Evolutionary interplay: virulence, endolysin-like hydrolases, and defense correlations in the Erwinia amylovora pangenome

Abstract

Erwinia amylovora, the causative agent of fire blight, poses a significant threat to global pome fruit production. This study presents a comprehensive genomic analysis of 317 E. amylovora strains and 227 Erwinia phages to elucidate virulence evolution, phage-host dynamics, and the genomic signatures of the co-evolutionary arms race. Our analysis suggests that a substantial portion of E. amylovora’s virulence factors (VFs) share evolutionary origins with diverse plant, human, and animal pathogens, underscoring widespread horizontal gene transfer. We identified bacterial phage hydrolases‑like proteins that share phylogenetic and domain-level similarities with phage endolysins. These observations are consistent with the possibility that some bacterial hydrolases originated from phage-derived ancestors, although functional repurposing remains to be experimentally validated. Crucially, our analysis identifies systematic, non-random associations between bacterial defense systems (e.g., RM, CRISPR-Cas, TA) and mobile anti-defense genes. Statistical correlations show strong patterns of co-occurrence and mutual exclusivity, which are consistent with an ongoing phage-bacteria arms race. These patterns provide a genomic basis for generating hypotheses about co-evolutionary dynamics. These findings may advance our understanding of E. amylovora pathogenicity and phage interactions, offering foundational insights for developing targeted phage-based biocontrol strategies against this devastating plant pathogen. Experimental validation of the predicted virulence factors and defense correlations is warranted to confirm their biological roles.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12866-026-05295-y.

Introduction

Bacteriophages or simply phages, viruses identified in the mid-1910s, are characterized by their ability to infect and reproduce solely within particular types of bacteria. Because of the antibacterial activity of phages, these viruses have been explored as antibacterial agents since their discovery and they are currently being reconsidered as an alternative to conventional antibiotics. These viruses have two main life cycles. Lysogenic (temperate) phages can either integrate their genome into the host chromosome (or maintain it as an episome) or, upon induction, enter a lytic cycle that produces new phage particles. On the other hand, virulent phages rapidly infect and kill bacteria. Temperate phages can change the characteristics of bacteria by expressing viral genes, a process known as “lysogenic conversion.” This phenomenon can also shield the bacteria from further phage infections [1]. Typically, temperate phages are not used for antibacterial applications due to their specific behavior while it has been shown that lytic phages are reliable tools to combat various types of hard-to-treat bacteria, especially antibiotic‑resistant ones, in different sectors such as medicine, agriculture, and food industry [2, 3].

Fire blight, the first bacterial plant disease ever identified (back in the 19th century), remains a major problem for apple and pear growers today, resulting in substantial financial losses. This disease, caused by the bacterium Erwinia amylovora, infects fruits, shoots, and flowers, leading to decreased harvests and even the death of trees [4]. Furthermore, E. amylovora strains can acquire antibiotic resistance genes (ARGs), conferring resistance to antibiotics such as streptomycin, which is currently considered the most potent management option for fire blight [5]. Additionally, virulence factors (VFs), particularly hrpN and dspE/dspA, enable this pathogen to invade host cells, necessitating molecular investigation of VFs and ARGs in E. amylovora strains [6, 7]. Regarding these characteristics of E. amylovora, there is an urgent need for anti-Erwinia agents to control this severe phytopathogen [8].

While phage-based biocontrol has shown promise against this pathogen [9–13], the fundamental evolutionary dynamics governing phage-host interactions in E. amylovora remain poorly understood. Understanding these dynamics, particularly how virulence genes evolve, how phage-derived elements are repurposed by bacteria, and how defense systems co-evolve with counter-defense mechanisms, is essential for developing effective and durable phage therapies.

Interaction between phages and their host is one of the most crucial aspects of phage studies. These interactions are essential for understanding the ecological, evolutionary, and practical functions of phages within microbial ecosystems [14]. Phages are able to reshape the microbial community of the target environment, influence the metabolism of bacteria, and transfer genes between bacteria by horizontal gene transfer [15]. Moreover, phages and bacteria are locked in an ongoing battle of evolution. As bacteria develop defenses against phages, the phages, in turn, evolve ways to overcome those defenses [16]. In addition to evolutionary dynamics, studying defense systems can yield critical insights for advancing biotechnology. These insights enable the modification of phages with broader host specificity or increased performance in medical and agricultural settings. Additionally, they facilitate the engineering of gene delivery platforms and synthetic biology tools by harnessing phage-driven processes such as lysogenic conversion [17].

One of the well-known defense systems of bacteria is CRISPR/Cas (clustered regularly interspaced short palindromic repeats/CRISPR-associated protein) system. These CRISPR arrays consist of repeating DNA sequences separated by unique spacer regions that originate from invading genetic elements like phages, serving as molecular “memory” of past infections [18, 19]. In addition to CRISPR arrays, to date, researchers have identified over 150 different defense mechanisms in prokaryotic organisms [20]. While some of these systems have unidentified functions, others have different types of anti-phage systems such as abortive infection (Abi), restriction-modification (RM), toxin-antitoxin (TA), cyclic oligonucleotide-based anti-phage signaling system (CBASS), and SspBCDE systems [21].

In order to counter bacterial host defense mechanisms, phages develop anti-defense systems that generally have anti-defense functions including modification of target host’s defense system, inactivation of some parts of host defense system by direct interaction, restoration of targets’ vital functions, and interference with immune-signaling pathways [20]. Despite valuable insights that unveiling phage anti-defense mechanisms can provide, in comparison to anti-phage strategies, investigations into anti-defense systems of phages have been gradual [22]. The development of high-throughput sequencing (NGS) technologies and the introduction of innovative bioinformatics tools have greatly sped up the process of discovering viruses and virus-host interactions by enabling efficient analysis of NGS data. Analyzing these data sets enables a cost-effective and efficient way to handle large datasets and discovery of virus-host interactions [23].

In this study, we address three specific knowledge gaps through comprehensive genomic analysis of 317 E. amylovora strains and 227 Erwinia phages. First, the evolutionary origins of E. amylovora VFs remain largely uncharacterized. We therefore identified VFs using three approaches: RASTtk annotation, PHI-base homology screening (Sect.  3.1.2), and machine learning (VirulentPred 2.0) on hypothetical proteins (Sect.  3.1.3). Second, we identified endolysins from prophages, free phages, and bacterial genomes, then compared their physicochemical properties, subcellular localization, domain architectures (Sect.  3.3–3.4), and phylogeny (Sect.  3.5) to investigate potential evolutionary relationships between phage endolysins and bacterial hydrolases. Third, we identified defense and anti-defense systems across bacterial genomes, phages, and prophages (Sect.  3.6) and applied Spearman’s rank correlation with Bonferroni correction to detect non‑random associations (Sect.  3.7–3.8). Our integrated analysis aims to elucidate the complex molecular interactions that govern phage-host dynamics and evolutionary relationships in this agriculturally significant pathosystem.

Methods

Utilized genomic data sets

Genomes of 317 E. amylovora strains and 227 Erwinia phage sequences were initially retrieved from the National Center for Biotechnology Information (NCBI) in September 2025. These datasets were subsequently updated on December 5 and December 7, 2025, from the NCBI GenBank and the NCBI Virus portal, respectively. Also, genome sequences of seven complete prophages including Ph-FC01, Ph-Ea7-3, Ph-Ea644, Ph-Ea6-96, Ph-Ea2-97, Ph-Ea3-97, and Ph-Ea4-96 and segmented prophage genomes (in total 290 prophage sequences) that were predicted from our previous work were used [24].

Identification of VFs and ARGs in E. amylovora strains

The identification of ARGs and VFs in the E. amylovora genomes was performed through a multi-step process. For ARGs, all bacterial genomes were first annotated using the RASTtk pipeline [25] accessible via the Bacterial and Viral Bioinformatics Resource Center (BV-BRC) [26], followed by the extraction of all ARG-related proteins from the annotation output.

For VFs, a three-tiered identification strategy was employed to improve confidence in functional assignment. First, VFs and their related proteins (including components) were similarly identified from the initial BV-BRC annotation data. Candidate VFs were selected based on functional annotations corresponding to systems previously implicated in E. amylovora pathogenesis, as supported by the scientific literature [27]. Specifically, we included components of the Type III secretion system (T3SS), amylovoran biosynthesis machinery, flagellar motility apparatus, fimbriae/pili, the AcrAB-TolC multidrug efflux system, and the metalloprotease PrtA and its associated type I secretion apparatus. In addition, we incorporated Type VI secretion system (T6SS) components and lipopolysaccharide (LPS) biosynthesis proteins, as both systems have been noted in recent studies as emerging or putative virulence determinants in E. amylovora [28, 29].

Second, all predicted proteins were screened against the Pathogen-Host Interactions database (PHI-base) [30] using BLASTP with stringent criteria (an E-value threshold of ≤ 1e-50, percent identity greater than 50%, and no gaps allowed); hits satisfying these parameters were annotated as putative VFs. Assignment of VFs to specific tissue and organism categories was based on the experimentally validated host tissue infection data reported in the PHI-base entries for each homologous gene.

Third, to assess proteins of unknown function, all sequences classified as hypothetical proteins were analyzed for their virulence potential using the VirulentPred 2.0 tool [31]. This tool is a machine learning-based program designed to rapidly identify bacterial virulence proteins. It operates by training its models on a large dataset of experimentally validated virulent protein sequences. This model achieved a prediction accuracy of 85.18% using a position‑specific scoring matrix profile‑based model, making it a suitable and accurate model to identify putative VFs (Fig. 1).

Fig. 1.

Systematic workflow for the identification of VFs from hypothetical proteins. The flowchart outlines the stepwise bioinformatic pipeline. The process begins with the analysis of hypothetical proteins from two strains using VirulentPred 2.0 to establish a core database. This is followed by the analysis of all hypothetical proteins from all strains, the removal of sequences identical to the core database, prediction of virulence potential for the remaining unique sequences, and finally, the elimination of internal duplicates to produce the final dataset of putative virulence factors

No experimental validation was performed in this study. The virulence factors reported here are computationally predicted candidates intended to guide hypothesis‑driven research. The full set of predicted sequences is provided in Supplementary Material 1 and is also publicly available in the VfHub GitHub repository (https://github.com/RahimianXphage/VfHub). We recommend that candidate VFs of interest be subjected to functional characterization to confirm their role in E. amylovora pathogenesis.

Prophage identification and annotation

Prophages were identified within the bacterial genomes using PhaMer [32], a tool accessible through the PhaBOX web application [33]. PhaMer employs a Transformer-based language model to detect phage sequences. It constructs a protein-cluster vocabulary from input contigs and utilizes a self-attention mechanism to learn contextual relationships and protein organization, enabling accurate classification. These prophage sequences were then functionally annotated using the Pharokka tool [34], accessed via the Phage Galaxy web portal (https://phage.usegalaxy.eu/) [35].

Endolysin identification and analysis

Putative endolysin protein sequences were identified from multiple sources for comprehensive analysis. Initial identification was based on functional annotation of prophage regions using Pharokka. This core set was supplemented with endolysin sequences retrieved from the NCBI Virus database and by extracting proteins annotated as “phage peptidoglycan hydrolase” from the E. amylovora genome datasets via the BV-BRC platform.

All compiled endolysin sequences were subsequently analyzed using InterProScan [36] to characterize and predict their constituent protein domains. Furthermore, the physicochemical properties of the identified endolysins, including amino acid composition, theoretical isoelectric point (pI), instability index, and molecular weight (M.W) were calculated using the ProtParam web server tool available on the ExPASy server (https://web.expasy.org/protparam/) [37]. Subcellular localization predictions were performed using the CELLO v.2.5 server (https://cello.life.nctu.edu.tw/) [38, 39].

To investigate evolutionary relationships, a maximum likelihood phylogenetic tree was constructed using the IQ-tree program v.2.4.0 [40]. The best-fit substitution model was selected according to ModelFinder prior to tree inference [41].

Defense and anti-defense system identification

The DefenseFinder tool [42] was used for the identification of any defense and anti-defense systems in bacterial strains, downloaded phage genomes, previously identified prophages, and prophages that were identified in the current study.

Genomic island and synteny analysis

To determine whether co‑occurring defense and anti‑defense systems form discrete genomic islands, we performed a synteny‑based proximity analysis. For each of the 317 E. amylovora genomes, all annotated defense and anti‑defense system components were mapped to their respective contigs and coordinates (protein_start, protein_end) using the genome annotation files. Systems located on the same contig were sorted by genomic position and assigned to the same “island” if the distance between the end of one system and the start of the next was ≤ 50 kb, following the threshold used in previous studies of defense‑related genomic islands [43, 44]. Islands containing only a single system were retained as singletons.

To assess whether these islands are associated with mobile genetic elements (MGEs), we examined the flanking regions (± 5 kb upstream and downstream) of each island. The per‑strain genome annotation files were parsed to extract all genes within those flanking intervals. A gene was considered a mobile genetic element if its functional annotation contained any of the following keywords (case‑insensitive): integrase, transposase, tnp, recombinase, tRNA, conjugative, mobilization, relaxase, or type IV secretion. The presence of at least one such gene in the upstream or downstream flank was recorded.

Statistical analysis and data visualization

Statistical analyses were performed to assess relationships between the identified bacterial and prophage genomic elements. Potential correlations among defense and anti-defense systems were evaluated using Spearman’s rank correlation analysis, with the statistical significance threshold adjusted via the Bonferroni correction to account for multiple comparisons. This analysis was conducted using STATA software, version 14.2 (StataCorp, College Station, TX, USA). The results of these correlations were visualized as heatmaps, which were generated using Morpheus available at https://software.broadinstitute.org/morpheus.

To account for phylogenetic non‑independence in the correlation analyses, we attempted phylogenetic logistic regression using the phyloglm function in the phylolm R package (method = “logistic_MPLE”) on all 741 defense‑system pairs, with the core‑genome tree as the phylogenetic covariance structure. However, due to the binary nature of the data and the large tree (317 tips), only 7 pairs (0.9%) converged successfully, precluding reliable phylogenetic correction. Therefore, the Spearman correlations reported herein are uncorrected for phylogeny, and this limitation is acknowledged in the Discussion.

GraphPad Prism software, version 8.4.3 (GraphPad Software, Boston, MA, USA) was used to generate bar charts. The phylogenetic tree was created by iTOL v6 [45]. Bubble charts were created using the Matplotlib package v3.10.8-7-g1957ba3918 in Google Colab. For the structural characterization of endolysin proteins, domain architectures were visualized using the Domain Graph tool, DOG 2.0 [46].

Results

Identification of Virulence Factors and Antimicrobial Resistance Genes in the E. amylovora genomes

Based on the analysis of all 317 E. amylovora genomes, which collectively encoded 1,135,051 predicted protein sequences, a comprehensive set of ARGs and VFs was identified through annotation.

Annotation-based identification of ARGs and VFs

Annotation of the E. amylovora pangenome via the BV-BRC RASTtk pipeline successfully identified a core repertoire of ARGs and VFs from the total proteome. For ARGs, a total of 9,312 protein sequences across all strains were associated with antimicrobial resistance functions, primarily belonging to multidrug efflux systems (Supplementary Fig. 1A). The most prevalent systems were the MdtABC-TolC complex (1,939 sequences; 20.82%) and the Drug/Metabolite Transporter (DMT) superfamily (1,917 sequences; 20.59%), followed by the MSF-type tripartite multidrug efflux system (1,599 sequences; 17.17%) and the Bcr/CflA family (1,286 sequences; 13.81%). Other efflux systems, including AcrAB-TolC (956 sequences; 10.27%), EmrAB-OMF (638 sequences; 6.85%), MATE family (321 sequences; 3.45%), and AcrAD-TolC (330 sequences; 3.54%), were also detected. The protein MprA accounted for 319 sequences (3.43%).

For VFs, 61,561 protein sequences were identified from the annotation data and categorized into seven major functional groups implicated in E. amylovora pathogenicity (Supplementary Fig. 1B). The flagellar apparatus was the most abundant system, comprising 24,194 sequences (39.30%), followed by the T6SS with 18,532 sequences (30.10%) and the T3SS with 12,111 sequences (19.67%). Proteins related to fimbriae/pili assembly constituted 3,839 sequences (6.24%), while efflux systems annotated as VFs (e.g., AcrAB-TolC components) accounted for 1,287 sequences (2.09%). Biosynthesis machinery for the exopolysaccharide amylovoran and for LPS represented 955 (1.55%) and 643 (1.04%) sequences, respectively.

Screening against the PHI-base database

To identify known virulence-associated proteins, the complete proteome of the 317 E. amylovora strains was screened against the PHI-base. This analysis identified 79,677 protein sequences belonging to 328 distinct functions in 314 strains.

These 328 functions were subsequently grouped into 13 broad functional categories for analysis (Supplementary Fig. 2). The largest category was “Metabolism - General Enzyme,” comprising 31,830 sequences (39.95%). This was followed by “Transcription/Regulation” (11,628 sequences, 14.59%) and “Transport/Membrane” (9,980 sequences, 12.53%). The “Other/Miscellaneous” category contained 9,436 sequences (11.84%), while “DNA/RNA metabolism” accounted for 4,069 sequences (5.11%). Proteins of “Unknown/Hypothetical” function represented 3,832 sequences (4.81%). Smaller proportions were identified in categories such as “Binding/Structural” (4.31%), “Translation/Ribosomal” (2.03%), “Chaperone/Folding” (1.58%), “Energy/Respiration” (1.18%), “Signal/Sensor” (0.89%), “Cell wall/Peptidoglycan” (0.39%), and “Phage-related” (0.79%).

Analysis of the taxonomic origin of the PHI-base homologs identified sequence similarities with 49 different bacterial isolates. The highest number of homologs originated from E. amylovora itself (20,440 sequences), followed by Salmonella enterica (18,912) and Escherichia coli (10,689). Other considerable sources included Klebsiella pneumoniae (5,340), Pseudomonas aeruginosa (3,424), and various species within the genera Yersinia, Dickeya, Pectobacterium, and Xanthomonas.

Identified E. amylovora VFs from PHI-base

Screening against the PHI-base database identified E. amylovora VFs, excluding those associated with unaffected pathogenicity (Fig. 2). This analysis resolved the VFs into 51 distinct functional groups, which were categorized based on their primary target plant tissues. The majority of VF groups were associated with fruit tissues (30 groups), followed by shoot tissues (11 groups). Multi-tissue VF groups were also identified, including those targeting both fruit and leaf (5 groups) and leaf, fruit, and shoot collectively (3 groups). Two VF groups were specific to leaf tissue, one group targeted both leaf and seedling, and one group was specific to flowers. One additional VF group affecting both flower and shoot tissues was identified and categorized separately. This tissue-specific distribution highlights the diverse and specialized virulence strategies employed by E. amylovora during plant infection.

Fig. 2.

Distribution of E. amylovora VFs identified through PHI-base homology screening. VFs are grouped by the plant tissue they affect, excluding factors associated with unaffected pathogenicity. A total of 51 distinct VF groups were identified and categorized as follows: fruit (30 groups), leaf (2 groups), both fruit and leaf (5 groups), leaf and seedling (1 group), flower (1 group), shoot (11 groups), and leaf, fruit, and shoot (3 groups). One VF group affecting both flower and shoot is displayed separately. Colors represent phenotypic outcomes, and shapes indicate the degree of sequence similarity to known VFs in the PHI-base database. RV: reduced virulence; IV: increased virulence; LP: loss of pathogenicity; UP: unaffected pathogenicity. In the top‑right corner of each panel, the letter “I” indicates the sequence identity level of the E. amylovora homolog to the reference VF in the PHI‑base database (based on BLASTP percent identity)

Comparative analysis of E. amylovora VFs with homologs from other plant pathogens

A comparative analysis of E. amylovora VFs was conducted by screening against VF homologs from 12 bacterial and one fungal plant pathogen, excluding factors associated with unaffected pathogenicity (Fig. 3). This analysis identified 21 distinct VF groups categorized by their target plant tissues: leaf (11 groups), seedling (1 group), tuber (2 groups), both tuber and leaf (1 group), both seedling and leaf (1 group), and onion foliage and bulb (1 group). Among the compared pathogens, Dickeya dadantii presented the highest number of matching VF proteins, comprising 1,260 sequences distributed across four VF groups.

Fig. 3.

Distribution of E. amylovora VF groups based on homology to VFs from other plant pathogens. VF groups are categorized by their target plant tissue (leaf, seedling, tuber, etc.), excluding factors associated with unaffected pathogenicity. A total of 21 VF groups are shown. The analysis includes homologs from 13 reference pathogens: Erwinia tracheiphila; D. dadantii, D. fangzhongdai, D. zeae; P. carotovorum, P. atrosepticum; X. oryzae, X. citri, X. campestris; Pantoea stewartii, P. ananatis; Paracidovorax citrulli; and Magnaporthe oryzae. Colors represent phenotypic outcomes, and shapes indicate the degree of sequence similarity to the reference VFs. RV: reduced virulence; IV: increased virulence; LP: loss of pathogenicity. In the top‑right corner of each panel, the letter “I” indicates the sequence identity level of the E. amylovora homolog to the reference VF in the PHI‑base database (based on BLASTP percent identity)

Similarity of E. amylovora proteins to VFs from human/animal pathogens

To explore potential evolutionary parallels in virulence strategies, the E. amylovora proteome was screened for similarity to VFs from bacteria pathogenic to humans and animals, excluding groups associated with unaffected pathogenicity (Fig. 4). This analysis identified sequence homology to VFs from 14 distinct bacterial species. The highest number of shared VF groups was identified with S. enterica (49 groups), followed by E. coli (27 groups). Other notable matches included P. aeruginosa (10 groups), K. pneumoniae (9 groups), Yersinia pseudotuberculosis (4 groups), and Staphylococcus aureus and Y. pestis (2 groups each). Single VF group similarities were detected with Acinetobacter baumannii, Borrelia burgdorferi, Burkholderia pseudomallei, Haemophilus ducreyi, Haemophilus influenzae, Yersinia enterocolitica, and Yersinia marcescens.

Fig. 4.

Similarity of E. amylovora proteins to VFs from human and animal bacterial pathogens. The figure illustrates the number of shared VF groups between E. amylovora and 14 reference pathogens, excluding factors associated with unaffected pathogenicity. Colors indicate the number of matching VF groups, and shapes denote functional categories. RV: reduced virulence; IV: increased virulence; LP: loss of pathogenicity; UP: unaffected pathogenicity. In the top‑right corner of each panel, the letter “I” indicates the sequence identity level of the E. amylovora homolog to the reference VF in the PHI‑base database (based on BLASTP percent identity)

Similarity of E. amylovora proteins to VFs from animal pathogens

Screening for VF homology was extended to pathogens primarily affecting animals. This analysis identified similarities between E. amylovora proteins and VFs from seven animal pathogens, excluding those associated with unaffected pathogenicity (Fig. 5). A total of 14 VF groups were identified across these pathogens. Citrobacter rodentium presented the highest number of matching VF groups (5 groups). Similarities were also found with Edwardsiella ictaluri (2 groups), Edwardsiella tarda (2 groups), Francisella tularensis (2 groups), Flavobacterium psychrophilum (1 group), Xenorhabdus nematophila (1 group), and Glaesserella parasuis (1 group).

Fig. 5.

Similarity of E. amylovora proteins to VFs from animal pathogens. The figure illustrates the number of shared VF groups between E. amylovora and seven reference animal pathogens, excluding factors associated with unaffected pathogenicity. Colors correspond to the number of VF groups, and shapes denote functional categories. RV: reduced virulence; IV: increased virulence; LP: loss of pathogenicity; UP: unaffected pathogenicity; CT: chemistry target. In the top‑right corner of each panel, the letter “I” indicates the sequence identity level of the E. amylovora homolog to the reference VF in the PHI‑base database (based on BLASTP percent identity)

Prediction of VFs from hypothetical proteins

To elucidate the potential virulence functions of uncharacterized proteins, a systematic, machine learning-based analysis was performed on hypothetical protein sequences. An initial set of 2,481 hypothetical proteins extracted from two representative E. amylovora strains was analyzed using VirulentPred 2.0, identifying 1,046 sequences as putative VFs. These high-confidence predictions formed the foundational core database for subsequent screening.

This database was then used to analyze the complete pool of 321,034 hypothetical proteins extracted from all 317 strains. From this total, 180,029 sequences were identical to those already present in the core database and were set aside. The remaining 141,005 unique sequences were subjected to VirulentPred 2.0 analysis, and 127,470 internal duplicate sequences were subsequently removed. This rigorous filtering procedure yielded a final, non-redundant set of 13,535 putative virulence protein sequences. This final dataset, which represents putative VFs discovered from previously unannotated hypothetical proteins, is provided in FASTA format in Supplementary Material 1.

The identification of 13,535 putative VFs from hypothetical proteins warrants careful interpretation. This set likely contains false positives. We deliberately adopted a sensitive approach to minimize false negatives, accepting that experimental validation (e.g., knockout or complementation studies) will be required to confirm genuine virulence functions. Researchers should use this dataset as a hypothesis‑generating resource rather than a definitive list of virulence determinants.

Identified prophages and endolysin

A total of 195 novel prophages were identified in this study. The prophage sequences exhibited an average length of 42.11 kbp, with the longest sequence measuring 192.52 kbp and the shortest 3.14 kbp. Putative endolysin proteins were identified from the prophage sequences. After removing exact duplicate sequences, 38 unique endolysin sequences were predicted from the prophages identified in this study. Additionally, 54 unique endolysin protein sequences were identified from the Erwinia phage genomes retrieved from the NCBI Virus portal, following the removal of identical sequences. Moreover, 313 proteins with sequence similarity to phage peptidoglycan hydrolases were identified in E. amylovora genomes and were categorized into three different types based on their sequence similarity (Table S1).

Endolysin and phage hydrolase‑like proteins characterization

The identified endolysins and phage hydrolase‑like proteins were characterized based on physicochemical properties and subcellular localization (Table 1). Prophage‑derived lysins (n = 38) were the largest (average M.W 42.3 kDa; 378 residues), with 13 predicted unstable. Lysins from free Erwinia phages (n = 54) were smaller (20.2 kDa; 181 residues) and predominantly stable (51/54). The phage hydrolase‑like proteins (Types 1–3, n = 313 total) were the smallest (average M.W ~ 14.8 kDa; 129 residues) and all stable.

Table 1.

General characteristics of the identified endolysins and phage hydrolase‑like proteins

Protein type			Instability
	Count	Average M.W (Da)	Stable	Unstable	Average amino acid composition
Prophage lysin	38	42339.21	25	13	378
Phage lysin	54	20150.4	51	3	181
Type 1	291	14814.96	291	0	129
Type 2	21	14798.96	21	0	129
Type 3	1	14831.02	1	0	129

Stability (stable/unstable) is predicted by the ProtParam instability index. Unstable proteins may have short half‑lives; this can be functionally relevant rather than an artifact

pI differed notably: prophage lysins had the lowest average pI (8.31), while hydrolase‑like proteins had the highest (range 9.46–9.63). Amino acid composition (Fig. 6) showed prophage lysins were enriched in Ala, Gly, and Leu relative to the other groups. These physicochemical properties are presented descriptively to enable cross‑group comparison; their biological significance is unknown.

Fig. 6.

Circular visualization (Circos plot) summarizing the average pI and amino acid composition of five protein groups: prophage lysins (n = 38), free phage lysins (n = 54), and three types of bacterial phage hydrolase‑like proteins (Type 1, n = 291; Type 2, n = 21; Type 3, n = 1). Each protein group is represented as a colored column (segment) around the circle. Inner arc length (thickness of each column) is proportional to the total number of proteins in that group (absolute count). Outer arc length for each amino acid (colored ribbons) represents the proportion of that amino acid’s count within the group relative to the total count of that amino acid across all five groups. The pI values (average per group) are displayed as numbers inside the colored blocks. pI and amino acid composition are presented descriptively to allow cross‑group comparison; their biological significance is unknown and may reflect neutral evolutionary divergence or adaptation to different subcellular environments. Standard three‑letter amino acid abbreviations are used (Ala, Arg, Asn, Asp, Cys, Gln, Glu, Gly, His, Ile, Leu, Lys, Met, Phe, Pro, Ser, Thr, Trp, Tyr, Val)

Subcellular localization predictions showed distinct patterns. All Type 1 hydrolase‑like proteins (n = 291) were predicted cytoplasmic, all Type 2 (n = 21) periplasmic, and the single Type 3 cytoplasmic. This contrasts with phage lysins, which were predominantly periplasmic (28/54) or cytoplasmic (11/54). The cytoplasmic prediction for Type 1 hydrolases is notable because canonical endolysins act in the periplasm or from within the cytoplasm to degrade peptidoglycan. If these proteins retain hydrolase activity, they might act on peptidoglycan precursors or be neo-functionalized for non‑lytic roles – but experimental validation is required. Prophage lysins showed diverse localization (16 periplasmic, 11 cytoplasmic). Full details for each protein are in Supplementary Material 2.

Predicted domains in endolysins and hydrolases

Domain architecture analysis of the identified endolysin and phage hydrolase‑like proteins revealed a diverse repertoire of catalytic and accessory domains associated with peptidoglycan degradation (Table 2). Notably, a subset of proteins (n = 10) lacked identifiable domains using the employed tools and thresholds.

Table 2.

Summary of predicted protein domains in identified endolysins and hydrolase‑like proteins

Domain Category	Count	Subtypes and Notes
No Identifiable Domains	10
Lysozyme-related Domains	16	Lysozyme-like alone (7), With signal peptide (3), With Ca²⁺-binding site & disorder (3)
Murein Transglycosylases	3	With prokaryotic transglycosylase domain (1), With transglycosylase & signal peptide (1), With transglycosylase & disorder (1)
Glycoside Hydrolases	5	Glycoside hydrolase alone (2), With Ca²⁺-binding site (3)
N-acetylmuramoyl-L-alanine Amidase	1
Cell Wall Hydrolases	11	Cell wall hydrolase alone (2), With signal peptide region (9)
Endopeptidase-like	2	Endopeptidase-like alone (1), With consensus disorder (1)
Autolysin (lyz_endolysin_autolysin)	6
N-acetylmuramidase	30	With peptidoglycan-binding domain (26), With peptidoglycan-binding domain & signal peptide (3), Alone (1)
L-Ala-D-Glu Peptidase-like	2*	*Also present in all 313 identified phage hydrolase‑like proteins.
T4-like Lysozyme	1
Multi-domain Architectures	6	Endopeptidase + Lysozyme + Ca²⁺-binding site (4); Endopeptidase-like + Glycoside hydrolase + Ca²⁺-binding site + disorder (1); Amidase with AMIN & LysM domains (1)

*This count refers to occurrences in the endolysin set; the domain is ubiquitous in the separate set of phage hydrolase‑like proteins

Lysozyme-like domains were the most prevalent, identified in 16 proteins. These were further classified: 7 contained a lysozyme-like domain alone, 3 featured this domain coupled with a signal peptide, 3 possessed a lysozyme-like domain along with a predicted calcium-binding site, and 3 contained a lysozyme-like domain along with a calcium-binding site and consensus disorder region.

Three proteins were predicted as murein transglycosylases. One contained a prokaryotic transglycosylase domain, another featured this domain along with a signal peptide region, and a third combined the transglycosylase domain with a consensus disorder region.

Five proteins were identified as glycoside hydrolases. Two contained the glycoside hydrolase domain alone, while three contained this domain together with a predicted calcium-binding site.

Other distinct catalytic domains included N-acetylmuramoyl-L-alanine amidase (1 protein) and L-Ala-D-Glu peptidase-like domains, the latter being identified in 2 endolysins and all 313 phage hydrolase‑like proteins. A single protein contained a T4-like lysozyme domain.

Proteins predicted as cell wall hydrolases were also common (n = 11), with the majority (n = 9) containing an associated signal peptide region. Two proteins were classified as endopeptidase-like, one alone and one with a consensus disorder region. Six proteins were annotated with autolysin (lyz_endolysin_autolysin) domains.

The N-acetylmuramidase domain was highly frequent, found in 30 proteins. Most of these (n = 26) were coupled with a peptidoglycan-binding domain, three contained both the peptidoglycan-binding domain and a signal peptide region, and one contained the catalytic domain alone.

Several proteins exhibited complex, multi-domain architectures. Four proteins contained a combination of endopeptidase, lysozyme, and calcium-binding site domains. One protein featured an architecture combining endopeptidase-like and glycoside hydrolase domains, along with calcium-binding site and consensus disorder regions. A particularly complex architecture was observed in one N-acetylmuramoyl-L-alanine amidase, which also contained AMIN and LysM domains, as well as a signal peptide region; the domain model for this protein is illustrated in Supplementary Fig. 3.

Phylogenetic analysis of endolysins and hydrolases

A maximum-likelihood phylogenetic tree was constructed to examine the evolutionary relationships among the identified endolysin and hydrolase-like proteins (Fig. 7). The resulting topology showed several well-supported clades and positioned LysFaunus at the root of the tree. Bootstrap support values are shown at internal nodes, with two major branches highlighted by support values ≥ 95%.

Fig. 7.

Maximum-likelihood phylogenetic tree of Erwinia phage endolysins and E. amylovora phage hydrolase‑like proteins. The tree was inferred using IQ‑TREE v2.4.0 with the best‑fit substitution model selected by ModelFinder. Ultrafast bootstrap approximation with 1000 replicates was used. Protein labels are color-coded according to origin: green, bacterial phage hydrolase‑like proteins (Hydrolase1–3); blue, endolysins from free Erwinia phages; orange, endolysins identified from prophages. Circles indicate ultrafast bootstrap support values. Two major branches with bootstrap support ≥ 95% are highlighted

A prominent clade contained all three E. amylovora phage hydrolase‑like groups (Hydrolase1–3) together with numerous endolysins derived from Erwinia phages. Hydrolase1 and Hydrolase2 formed a sister grouping supported by a bootstrap value of 95, consistent with their high sequence similarity. Hydrolase3 clustered within the same major lineage but occupied a distinct position adjacent to a diverse assemblage of characterized endolysins, including LysvB_EamM-Bue1 and LysphiEa2809.

A separate lineage comprised LysvB_EamM-Y2, which was sister to a large clade containing the Hydrolase1-3 groups and LysEa-type endolysins. This clade was clearly separated from the Hydrolase1–3 lineage by long internal branches and multiple highly supported nodes. Several additional endolysin clusters were observed throughout the tree, reflecting the substantial sequence diversity present among Erwinia phage lysins.

Overall, the phylogeny suggests that the bacterial phage hydrolase‑like proteins may be more closely related to specific subsets of phage endolysins than to other lysin groups.

Identification of defense and anti-defense systems in bacterial, prophage, and phage genomes

Analysis of defense and anti-defense systems across phages, prophages, and bacterial genomes revealed distinct prevalence patterns (Fig. 8). In bacterial genomes, defense systems were dominated by a core set of highly prevalent mechanisms. The most widespread systems included RM Types IV, II, and III (identified in 607, 469, and 301 genomes, respectively), Class1-Subtype-I-E (in 316 genomes), and toxin-antitoxin systems such as Mok_Hok_Sok (316 genomes) (Fig. 8A). In contrast, a larger group of 23 defense systems exhibited low prevalence (found in 1–18 genomes), including Gabija, AbiE, Retron_VI, Dnd_ABCDE, and several recently discovered systems like Lamassu and RosmerTA.

Fig. 8.

Prevalence of defense and anti-defense systems in bacterial and phage genomes. A Bar chart showing the high prevalence (top, counts ≥ 42) identified defense systems in bacterial genomes. B Prevalence of anti-defense systems in bacterial genomes. C Prevalence of anti-defense systems identified in phage genomes. System ‘gnarl1’ is categorized as ‘other’. D Identified defense systems and the single anti-defense system within prophage regions

Bacterial anti-defense systems were primarily represented by the dam (581 genomes), with other anti-restriction systems like ardC and ardK present at lower frequencies (75 and 53 genomes, respectively) (Fig. 8B). Phage-encoded anti-defense systems showed a different profile. The most common were anti-CBASS (acb1; 61 genomes) and NARP1 (25 genomes), followed by the phage-encoded dam (48 genomes) (Fig. 8C).

Prophages carried a limited repertoire of both defense and anti-defense systems. The most frequently identified defense system in prophages was Mok_Hok_Sok (7 occurrences). Anti-defense was rarely observed, with only a single instance of ardC detected (Fig. 8D).

Co-occurrence and relationships among defense and anti-defense systems in E. amylovora

Spearman’s rank correlation analysis, adjusted for multiple comparisons using the Bonferroni correction, was performed to evaluate potential associations between the identified defense and anti-defense systems across the E. amylovora genomes. In total, 91 pairs of systems exhibited p-values equal to zero, 1 pair with p ≤ 0.0001, 13 pairs with p ≤ 0.001, 9 pairs with p ≤ 0.01, and 2 pairs with p ≤ 0.05. These correlations included both positive associations (suggesting potential synergy or co-deployment) and negative associations (indicating possible mutual exclusivity or antagonistic relationships). The full correlation matrix is visualized as a heatmap (Fig. 9), with significant pairs highlighted for clarity.

Fig. 9.

Heatmap of Spearman’s rank correlation coefficients (ρ) among defense and anti-defense systems identified in E. amylovora genomes. Color intensity represents the strength and direction of the correlation (red for positive, blue for negative), with ρ values ranging from − 1 to + 1. Significant correlations after Bonferroni correction are marked as follows: **** p ≤ 0.0001, *** p ≤ 0.001, ** p ≤ 0.01, * p ≤ 0.05, ▲ p = 0. Systems are clustered hierarchically on both axes to highlight patterns of co-occurrence and mutual exclusivity. To avoid redundancy, only the upper triangle of the correlation matrix is displayed; the lower triangle is identical

Among the pairs with the highest significance (p = 0), 4 relationships with ρ = 1 were identified: DS-36 with Rst_3HP and DarTG, Rst_3HP with DarTG, and HEC-06 with Pycsar were identified. Other notable positive correlations included Menshen with Gabija (ρ = 0.7026), PD-T7-5 with ardK (ρ = 0.8843), Kiwa with RosmerTA (ρ = 0.706), ardC with PD-T7-5 (ρ = 0.6828), ardC with ardK (ρ = 0.7766), Gao_Her_DUF with SDIC1 (ρ = 0.8632), Menshen with RM_Type_III, and RM_Type_III with Dnd_ABCDEFGH (ρ = 0.5923), AbiE with RM_Type_I (ρ = 0.7497), and Gabija with RM_Type_I (ρ = 0.5232). Conversely, prominent negative correlations at p = 0 included RM_Type_II with MazEF (ρ = -0.3456), Lamassu-Cap4_nuclease with CAS_Class1-Subtype-I-E (ρ = -0.4037), Gao_Tmn with CAS_Class1-Subtype-I-E (ρ = -0.706), RM_Type_III with RM_Type_I (ρ = -0.4749), Gabija with RM_Type_III (ρ = -0.4996), AbiE with RM_Type_III (ρ = -0.4081), Hachiman with RM_Type_III (ρ = -0.6487), Divona with RM_Type_III (ρ = -0.5979), Dnd_ABCDEFGH with RM_Type_I (ρ = -0.8911), and AbiE with Dnd_ABCDEFGH (ρ = -0.7713).

For the pair with p ≤ 0.0001, a positive correlation was observed between RM_Type_II and Retron_IV (ρ = 0.2916). At p ≤ 0.001, all identified correlations were negative. Key examples included Mok_Hok_Sok with ardb_klca_acric11 (ρ = -0.2918), Retron_IV with SDIC1 (ρ = -0.2724), Prometheus with Lamassu-Cap4_nuclease (ρ = -0.2808), Gao_Her_DUF with RM_Type_II (ρ = -0.2852), RM_Type_I with RM_Type_II (ρ = -0.2862), and a cluster of consistent negative associations between Dnd_ABCDEFGH and several systems: Divona (ρ = -0.2666), HEC-06 (ρ = -0.2666), Pycsar (ρ = -0.2666), Rst_3HP (ρ = -0.2666), DS-36 (ρ = -0.2666), DarTG (ρ = -0.2666), RosmerTA (ρ = -0.2666), and CBASS_I (ρ = -0.2666).

Correlations at p ≤ 0.01 were mixed in direction. Negative associations included RM_Type_III with Dazbog (ρ = -0.2363), acrie9 (ρ = -0.2363), and Gao_Tmn (ρ = -0.2363); ardC with Dnd_ABCDEFGH (ρ = -0.2806); RM_Type_II with Dnd_ABCDE (ρ = -0.2548) and Gao_Tmn (ρ = -0.244); while positive ones comprised ardC with RM_Type_I (ρ = 0.2524) and AbiE with acrie9 (ρ = 0.244) and Gao_Tmn (ρ = 0.244). Finally, at the p ≤ 0.05 threshold, two additional correlations were identified: a positive association between Prometheus and RM_Type_II (ρ = 0.232), and a negative correlation between Retron_IV and Gao_Her_DUF (ρ = -0.2351).

Correlations among phage anti-defense systems

To explore potential functional relationships and co-evolutionary patterns in phage-encoded anti-defense mechanisms, we applied Spearman’s rank correlation analysis with Bonferroni correction for multiple testing on the identified systems within the Erwinia phage genomes. Out of the analyzed pairs, five demonstrated p-values of 0, while one exceptional pair reached p = 0.0508—a value that, although not as definitive, borders on conventional significance (p ≤ 0.05) and may represent a subtler or context-specific interaction deserving of special consideration. These findings are depicted in a comprehensive heatmap of correlation coefficients (Fig. 10).

Fig. 10.

Heatmap of Spearman’s rank correlation coefficients (ρ) among anti-defense systems identified in Erwinia phage genomes. Color intensity denotes the strength and direction of correlations (red for positive, blue for negative), with ρ values ranging from − 1 to + 1. Significant pairs after Bonferroni correction are marked as follows: ▲ p = 0; the marginally significant pair (p = 0.0508) is marked by ♦. Systems are hierarchically clustered on both axes to reveal patterns of synergy and antagonism. To avoid redundancy, only the upper triangle of the correlation matrix is displayed; the lower triangle is identical

The highly significant correlations (p = 0) encompassed a mix of positive and negative ρ values. Positive associations were particularly striking for gad2 and NARP2 (ρ = 1), dam and acb1 (ρ = 0.7426), and gp4.5 and samase (ρ = 0.8119). On the negative side, NARP1 showed strong inverse relationships with acb1 (ρ = -0.8508) and dam (ρ = -0.67).

Genomic island and synteny analysis of defense and anti-defense systems

To examine the genomic context and potential for horizontal transfer of defense and anti-defense systems, a synteny‑based proximity analysis was performed across all 317 E. amylovora genomes. A total of 9,026 individual system components were mapped to their respective contigs and, using a proximity threshold of ≤ 50 kb between adjacent systems, these components were clustered into 2,336 distinct genomic islands. The analysis of flanking regions (extending 5 kb upstream and downstream of each island) revealed a close association with MGEs. Specifically, an MGE (e.g., integrase, transposase, or recombinase) was identified in the upstream flank of 1,094 islands (46.8%) and in the downstream flank of 1,055 islands (45.2%). Overall, 1,550 islands (66.4%) possessed at least one MGE in either flank, and 599 islands (25.6%) had MGEs on both sides. The types of MGEs observed included transposases (e.g., InsN/InsO of IS911), phage integrases, prophage integrases, tyrosine recombinases, and occasionally CRISPR‑associated proteins. The presence of these elements in close proximity to defense and anti‑defense gene clusters suggests that a majority of these islands are situated in regions with potential for mobility, possibly facilitating their dissemination within the pangenome. The complete dataset of genomic islands, including their constituent systems, coordinates, and flanking MGE annotations, is provided in Supplementary Material 3.

Discussion

Our genomic analysis identified extensive homology between E. amylovora proteins and virulence determinants from diverse bacterial species. PHI-base screening identified 79,677 protein sequences belonging to 328 distinct functions across 314 strains, with matches to 49 different bacterial species. The highest number of homologs originated from S. enterica (18,912 sequences) and E. coli (10,689 sequences), despite their different host ranges. Further analysis resolved 51 distinct VF groups in E. amylovora with specific tissue-targeting patterns, where 30 groups specifically affected fruit tissues. Comparative analysis showed that the plant pathogen D. dadantii shared the highest number of matching VF proteins (1,260 sequences across four groups), while surprisingly, human pathogens like S. enterica shared 49 VF groups and E. coli shared 27 VF groups with E. amylovora.

These findings are consistent with our first hypothesis that VFs from diverse bacterial species can illuminate evolutionary pathways. Many E. amylovora virulence genes have homologs in other pathogens, suggesting possible horizontal gene transfer or functional conservation. For example, the E. amylovora type III effector DspE is a near-identical functional homolog of the AvrE effector from Pseudomonas syringae [47]. Indeed, the dspE gene encodes a 198-kDa pathogenicity protein shares homology with avrE of P. syringae, and it is co-located with dspF, which encodes a small chaperone related to Yersinia/Shigella Yop chaperones (Syc proteins) [48]. In other words, the entire dspEF locus in E. amylovora mirrors the avrE locus of P. syringae (and even the secretion chaperone avrF) [47]. Similarly, filamentous adhesins like HecA are broadly conserved. The D. dadantii (called Erwinia chrysanthemi in the cited reference) hecA gene (encoding a 3,850-amino-acid Bordetella-like adhesin) is required for full virulence, and HecA homologs are present in both plant and animal bacterial pathogens and appear to be unique to pathogens, being universally found in necrogenic plant pathogens [49]. This is consistent with the possibility that HecA-like adhesins may have been horizontally acquired, although alternative explanations such as vertical inheritance with differential retention cannot be excluded without formal phylogenetic testing. Notably, even genes of non-bacterial origin can become virulence factors; for instance, an operon encoding a microbial expansin (plant cell-wall loosening protein) was horizontally acquired by E. tracheiphila and significantly enhanced its ability to wilt cucurbits [50]. Rocha et al. found that an expansin–GH5 operon is required for systemic xylem colonization in E. tracheiphila, and they point out that expansin genes occur in many bacterial and fungal wilt pathogens, suggesting that microbial expansins may represent an under‑recognized class of VFs [50]. By extension, the presence of expansin-like or other shared virulence genes in E. amylovora could, if functionally confirmed, provide insights into its evolution and host adaptation.

In sum, the detection in our analysis of E. amylovora genes closely related to known VFs from diverse bacteria (plant and animal pathogens alike) is consistent with a pattern that could reflect cross‑species exchange and divergence of pathogenicity determinants, but this inference is based solely on sequence homology. Such comparative insights are consistent with the hypothesis that many E. amylovora virulence traits have been inherited from or shared with other bacteria, and that tracing these shared factors can clarify their evolutionary history and functional roles in this plant pathogen.

Moreover, the prediction of VFs from hypothetical proteins via VirulentPred 2.0 expands the known repertoire, yielding 13,535 putative factors. This underscores the untapped potential in unannotated genomes, where machine learning suggests hidden virulence determinants, potentially involved in novel infection strategies. Functional validation of these predicted VFs – for example, through gene knockout or complementation studies in E. amylovora, followed by plant infection assays – will be required to determine whether they contribute to pathogenesis. The large numbers of protein sequences reported in this study should be interpreted in the context of the E. amylovora pangenome. These counts represent occurrences of protein families across 317 strains, not functionally independent proteins. Therefore, the numbers reported primarily illustrate the abundance and conservation of these systems across the species, not a vast repertoire of unique virulence determinants. When discussing distinct functional groups (e.g., VF groups in Fig. 1), we refer to categories of experimentally characterized virulence functions, which provide a more meaningful measure of virulence repertoire diversity.

Our second hypothesis proposes that some bacterial phage hydrolase‑like proteins may share a common evolutionary origin with phage endolysins. The phylogenetic analysis suggests a close evolutionary association: the three types of E. amylovora phage hydrolase‑like proteins form a distinct clade closely related to phage endolysins, with LysFaunus positioned as a basal outgroup. Critically, the L-Ala-D-Glu peptidase-like domain that characterizes all 313 phage hydrolase‑like proteins is also present in LysFaunus and LysvB_EamM-Y2, providing preliminary correlative evidence consistent with a possible evolutionary relationship. The shared domain architecture may reflect evolutionary conservation of catalytic functions. While canonical endolysins primarily contain lysozyme-like or N-acetylmuramidase domains for peptidoglycan degradation, the shared L-Ala-D-Glu peptidase-like domain suggests that these proteins may have retained peptidoglycan-processing capabilities despite sequence diversification. Supporting this, subcellular localization predictions show a dramatic functional shift: while phage lysins are predominantly periplasmic (28 of 54 sequences) where they access peptidoglycan, all 291 Type 1 hydrolases are cytoplasmic, indicating that if these proteins are functional hydrolases, their cellular roles would differ from those of classical phage endolysins. However, the predicted cytoplasmic localization could also indicate that these proteins have been entirely repurposed for non-lytic functions. The physicochemical properties further indicate evolutionary streamlining: phage hydrolase‑like proteins (average M.W. ~14.8 kDa) are markedly smaller than prophage-derived lysins (42.3 kDa) yet maintain stability, with uniformly high pI values (9.46–9.63) compared to prophage lysins (8.31). Currently, the biological significance of this difference is unknown; it may reflect adaptation to different cellular compartments or simply neutral evolutionary divergence. This conservation of key molecular features between specific endolysins (LysFaunus, LysvB_EamM-Y2) and bacterial phage hydrolase‑like proteins is consistent with the possibility that some bacterial hydrolases originated from phage-derived ancestors, although additional experimental studies are required to determine their biological functions.

Previous studies have demonstrated that phage-derived lytic proteins can be co-opted for alternative functions in bacterial systems, providing biological precedents that may help interpret the patterns observed in our dataset. For instance, in Y. enterocolitica strain W22703, a phage-related holin/endolysin (holY-elyY) cassette resides on an “insecticidal pathogenicity island” and is essential for secretion of a high-molecular-weight toxin complex [51]. Sänger et al. showed that this dual holin/endolysin pair (now considered a type 10 secretion system) releases toxin subunits without causing host-cell lysis, representing a novel adaptation of a phage‑related function for bacterial toxin release [51]. Likewise, Salmonella Typhi uses a phage-like N-acetyl-β-d-muramidase (TtsA) to export typhoid toxin. TtsA retains homology to phage endolysins and cleaves peptidoglycan for toxin translocation [51]. Serratia marcescens has an analogous system (ChiW/ChiX) to secrete chitinases. Moreover, in Clostridioides difficile, the phage-like holin TcdE and its partner endolysin TcdL are required to export the large toxins TcdA and TcdB. Importantly, C. difficile TcdE/TcdL does not lyse the cell but instead mediates toxin release through a nonlytic mechanism [51]. These examples show that phage-encoded lytic enzymes can be co-opted by bacteria as specialized hydrolases or secretion factors that modulate peptidoglycan to secrete virulence molecules rather than destroy the host cell. The identification in our study of peptidoglycan-hydrolase homologs in E. amylovora genomes is consistent with this trend. Analogous Yersinia and Salmonella systems indicate that phage-derived peptidoglycan hydrolases can acquire non-lytic cellular functions. Whether the E. amylovora hydrolase‑like proteins identified here perform comparable roles remains unknown. It is also possible that these proteins have no hydrolase activity at all and instead serve unrelated functions. The sequence similarity to phage endolysins could reflect distant homology with divergent function. Experimental validation is required to determine whether these proteins indeed cleave peptidoglycan. In particular, the phylogenetic proximity between some endolysins and bacterial hydrolases is compatible with a historical co-option event, although direct evidence for such a transition is currently lacking.

Our analysis of E. amylovora genomes identified systematic associations between host defense systems (RM, CRISPR-Cas, toxin-antitoxin) and mobile anti-defense genes. Among 91 significant pairs, correlations showed strong co-occurrence and mutual exclusion. Positive associations included near-perfect linkage between DS-36, Rst_3HP and DarTG, and between HEC-06 and Pycsar. Anti-restriction genes ardC and ardK also strongly co-occurred with specific host defenses, suggesting linked function. Conversely, antagonistic patterns emerged: RM_Type_III frequently excluded RM_Type_I and other systems (e.g., Gabija, Hachiman; ρ ≈ − 0.6 to − 0.89), and the dnd system was almost never found with certain RMs. In phage genomes, similar trends appeared, with strong positive correlations between anti-defense pairs (e.g., dam/acb1; gad2/NARP2) and strong negative ones (e.g., NARP1 versus dam/acb1).

The observed co-occurrence patterns between bacterial defense systems and phage-encoded anti-defense genes are consistent with the expectations of a co-evolutionary arms race, but also alternative explanations must be considered. First, bacteria frequently encode multiple anti-phage systems, and some co-occurring pairs have been shown experimentally to exhibit synergistic antiphage activity [52]. Wu et al. found that certain defense system pairs (e.g., Zorya II with Druantia III, or tmn with Gabija) co-occur and act cooperatively to improve phage resistance [52], suggesting that selection can favor compatible defense repertoires. However, co-occurrence in genomic data does not prove synergy; it may also result from genetic linkage (e.g., systems located on the same mobile element) or neutral processes such as preferential attachment in defense islands. In this context, anti-defense genes also play potential roles. Phages and other mobile elements often carry counter-defense tools (anti-CRISPRs, anti-RM, etc.), and some have been hypothesized to be co-opted by bacteria [53]. Samuel et al. showed that conjugative plasmids’ leading regions are hotspots that encode a wide range of anti‑defense systems, including anti‑CRISPR, anti‑restriction, and anti‑SOS proteins [54].

In our data, we observed positive correlations between certain defense systems and anti-defense genes (e.g., dam with RM_Type_IV; acb1 with CBASS_I). While these patterns are consistent with the hypothesis that anti-defense loci co-occur with their cognate defense systems, we cannot exclude that these correlations arise from shared evolutionary history or sampling biases. Similarly, prophages in our dataset carried defense and anti-defense systems, but the low frequency of these events (e.g., Mok_Hok_Sok in 7 prophages, ardC in 1 prophage) precludes strong evolutionary conclusions. Therefore, we interpret these correlations as genomic signatures that are compatible with, but not definitive proof of, an ongoing arms race. Functional assays and longitudinal studies are needed to establish causality. We attempted to correct for phylogenetic non‑independence using phylogenetic logistic regression, but convergence failures (only 7 of 741 pairs converged) prevented reliable correction. Consequently, the Spearman correlations reported in Sect.  3.7 are uncorrected for shared ancestry. Thus, some of the observed correlations may be influenced by the evolutionary relationships among strains rather than functional interactions between systems. Future studies with larger sample sizes or alternative phylogenetic methods (e.g., phylogenetic generalized least squares for continuous traits) are needed to disentangle these effects.

Co-occurrence patterns hint at mutual reinforcement or trade-offs; clusters of defense genes often appear near anti-defense genes, and some authors describe genomic anti-defense islands where multiple phage countermeasures cluster [53]. For example, Bernheim et al. report that some anti-defense genes co-localize in islands in phages (e.g. E. coli T4 and lambda), although many anti-defense genes are found alone [53]. In practice, this means that a bacterial genome or its prophages may simultaneously encode a suite of antagonistic systems (e.g. a CRISPR array alongside a phage-derived anti-CRISPR) whose interplay can tune the outcome of infection. The patterns we observe in E. amylovora (correlations between the presence of defense loci (CRISPR, RM, toxin–antitoxin, etc.) and prophage-carried counter-defense genes) mirror these broader trends. Collectively, the literature shows that such co-occurrence is not random but shaped by selection; defense systems cooperate to enhance immunity [55], while anti-defense systems hitchhike on mobile elements to facilitate gene flow [54, 55]. In the ecological context of a plant pathogen like E. amylovora, these dynamics likely influence phage–host interactions on the plant surface and within plant tissues, impacting virulence and fitness.

At the same time, we note important caveats. Not every co-occurrence necessarily indicates a functional partnership. Some patterns might arise from neutral processes or shared genetic linkage rather than selection. For instance, Wu et al. themselves cautioned that defense islands could evolve via “preferential attachment,” accumulating multiple systems simply in genomic regions permissive to insertions [52]. In this view, a cluster of defense and anti-defense genes might not interact directly but happen to transpose together. In our dataset, a few associations had modest rho values (ρ ≈ 0.2–0.3), which could arise by chance in a large pangenome. Moreover, very strong statistical correlation in presence–absence does not prove mechanistic synergy. It will be important to follow up with functional assays to confirm that specific gene pairs do in fact cooperate or counteract one another. Thus, while our E. amylovora co-occurrence patterns broadly reflect the expectations of a co-evolutionary arms race, they are not absolute rules.

Conclusion

This study provides a comprehensive genomic analysis of the phytopathogen E. amylovora, identifying intricate evolutionary and functional relationships in its virulence arsenal and its dynamic conflict with phages. Our findings suggest that E. amylovora’s VFs share considerable evolutionary pathways with diverse bacterial pathogens, highlighting extensive horizontal gene transfer. Furthermore, we identified bacterial phage hydrolase‑like proteins that exhibit phylogenetic, structural, and domain-level similarities to phage endolysins. These findings are consistent with a potential evolutionary relationship between the two protein groups, although experimental validation is required to determine their origins and biological functions. Crucially, our analysis identifies systematic, non-random associations between bacterial defense systems and phage-encoded anti-defense genes, characterized by strong patterns of co-occurrence and mutual exclusivity. These patterns are consistent with genomic signatures that could reflect an ongoing co-evolutionary arms race, though correlational data cannot establish causation. Our findings provide testable hypotheses for future functional and longitudinal studies. Collectively, these findings deepen our understanding of E. amylovora’s pathogenicity and phage interactions, offering new insights for developing targeted biocontrol strategies.

Authors’ contributions

M.R. and M.A.S. conceptualized the study. M.R. developed the methodology and software, conducted the investigation and data curation, and performed validation and visualization. M.R. and M.A.S. contributed to the methodology and software. M.R. wrote the original draft, and all authors reviewed and edited the manuscript.

Funding

The author declares that no funds, grants, or other support were received during the preparation of this manuscript.

Data availability

All genomic sequence data analysed in this study are available from the National Center for Biotechnology Information (NCBI). The curated datasets of predicted virulence factors and negative controls are deposited in the VfHub repository (https://github.com/RahimianXphage/VfHub). This repository contains species‑specific FASTA files of VirulentPred2‑predicted virulence factors for E. amylovora and Agrobacterium burrii, as well as a combined non virulent (negative control) file.

Code Availability

Any custom code used in this study is available in the public GitHub repository: https://github.com/MohErfan/Erwinia-amylovora-virulence-factors-and-prophage-co-evolution. The repository includes a step‑by‑step configuration guide and complete environment setup to run the standalone version of VirulentPred 2.0 on Ubuntu, as well as the DiamondX component used for homology‑based screening against the PHI‑base database.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.