Spatiotemporal genomic analysis and risk assessment of the plasmids carrying blaOXA-48-like genes based on a large-scale international dataset Authors: Qiheng Yuan, Junlin Wang, Xiaowei Liu, Xumeng Yu, Jiatao Li, Du Guo, Qinru Jing, Yongliang Lou, Yutong Kang, Meiqin Zheng Published: 2026 ABSTRACT AbstractBackgroundThe spread of OXA-48-like carbapenemases represents a major public health challenge. Although previous studies have investigated OXA-48-like carbapenemases risk factors, nosocomial dissemination, and plasmid dynamics, an integrated plasmid-centered framework combining complete plasmid mining, transmission-unit analysis, phylogenetic reconstruction, and machine learning–based risk assessment remains limited.MethodsWe systematically collected 747 complete plasmid sequences carrying blaOXA-48-like genes from the NCBI database, establishing the largest collections of complete plasmid sequences to date. Using an integrative framework of population genomics, phylogenetic dating, and machine learning, this study aimed to characterize the dissemination patterns, plasmid replicon diversity, transmission units, mobile genetic elements, co-resistance profiles, and risk classification of these plasmid.ResultsPlasmids carrying blaOXA-48-like genes were detected across 50 countries on six continents, with blaOXA-48 predominating in Europe, blaOXA-181 in South Asia, and blaOXA-232 largely in Asia. IncL and ColKP3/IncX3 replicons, together with Tn1999.2 and other MGEs, were central drivers of plasmid maintenance and spread. Sixteen transmission units were defined, with AA068_Cluster3 estimated to have originated in the Netherlands around 2005 before expanding to Europe, the Middle East, Asia, and North America. Co-resistance analyses revealed frequent modules involving aminoglycoside and quinolone resistance, with qnrS1 and aph(3’’)-Ib most prevalent. Notably, high-risk transposon structures were often identified in non-clinical environments, underscoring their cross-ecological transmission potential. Machine learning-based classification models showed good internal performance for predefined composite-risk categories, with plasmid mobility, clinical/non-clinical source composition, and host background contributing to the classification results.ConclusionsThis study provides a large-scale plasmid-centered genomic analysis of publicly available complete plasmid sequences carrying blaOXA-48-like genes, integrating transmission-unit inference, phylogeographic reconstruction, mobile genetic element and co-resistance profiling, and composite genomic risk stratification. This gene-centered framework may support future One Health–oriented antimicrobial resistance surveillance and prioritization of plasmids with higher dissemination and resistance potential.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12866-026-05340-w.Keywords: Plasmids carrying blaOXA-48-like genes, Antimicrobial resistance, Spatiotemporal dynamics, Horizontal gene transfer, One Health, Machine learning risk assessment Introduction The World Health Organization, the European Union, the United States government, and the U.S. Centers for Disease Control and Prevention have identified the global dissemination of antibiotic-resistant bacteria as one of the most serious threats to human health [1]. Carbapenems are often regarded as the last resort for treating infections caused by multidrug-resistant Gram-negative bacteria; however, their effectiveness is being undermined by the emergence and worldwide spread of carbapenemase-producing organisms [2, 3]. Among these, class D carbapenemases of the OXA-48 family represent a particularly critical threat [4, 5]. OXA-48, the most prevalent member of the OXA-48-like enzymes worldwide, was first reported in Klebsiella pneumoniae in Istanbul, Turkey, in 2004 [6]. Since then, OXA-48-like enzymes have been reported predominantly in Enterobacterales, particularly Klebsiella pneumoniae and related Klebsiella spp., as well as Escherichia coli, members of the Enterobacter cloacae complex, Citrobacter freundii, Serratia marcescens, and other clinically important species [7–10]. To date, bacteria carrying the gene have achieved a truly global distribution, being isolated not only from humans but also from animals and diverse environmental sources across multiple continents [4, 11–13]. The dissemination of blaOXA-48 is driven by both horizontal gene transfer and vertical clonal expansion, primarily associated with various Tn1999 elements and their derivatives. Among these, plasmids belonging to the IncL family are considered to play a central role in the stable maintenance of blaOXA-48 [14, 15]. In addition, other plasmid types such as IncF, IncX3, and IncR have also been reported to harbor blaOXA-48-like genes. These mobile genetic elements enable the genes to spread efficiently across different bacterial populations through horizontal transfer, thereby facilitating their broad dissemination [16, 17]. To better understand the transmission trajectories, molecular mechanisms, and associated risks of horizontal transfer of OXA-48-like carbapenemases, further investigation into the epidemiological characteristics of plasmids carrying blaOXA-48-like genes is essential. This includes plasmids carrying blaOXA-48, blaOXA-181, blaOXA-232, blaOXA-204, blaOXA-162, blaOXA-484, and blaOXA-244 genes. Previous studies have substantially advanced our understanding of the genetic features, plasmid backgrounds, mobile genetic elements, and transmission dynamics of OXA-48-like carbapenemases and other carbapenemase-encoding plasmids across different clinical, bacterial, and geographic contexts [14, 18–23]. An integrated plasmid-centered framework that combines complete plasmid mining, transmission-unit definition, Bayesian phylogenetic dating, phylogeographic reconstruction, mobile genetic element and co-resistance analysis, and interpretable machine learning–based classification or stratification of predefined plasmid risk categories remains limited. This gap restricts our ability to systematically evaluate how these plasmids disseminate across regions, hosts, and ecological sources, and to identify plasmids with high transmission risk. In this study, we curated a large-scale collection of publicly available complete plasmid sequences carrying blaOXA-48-like genes and performed comparative analyses from a population genomics perspective. We investigated plasmid diversity, local genetic contexts, putative transmission units, and model-based phylogeographic patterns of these plasmids based on currently available complete plasmid data. We also explored their distribution across clinical and non-clinical sources, while recognizing the potential influence of source-related sampling bias. Furthermore, we developed a composite plasmid-level genomic risk stratification framework and evaluated internal supervised classification performance for identifying plasmids assigned to predefined upper risk categories. Building upon these analyses, we propose a gene-centered analytical framework for tracking and prioritizing antimicrobial resistance gene-carrying plasmids using complete plasmid sequence data. This framework may be adaptable to other antimicrobial resistance genes and may inform future genomic surveillance and plasmid-mediated resistance risk assessment. Method Collection of plasmids carrying blaOXA-48-like genes A total of 747 completed sequences of these plasmids carrying blaOXA-48, blaOXA-181, blaOXA-232, blaOXA-204, blaOXA-162, blaOXA-484, and blaOXA-244 genes [4, 17], were retrieved from the NCBI database in April 2025 using gene names as search terms. Retrieved records were manually curated, and only sequences annotated as plasmids, marked as complete, and accompanied by complete temporal and geographic metadata were retained. Incompatibility (Inc) groups were determined using PlasmidFinder v2.1 [24]. Plasmid mobility was determined using Mob-suite [25]. MOB-suite classifies plasmids as conjugative, mobilizable, or non-mobilizable according to the presence of mobility-associated modules, including relaxase genes, mate-pair formation/type IV secretion system components, and origin-of-transfer elements. Metadata, including geographic origin, sampling year, host species, isolation source, length and plasmid closed status are summarized in Table S1. Animal and environmental sources were grouped as non-clinical sources for downstream analyses. Host background was assigned according to whether the annotated bacterial host was included in the Catalogue of Pathogenic Microorganisms Infectious to Humans issued by the National Health Commission of China [26]. Annotation of antimicrobial resistance genes Open reading frame (ORF) sequences were annotated using Prokka v1.12 [27] to facilitate downstream antibiotic resistance gene (ARG) identification. ARG prediction was performed by integrating results from AMRFinderPlus v4.0.22 (database 2025–03–25.1; parameters: amrfinder –report_all_equal –plus –ident_min 0.95 –coverage_min 90) [28] and RGI v6.0.4 (CARD v4.0.1; parameters: rgi main –local –clean –low_quality –include_nudge) [29] via the hAMRonization pipeline v1.1.4 [30]. To avoid redundant annotations, entries with identical gene symbols, strand orientation, and genomic coordinates (within ± 10 nt) were considered duplicates, and only the AMRFinderPlus annotations were retained [31]. Construction of transmission units Upstream and downstream sequences of the genes were extracted using SeqKit [32]. The 50-kb upstream and 50-kb downstream flanking regions of each blaOXA-48-like genes were subsequently clustered with CD-HIT [33] at 99.5% sequence identity to characterize their genetic contexts [34]. A transmission unit was defined by combining secondary clusters predicted by Mob-suite [25] with the clustering results of the genes genetic environments [34]. For transmission network construction, isolates sharing an identical transmission unit were considered genetically related and putatively linked through a closely related plasmid-mediated resistance-gene context. Phylogenetic and phylogeographic analysis of the largest transmission unit For the largest transmission unit, sequences were clustered with CD-HIT [33] at 99.5% identity, and the most abundant cluster representative was selected as the reference. Snippy [35] with default settings was then used to detect SNPs within this transmission unit. Phylogenetic relationships among representative plasmids were reconstructed using BEAST v1.10.5 [36] under the HKY substitution model and an uncorrelated relaxed clock to estimate divergence times. HKY is suitable for handling heterogeneous base frequencies and differences in transition/transversion rates, while the relaxed clock model is appropriate for situations where evolutionary rates may vary across different plasmid lineages. A Yule process was applied as the tree prior. The resulting phylogeny was interpreted as an approximation of SNP-based genetic relatedness and temporal diversification [37]. Markov chain Monte Carlo (MCMC) analyses were run for 2 × 10⁸ iterations with sampling every 20,000 steps, and convergence was confirmed using Tracer v1.7.2 [38], ensuring effective sample sizes > 200 for all major parameters. The final Maximum Clade Credibility (MCC) tree was generated using TreeAnnotator v10.5.0 with a burn-in fraction set to 0.1. Figtree v1.4.4 [39] was used for tree visualization. In addition, SpreaD3 [40] was employed to generate a geographic diffusion map from the MCC tree using discrete traits, with latitude and longitude coordinates assigned to the capitals of the corresponding countries or regions. Screening of insertion sequences in AMR-gene flanking regions To examine the potential linkage between insertion sequences (ISs) and ARGs on plasmids, genomic regions spanning 5 kb upstream and downstream of each ARG were retrieved and analyzed for the presence of ISs using an identical workflow. IS elements were detected via BLASTN homology searches against the ISfinder database [41], with an E-value cutoff of ≤ 1 × 10–10. Only matches exhibiting ≥ 80% nucleotide identity and ≥ 80% alignment coverage were considered [42]. When multiple hits satisfied these criteria, the match with the highest bit score was selected for further analysis. Detection of presumably recent horizontal transfer events of ARGs In line with the criteria described by Wang et al. [43], ARGs were classified as putative recent transfer events when they showed > 99% nucleotide identity and complete (100%) sequence coverage across plasmids originating from different bacterial hosts, as determined by BLASTn (E-value ≤ 1 × 10–5). To avoid misclassification due to entire plasmid transfer, plasmids were considered independent entities only if their pairwise alignments demonstrated less than 80% identity and coverage. Subsequently, all candidate ARG sequences were extracted from the plasmid datasets and subjected to pairwise sequence comparison. To strengthen the evidence for horizontal transfer, the genomic context of each candidate ARG was examined. Only ARGs flanked within 5 kb by at least one same IS element were retained. This two-step filtering strategy—combining high sequence conservation across plasmids with physical linkage to mobile IS elements—yielded a subset of conserved ARGs most likely to represent recent transfers between environmental and clinical plasmid pools. Applying these criteria allowed systematic identification of ARGs potentially involved in such transfers. As an illustrative case, we selected ten representative plasmids and employed genoPlotR [44] to detect insertion events involving gene and their ± 5 kb flanking regions through collinearity analysis. Composite risk stratification of plasmids Each plasmid was assigned a composite risk score based on multiple scoring items grouped into three major domains: host pathogenicity, transmission potential, and accessory burden. The detailed scoring criteria and individual scoring items are provided in Table S4. Composite scores were ranked and stratified by percentiles: 1–20% (low risk), 20–40% (low-middle risk), 40–60% (middle risk), 60–80% (middle-high risk), and 80–100% (high risk) (Table S5). This percentile-based framework allowed standardized relative risk comparison across the plasmid dataset. Machine learning-based classification of plasmids assigned to upper composite-risk categories A machine learning framework incorporating 11 algorithms—including Random Forest, Gradient Boosting, SVM, Logistic Regression, k-NN, PLS, Bayesian methods, Lasso regression, Neural Networks, Discriminant Analysis, and Boosting—was used to classify plasmids into High-risk (top 40%) and non-High-risk categories. Input features included: (i) assignment to a transmission unit, (ii) temporal span of the unit, (iii) number of host genera involved, (iv) geographic diversity of countries, (v) coexistence of clinical and non-clinical plasmids, (vi) plasmid mobility, (vii) number of AMR genes carried, (viii) presence of IS elements flanking blaOXA-48-like genes, and (ix) host background. Models were trained (70%) and tested (30%) on a stratified dataset. Performance was evaluated by ROC curves, confusion matrices, decision curve analysis (DCA), trend plots, and AUC forest plots, with 95% confidence intervals reported. SHAP analysis was applied to interpret feature contributions. This workflow provided a robust and interpretable framework for the identification of high-risk plasmids carrying blaOXA-48-like genes. Result Large-scale international dataset reveals extensive but uneven geographical distribution and genetic diversity of plasmids carrying blaOXA-48-like genes After quality control and manual curation, we constructed a large-scale international dataset comprising 747 complete plasmids carrying blaOXA-48-like genes (harboring a total of 756 blaOXA-48-like genes, as a small number of plasmids carried multiple copies). Figure 1 illustrates the global distribution and sources of these plasmids, with countries reporting fewer than 10 sequences grouped as “Others.” blaOXA-48-like genes has been detected across Europe, Asia, North America, South America, Africa, and Oceania, with a total of 49 countries and regions reporting these plasmids. The highest numbers were found in European countries such as the Netherlands, Switzerland, and Spain, as well as in China from Asia. Notably, India, the United States, and countries classified as “Others” also contributed considerable numbers of blaOXA-48-like genes. Most plasmids originated from clinical sources; however, non-clinical plasmids were first observed in 2015 and appeared to increase notably during 2017–2021 (Figure S1). Figure 1A further shows the genus-level distribution, with Klebsiella and Escherichia representing the predominant hosts, while Enterobacter in the Netherlands and Citrobacter in Spain were also frequent carriers. This observation is broadly consistent with previous epidemiological studies in which Klebsiella pneumoniae and Escherichia coli have been identified as major hosts of OXA-48-like carbapenemases [17, 22]. The detection of these plasmids in Enterobacter, Citrobacter, and other genera further supports the potential for cross-genus dissemination among Enterobacterales. Temporally, reports of these plasmids have steadily increased from 2010 to 2014, with a particularly sharp rise observed after 2014. An apparent decline was observed after 2020; However, this trend should be interpreted cautiously because recent records in public databases may be affected by sequencing, curation, and submission delays rather a true epidemiological decline. Different regions exhibited distinct distributions of these plasmids variants: plasmids carrying blaOXA48 predominated in Europe, whereas plasmids carrying blaOXA232 and blaOXA-181 were more common in South Asia (Fig. 1B). Temporal, source, bacterial genus, and geographical distribution of the plasmids carrying blaOXA-48-like genes. A Distribution of isolates by collection time, source, bacterial genus, and country. B Global distribution map of different OXA-48-like variants Meanwhile, we found that 560 plasmids (74.97%) were identified as conjugative, among which 407 carried blaOXA-48 and 139 carried blaOXA-181. In addition, 140 plasmids (18.74%) were classified as mobilizable, of which 103 carried blaOXA-232 (Figure S2, Table S1). These findings underscore both the extensive geographical dissemination and the genetic diversity of these plasmids, highlighting their broad transmission potential. Plasmid length analysis further revealed structural variability among different plasmid variants (Figure S2). Although plasmids carrying blaOXA-48, blaOXA-181, blaOXA-232, blaOXA-204 generally showed relatively concentrated length distributions within each variant group, a few plasmids displayed markedly divergent sizes. Notably, plasmids carrying blaOXA-232 were mainly small plasmids, whereas plasmids carrying blaOXA-204 tended to be substantially larger. These results suggest that the plasmids carrying blaOXA-48-like genes include diverse plasmid backbones with variant-specific size patterns (Figure S3). Determination of Inc types A total of 58 Inc types were identified among all plasmids. The three most prevalent types were IncL (382, 51.1%), ColKP3/IncX3 (120, 16.1%), and ColKP3 (118, 15.8%). The top 15 most frequent Inc types were displayed individually, whereas the remaining less frequent Inc types were grouped as “Others” (Fig. 2, Table S1). Among these, 183 plasmids (24.5%) were fusion plasmids containing at least two Inc types, the majority of which were associated with the ColKP3 replicon (143, 78.1%). In addition, 0.4% of plasmids had Inc types that remained unclassified. Distribution of incompatibility (Inc) types among the plasmids carrying blaOXA-48-like genes. The Sankey diagram shows the relationships among OXA-48-like variants, Inc types, and isolation sources. Colors on the left represent different OXA-48-like variants, colors in the middle represent different Inc-type categories, and colors on the right represent clinical and non-clinical sources. The width of each flow is proportional to the number of plasmids in the corresponding category. The top 15 most frequent Inc types are shown individually, whereas the remaining low-frequency Inc types are grouped as “Others” We further examined the correspondence between different plasmids and Inc types. Fisher’s exact test with Monte Carlo simulation showed a significant association between OXA variants and Inc-type categories (P < 0.001). Cramer’s V indicated a strong association between OXA variants and Inc-type categories (Cramer’s V = 0.64). A total of 26 Inc types were associated with plasmids carrying blaOXA-48, with the majority corresponding to IncL (83.6%). IncL was also identified as the characteristic replicon type specific to plasmids carrying blaOXA-48, while IncM1 and IncHI1B(pNDM-MAR) were also observed at notable frequencies. In contrast, plasmids carrying blaOXA-181 was associated with 23 Inc types, of which 70.7% corresponded to ColKP3/IncX3. Notably, among the 120 ColKP3/IncX3 plasmids, only two carried plasmids carrying blaOXA-484, whereas all others were linked to plasmid carrying blaOXA-181. The correspondence was even more pronounced for plasmids carrying blaOXA-232, which was associated with five Inc categories, with 95.4% belonging to ColKP3. Other plasmids variants were linked to distinct Inc types, for example, plasmids carrying blaOXA-204 was associated with IncC, IncFIB(pB171), IncFII(S), and IncFII(SARC14). Across all Inc categories, seven were detected in both clinical and environmental sources. Among these, plasmids carrying blaOXA-48 was associated with IncL, Col156, IncM1, and IncFII(pSE11); plasmids carrying blaOXA-181 with ColKP3, IncX3, and IncR (in combination with ColKP3); and plasmids carrying blaOXA-204 exclusively with ColKP3. Spatiotemporal dynamics of the plasmids carrying blaOXA-48-like genes transmission units To better infer the transmission events of these plasmids, we integrated the secondary cluster IDs identified by Mob-Suite with clusters of the 50-kb upstream and downstream flanking regions of blaOXA-48-like genes, thereby defining a new variable termed “transmission unit.” Plasmids assigned to the same transmission unit were considered to share an epidemiological relationship. As shown in Fig. 3 and TableS1, a total of 16 transmission units were identified. The three most prevalent were AA068_Cluster3, AA055_Cluster27, and AA041_Cluster10, comprising 159, 100, and 93 plasmids, respectively. Notably, 157 of the 159 plasmids in AA068_Cluster3 carried the IncL replicon; all 100 plasmids in AA055_Cluster27 carried ColKP3; and all 93 plasmids in AA041_Cluster10 carried ColKP3/IncX3, with every plasmid classified as a fusion plasmid. These findings highlight the critical role of specific replicon types in the transfer and maintenance of these plasmids. Distribution characteristics of different transmission units, including source (clinical or non-clinical), year of isolation, geographic region, bacterial genus, incompatibility (Inc)type, and host background. Values represent absolute counts of isolates In terms of OXA variants, AA068_Cluster3 corresponded to OXA-48, AA055_Cluster27 to OXA-232, and AA041_Cluster10 to OXA-181. Interestingly, in the case of OXA-232, 100 out of 109 plasmids (91.74%) were clustered into the same transmission unit, further underscoring the high dissemination risk of these plasmids. Geographically, AA068_Cluster3 was predominantly distributed in Europe, followed by Asia and North America; AA055_Cluster27 was mainly detected in Asia; and AA041_Cluster10 showed comparable activity in both Asia and Europe. With respect to host diversity, most transmission units spanned multiple genera: for instance, AA068_Cluster3 encompassed seven genera, AA041_Cluster10 six genera, and AA068_Cluster4 five genera. By contrast, AA055_Cluster27 was identified in only three genera, suggesting a more limited interspecies transmission potential. The majority of strains within transmission units originated from clinical sources. For example, clinical isolates accounted for 83.6% (133/159) of AA068_Cluster3 and 97% (97/100) of AA055_Cluster27. The majority of plasmids within transmission units originated from clinical sources. However, this pattern should be interpreted cautiously because complete plasmids are more frequently recovered and sequenced from clinical settings than from animal or environmental sources. Thus, the limited non-clinical representation within some transmission units may reflect source-related sampling bias rather than true absence of non-clinical dissemination. Nevertheless, non-clinical plasmids were detected in several transmission units, suggesting their potential contribution to the broader dissemination of these plasmids. Strikingly, non-clinical sources contributed to 31% (26/84) of AA068_Cluster4 and 63.6% (7/11) of AA068_Cluster14, indicating the presence of environmental transmission pathways. Importantly, each representative transmission unit was strongly associated with a specific replicon type, such as IncL for AA068_Cluster3, ColKP3 for AA055_Cluster27, and ColKP3/IncX3 for AA041_Cluster10. From a temporal perspective, only a few transmission units were annotated during the early stage (2010–2012), consistent with initial clinical outbreaks and clonal dissemination of blaOXA-48-like genes. The number of identified units increased from 2013 to 2017, peaking in 2018–2019, and subsequently declined after 2020. Of note, multiple units, including AA068_Cluster3, AA055_Cluster27, AA041_Cluster10, and AA068_Cluster4, were detected over extended periods, suggesting their sustained role in the dissemination of these plasmids. While some units exhibited limited spread, their localized circulation in specific regions underscores their adaptive potential. Increasing genetic variability may further drive the emergence of transmission units with enhanced adaptability. Inference of OXA-48-like transmission trajectories using Bayesian analysis To statistically infer the origin and cross-country dissemination pathways of the plasmids carrying blaOXA-48-like genes, we performed Bayesian phylogenetic dating and ancestral state reconstruction using BEAST on AA068_Cluster3, the largest transmission unit. The Maximum Clade Credibility (MCC) tree based on SNPs from AA068_Cluster3 is shown in Fig. 4A. The reconstruction indicated that AA068_Cluster3 most likely originated in the Netherlands around 2005, followed by transmission to Switzerland and subsequent widespread dissemination across Europe. In 2008, the lineage was transmitted from Spain to Lebanon and other Middle Eastern countries, and in 2011 from Lebanon to the United States. Further transmission events were inferred from Germany to China in 2013 and from the Netherlands to India in 2017, ultimately forming a transcontinental dissemination network spanning Europe, Asia, and North America. Phylogenetic analysis based on SNPs of the largest transmission unit (AA068_Cluster3). A Maximum clade credibility (MCC) tree. The x-axis of branch nodes indicates the estimated divergence time, and each branching process from the ancestral node to the descendant node represents a transmission event. B Transmission map of plasmids in the AA068_Cluster3 cluster. Lines connecting countries represent different transmission events. Countries with a major impact on transmission are highlighted in red, and the estimated timing of key transmission events is also indicated Of note, after 2016, evidence suggested bidirectional transmission between clinical and non-clinical environments in Europe, accompanied by frequent interspecies transfers. The geographic distribution map highlighted transmission hotspots primarily in Europe, with the Netherlands, Switzerland, Spain, and Germany making major contributions to dissemination (Fig. 4B). Collectively, these results provide theoretical support for the inferred trajectories of key transmission units and further confirm the rapid and frequent interregional and interspecies spread of these plasmids over the past two decades. Co-occurrence of ARGs in the plasmids carrying blaOXA-48-like genes A total of 2,097 antimicrobial resistance genes (ARGs) were identified by integrating AMRFinderPlus and RGI annotations. After excluding the index blaOXA-48-like genes used to define the plasmid dataset, 1,309 additional ARGs remained for co-resistance analysis, including 38 ARGs uniquely detected by RGI. In total, 96 unique ARGs were detected, spanning 15 major antibiotic classes, including β-lactams, aminoglycosides, quinolones, macrolides, sulfonamides, among others. Genes conferring resistance to multiple antibiotic classes were categorized into a separate class, designated as MULTIDRUG (Fig. 5, Table S2). Coexistence of OXA-48-like and other resistance genes and classes. A Number of resistance genes (excluding blaOXA-48-like genes itself) and their corresponding plasmid counts (line represents the number, bar chart represents the proportion). B Distribution of resistance classes (excluding blaOXA-48-like genes) and their corresponding plasmid counts (line represents the number, bar chart represents the proportion). C Correlation between different plasmid variants and coexisting resistance classes, with a distinction between clinical and non-clinical sources. D Different resistance modules composed of various coexisting resistance classes and their corresponding counts The number of resistance classes annotated per plasmid ranged from 1 to 10. Among them, 499 plasmids carried only one resistance class, comprising 430 plasmids from clinical sources (66.56% of all clinical isolates) and 69 from non-clinical sources (68.32% of all non-clinical isolates). Additionally, 146 plasmids harbored two resistance classes (Fig. 5A). When excluding blaOXA-48-like genes, the most prevalent resistance class was quinolone resistance genes, detected in 160 plasmids (135 from clinical and 25 from non-clinical sources). Other frequent classes included β-lactam (non-OXA-48-like) and aminoglycoside resistance genes (Fig. 5B). To further characterize co-resistance patterns, we classified the ARG combinations into 59 distinct resistance modules. Among them, 50 modules were found exclusively in clinical isolates, 5 exclusively in environmental isolates, and 4 were shared between clinical and environmental sources (Fig. 5D). The three most common modules were: module 1 (quinolone), module 6 (aminoglycoside + β-lactam), and module 5 (aminoglycoside + β-lactam + quinolone). We also examined the association of different these plasmids variants with co-resistance classes and genes. Most plasmid types were predominantly associated with clinical isolates, although some exceptions were notable. For example, plasmids carrying blaOXA-162 were primarily associated with environmental sources (green links), whereas plasmids carrying blaOXA-181 were detected in both clinical and non-clinical sources (grey links) (Fig. 5C). Among co-occurring resistance determinants, aminoglycoside and β-lactam resistance genes were the most diverse and abundant, with qnrS1, aph(3’’)-Ib, and aph(6)-Id being the most frequently detected co-resistance genes (Figure S4, Table S2). Local genetic contexts and IS-associated mobilization of blaOXA-48-like genes To distinguish gene-level mobilization from plasmid-level conjugative dissemination, we analyzed the local genetic contexts of blaOXA-48-like genes by examining IS elements within ± 5 kb flanking regions. This analysis was intended to identify IS- or transposon-associated mobilization platforms, rather than to infer plasmid conjugation events. Analysis of the genetic contexts of blaOXA-48-like genes (± 5 kb flanking regions) revealed remarkable diversity, with a total of 2,103 insertion sequences (ISs) belonging to 28 families being annotated (Fig. 6, Table S3). The most prevalent elements were IS10A (40.85%), ISKpn91 (27.53%), and IS1R (20.21%), with distinct OXA-48-like variants associated with specific IS families. Notably, the elements annotated as IS10A in our IS annotation workflow correspond to the IS1999 elements described in the classical Tn1999 platform, reflecting nomenclature differences among IS annotation databases. The primary IS elements mediating OXA-48 transfer were IS10A and IS1R, whereas OXA-181 was predominantly associated with ISKpn19 and IS15DIV. In contrast, OXA-232 was linked to ISEc9 and ISKpn21, albeit at low frequencies (five ISs detected in four plasmids). Interestingly, a substantial proportion of these IS elements were detected in plasmids from non-clinical environments. In particular, IS elements associated with blaOXA-48 and blaOXA-181 were identified across both clinical and environmental sources, underscoring the cross-ecosystem transfer potential of resistance determinants mediated by mobile genetic elements such as IS10A and ISKpn19 (Fig. 6A). Major mobilization elements mediating the plasmids carrying blaOXA-48-like genes transfer and their genetic contexts. A IS elements annotated within the 5-kb upstream and downstream regions of different OXA-48-like variants and their associations with various sources. B Representative genetic environments of OXA-48-like variants To further assess the risks posed by different MGEs, we conducted synteny analysis of the 10 plasmids with the highest degree of connectivity in the ARG co-occurrence network (Figure S5). The analysis revealed that plasmids carrying blaOXA-48 represented the highest risk for resistance dissemination. The key MGE structures mediating the transfer of blaOXA-48 and its co-residing resistance genes were composite transposons formed by upstream IS10A and IS1R together with downstream IS10A. Notably, we also identified transposon units comprising upstream IS1R and downstream IS10A, which mediated similar transfer events (Fig. 6B). Importantly, all of the 10 plasmids analyzed originated from non-clinical environments, highlighting the significant risk of resistance dissemination driven by these transposon units in environmental settings. Risk assessment of the plasmids carrying blaOXA-48-like genes To further evaluate the risks associated with the plasmids carrying blaOXA-48-like genes, we developed a plasmid risk assessment model. Comprehensive benchmarking of 11 classification machine learning models demonstrated robust and reliable performance. In the training dataset, although the BayesMethod and PLSModel performed relatively poorly, all other models achieved outstanding results across multiple evaluation metrics, including accuracy, F1 score, negative predictive value (NPV), positive predictive value (PPV), sensitivity, specificity, and Youden’s index, with most values consistently above 0.90 (Fig. 7A). The NeuralNet and Lasso models achieved the highest area under the receiver operating characteristic curve (AUC = 0.999), followed by Gradient Boosting (AUC = 0.998), Random Forest, SVM_Kernel, BoostingMethod, and DiscriminantModel (all AUC = 0.996). Both LogisticModel (AUC = 0.992) and NeighborMethod (AUC = 0.989) also performed strongly, whereas PLSModel (AUC = 0.933) and BayesMethod (AUC = 0.947) showed slightly lower but still acceptable performance (Fig. 7C). Evaluation of the plasmids carrying blaOXA-48-like genes transmission risk using multiple models. Performance metrics of different models in the training set and test set (A, training set; B, test set). Receiver operating characteristic (ROC) curves and area under the curve (AUC) scores for different models in the training and test sets (C, training set; D, test set). Decision curve analysis (DCA) curves for different models in the training and test sets, where the gray dashed line represents the strategy of treating all, the black dashed line represents treating none, and curves positioned higher indicate greater net benefit (E, training set; F, test set) As expected, model performance declined in the validation dataset, yet the vast majority of metrics (62/63, 98.4%) for models other than BayesMethod and PLSModel remained above 0.85, indicating good internal classification performance in the held-out test dataset (Fig. 7B). AUC analysis confirmed this trend: the LogisticModel achieved the highest AUC (0.947), followed by SVM_Kernel (0.945), Lasso and NeighborMethod (both 0.940), DiscriminantModel and BoostingMethod (both 0.939), Random Forest and Gradient Boosting (both 0.935), NeuralNet (0.928), BayesMethod (0.917), and PLSModel (0.865) (Fig. 7D). Exploratory decision curve analysis (DCA) showed that the model-based strategy provided higher net classification benefit than the “classify-all” and “classify-none” strategies across a range of thresholds for identifying predefined high-risk plasmids. This analysis was based on composite-risk labels rather than clinical outcomes. (Fig. 7E, F), supporting their potential clinical utility. Overall, while all models performed well in the training dataset, LogisticModel, SVM_Kernel, and Lasso exhibited particularly strong and consistent performance in the validation dataset, underscoring their potential reliability in real-world applications. To further examine the contribution of individual risk factors to overall plasmid risk, we selected the SVM_Kernel model, which achieved the highest average AUC across training and test sets, for subsequent analysis. The model attained an accuracy of 99.2% in the training set and 94.7% in the held-out test set, demonstrating excellent classification performance and generalization ability (Fig. 8A, B). To interpret the model predictions, we applied SHAP (SHapley Additive exPlanations) analysis, which quantified the relative importance of predictor variables. The most influential feature was plasmid mobility (mean |SHAP|= 0.0995), followed by clinical versus non-clinical source within transmission units (0.0932) and host background (0.0791) (Fig. 8C). The SHAP summary plot further illustrated how each feature contributed to the model classification output: positive SHAP values increased the probability of assignment to the predefined high-risk category, whereas negative SHAP values shifted the output toward the non-high-risk category; the color gradient indicates the relative value of each feature (Fig. 8D). These findings were consistent with the biological and epidemiological context of plasmid-mediated resistance. Performance of the representative model and SHAP (SHapley Additive exPlanations) analysis. Model performance in the training and test sets (A, training set; B, test set). C Feature importance of the representative model (mean absolute SHAP values). D SHAP value distribution (bee swarm plot) of the representative model Discussion In this study, we systematically collected and analyzed 747 complete plasmids sequences carrying blaOXA-48-like genes, thereby constructing large-scale international dataset to date. By integrating population genomics, phylogenetics, and machine learning approaches, we comprehensively elucidated their transmission patterns and resistance risks. Compared with previous fragmented epidemiological reports, our dataset’s scale and analytical depth substantially enhance the understanding of the evolutionary landscape of these plasmids. Our findings further confirm their widespread dissemination across Europe, Asia, and North America, and reconstruct their cross-national and cross-host evolutionary trajectories. Notably, whereas earlier studies emphasized the predominance of these plasmids in the Middle East, Africa, and the Indian subcontinent [4, 45, 46], our results highlight a stronger expansion trend in Europe (particularly in the Netherlands and Switzerland), revealing frequent cross-national, intercontinental, interspecies, and cross-ecological transmission events. Importantly, we observed a sharp increase in these plasmids dissemination after 2014, reflecting the escalating severity of its spread, whereas the subsequent decline in numbers after 2020 may, at least in part, reflect delays in data submission to public databases rather than a true reduction in transmission. Importantly, these plasmids were not only highly prevalent in clinical isolates but were also increasingly detected in non-clinical sources, suggesting that cross-ecological transmission pathways may accelerate the spread of resistance, thereby posing a potential public health threat. This finding underscores the importance of coordinated surveillance across clinical, human, animal, and environmental settings under the One Health framework, in order to achieve a more comprehensive understanding and containment of the global dissemination of resistance plasmids. This trend is consistent with observations reported in several recent epidemiological studies [47–49]. The study revealed that specific replicons, such as IncL and ColKP3/IncX3, may play important roles in the maintenance and dissemination of these plasmids. This finding is consistent with previous reports highlighting the importance of IncL plasmids in the stable maintenance of blaOXA-48 [50]. However, the presence of blaOXA-48-like genes, including on IncL plasmids, does not necessarily indicate high-level phenotypic resistance, because resistance levels may be influenced by plasmid-host adaptation, fitness cost, gene expression, and other host genomic factors [51]. Furthermore, we identified distinct associations between gene variants and replicon types: for example, blaOXA-181 was predominantly carried by ColKP3/IncX3 plasmids, while blaOXA-232 was mainly associated with ColKP3 plasmids. These OXA variant–Inc type associations may reflect biological advantages of certain plasmid backbones, such as stability, host compatibility, or transfer potential [52], but they may also be influenced by database sampling bias. We defined transmission units to more accurately infer the dissemination trajectories of these plasmids. By combining plasmid grouping with transmission unit analysis, this study not only elucidates the evolutionary dynamics of plasmid dissemination but also provides new insights into the mechanisms driving the spread of antimicrobial resistance plasmids. Through BEAST-based Bayesian phylogenetic analysis of AA068_Cluster3, we inferred that its origin could be traced back to the Netherlands around 2005. Following its emergence, it rapidly disseminated across Europe before spreading to the Middle East, Asia, and North America, ultimately forming a transcontinental transmission network spanning three continents. This analysis not only provided temporal evidence through phylogenetic dating but also revealed the possibility of bidirectional transmission between clinical and non-clinical environments in Europe, along with multiple cross-species transfer events. These findings underscore that the spread of the plasmids carrying blaOXA-48-like genesis not driven by clonal expansion alone but rather by a combination of horizontal plasmid transfer and multi-host adaptation [53]. Importantly, this highlights the need for surveillance strategies that extend beyond clinical outbreak investigations to include environmental sources and cross-species transmission pathways, thereby enabling a more comprehensive understanding and effective containment of the global dissemination of these plasmids [4, 11]. We found that these plasmids frequently co-harbor multiple antimicrobial resistance genes, with qnrS1 and aph(3’’)-Ib being particularly prevalent, indicating that these plasmids not only confer carbapenem resistance but also commonly accumulate additional resistance to aminoglycosides, quinolones, and other antibiotic classes, thereby significantly increasing clinical treatment complexity and the risk of therapeutic failure [54, 55]. The formation of such multidrug resistance modules is consistent with previously reported mechanisms of resistance gene accumulation mediated by MGEs [43, 56], and the detection of modules shared across clinical and environmental sources further underscores the cross-ecological dissemination risk [47]. Importantly, we identified recurrent insertion sequences, including IS10A (IS1999 in blaOXA-48-like genes), IS1R, and ISKpn19, in the genetic backgrounds of these plasmids, highlighting their central role in mediating horizontal transfer, with their presence in both clinical and environmental isolates emphasizing the potential for cross-ecological spread. These findings are consistent with prior studies by João Botelho, You Che, and others, reinforcing the role of MGEs as key drivers of plasmid evolution and global dissemination [42, 57]. Further genetic context analysis revealed that the highest-risk transmission events were consistently associated with MGEs composed of IS10A (upstream and downstream) together with IS1R (upstream), corresponding to the Tn1999.2 variant among the many Tn1999 variants [17, 58], whereas alternative arrangements such as IS91–IS1R–IS10A, although not belonging to the Tn1999 family, still posed notable transmission risks. Strikingly, the ten plasmids all originated from non-clinical environments yet were conjugative plasmids carried by pathogenic bacteria, providing compelling evidence of their high-risk nature and highlighting the potential for cross-ecological and interspecies dissemination that poses a serious global public health concern. This study further developed a machine learning-based classification pipeline to evaluate whether plasmids could be assigned to predefined upper composite-risk categories. The models showed good internal classification performance, with most algorithms achieving high AUC values in the held-out test dataset. Among them, LogisticModel, SVM_Kernel, and Lasso showed relatively consistent performance across the training and held-out test datasets, suggesting that the predefined composite-risk categories could be internally classified using the selected genomic and metadata-derived features. However, these results should be interpreted as internal classification performance rather than external validation or prediction of future plasmid behavior. SHAP analysis indicated that plasmid mobility, coexistence of clinical and non-clinical sources within transmission units, and host background contributed to the classification results. Importantly, these variables should not be interpreted as direct evidence of experimentally confirmed transmissibility, pathogenicity, or clinical outcome risk. In particular, host background was assigned according to host annotation and a catalogue-based host classification, rather than virulence-gene profiling or wet-lab pathogenicity assays. Therefore, this framework should be regarded as an exploratory tool for prioritizing plasmids within the current dataset. Future studies incorporating virulence-gene profiling, phenotypic resistance data, host genomic background, and experimental validation will be needed to further refine plasmid risk assessment. In summary, by systematically collecting and analyzing 747 complete plasmids carrying blaOXA-48-like genes, we comprehensively delineated their global evolutionary and dissemination patterns. Our study identified IncL, ColKP3/IncX3, and other replicons, as well as mobile genetic elements such as Tn1999.2, as central drivers of plasmid maintenance and spread, while also uncovering cross-species and cross-ecological transmission pathways that highlight the critical role of environmental sources in resistance dissemination. By integrating the definition of transmission units with phylogenetic dating, we reconstructed a cross-ecological, cross-national, and cross-species transmission network of these plasmids, demonstrating that plasmid horizontal transfer and multi-host adaptation jointly drive their global prevalence. Furthermore, we evaluated a machine learning-based classification framework to internally prioritize plasmids assigned to predefined upper composite-risk categories, illustrating the potential utility of interpretable classification approaches in resistance plasmid surveillance. Overall, this study proposes a gene-centered transmission-tracing framework based on fully sequenced plasmids, combined with machine learning–based risk assessment for high-risk plasmids. This framework is both generalizable and scalable, enabling precise monitoring of these plasmids and providing a cost-effective solution for studying the evolution and dissemination of other antimicrobial resistance genes, ultimately laying a solid foundation for the development of efficient and forward-looking antimicrobial resistance control strategies. Despite leveraging a large-scale dataset and multidimensional analytical approaches to elucidate the dissemination patterns and risks of these plasmids, this study has several limitations. First, the geographic distribution of plasmid sources in public databases is uneven, particularly because our analysis focused exclusively on completely sequenced plasmids. While this strategy ensured reliable plasmid-level analyses, many plasmids available only in draft assemblies or incomplete contigs may have been excluded, potentially introducing geographic and sampling bias and affecting the global representativeness of the dataset. This may have led to underestimation in certain under-sampled regions, such as Africa and the Middle East. Therefore, the observed geographic distribution and inferred transmission trajectories should be interpreted as patterns derived from currently available complete plasmid sequences rather than as an unbiased estimate of the true worldwide prevalence of these plasmids. Second, because most public records provided only country-level geographic metadata, we assigned coordinates to country capitals as standardized spatial proxies for phylogeographic reconstruction. This approach may oversimplify within-country spatial heterogeneity and could bias inferred diffusion distances or routes. In addition, our conclusions are primarily based on genomic evidence, and patient-level epidemiological data were unavailable. Therefore, plasmids assigned to the same transmission unit should be interpreted as genetically related or putatively linked rather than as evidence of direct epidemiological transmission. Further validation through clinical outcomes and experimental models is warranted to substantiate the inferred transmission risks. Although model performance was evaluated using a held-out 30% test set, this represents internal validation rather than external validation; therefore, independent plasmid datasets will be required to further assess model generalizability and potential overfitting. Conclusion By systematically analyzing 747 complete the plasmids carrying blaOXA-48-like genes, this study provides a large-scale overview of their genetic diversity, plasmid backgrounds, local genetic contexts, dissemination-related genomic features and associated risk factors. We confirmed and quantified the predominant associations of major blaOXA-48-like gene variants with replicon backgrounds such as IncL and ColKP3/IncX3, and identified Tn1999.2 and other IS-associated structures as important genetic contexts surrounding blaOXA-48-like genes. Transmission-unit analysis and phylogenetic reconstruction further revealed putative cross-regional, cross-host, and cross-source dissemination patterns. The frequent co-occurrence of additional resistance determinants and the detection of related plasmids in both clinical and non-clinical sources highlight the importance of integrated surveillance across clinical, animal, and environmental reservoirs. In addition, the machine learning analysis provided an internal classification framework for prioritizing plasmids assigned to predefined upper composite-risk categories. Overall, this study supports a gene-centered plasmid genomic framework that may inform future One Health–oriented surveillance and prioritization of antimicrobial resistance plasmids. Supplementary Information Supplementary Material 1: Figure S1. Temporal changes in the proportion of clinical and non-clinical sources among the plasmids carrying blaOXA-48-likegenes. Stacked bar plot showing the yearly proportion of plasmids derived from clinical and non-clinical sources between 2010 and 2023. Clinical sources are shown in blue, and non-clinical sources are shown in red. Supplementary Material 2: Figure S2. Mobility analysis of the plasmids carrying blaOXA-48-likegenes. (A) Mobility classification of all plasmids carrying blaOXA-48-likegenes. (B) Mobility classification of different plasmids variants. Supplementary Material 3: Figure S3. Plasmid length variation among the plasmids carrying blaOXA-48-likegenes. Median plasmid lengths are labeled above each box. Supplementary Material 4: Figure S4. Associations between different OXA-48-like variants and their coexisting resistance genes, with a distinction between clinical and non-clinical sources Supplementary Material 5: Figure S5. Network diagram illustrating the dissemination of coexisting resistance genes among plasmids Supplementary Material 6. Supplementary Material 7. Supplementary Material 8. Supplementary Material 9. Supplementary Material 10. Acknowledgements Not applicable. Abbreviations Antimicrobial resistance Antimicrobial resistance gene Incompatibility Insertion sequence Mobile genetic element National Center for Biotechnology Information SHapley Additive exPlanations Single-nucleotide polymorphism World Health Organization Authors’ contributions MZ, QY and YK designed the study, QY, JW, XL and XY wrote the manuscript. QY, YK, JL, DG and QJ collected the data and conducted the analyses, MZ and YL edited and revised the manuscript. All authors have approved the submitted version and agreed with the contribution declarations. Funding This project is funded by Science and Technology Plan Project of Wenzhou Municipality (No. ZY2022018). Data availability All plasmid sequences analyzed in this study were obtained from the NCBI database (available at: https://www.ncbi.nlm.nih.gov). The accession numbers and associated metadata for these sequences are provided in Table S1. Declarations Ethics approval and consent to participate This study used publicly available data and did not involve human participants, animal subjects, or any intervention requiring ethical approval. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Yongliang Lou, Email: lyl@wmu.edu.cn. Yutong Kang, Email: kangyutong0601@163.com. Meiqin Zheng, Email: zmq@eye.ac.cn. References Associated Data Supplementary Materials Supplementary Material 1: Figure S1. Temporal changes in the proportion of clinical and non-clinical sources among the plasmids carrying blaOXA-48-likegenes. Stacked bar plot showing the yearly proportion of plasmids derived from clinical and non-clinical sources between 2010 and 2023. Clinical sources are shown in blue, and non-clinical sources are shown in red. Supplementary Material 2: Figure S2. Mobility analysis of the plasmids carrying blaOXA-48-likegenes. (A) Mobility classification of all plasmids carrying blaOXA-48-likegenes. (B) Mobility classification of different plasmids variants. Supplementary Material 3: Figure S3. Plasmid length variation among the plasmids carrying blaOXA-48-likegenes. Median plasmid lengths are labeled above each box. Supplementary Material 4: Figure S4. Associations between different OXA-48-like variants and their coexisting resistance genes, with a distinction between clinical and non-clinical sources Supplementary Material 5: Figure S5. Network diagram illustrating the dissemination of coexisting resistance genes among plasmids Supplementary Material 6. Supplementary Material 7. Supplementary Material 8. Supplementary Material 9. Supplementary Material 10. Data Availability Statement All plasmid sequences analyzed in this study were obtained from the NCBI database (available at: https://www.ncbi.nlm.nih.gov). The accession numbers and associated metadata for these sequences are provided in Table S1.