encompasses 117,698 lysins precisely targeting 29 disease-related gut bacterial families (Fig. 1E, Supplementary Data 3). Of these, 35.20% (n = 41,429) can be identified through both VirHost Hunter and the CRISPR-based method, 13.27% (n = 15,617) were identified using the CRISPR-based method, and 51.53% (n = 60,652) were exclusively identified by VirHost Hunter. Hydrolases, holins, and endolysins were the predominant functional categories (Fig. 6A). To better understand the functionality, stability, and potential applications of lysin proteins, we conducted various analyses focusing on their physicochemical properties. Their secondary structures were composed predominantly of turn, sheet, and helix (Fig. 6B), varying in length from 30 bp to 5811 bp, with a mean length of 195 bp. The molecular weight ranged from 2.8 kDa to 65 kDa, with a mean of 21.6 kDa (Fig. 6C), suggesting favorable attributes for efficient synthesis and manipulation. A majority (81.98%) were stable with an instability index below 40 (Fig. 6C). Amino acid frequency analysis indicated a prevalence of hydrophobic alanine, leucine, and isoleucine, potentially enhancing protein stability and function.

Fig. 6. Characteristics of the Gut Phage Lysin Database.

A Composition of the database. Holin, endolysin, hydrolase, lysis protein, tail-associate lysin, and transferase are represented by different colors. B Secondary structure fractions of lysins within the database. Red: helix. Green: turn. Blue: sheet. C Basic properties of lysins within the database(n = 117,698): protein length, molecular weight, isoelectric point, instability index, and amino acids. Data are presented as box plots, where the center line indicates the median, the box boundaries represent the 25th and 75th percentiles, and whiskers extend to 1.5 times the interquartile range (IQR)(Source Data).

To analyze the functional diversity and sequence-function relationships within the gut lysin protein family, we employed the sequence similarity network (SSN) tool^86,87, which provided insights into their sequence and functional divergence. Lysins were grouped into 603 clusters based on sequence similarity, with nodes colored according to their host taxonomical phylum (Fig. 7A). The SSN was predominantly populated by proteins from Firmicutes (n = 4559), followed by Proteobacteria (n = 996), Bacteroidetes (n = 906), and Actinobacteria (n = 600) (Fig. 7A). The protein clusters in the SSN were categorized according to their respective protein types, indicating that proteins with similar functions might possess conserved domains that confer these functions (Fig. 7A). Notably, holins, differing from other protein types in the network, exhibit high diversity in their sequences and structures (Fig. 7A). Furthermore, proteins targeting the same host phylum tend to cluster together, implying that lytic proteins, much like phages, exhibit a high degree of host specificity. This host specificity is likely mediated by conserved regions within the lytic proteins that are essential for identifying and binding to host cells.

Fig. 7. Sequence similarity network of Gut Phage Lysin Database and conserved motifs within major clusters.

A Sequence similarity network of lysins within the database. The cutoff was 35% identity and 50% coverage. B Conserved motif within Cluster 1–3 from the sequence similarity network.

To uncover the conserved functional motifs and the underlying mechanisms, we generated sequence logos⁸⁸ for three representative clusters. Cluster 1, the largest cluster, contained 1735 protein sequences targeting Firmicutes, Actinobacteria, Bacteroidetes, and Proteobacteria, including holins (n = 3170), hydrolases (n = 3098), endolysins (n = 1125) and lysis proteins (n = 796). The sequence logo analysis revealed three conserved motifs (RHTKAPAVLIECCFVDNKDD, NVTVHRDFANKSCPG, and RSWCSSSAANDNRAITIEVA), all located in the N-acetylmuramoyl-L-alanine amidase domain, which is crucial for phage-mediated bacterial lysis (Fig. 7B). Cluster 2 comprised 1213 representative holins sequences, and two conserved motifs were identified in the toxin secretion domain, which could facilitate the release of lytic enzymes to lyse bacterial cells (Fig. 7B). Cluster 3 consisted of 296 hydrolase representative sequences, and its motif was mainly associated with the N-(deoxy)ribosyltransferase-like domain, which functions in degrading bacterial cell walls during phage infection (Fig. 7B). Functionally important residues were conserved within putative isofunctional clusters, while motif and domain analyses revealed structural and functional differences between different types of lysin proteins. The findings have valuable implications for the design and engineering of lysins and their application in lysin therapy.

Lysin Ply491_6 effectively and specifically inhibits an obesity-promoting bacterium

Obesity has emerged as a significant global health concern, with the gut microbiome implicated in its onset and progression⁸⁹. Comparative analyses have revealed distinct microbiome profiles between obese and non-obese individuals, suggesting an association between certain bacterial genera and obesity, including Bacteroides, Megamonas, Ruminococcus, Dorea, Coprococcus, Fusobacterium, Blautia, and Eubacterium^72,90,91. While phage therapy holds promise for modulating the gut microbiota, the lack of reported phages targeting Megamonas and our failure in repetitive attempts to isolate Megamonas phages prompted our investigation into the therapeutic potential of lysins from the Gut Phage Lysin Database (GPLD) against this bacterial genus (Fig. 1F).

We identified 505 unique lysin sequences specific to Megamonas from GPLD, which were clustered into 167 distinct clusters (Fig. 8A). Among these, Ply491_6 (ivig_491_6) is the representative sequence of the protein cluster with the highest number of proteins (Figs. 8A, S9). The cDNA sequence encoding Ply491_6 spans 561 base pairs. Ply491_6 comprises 187 amino acids, with a molecular weight of 20.8 kDa and a theoretical isoelectric point (pI) of 5.37. Ply491_6 exhibits hydrophilicity, with a grand average of hydropathicity (GRAVY) value of -0.207. The instability index is 25.93, indicating that Ply491_6 is a stable protein. Additionally, Ply491_6 is devoid of signal peptides and transmembrane regions and is structurally characterized by four predominant α-helices alongside multiple β-sheets (Fig. 8B). Ply491_6 shares high sequence identity (99.46%) with QIW89318.1, a cell wall hydrolase autolysin from Caudoviricetes sp., and contains a conserved N-acetylmuramoyl-L-alanine amidase domain.

Fig. 8. Screening and functional verification of lysin Ply491_6.

A Decision-making process of selecting Ply491_6. B Structural prediction of Ply491_6. C Lytic assay of Ply491_6 on M. rupellensis. M. rupellensis strain 150922 was grown overnight, diluted 1:100, and grown to the mid-logarithmic phase. The bacterial cells were harvested, washed, and resuspended in PBS to an OD₆₀₀ of 0.9. The bacterial resuspensions were then plated in a 96-well plate with either buffer only (no Ply491_6 control) or Ply491_6 at a concentration of 20 µg/mL. The OD₆₀₀ was measured at 10-min intervals over a 240-min period at 37 °C. Values represent means ± SEM with three independent experiments. D Lytic assay of Ply491_6 on common gut commensal bacteria. B. fragilis bf2 (BF1), B. fragilis bf5 (BF2), C. perfringens 0840 (CP1), C. perfringens 0812 (CP1), R. gnavus 1177 (RG1), R. gnavus 1186 (RG2). Ply491_6 was used at 20 µg/ml. Data are presented as mean ± SD. BF1 + Ply491_6 was measured based on four technical replicates, while all other conditions were measured based on three technical replicates (Source Data). E Lytic assay of Ply491_6 on probiotic bacteria. B. longum 4486 (BL1), B. longum 2366 (BL2), L. paracasei LAC-F (LP1), L. paracasei LAC-J (LP2) and L. plantarum SZHD0015 (L. plantarum). Ply491_6 was used at 20 µg/ml. Data are represented as mean ± SD from three technical replicates(Source Data).

Therefore, we synthesized and purified the Ply491_6 protein for in vitro assays to verify its lytic activity against Megamonas. We incubated Ply491_6 with Megamonas rupellensis and monitored the bacterial turbidity over time. It was shown that Ply491_6 effectively lysed bacterial cells at concentrations as low as 20 µg/mL, with a significant reduction in bacterial turbidity observed within 150 min (Fig. 8C). To further assess the specificity of Ply491_6, we measured its lytic activity against other high-abundance gut bacteria and common probiotics, including Bacteroides fragilis, Clostridium perfringerns, Ruminococcus gnavus, Bifidobacterium longum, Lacticaseibacillus paracasei, and Lactiplantibacillus plantarum. Ply491_6 demonstrated minimal impact on the viability of these bacteria (Fig. 8D, E). These results underscore the efficacy and specificity of Ply491_6 against Megamonas, positioning it as a promising candidate for targeted bactericidal therapy against obesity-associated dysbiosis. These findings provide valuable insights into phage-bacteria interactions in the gut and offer essential data to support the development of precision therapies targeting intestinal pathobionts.

Discussion

The VirHost Hunter framework presented here integrates three highlights. By conducting control analyses, we verified that each highlight enhanced the prediction performance. By comprehensive comparison with other methods across multi-taxonomic levels, VirHost Hunter demonstrated superior precision and recall, and it also showed a higher resolution as it reached accurate species-level host prediction. This can be attributed to three key factors: 1) the integration of a large language model, specifically ProtT5, allows for advanced contextual understanding of protein sequences, enabling VirHost Hunter to capture functional homology effectively; 2) by focusing on phage tails and lysins, VirHost Hunter can directly relate to the functional roles of these key proteins in phage-host interactions and make accurate, high-resolution predictions even in cases of incomplete genomic data; 3) the incorporation of DNA sequence features, such as codon usage and nucleotide composition, as complementary to protein features, further enriches the predictive capabilities of VirHost Hunter. These results provided insights into how to leverage machine learning to predict protein function and mine sequencing data in the future.

Because the CRISPR-based method has been the most widely used tool to assign bacterial host, we also compared the performance of VirHost Hunter and the CRISPR-based method using two independent datasets with biological experimental evidence: a collection of 156 cultivated gut phages, and another collection of 31 lysins (Supplementary Results). We demonstrated that both methods had similar recall when precision was set at 84% and 69%, but VirHost Hunter had higher recall than the CRISPR-based method when the precision was set at 95%. Interestingly, combining both methods resulted in an improved host assignment ratio compared to either method used alone, likely due to the differences in training datasets. The synergy between VirHost Hunter and the CRISPR-based predictions allowed us to expand the host assignment ratio of the GPD from 28.66% to 62.66%. Therefore, we propose a guideline for users: we recommend prioritizing VirHost Hunter if aiming for highly precise or species-level prediction, and employing both VirHost Hunter and the CRISPR-based method in parallel for general purposes.

Using the calibrated model, we greatly improved the host assignment ratio of the gut phage database, particularly for phages associated with chronic diseases. We also identified dozens of previously uncharacterized phages targeting Akkermansia muciniphila and Prevotella copri, whose phages have hardly been characterized before. To further facilitate the application of this resource, we established the Gut Phage Lysin Database, cataloging 117,698 host-specific lysins targeting various gut bacteria. This database is pivotal for identifying and engineering lysins, particularly against bacteria linked to chronic diseases. As a proof of concept, we synthesized a lysin, Ply491_6 and verified its efficacy and specificity against Megamonas, an obesity-promoting bacterium. Additionally, we initially screened several holin proteins targeting Megamonas; however, we encountered challenges in expressing these holin proteins in prokaryotic systems due to their transmembrane domains, preventing us from verifying their antibacterial activity. Notably, we have not seen any reported means targeting Megamonas before, and we have failed to isolate Megamonas phages in our repetitive efforts during the past few years. In fact, it remains technically difficult to isolate phages targeting all obligate anaerobic bacteria. Under such circumstances, VirHost Hunter can be exceptionally useful, deciphering previously uncharacterized phages to reveal biological insights and discovering previously uncharacterized lytic proteins to inform therapeutic potentials. Furthermore, we advocate that effective management of complex diseases such as obesity requires a multifaceted strategy. Phage therapy, with its high specificity and targeted bactericidal activity, may serve as one component of a combinatorial approach, for instance, in conjunction with dietary modifications and probiotic supplementation.

Fujimoto et al. have shown that E. faecalis phage-derived endolysin worked effectively in humanized gnotobiotic acute graft-versus-host disease (GVHD) mice, as it reduced levels of intestinal cytolysin-positive E. faecalis and significantly improved survival⁹². Compared to the 7-log reduction demonstrated by Fujimoto et al., lysin Ply491_6 inhibited the bacterial growth by only 1- to 2-log, which is a promising start but requires further engineering for downstream applications. Some possible directions for engineering include: 1) fusing lytic proteins with functional peptides to form nanoparticles, which can enhance both lytic efficacy and stability⁹³; 2) integrating the enzymatic active domains (EAD) and cell wall binding domains (CBD) from different lysins, particularly for endolysins, to boost lytic activity and broaden the host range⁹⁴; 3) introducing targeted mutations at active sites or increasing positive charges to enhance lytic activity and binding efficiency⁹⁵; 4) fusing lysins with receptor-binding proteins to improve the targeting specificity⁹⁶.

While VirHost Hunter demonstrated strong predictive performance, there are some limitations. Firstly, we only utilized phage tails and lysins for model training and host prediction. Although our data ruled out the possibility of using two structural proteins, other proteins might also confer different levels of host specificity. Secondly, VirHost Hunter should be robust when calibrating with any dataset, but due to the focus of this work, we only verified the scenario of searching for phages and lysins targeting disease-associated gut bacteria. Future research should aim to refine VirHost Hunter, which is intended for use across different scenarios, by incorporating a broader range of datasets. Because merging datasets from different sources may skew the model by overrepresenting well-characterized groups (e.g., gut phages) while neglecting others (e.g., phages of extreme environments), we recommend customizing the dataset for specific application scenarios. In this way, each model is optimized for its context. In this study, our model is implemented in two versions: one is a general version, and the other is specifically optimized for targeting gut bacteria. A great advantage is that VirHost Hunter only requires input of key proteins, which can be extracted from prophages integrated within bacterial genomes and fragmented phage genomes from metagenomic sequencing, vastly expanding the scale of datasets. The type of data used for calibration can be flexibly selected based on the user’s research focus, allowing adaptation to specific application scenarios.

For instance, VirHost Hunter can be calibrated to target other gut bacteria that are not necessarily disease-associated, further improving the host assignment ratio of the gut phage database. The implications of this study also extend to the broader field of environmental microbiology beyond the gut microbiome, as environmental microbiologists encounter an even worse situation in phage host assignment. With an estimated 10³¹ particles globally, phages are a key component of Earth’s ecosystems and play crucial roles in regulating microbial populations, nutrient cycling, and ecosystem dynamics^97,98. VirHost Hunter can then be calibrated to target environmental bacteria, shedding light on the “viral dark matter” and their interactions with bacteria in various ecosystems, out of which extreme environments will be of special interest. Identifying the hosts of environmental phages will enhance our understanding of virus-host-environment interactions, their role in microbial community structures, and their influence on biogeochemical processes. These insights can inform conservation efforts, bioremediation strategies, and the management of microbial communities in natural and engineered environments.

Methods

Establishment of the VirHost Hunter framework

VirHost Hunter consists of two primary components: a feature extractor and a classifier (Fig. 1A). When extracting features for VirHost Hunter, we utilize three distinct tools to process phage-specific proteins and their corresponding DNA sequences, resulting in three types of features: protein sequence embeddings from the pre-trained ProtT5 model, physical-chemical characteristics of DNA sequences, and k-mer features of DNA sequences extracted via a DNN network. These features will be elaborated upon in detail below.

For protein sequence representation, we leveraged the capabilities of the pre-trained protein language model ProtT5 to generate dense vector representations (embeddings) of protein sequences. Specifically, we utilized only the encoder portion of the ProtT5 model. The encoder integrates essential components such as a multi-head attention mechanism and feedforward layers, enabling it to capture intricate relationships between amino acid residues in the input protein sequence. This process yields rich embedding vectors containing valuable information regarding protein structure and functionality. We extract the average embedding vector from the last layer of the pre-trained model to generate the embedded feature vector, resulting in a 1024-dimensional feature vector for each protein sequence.

The physical-chemical features employed to represent DNA sequences align with the methodology proposed by Boeckaerts et al.³⁹. These features encompass nucleotide frequency, GC content, codon frequency, and codon usage bias, amounting to a total of 133 dimensions for the representation of DNA sequences.

To preserve the intrinsic sequence information of DNA sequences, we encoded them following the approach outlined by Wang et al. in their study DeepHost²⁹. This method represents DNA sequences through K-mer frequency. Subsequently, we construct a deep neural network (DNN). The DNN incorporates a convolutional neural network with three paths, each outputting a different number of channels, facilitating the capture of feature information at varying scales (Fig. 1B). By using multiple channels in parallel and fusing their outputs, the model can simultaneously learn abstract features at different levels. Subsequently, we leverage the Vision Transformer (ViT)⁹⁹, utilizing the self-attention mechanism of the ViT model to capture global relationships and multi-channel feature representations, yielding richer original sequence feature embeddings.

Subsequently, we concatenate the three types of features learned by the model into a unified vector, which serves as input for the classifier ensemble. This ensemble includes an MLP neural network, an autoencoder, and a random forest. The training process follows four sequential phases: (1) The MLP and feature extractor are trained end-to-end using Adam optimization with cross-entropy loss; (2) The feature extractor parameters are then frozen (“stabilized”) to maintain consistent feature representations; (3) The autoencoder is trained on these fixed features using reconstruction loss; (4) Finally, the random forest is trained on the processed features through an exhaustive grid search that optimizes tree depth, node splitting criteria, and ensemble size to maximize prediction accuracy. Throughout this process, the MLP neural network updates parameters while the feature extractor remains fixed after initial training. We employ the softmax function as the activation function, cross-entropy as the loss function, and utilize the Adam algorithm to optimize the loss function.

Due to the inherent characteristics of the softmax function, which may produce overconfident predictions in incorrect categories, we implement a blended prediction approach combining MLP and random forest outputs. This integration occurs through three steps: First, we train the MLP and feature extractor, then stabilize the feature extractor parameters before training the autoencoder and RF. During testing, the model executes a blending operation represented by the “×” symbol in Fig. 1A. The MLP generates softmax-normalized class probabilities while the RF produces independent probability estimates. These outputs are linearly combined with equal weighting (0.5) to yield the final prediction according to the equation:

$$FinalPrediction=0.5\times MLPOutput+0.5\times RFOutput$$

This design enables the RF to calibrate potential MLP overconfidence while preserving accurate predictions. Implementation details are provided in Supplementary Table S7.

Bioinformatics pipeline for phage genome annotations

A bioinformatics pipeline was developed to enable the rapid and efficient annotation of phage tails and lytic proteins. The pipeline involved several steps. Firstly, proteins predicted from phage genomes using Prodigal v2.6.3 (-f gff -c -p meta)¹⁰⁰. Secondly, the predicted proteins were aligned against multiple databases, including 1) the NR phage protein database using Blastp v2.3.0 (-evalue 1e-5 -max_target_seqs 1 -outfmt ‘6 qseqid sseqid stitle pident length mismatch gapopen qstart qend sstart send evalue bitscore’)⁴⁷, 2) Uniref phage protein database using phmmer v3.1b2 (-E 1e-5), 3) Uniprotkb phage protein database using phmmer v3.1b2 (-E 1e-5) and 4) TIGRFAM, SMART, CDD, ProSiteProfiles, SUPERFAMILY, PRINTS, PANTHER, Gene3D, PIRSF, Pfam, Coils, and MobiDBLite database using hmmscan v3.1b2 (-E 1e-5). BLASTp results were prioritized; if no confident match was found, we used a consensus from phmmer and hmmscan.

To ensure annotation reliability, alignment results were further filtered to retain only those with ≥50% sequence coverage and ≥50% identity, and hits were considered valid only when supported by one or more curated domain databases^101–105. The final annotation was generated by integrating results across all tools and databases, emphasizing high-confidence matches^106,107. This conservative strategy prioritizes specificity over sensitivity, which is particularly important when constructing datasets for downstream analyses that rely on accurate functional labels.

Phage tail proteins were identified using the keyword “tail”, and lysin proteins were extracted using keywords such as “lysis,” “lyase,” “lysin,” “holin,” “hydrolase,” “spanin,” and “endolysin” from final annotations via Seqkit v0.16.0¹⁰⁸.

Construction of benchmark datasets

Complete phage genomes were collected from NCBI using specific keywords related to bacterial hosts, including ‘Staphylococcus’,’ Acinetobacter’, ‘Escherichia’, ‘Clostridium’, ‘Klebsiella’, ‘Pseudomonas’, and ‘Salmonella’. A total of 3116 phage genomes were collected. Protein annotation was performed using the bioinformatics pipeline, resulting in 22,151 phage tail proteins (21,264 from the pipeline and 887 from a published paper by Boeckaerts et al.³⁹). From these, 7493 RBPs were screened out using specific keywords related to the tail protein functions, including ‘fiber’, ‘fiber’, and ‘spike’.

Three filters were applied to clean the tail proteins and RBP datasets: 1) sequences with lengths shorter than 50 amino acids or longer than 1500 amino acids were removed, 2) sequences containing undetermined amino acid ‘X’ in protein sequences or undetermined nucleotides ‘N’ in CDS were excluded, and 3) identical protein sequences with different hosts were discarded to remove redundancy. The final benchmark datasets consisted of 4845 RBPs in DRRBP and 12,509 tail proteins in DRTail, respectively.

Construction of tail protein and lysin datasets at multi-taxonomic levels

Phage genomes from the viral category were screened in the NCBI RefSeq database as of December 29, 2021 (https://www.ncbi.nlm.nih.gov/genome/browse/#!/viruses). Those containing partial genomes and coding sequences were excluded from the dataset. Genomic sequences in FASTA format and annotation files in GBFF format were downloaded from the corresponding table on the NCBI FTP site. This screening process resulted in a total of 7598 phage genomes for further analysis. Next, information related to the host organism was extracted from the annotation files (GBFF format) using a custom script. If the ‘host = ’ field was empty, the species information mentioned in front of the phage in the GenBank tile (ORGANISM) was selected as the host information. For instance, if the host information of phage AF234172 was empty, we selected ‘Escherichia’ as the host based on the record ‘ORGANISM: Escherichia virus P1’. Then, the NCBI taxonomy toolkit, TaxnoKit, was used to obtain the taxonomy ID and taxonomy level of the host organism (taxnokit name2taxid –show-rank). The host taxonomic information was transformed into a standard format, including phylum, class, order, family, genus, species, and strain (taxonkit lineage | taxnokit reformat | cut -f 1,3). This process resulted in the compilation of phage-host taxonomic rank information.

We counted the number of RBPs and tail proteins in the datasets and observed an average of 1.33 RBPs and 15.24 tail proteins per phage (Fig. S1). We also found that 53.10% of phages lacked RBPs, prompting us to construct a multi-taxonomic-level dataset at different taxonomic ranks, enabling the establishment of a tail protein-based VirHost Hunter (VirHost Hunter-tail) for broader applications. We filtered the phage data and created a phage tail protein dataset, including 37 families, 54 genera, and 57 species. Additionally, we trained VirHost Hunter on lysins—another type of host-specific protein—using 37,469 lysin protein sequences from the same 7598 phages to construct lysin-based VirHost Hunter (VirHost Hunter-lysin). The lysin dataset comprised 37 families, 42 genera and 47 species.

Family, genus, and species datasets for tail proteins were constructed based on the taxonomic ranks obtained in the previous step. The three datasets were filled using the same three filters as used in constructing the benchmark datasets. Category with fewer than 50 counts in each taxonomic dataset were discarded. After filtering, there were 47 families, 72 genera, and 120 species remaining in the tail protein datasets. These three datasets were used to train the VirHost Hunter-tail model. To address bias issues observed in certain taxa, taxa with precision lower than 0.7 were eliminated from the datasets. For example, Enterobacteriaceae has a recall of 0.9098 and precision of 0.6941 at the family level, Escherichia has a recall of 0.7125 and precision of 0.3526 at the genus level, and Escherichia coli has a recall of 0.7373 and precision of 0.3199 at the species level. As a result, the final set of taxa includes 37 families, 54 genera, and 57 species left. Each dataset was randomly split into training, validation, and testing sets with a proportion of 6:2:2.

The original lysin dataset contained a total of 37,469 protein sequences. The same three filters as previously applied to the benchmark datasets were used to clean the lysin dataset. However, the maximum allowed sequence length was set to 1000 amino acids since protein sequences with a length of over 1000 amino acids accounted for less than 2%. The screening process and building procedures for the VirHost Hunter-lysin dataset followed a similar approach to VirHost Hunter-tail. Taxa with precision lower than 0.62 were eliminated from the family, genus, and species taxonomy datasets based on the training results. This step ensured reliable predictions for the remaining taxa. After eliminating low-precision taxa, the final lysin datasets consisted of 37 families, 42 genera, and 47 species. Each dataset was randomly split into training, validation, and testing sets with a proportion of 6:2:2.

Gut prophage dataset construction for model calibration

To calibrate the VirHost Hunter model for improved identification of gut phages, particularly those targeting disease-associated gut bacteria, we first identified 60 gut bacterial species associated with diseases. These associations were based on positive correlations from large-scale cohort metagenomic analyses, validation of pathogenicity through animal experiments, or studies on pathogenic mechanisms (Table S5). We then downloaded 13,483 high-quality genomes of these bacterial species from NCBI RefSeq.

Prophages were identified within these bacterial genomes using ProphageHunter¹⁰⁹, a reliable tool designed to detect prophages with high precision. Subsequently, tail and lysin proteins were identified through stringent sequence alignment, resulting in a dataset of 328,701 unique tail proteins and 312,565 lysins. Since these prophages integrate into the bacterial genome rather than causing immediate lysis, their phage-host relationships are inherently accurate, providing definitive host information. We used this information as the host data for the tail and lysin proteins of prophages to calibrate our model.

Additional filter for higher precision at multi-taxonomic ranks

Since the range of categories that our model can cover is limited, an additional filter was implemented to VirHost Hunter trained in the multi-taxonomic levels’ dataset and the gut prophages dataset, to generate an ‘Unknown’ output for any given input that exceeded the prediction range. To determine the appropriate cutoff for this filter, two datasets were constructed: Positive Control, which comprised samples from the test dataset, and Negative Control, containing samples not belonging to any predefined classes in the training dataset. The recall and precision on the Positive Control and the specificity on the Negative Control were illustrated in Figure S10-11. These figures showed that more stringent cutoffs resulted in higher precision and lower recall. This phenomenon occurred because, as the cutoff increased, more data were classified as ‘Unknown’, and the remaining data were considered more reliable by VirHost Hunter.

To benchmark VirHost Hunter’s performance against other methods, we considered the work of Dion et al.²². where they evaluated the precision and recall of a CRISPR spacer-based method under different cutoffs of mismatch numbers or e-value at the genus level. They found that with an e-value of 10⁻⁹, the method achieved the highest precision of 95% but the lowest recall of 2.5%. With zero mismatches, the method achieved 84% precision and 31% recall. By tolerating two mismatches, the method obtained a balanced performance of 69% precision and 49% recall. Accordingly, several probability cutoffs were selected at the family, genus, and species levels to achieve the same precision values of 95%, 84%, and 69%, respectively (Table S8). Consequently, when the precision on the Positive Control and the specificity on the Negative Control surpassed 95%, VirHost Hunter demonstrated a precision of 95%.

Application of synthetic phage lysin data for host prediction

The latest SQL file (v2021_04) was downloaded from the largest available Phage Lytic Protein database (PhaLP)¹⁷. The SQL file provided lysin IDs, corresponding phage genome IDs, lysin annotation information, host taxonomy information, and experimental support information. Phage genome IDs that were not used in VirHost Hunter construction were marked in the dataset collected from NCBI, resulting in 3448 phage genomes. Lysins that were synthesized and experimentally validated, and their corresponding 31 phages were screened out from the dataset. A total of 138 tail proteins were annotated using the custom bioinformatics pipeline. Three phages could not be annotated with tail proteins, leaving final real-world evidence of 31 phage genomes and 138 phage tail protein sequences.

Application of VirHost hunter in GPD

Gut Phage Database (GPD) and the corresponding taxonomy information table by Camarillo-Guerrero et al. were downloaded⁴³. Based on the ‘Host_range_taxo’ field in the information table, phages were categorized into two groups: those with host information and those without host information. Phage tail and lysin proteins were annotated using a custom annotation pipeline. A total of 163,590 lysin sequences and 388,894 tail protein sequences were obtained from 111,355 phages, which accounted for 77.97% of the total 142,809 phage genomes. A comprehensive dataset of 42,586 proteins was downloaded from NCBI (as of February 22, 2024) using keywords(lysis protein, lysin, lyase, holin, hydrolase, and endolysin AND phage). Lysins encoded by gut phages were identified by comparing with the dataset using BLASTP with a threshold of 60% identity and 50% coverage. VirHost Hunter-lysin model (95%, 84%), VirHost Hunter-tail model (95%, 84%), and CRISPR-based method (84%) were used to construct the Gut Phage Lysin Database (GPLD) targeting human gut commensal bacteria.

Sequence similarity network for the Gut Phage Lysin Database (GPLD)

Biopython was employed to conduct descriptive statistical analysis of the GPLD database, which included aspects such as protein categories, secondary structure proportions, length, amino acid composition, molecular weight, isoelectric point, and stability index. The results were visually represented using ggplot2¹¹⁰. A sequence similarity network was established using a tool developed by Miguel M. Sandin (available at https://github.com/MiguelMSandin/SSNetworks), which was based on lysin sequences clustered using CD-HIT with a 70% similarity threshold. The network construction parameters were set at an identity level of 35% and a coverage level of 50%. The resulting networks were visualized using Cytoscape. Furthermore, MEME was utilized to identify conserved motif sites within the three primary clusters of the network, employing default parameters.

Identification of Megamonas-targeting lysin from GPLD

A total of 505 unique lysin sequences specific to the genus Megamonas were identified from the GPLD. These sequences were subsequently clustered into 167 distinct groups using the CD-HIT with a sequence similarity threshold of 95% and a coverage threshold of 90% (-c 0.95 -aL 0.9). Ply491_6 (ivig_491_6), representing the largest cluster among these groups, was chosen for in-depth characterization and experimental validation of lytic activity. This process involved the prediction of signal peptides using SignalP (https://services.healthtech.dtu.dk/services/SignalP-6.0/), identification of transmembrane regions with HMMTOP (https://services.healthtech.dtu.dk/services/TMHMM-2.0/), and assessment of physicochemical properties via ProtParam (https://web.expasy.org/protparam/).

Synthesis and purification of Ply491_6

To synthesize and purify Ply491_6, the gene encoding Ply491_6 was synthesized and subcloned into the pET-30a(+) plasmid using NdeI and XhoI restriction sites, which was obtained from General Biosystems (China) company. The plasmids were constructed and transformed into BL21 (DE3) competent cells. These transformed cells were cultured on agar plates containing kanamycin at a final concentration of 50 µg/mL at 37 °C. Colonies were picked from the plates and cultured until the optical density at 600 nm (OD₆₀₀) reached 0.6-0.8. Protein expression was induced by adding IPTG to a final concentration of 0.5 mM, followed by incubation of the cultures for an additional 4 h at 37 °C. Cells were then harvested, lysed, and the lysates were subjected to SDS-PAGE to verify protein expression. Then the proteins were purified using Ni-NTA affinity chromatography. The purified proteins were dialyzed into phosphate-buffered saline (PBS) containing 300 mM NaCl, 10% glycerol, and adjusted to pH 7.4, followed by filter sterilization.

Bacterial strains

Megamonas rupellensis strain 150922 was used for the lysin activity assay of Ply491_6 in vitro. Bacteroides fragilis bf2 (BF1), B. fragilis bf5 (BF2), Clostridium perfringerns 0840 (CP1), C. perfringerns 0812 (CP1), Ruminococcus gnavus 1177 (RG1), R. gnavus 1186 (RG2), Bifidobacterium longum 4486 (BL1), B. longum 2366 (BL2), Lacticaseibacillus paracasei LAC-F (LP1), L. paracasei LAC-J (LP2) and Lactiplantibacillus plantarum SZHD0015 (L. plantarum) were used for comparing lysin activity of Ply491_6 in vitro. All bacterial strains were isolated from human feces. All bacterial strains were grown overnight in BHI-YH (Brain Heart Infusion medium supplemented with 5 g/L yeast extract, 5 mg/L hemin). To maintain anaerobic conditions, all media and buffers were additionally supplemented with 0.5 g/L L-cysteine hydrochloride and 0.25 g/L anhydrous sodium sulfide, serving as reducing agents.

Lytic activity and specificity of Ply491_6

M. rupellensis strain 150922 was grown overnight, diluted 1:100, and grown to the mid-logarithmic phase. The bacterial cells were centrifuged, washed, and resuspended in phosphate-buffered saline (PBS, pH 7.4) to an OD₆₀₀ of 0.9. Phage lysin Ply491_6 was added to the bacterial suspension with a final concentration of 20 μg/mL. Each concentration was plated in a U-bottomed 96-well plate in triplicate technical replicates. Ply491_6 dilution plates were then incubated at 37 °C in a BioTek Epoch2 Microplate Spectrophotometer(BioTek Instruments, Inc., USA) for 240 min. The OD₆₀₀ was measured every 10 min.

To verify the specificity of Ply491_6, B. fragilis (n = 2), C. perfringerns (n = 2), R. gnavus (n = 2), B. longum (n = 2), L. paracasei (n = 2) and L. plantarum (n = 1) strains were each grown overnight. The bacterial cells were then centrifuged, washed, and resuspended in PBS, and were then incubated with 20 µg/mL Ply491_6 or PBS at 37 °C for 240 min. The OD₆₀₀ was measured every 10 min.

Statistics & reproducibility

No statistical method was used to predetermine sample size. No data were excluded from the analyses. The experiments were not randomized. The Investigators were not blinded to allocation during experiments and outcome assessment. Statistical analysis was performed using R (4.0.2) and OriginPro 2024 (Learning Edition). The Permutational Multivariate Analysis of Variance (PERMANOVA) was used to assess the statistical significance of the relative abundance of the most abundant host families and different continents.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

M.X. conceived the study. Z.D., K.L., and Y.O. developed the tool. M.L. and B.X. compiled the training, validation, and test sets. K.L., M.L., B.X., and M.X. analyzed the viral dark matter. K.L., M.L., B.X., and Y.O. drafted the manuscript and made the figures. Z.L. and M.L. maintained and improved the code. Z.D., M.X., and Junhua L. revised the manuscript. W.S., J.C. and Jianqiang L. provided consultation. B.X., Y.O., and Z.L. contributed equally to this work. All authors read, edited, and approved the final manuscript.

Peer review

Peer review information

Nature Communications thanks Karthik Anantharaman and the other anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The Gut Phage Lysin Database (GPLD) protein sequence data generated in this study have been deposited in the CNGB Sequence Archive (CNSA)¹¹¹ of the China National GeneBank DataBase (CNGBdb)¹¹² with accession number CNP0005794. The benchmark datasets, derived training and testing data generated in this study, are available in the Zenodo repository under 10.5281/zenodo.17340915. All accession codes are from previously published datasets and are provided in Supplementary Data 5. Source data supporting the findings of this study are provided in the Source Data. All these bacterial strains and plasmids can be requested by direct correspondence with the lead contact, Dr. Minfeng Xiao (xiaominfeng@genomics.cn). These resources are strictly limited to academic research use, and this publication must be cited accordingly. Source data are provided with this paper.

Code availability

The models constructed in this study, together with the corresponding scripts, are publicly available on GitHub at https://github.com/YuehuaOu/Viral-Host-Hunter and have been archived on Zenodo under 10.5281/zenodo.18399248. The pre-trained model weights generated in this study are available in the Zenodo repository under 10.5281/zenodo.17340381.

Competing interests

Authors affiliated with BGI Research (M. L., M. X., and Junhua. L) and authors affiliated with Shenzhen University (Z. D., K. L., and Jiangqiang. L) have filed a patent application titled “ Method and System for Predicting Host Range of Bacteriophages, and Corresponding Computer Device or Medium”(Application No. PCT/CN2025/075560). It describes predicting algorithm in this study that predicts host specificity of phages based on tail and lysin proteins. Other authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-70613-x.