PlasmidRiskNet: An explainable machine-learning framework for antimicrobial resistance plasmid risk stratification Authors: Oluwaseun E. Agboola, Samuel S. Agboola, Oluwaseun Ruth Olasehinde, Temitope C. Aribigbola, Anuoluwapo Bukola Shaleye, I. Adefolake Alebiosu, Babatunji E. Oyinloye, Idowu O. Omotuyi, Abel Kolawole Oyebamiji, Olusola Tolulope Fakoyede Published: 2026 ABSTRACT BackgroundPlasmids are the principal vehicles of horizontal antimicrobial resistance (AMR) gene transfer, yet risk analyses rarely combine gene content, mobility, and network topology. We asked whether explainable machine learning over these dimensions can stratify plasmid dissemination risk, and tested rigorously where it succeeds and fails.MethodsFrom 72,556 PLSDB 2025 plasmids we integrated 251,138 AMRFinderPlus gene records with CARD v3 ontology and MOBsuite typing, built a co-resistance network, and derived a composite PlasmidRisk score from five features. Three classifiers were evaluated by five-fold cross-validation; external validation used WHO and ECDC 2024 to 2025 carbapenemase designations as a feature-independent reference. We added length- and host-adjusted burden models, phylum-normalized enrichment, and feature-category ablation.ResultsAMR genes occurred in 41.0% of plasmids across 85 drug classes. The network (29,758 nodes) was heterogeneous rather than scale-free. Internal cross-validation AUCs exceeded 0.999, but because labels derived from the scored features this reflects internal consistency, not generalization. The feature-independent external AUC was modest (0.607): strong for the metallo-beta-lactamases blaNDM, blaVIM, and blaIMP (0.72 to 0.73) but at or below chance for blaKPC and blaOXA-48 (0.45 to 0.51). The conjugative burden advantage did not survive adjustment for length and host phylum (adjusted incidence rate ratio 0.93), with length dominant.ConclusionsPlasmidRiskNet offers a useful pre-screening layer for MBL-bearing plasmids but not for the compact serine-carbapenemase backbones (blaKPC, blaOXA-48), which require replicon typing. Honest external and confounder-adjusted evaluation, not internal metrics, defines its class-specific surveillance value. Introduction The global burden of antimicrobial resistance demands urgent attention. Antimicrobial-resistant infections killed an estimated 1.27 million people directly in 2019, and projections reaching 10 million deaths annually by 2050 have been widely cited since the O'Neill review [1]. Addressing resistance is no longer a niche microbiological concern but an explicit global health emergency. The World Health Organization (WHO) 2024 Bacterial Priority Pathogens List places carbapenem-resistant Enterobacterales, Acinetobacter baumannii, and Pseudomonas aeruginosa in the most critical tier, and the European Centre for Disease Prevention and Control (ECDC) has separately flagged these same organisms as the central focus of European AMR surveillance [2,3]. What makes them dangerous is not just their resistance phenotype in isolation but the remarkable ease with which they share resistance genes with other bacteria. Plasmids sit at the heart of that sharing process. These extrachromosomal, self-replicating DNA molecules can cross species barriers in hours through conjugation, a direct cell-to-cell transfer process that bypasses the usual barriers to DNA exchange [4,5]. Conjugative plasmids encode the complete transfer apparatus: an origin of transfer (oriT), a relaxase that nicks oriT to initiate strand transfer, a type IV coupling protein, a type IV secretion system, and mating-pair formation machinery. Mobilizable plasmids carry an oriT and usually a cognate relaxase but lack a complete secretion system, so they are transferred in trans by the conjugation machinery of a co-resident conjugative element [6,7]. This layered transfer hierarchy means the risk a given plasmid poses is not simply a function of how many resistance genes it carries. A plasmid that transfers efficiently, even when it carries only one carbapenemase gene, can seed resistance across an entire ward or ecosystem within a short timeframe. Resistance gene accumulation on plasmids is itself an active and dynamic process. Class 1 integrons capture gene cassettes from environmental pools. Composite transposons such as Tn3 and Tn10 move resistance modules between replicons. Insertion sequences generate deletions, inversions, and co-integrates that shuffle cassette arrangements [8]. The practical consequence is that clinically relevant plasmids tend not to carry single resistance genes but arrays of them, encoding resistance to aminoglycosides, beta-lactams, sulfonamides, tetracyclines, and metal biocides simultaneously. Co-selection by metals and biocides in agricultural soils and hospital environments then maintains these elements at high frequency even in the absence of direct antibiotic pressure [9]. The global epidemiology of pandemic carbapenemase plasmids, particularly those carrying blaNDM, blaKPC, and blaOXA-48 variants in high-risk clonal lineages of Klebsiella pneumoniae and Escherichia coli, illustrates how quickly such elements can achieve international spread through routine patient movement and healthcare linkages [2,10]. Against this backdrop, researchers and public health agencies currently have surprisingly limited analytical tools for risk-stratifying plasmids. Individual plasmids are typically assessed by their gene content alone. The structural relationships between plasmids, whether they share gene repertoires with many other mobile elements or occupy isolated positions in the mobilome, are rarely considered. Network-based methods have proven genuinely informative in adjacent biological domains, including protein interaction networks, co-occurrence ecology, and studies of resistance gene co-selection [[11], [12], [13]], and there is a reasonable expectation that similar approaches would add value when applied to the plasmid mobilome. A plasmid that shares resistance gene families with hundreds of other plasmids across diverse host taxa occupies a qualitatively different epidemiological position from one carrying the same genes in isolation. Graph-theoretic metrics such as betweenness centrality can identify the bridge elements that connect otherwise separate resistance communities, which represents exactly the kind of structural intelligence that traditional gene-content cataloging misses. Machine learning now underpins much of computational drug discovery and AMR analytics, and our group has applied supervised and multimodal models across regulatory safety assessment, drug-target interaction, pharmacokinetic modeling, and phytochemical screening [[14], [15], [16], [17], [18], [19], [20]]. That experience carries a methodological warning central to the present work: when evaluation labels are derived from the same features a model is given, or when biologically similar items leak between training and test partitions, performance inflates and generalization is overstated [14,15]. We therefore designed PlasmidRiskNet as an explainable framework in which network topology, gene content, and mobility contribute interpretable, individually attributable features, and in which the decisive performance estimate comes from a feature-independent external standard rather than from internally derived labels. PLSDB 2025 now makes this kind of analysis tractable at scale. The database contains 72,556 verified plasmid sequences with integrated AMRFinderPlus annotations, MOBsuite and PlasmidFinder mobility predictions, and NCBI host taxonomy classifications [21]. That breadth spans clinical, agricultural, environmental, and commensal isolates from across the globe, giving a genuinely population-level view of the mobilome. In this study, we used that resource to build PlasmidRiskNet: a computational framework that constructs a plasmid co-resistance network from the full PLSDB 2025 mobilome, derives a composite PlasmidRisk score by combining network position with genomic and mobility features, evaluates machine learning classifiers for risk tier prediction, benchmarks the composite score against simpler alternatives, and critically tests external validity using WHO and ECDC priority carbapenemase gene designations as an independent clinical reference standard. Materials and methods Data sources and acquisition Plasmid metadata, resistance gene annotations, mobility predictions, and host taxonomy were downloaded from PLSDB 2025 (https://ccb-microbe.cs.uni-saarland.de/plsdb2025; accessed January 2026), a curated repository of complete bacterial plasmid sequences described by Galata and colleagues [21]. The distribution package contained four structured files: nuccore.csv, covering sequence identifiers, plasmid length, GC content, and topology; amr.tsv, containing per-gene AMR annotations from NCBI AMRFinderPlus (database version 2024-01-31.1) as described by Feldgarden et al. [22]; typing.csv, holding in silico mobility classifications derived from MOBsuite [23], which predicts conjugative capacity from the presence of MOB relaxase genes, type IV secretion system (T4SS) components, and mating pair formation (MPF) gene clusters, and replicon typing from PlasmidFinder [24]; and taxonomy.csv, which maps each plasmid to its NCBI Taxonomy host entry. Drug class and resistance mechanism annotations were drawn from CARD version 3 (https://card.mcmaster.ca; aro_index.tsv; 6445 Antibiotic Resistance Ontology (ARO) entries; accessed January 2026) as documented by Alcock and colleagues [25]. Data integration All four PLSDB files were merged on the NUCCORE_ACC primary key using pandas (version 2.x) [26]. We computed per-plasmid AMR gene count as the total number of AMRFinderPlus hits and drug class diversity as the number of unique CARD drug class labels per plasmid, after splitting semicolon-delimited multi-class entries. Plasmids with at least one annotated resistance gene were classified as AMR-positive. Numerical operations used NumPy [27]. All processing ran under Python 3.12. Co-resistance network construction We used NetworkX (version 3.x) [28] to build an undirected plasmid-plasmid co-resistance graph. Nodes were AMR-positive plasmids (n = 29,758). An edge was placed between two plasmids whenever they shared two or more AMR gene families at the gene_symbol level, with edge weight equal to the count of shared families. Before building edges, we excluded gene symbols appearing in more than 500 plasmids. Genes at that prevalence, such as sul1, blaTEM-1, and qacEdelta1, are so widespread across ecologically unrelated plasmids that including them would create spurious connections between elements that have no meaningful epidemiological relationship; this exclusion criterion follows approaches used in prior microbial network analyses [13]. The threshold of two shared gene families was chosen to require a minimum level of shared resistance repertoire rather than coincidental single-gene overlap. Network topology analysis We evaluated degree distribution in log-log space and estimated the power-law scaling exponent by ordinary least-squares regression of log-transformed degree frequency on log-transformed degree, interpreted cautiously, since strict scale-free structure is rare [29]. Betweenness centrality was computed on the largest connected component using the NetworkX implementation with k = 500 randomly sampled source nodes (seed = 42), an approximation that balances accuracy with computational feasibility consistent with the algorithm of Brandes [30]. Local clustering coefficients were computed for all nodes. Spearman rank correlation between AMR gene count and network degree was calculated using scipy.stats [31]. Plasmid risk score For each AMR-positive plasmid, we computed a composite risk score as a weighted sum of five MinMax-normalized features using scikit-learn's MinMaxScaler [32]. The features and weights were AMR gene count (0.25), drug class diversity (0.25), network degree (0.20), betweenness centrality (0.15), and mobility class risk weight (0.15). Mobility weights were set at conjugative = 1.0, mobilizable = 0.6, and non-mobilizable = 0.1, reflecting empirical conjugation frequency data from Smillie and colleagues [7]. Plasmids at or above the 90th percentile of risk score were designated high-risk (n = 2976). Machine learning classifiers We trained three supervised binary classifiers to distinguish high-risk plasmids (top decile) from the remaining AMR-positive plasmids, using eight features: AMR gene count, drug class diversity, network degree, betweenness centrality, mobility risk weight, log10-transformed plasmid length, GC content, and a binary indicator for circular topology. The classifiers were Random Forest [33] (200 trees, maximum depth 12, minimum five samples per leaf, balanced class weights), Gradient Boosting [34] (200 estimators, depth 4, learning rate 0.05, subsample ratio 0.8), and Logistic Regression (L2 regularization, C = 1.0, balanced class weights, 1000 iteration maximum, applied to MinMax-normalized features). All three were evaluated by five-fold stratified cross-validation (seed = 42) using scikit-learn 1.8 [32], reporting the mean and standard deviation of ROC-AUC, average precision (PR-AUC), and F1 score. Feature importances were extracted from the final fitted Random Forest as the mean Gini impurity decrease. External validation For external validation, we needed a reference set of high-risk plasmids that was defined entirely independently of our scoring framework. We used the WHO Bacterial Priority Pathogens List 2024 [2] and the ECDC Carbapenem-Resistant Enterobacterales Risk Assessment, Third Update, February 2025 [3], which together identify blaNDM, blaKPC, blaOXA-48, blaVIM, blaIMP, and blaOXA-23 as the six principal plasmid-borne carbapenemase families requiring continuous public health surveillance. Additional epidemiological characterization of pandemic plasmid lineages was drawn from Getino et al. [35] and Vezina et al. [36]. Any PLSDB plasmid carrying at least one AMRFinderPlus hit matching one of these six gene-family prefixes was assigned to the external validation positive set (n = 5479). We then computed three metrics: (i) the recovery rate, defined as the fraction of these WHO/ECDC priority plasmids that fall in our top risk decile; (ii) a Mann-Whitney U test comparing risk scores of the critical gene carriers against non-carriers, with rank-biserial correlation r as the effect size [37]; and (iii) the ROC-AUC for using the continuous PlasmidRisk score to separate critical gene carriers from non-carriers. We ran these analyses both pooled across all six carbapenemase families and separately for each gene family. Benchmarking To put the composite score in context, we compared it against six simpler alternatives as single-feature scoring approaches: AMR gene count, drug class count, network degree, mobility risk weight, log10-transformed plasmid length, and a random score generated from a fixed pseudorandom number generator (NumPy RandomState, seed = 42) [27]. Performance was measured by ROC-AUC and average precision for the binary high-risk classification task. Bootstrap 95% confidence intervals for AUC used 1000 resampled replicates with the percentile method of Efron and Tibshirani [38]. Statistical analysis One-way analysis of variance (ANOVA; scipy.stats [31]) tested differences in AMR gene burden across mobility classes, conducted across all 72,556 plasmids with non-AMR plasmids contributing a count of zero. Given the large sample sizes in this dataset, we focus on effect magnitudes alongside p-values when interpreting group differences. Spearman rank correlation examined the relationship between AMR gene count and network degree. A significance threshold of p < 0.05 was applied throughout. The complete reproducible pipeline is provided as Supplementary File S1. Results How common is resistance in the plasmid mobilome? The broader cross-tabulation of drug class distribution across all host phyla in the full mobilome is depicted in Fig. 1A.Fig. 1Drug class and host range co-occurrence. (A) Heatmap of drug class versus host phylum (row-normalized to percentage). (B) Bubble chart of mobility class versus drug class; bubble area represents gene occurrence count. Drug class and host range co-occurrence. (A) Heatmap of drug class versus host phylum (row-normalized to percentage). (B) Bubble chart of mobility class versus drug class; bubble area represents gene occurrence count. When we ran the analysis across all 72,556 plasmid sequences in PLSDB 2025, the first finding that stood out was the sheer prevalence of resistance gene carriage. Just over 41% of plasmids (29,758) harbored at least one AMRFinderPlus-annotated resistance gene, generating 251,138 gene records in total. Those records cover 85 distinct drug classes, which makes it clear that this is not a phenomenon confined to clinical pathogens carrying a handful of well-known genes but a characteristic distributed broadly across the mobilome. In terms of which drug classes were most represented, aminoglycosides came first with 28,577 gene occurrences, followed by beta-lactams at 25,402. Perhaps less expected was the volume of metal tolerance genes: copper and silver resistance accounted for 15,638 occurrences, and mercury, arsenic, and tellurium resistance collectively added tens of thousands more. Metal resistance gene abundance is consistent with co-selection reported in agricultural and wastewater environments [9]. At the individual gene level, sul1 appeared on 7827 plasmids, more than any other single gene, followed by qacEdelta1 (6381), blaTEM-1 (5458), aadA2 (5111), and tet(A) (4620). All five are classic markers of class 1 integron-associated resistance cassette arrays [8] (Fig. 1B). The mobility data revealed a clear gradient that we had expected to find but whose magnitude was still striking. Of the 22,588 conjugative plasmids in the dataset, 71.5% carried resistance genes. Computed across all plasmids in the dataset (including the 42,798 without annotated AMR genes, which contribute zero to the mean), conjugative plasmids had a mean AMR gene burden of 6.91 genes per plasmid, mobilizable plasmids averaged 2.10, and non-mobilizable elements averaged 1.76, giving a nominal 3.9-fold difference between the extremes when AMR-negative plasmids are pooled in as zero counts (ANOVA F = 15,713; p < 2.2 x 10-308). This pooled gradient is confounded: within the AMR-positive set the conjugative advantage is only 1.27-fold, and in a negative binomial model adjusting for plasmid length and host phylum the conjugative effect disappears (adjusted incidence rate ratio 0.93), with plasmid length the dominant driver (Section 3.9). The mobility gradient in Fig. 2 should therefore be read as largely a plasmid-size and taxonomic effect rather than an independent effect of conjugative capacity [6,8].Fig. 2AMR landscape across PLSDB 2025. (A) Mobility class distribution (n = 72,556 plasmids). (B) AMR gene prevalence by mobility class. (C) Top 12 drug class categories by AMRFinderPlus occurrence count. (D) AMR gene burden per plasmid among AMR-positive sequences (mean = 8.44; median shown). AMR landscape across PLSDB 2025. (A) Mobility class distribution (n = 72,556 plasmids). (B) AMR gene prevalence by mobility class. (C) Top 12 drug class categories by AMRFinderPlus occurrence count. (D) AMR gene burden per plasmid among AMR-positive sequences (mean = 8.44; median shown). Network structure We built the co-resistance network from 29,758 AMR-positive plasmids and obtained 877,272 edges after applying the shared-gene-family criteria described in Methods. The largest connected component contained 10,878 nodes and 862,492 edges, meaning that just over one-third of AMR-positive plasmids (36.6%) form a single densely wired supercluster sharing resistance gene repertoires. The remaining plasmids either formed small separate components or sat as isolated nodes, indicating resistance gene profiles that are largely unique or restricted to narrow ecological niches. The degree distribution followed a pattern broadly consistent with a power law in log-log space, with an estimated exponent of 1.001 and R2 of 0.581 (Fig. 3A). We would not claim this is a strict scale-free network; the moderate R2 tells us the fit is imperfect and the distribution is probably better described as broadly heterogeneous. What is clear is the heavy tail: a handful of plasmids had degrees exceeding 700 while the median degree was zero, a pattern that indicates a small number of highly connected hubs co-existing with a large pool of sparsely connected elements. This kind of hub-spoke structure, seen in other biological sharing networks [29], has practical consequences because the hubs are where co-resistance communities converge.Fig. 3Co-resistance network topology. (A) Degree distribution in log-log space; line shows power-law fit (exponent = 1.001; R2 = 0.581). (B) Clustering coefficient distribution across all nodes; dashed line = mean (0.288). (C) Betweenness centrality distribution for the largest connected component; y-axis log-scaled. Co-resistance network topology. (A) Degree distribution in log-log space; line shows power-law fit (exponent = 1.001; R2 = 0.581). (B) Clustering coefficient distribution across all nodes; dashed line = mean (0.288). (C) Betweenness centrality distribution for the largest connected component; y-axis log-scaled. The mean clustering coefficient was 0.288, which tells us that neighbors of a given plasmid in the network are themselves more likely to be connected to each other than random chance would predict. This clustering pattern is consistent with resistance cassette co-arrival in genomic islands, integrons, and transposons [8]. Betweenness centrality was strongly skewed, with most plasmids having near-zero values but a small number reaching as high as 0.082, marking them as structural bridges between otherwise disconnected resistance communities (Fig. 3C). There was a moderate positive correlation between AMR gene count and network degree (Spearman rho = 0.505; p < 10−300), though the correlation is far from perfect, which means gene count and network position capture related but distinct information. Risk scores and what they reveal PlasmidRisk scores ranged from 0.000 to 0.566 across the 29,758 AMR-positive plasmids, with a mean of 0.162 and a strong positive skew. The separation across mobility classes was pronounced: conjugative plasmids averaged a risk score of 0.215, mobilizable elements 0.137, and non-mobilizable ones 0.058. The scores for conjugative plasmids formed a clearly shifted distribution with a longer right tail, reflecting the compound effect of higher gene counts, greater drug class diversity, and elevated network connectivity in this group (Fig. 4A).Fig. 4PlasmidRisk score analysis. (A) Score distributions by mobility class. (B) Risk score versus AMR gene count for a random sample of 5000 plasmids; colour reflects network degree. (C) Host phylum distribution of 2976 high-risk plasmids. (D) Mean drug class diversity per risk quintile. PlasmidRisk score analysis. (A) Score distributions by mobility class. (B) Risk score versus AMR gene count for a random sample of 5000 plasmids; colour reflects network degree. (C) Host phylum distribution of 2976 high-risk plasmids. (D) Mean drug class diversity per risk quintile. Plotting the risk score against the AMR gene count and coloring by network degree (Fig. 4B) reveals the 'double-threat' subset the analysis assessed in: plasmids in the upper right with many resistance genes and high network centrality. The 2976 high-risk plasmids (top decile) had a median length of 217,193 bp, compared to 92,958 bp across all 29,758 AMR-positive plasmids (the median across all 72,556 PLSDB plasmids, including those without annotated resistance genes, was 53,666 bp). It tracks with what we know about large conjugative resistance plasmids in clinical Gram-negatives, where the backbone of a plasmid like pOXA-48 or pNDM-1 is already tens of kilobases and resistance gene insertions add further. Host taxonomy reflected this clinical context: 87.6% of high-risk plasmids (2608 of 2976) were associated with Pseudomonadota hosts. Bacillota accounted for most of the remainder (365 plasmids). However, Pseudomonadota also make up 81.8% of the AMR-positive background, so this predominance is only a modest enrichment over the base rate; within-phylum only 10.7% of Pseudomonadota plasmids are high-risk, and balanced resampling reduces the Pseudomonadota share of the high-risk tier to 59.6% (Section 3.9). The concentration therefore substantially reflects database composition and plasmid size rather than a phylum-intrinsic risk (Fig. 4C). At genus level, Klebsiella predominated among high-risk plasmids (1055 plasmids; 35.5%), followed by Escherichia (649; 21.8%), Salmonella (316; 10.6%), Enterobacter (291; 9.8%), and Staphylococcus (252; 8.5%). The presence of Staphylococcus in the high-risk tier, representing 8.5% of top-decile plasmids, reflects large conjugative elements carrying multiple drug class arrays in Gram-positive hosts and warrants further investigation given the known role of plasmid exchange between Gram-negative and Gram-positive reservoirs. Network degree and AMR gene count are therefore complementary rather than redundant dimensions. The drug class diversity gradient across risk quintiles confirmed this was not an artifact of any single feature: mean drug class count rose from 0.8 in the lowest quintile to 5.2 in the highest (Fig. 4D) when computed across all 72,556 plasmids; among the 29,758 AMR-positive plasmids, the equivalent gradient was 2.0 to 10.6, confirming genuine multi-drug resistance co-accumulation with increasing risk score. External validation: How does the score fare against WHO and ECDC priority genes? External validation assessed whether the PlasmidRisk score, built entirely from PLSDB-derived features, independently predicts carriage of genes designated by WHO and ECDC as the most urgent clinical threats. Across all 5479 plasmids carrying at least one priority carbapenemase, 744 (13.6%) landed in the top risk decile. That is 1.36 times the 10% we would expect by chance. Risk scores were significantly higher for the critical gene carriers (mean 0.189 versus 0.156 for non-carriers; Mann-Whitney U = 8.07 x 107; p = 4.54 x 10−136; effect size r = 0.607), and the overall ROC-AUC for discrimination was 0.607 (PR-AUC = 0.240) (Fig. 5).Fig. 5External validation against WHO and ECDC critical carbapenemase genes. (A) PlasmidRisk score distributions for priority carbapenemase carriers versus non-carriers. (B) Top-decile recovery rates per gene family; dashed line = 10% expected by chance; AUC values annotated. (C) ROC curve for PlasmidRisk versus all six critical gene families combined (AUC = 0.607; PR-AUC = 0.240). External validation against WHO and ECDC critical carbapenemase genes. (A) PlasmidRisk score distributions for priority carbapenemase carriers versus non-carriers. (B) Top-decile recovery rates per gene family; dashed line = 10% expected by chance; AUC values annotated. (C) ROC curve for PlasmidRisk versus all six critical gene families combined (AUC = 0.607; PR-AUC = 0.240). Per-gene results provided additional resolution (Table 1). Plasmids carrying blaNDM, blaVIM, and blaIMP showed genuinely strong enrichment in the high-risk stratum: top-decile recovery rates of 21.6%, 30.9%, and 30.9%, respectively, and external AUC values of 0.717, 0.717, and 0.734. These three carbapenemases typically reside on large, multi-resistance, integron-containing conjugative plasmids [35,36], the exact profile our scoring was designed to capture. Finding them at two to three times the expected frequency in the top decile is a useful validation of the framework's biological coherence.Table 1Per-gene external validation results. Expected top-decile rate by chance is 10.0%.Table 1Gene FamilyN PlasmidsTop Decile (%)EnrichmentMean Risk ScoreExternal AUCblaNDM191821.62.16x0.22320.717blaVIM29830.93.09x0.23120.717blaIMP42430.93.09x0.23690.734blaKPC19995.60.56x0.15920.506blaOXA-486750.60.06x0.15390.449blaOXA-232073.40.34x0.11730.313 Table 1Per-gene external validation results. Expected top-decile rate by chance is 10.0%.Table 1Gene FamilyN PlasmidsTop Decile (%)EnrichmentMean Risk ScoreExternal AUCblaNDM191821.62.16x0.22320.717blaVIM29830.93.09x0.23120.717blaIMP42430.93.09x0.23690.734blaKPC19995.60.56x0.15920.506blaOXA-486750.60.06x0.15390.449blaOXA-232073.40.34x0.11730.313 Per-gene external validation results. Expected top-decile rate by chance is 10.0%. Mechanistic stratification by carbapenemase class: A biologically coherent dichotomy Partitioning the six priority carbapenemase families by their biochemical mechanism reveals a pattern that is biologically coherent and has direct surveillance implications. The three metallo-β-lactamase (MBL) families, blaNDM, blaVIM, and blaIMP, achieved a mean external AUC of 0.723 (range 0.717-0.734) with mean top-decile recovery of 27.8%, representing 2.1-3.1× enrichment over the 10% chance baseline. The three serine carbapenemase and oxacillinase families, blaKPC, blaOXA-48, and blaOXA-23, achieved a mean AUC of 0.423 (range 0.313-0.506) with mean top-decile recovery of only 3.2%, below chance. The AUC difference between MBL and serine classes is 0.300, a magnitude that is not a statistical fluctuation but a reflection of fundamentally different plasmid biology. To quantify the practical surveillance value, we computed the Number Needed to Screen (NNS): the number of top-decile plasmids that must be examined to identify one priority carbapenemase carrier, compared to the baseline of screening all 29,758 AMR-positive plasmids at random. For blaNDM, the NNS was 7.2 (baseline 15.5; 2.2× efficiency gain). For blaVIM and blaIMP, NNS was 32.3 and 22.7 respectively, against baselines of 99.9 and 70.2 (3.1× gain for both). Combined, the 2640 MBL-carrying plasmids (blaNDM + blaVIM + blaIMP) had 638 representatives in the top decile, 24.2% recovery against a 10.0% chance expectation, a 2.42× enrichment (Table 2). By contrast, blaKPC NNS of 26.6 is worse than the baseline of 14.9, and blaOXA-48 NNS of 744.0 versus a baseline of 44.1 confirms near-complete anti-enrichment for this family.Table 2Surveillance efficiency of PlasmidRiskNet for six WHO/ECDC priority carbapenemase families. NNS = Number Needed to Screen to identify one priority carbapenemase carrier. PPV = Positive Predictive Value at top-decile threshold. Expected top-decile rate by chance = 10.0%; expected NNS by chance = N_total/N_gene.Table 2Gene FamilyN PlasmidsTop Decile (%)NNS (PlasmidRiskNet)NNS (Baseline)Efficiency GainPPV (%)ClassblaNDM191821.67.215.52.2×13.9MBLblaVIM29830.932.399.93.1×3.1MBLblaIMP42430.922.770.23.1×4.4MBLblaKPC19995.626.614.90.6×3.8SerineblaOXA-486750.6744.044.10.1×0.1SerineblaOXA-232073.4425.1143.80.3×0.2OXA Table 2Surveillance efficiency of PlasmidRiskNet for six WHO/ECDC priority carbapenemase families. NNS = Number Needed to Screen to identify one priority carbapenemase carrier. PPV = Positive Predictive Value at top-decile threshold. Expected top-decile rate by chance = 10.0%; expected NNS by chance = N_total/N_gene.Table 2Gene FamilyN PlasmidsTop Decile (%)NNS (PlasmidRiskNet)NNS (Baseline)Efficiency GainPPV (%)ClassblaNDM191821.67.215.52.2×13.9MBLblaVIM29830.932.399.93.1×3.1MBLblaIMP42430.922.770.23.1×4.4MBLblaKPC19995.626.614.90.6×3.8SerineblaOXA-486750.6744.044.10.1×0.1SerineblaOXA-232073.4425.1143.80.3×0.2OXA Surveillance efficiency of PlasmidRiskNet for six WHO/ECDC priority carbapenemase families. NNS = Number Needed to Screen to identify one priority carbapenemase carrier. PPV = Positive Predictive Value at top-decile threshold. Expected top-decile rate by chance = 10.0%; expected NNS by chance = N_total/N_gene. This dichotomy is mechanistically explicable. MBL carbapenemases, particularly blaNDM and blaVIM, are characteristically embedded within complex class 1 integron arrays on large IncF, IncA/C, or IncHI conjugative backbones carrying multiple co-resistance cassettes [35,36]. This is precisely the genomic signature PlasmidRiskNet is designed to detect: high gene count, high drug class diversity, and high network connectivity reflecting shared resistance gene repertoires. Serine carbapenemases, by contrast, achieve pandemic spread through a fundamentally different strategy. KPC predominantly propagates on IncF pKpQIL and IncN plasmids that are gene-sparse and metabolically inexpensive [35]; OXA-48 spreads primarily via IncL/M replicons of similar compact architecture. Their transmissibility is encoded in transfer kinetics and conjugation efficiency, not gene density, features entirely invisible to a content-weighted network score. The framework's differential performance is therefore not a weakness but an empirically validated characterisation of the two biological strategies by which carbapenemase plasmids achieve global spread. The relatively weak external AUC for blaKPC (0.506) and blaOXA-48 (0.449) reflects biological differences in the plasmid backbones these genes preferentially occupy rather than a scoring artifact. KPC and OXA-48 plasmids, predominantly IncN and IncL/M backbones, are typically compact and gene-sparse replicons that achieve pandemic spread through high conjugation rates at minimal metabolic cost to the bacterial host [35]. They are dangerous precisely because they are not gene-dense; they carry just enough to cause harm and not enough to be detected by a content-based risk screen. Replicon typing alongside the PlasmidRisk score is indicated for surveillance of this carbapenemase subset. The low AUC for blaOXA-23 (0.313) is noted here; the mechanistic basis is addressed in the Discussion. These findings suggest a two-track genomic surveillance strategy. For MBL carbapenemases (blaNDM, blaVIM, blaIMP), which together account for 2640 of the 5479 WHO/ECDC priority plasmids in PLSDB, PlasmidRiskNet top-decile prioritisation achieves 24.2% recovery at 2.42× enrichment, providing a practically useful pre-screening layer that reduces the number of sequences requiring detailed clinical investigation by more than half compared to unranked screening. For serine carbapenemases (blaKPC, blaOXA-48), replicon typing to identify IncN, IncL/M, and pKpQIL-type backbones should be applied as the primary detection modality, complementary to but independent of the PlasmidRisk score. Embedding both layers into a unified surveillance workflow, content-network scoring for MBLs, replicon typing for serines, would address the full spectrum of priority carbapenemase plasmids that current single-metric approaches cannot comprehensively cover. Machine learning performance All three classifiers performed very well in cross-validation, and their agreement was the most convincing aspect of these results. Random Forest achieved a mean ROC-AUC of 0.9995 (SD 0.0002), PR-AUC of 0.9956 (0.0013), and F1 of 0.953 (0.008). Gradient Boosting was marginally better at AUC 0.9998 (0.0001) and F1 0.976 (0.006). Logistic regression also crossed AUC 0.999 but showed a noticeably lower F1 of 0.849 (0.009), which reflects the genuinely non-linear boundaries in this problem space that a linear model cannot fully capture (Fig. 6A–C).Fig. 6Machine learning results. (A) Five-fold CV ROC curves for all three classifiers; individual folds shown as thin lines. (B) Random Forest Gini feature importances. (C) Model comparison of ROC-AUC, PR-AUC, and F1 (mean plus/minus SD). (D) Benchmarking of PlasmidRisk against six single-feature baselines. Machine learning results. (A) Five-fold CV ROC curves for all three classifiers; individual folds shown as thin lines. (B) Random Forest Gini feature importances. (C) Model comparison of ROC-AUC, PR-AUC, and F1 (mean plus/minus SD). (D) Benchmarking of PlasmidRisk against six single-feature baselines. The fact that three architecturally different classifiers all converge to AUC above 0.999 is reassuring and interpretable: it means that high-risk tier membership is genuinely encoded in the eight input features, not an artifact of any one model's flexibility. The train-to-test AUC gap for Random Forest was only 0.0005 (train = 1.000; test = 0.9995 mean), which rules out serious overfitting. A critical interpretive caveat applies to these cross-validation AUC values: the high-risk tier labels were defined as the top decile of the PlasmidRisk score, which is itself a weighted combination of AMR gene count (0.25), drug class diversity (0.25), network degree (0.20), betweenness centrality (0.15), and mobility weight (0.15), the same features used as classifier inputs. Consequently, the near-perfect internal AUC reflects consistency between the feature set and the labels derived from it, not independent generalizability. The external validation AUC of 0.607, where the reference standard (WHO/ECDC carbapenemase designation) was defined entirely independently of the scoring framework, is the operationally meaningful performance estimate and should be interpreted as such. Feature importance analysis showed (Fig. 6B), AMR gene count contributed the most to Random Forest decisions (Gini importance 0.291), followed by drug class diversity (0.281) and network degree (0.184). Together those three features account for 75.6% of total predictive power. The remaining five, plasmid length (0.080), mobility weight (0.075), betweenness (0.062), GC content (0.026), and circular topology (0.0003), each added something, though circular topology barely registered and could be dropped in a simplified version of the model without meaningful loss. Benchmarking Compared against single-feature alternatives, the composite score did what a composite score should: it outperformed each individual component. AMR gene count alone gave AUC = 0.944 (bootstrap 95% CI: 0.941 to 0.947), and drug class count alone gave an identical AUC of 0.944, which makes sense because those two features are correlated. The network degree added something different, reaching AUC = 0.845 on its own. Mobility class alone managed 0.695, and plasmid length alone reached 0.786. The random baseline confirmed the absence of any systematic bias at AUC = 0.501 (95% CI: 0.490 to 0.512). The composite score incorporates all these dimensions simultaneously, which is why it separates from any of them individually. Summary of quantitative findings Table 3 below consolidates the key quantitative outputs from the PlasmidRiskNet pipeline for rapid reference.Table 3Summary of key quantitative findings from PlasmidRiskNet analysis of PLSDB 2025.Table 3MetricValueSource or MethodTotal plasmids analyzed72,556PLSDB 2025 [21]AMR-positive plasmids29,758 (41.0%)AMRFinderPlus [22]Total AMR gene records251,138AMRFinderPlus [22]Unique AMR gene symbols2274AMRFinderPlus [22]Drug classes represented85CARD v3 [25]Network edges877,272NetworkX [28]Largest component nodes10,878 (36.6%)NetworkX [28]Power-law exponent gamma1.001 (R2 = 0.58)scipy.stats [31]Mean clustering coefficient0.288NetworkX [28]Conjugative AMR prevalence71.5%This studyAMR genes per conjugative plasmid6.91 (ANOVA F = 15,713)This studyHigh-risk plasmid median length217,193 bp vs 53,666 bp overallThis studyExternal AUC (all 6 critical genes)0.607; p = 4.5 x 10-136WHO [2]; ECDC [3]blaNDM external AUC0.717; 21.6% in top decileGetino et al. [35]blaIMP external AUC0.734; 30.9% in top decileVezina et al. [36]RF ROC-AUC 5-fold CV0.9995 (SD 0.0002)scikit-learn [32]GB ROC-AUC 5-fold CV0.9998 (SD 0.0001)scikit-learn [32]PlasmidRisk vs best single feature1.000 vs 0.944 (AMR count)Bootstrap CI [38] Table 3Summary of key quantitative findings from PlasmidRiskNet analysis of PLSDB 2025.Table 3MetricValueSource or MethodTotal plasmids analyzed72,556PLSDB 2025 [21]AMR-positive plasmids29,758 (41.0%)AMRFinderPlus [22]Total AMR gene records251,138AMRFinderPlus [22]Unique AMR gene symbols2274AMRFinderPlus [22]Drug classes represented85CARD v3 [25]Network edges877,272NetworkX [28]Largest component nodes10,878 (36.6%)NetworkX [28]Power-law exponent gamma1.001 (R2 = 0.58)scipy.stats [31]Mean clustering coefficient0.288NetworkX [28]Conjugative AMR prevalence71.5%This studyAMR genes per conjugative plasmid6.91 (ANOVA F = 15,713)This studyHigh-risk plasmid median length217,193 bp vs 53,666 bp overallThis studyExternal AUC (all 6 critical genes)0.607; p = 4.5 x 10-136WHO [2]; ECDC [3]blaNDM external AUC0.717; 21.6% in top decileGetino et al. [35]blaIMP external AUC0.734; 30.9% in top decileVezina et al. [36]RF ROC-AUC 5-fold CV0.9995 (SD 0.0002)scikit-learn [32]GB ROC-AUC 5-fold CV0.9998 (SD 0.0001)scikit-learn [32]PlasmidRisk vs best single feature1.000 vs 0.944 (AMR count)Bootstrap CI [38] Summary of key quantitative findings from PlasmidRiskNet analysis of PLSDB 2025. Discussion Robustness, confounding, and feature-category ablation (new analyses) Because the headline mobility contrast could be confounded by plasmid size and host taxonomy, we re-examined it with adjusted models on the 29,607 AMR-positive plasmids with an assigned host phylum. The unadjusted conjugative advantage was modest within the AMR-positive set (1.27-fold over non-mobilizable plasmids) and reached the 3.9-fold figure only when AMR-negative plasmids were pooled in as zero counts. In a negative binomial model of AMR gene count adjusting for log10 plasmid length and host phylum, the conjugative effect did not merely shrink but slightly reversed (adjusted incidence rate ratio 0.93, 95% confidence interval 0.90 to 0.95), whereas plasmid length was the dominant driver (incidence rate ratio 3.74 per log10 base pairs, 95% confidence interval 3.64 to 3.84). Length alone explained 25.5% of burden deviance versus 2.9% for mobility class. The apparent mobility gradient is therefore largely a plasmid-size and taxonomic effect rather than an independent property of conjugative transfer. We applied the same scrutiny to the taxonomic composition of the high-risk tier. Pseudomonadota accounted for 87.7% of high-risk plasmids, but they also constitute 81.8% of the AMR-positive background, so the raw predominance is only a modest enrichment over the base rate. Within-phylum, 10.7% of Pseudomonadota plasmids were high-risk versus 7.6% of Bacillota, both close to the 10% definitional expectation. Under balanced resampling that equalizes Pseudomonadota and Bacillota representation, the Pseudomonadota share of the high-risk tier fell from 87.7% to 59.6%, and after adjusting for plasmid length and mobility the independent Pseudomonadota effect was not positive (adjusted odds ratio 0.43, 95% confidence interval 0.38 to 0.49). The taxonomic concentration thus substantially reflects database composition and plasmid size rather than a phylum-intrinsic risk. Finally, we quantified the internal circularity directly. Because the top-decile label is a deterministic function of the five scored features, a hand-coded weighted sum reproduced the label with an AUC of 1.000 and correlated perfectly with the published PlasmidRisk score, confirming that the near-perfect classifier AUCs and the apparent superiority of the composite over single features are tautological. On this internal label, adding network degree and betweenness raised cross-validated AUC from 0.965 (content only) to 0.9997, but that gain is mechanically guaranteed because those metrics are themselves score components, and network degree is partly a projection of gene burden (Spearman rho 0.505; 25.5% of degree rank variance). The only non-circular test of whether topology and mobility add value is against a feature-independent endpoint; we therefore provide, as Supplementary File S2, a turnkey ablation that predicts WHO/ECDC carbapenemase carriage from content-only, content-plus-mobility, and content-plus-network feature blocks with DeLong comparisons. Consistent with the per-family external results in Table 1, any contribution of topology beyond gene content on the independent endpoint is small and class-dependent, and we have rewritten the network claims accordingly. Full code and adjusted estimates are in Supplementary File S2 and Fig. S7. The central question motivating this work was whether the structural position of a plasmid in the co-resistance network, combined with its genomic features and mobility capacity, meaningfully predicts which plasmids carry the resistance genes that clinical and public health authorities consider the most urgent threats. Our answer, based on the external validation results, is a qualified yes. The qualification matters: the score works well for some carbapenemase families and not others, and understanding why it fails is arguably as informative as understanding why it succeeds. Starting with the prevalence findings: 41.0% of PLSDB 2025 plasmids carry demonstrable resistance gene content across 85 drug classes. This prevalence is unlikely to reflect sampling bias toward clinical isolates. PLSDB draws on GenBank across environmental, agricultural, and clinical sources, and a 41% resistance prevalence across that breadth suggests that resistance gene carriage is a normal feature of the plasmid mobilome rather than a property restricted to pathogens under antibiotic pressure. The dominant role of aminoglycosides and beta-lactams at the top of the drug class prevalence list is unsurprising, but the volume of metal resistance co-carriage is worth noting. Co-selection by metals in agricultural and hospital environments is a well-documented mechanism for maintaining AMR plasmids without direct antibiotic pressure [9], and our data suggest this is not a minor phenomenon in the mobilome. The external validation results deserve careful interpretation. The pooled AUC of 0.607 for detecting WHO and ECDC priority carbapenemase plasmids could be read as modest, but we think that framing misses the more interesting story in the per-gene data. The strong discrimination for blaNDM, blaVIM, and blaIMP plasmids (AUC 0.717 to 0.734; top-decile recovery two to three times the chance expectation) tells us that the scoring framework is genuinely capturing something real about these genes and the plasmids they tend to inhabit. These carbapenemases have an established association with large, multi-drug-resistant, conjugation-competent backbones that carry integrons and multiple cassette arrays [35,36]. That is exactly the profile our feature set is designed to detect, and detecting them at 30.9% top-decile recovery (against a 10% baseline) represents a meaningful practical enrichment for surveillance purposes. The weak performance for blaKPC and blaOXA-48 is a finding that we think enriches rather than undermines the framework. KPC and OXA-48 plasmids have achieved pandemic spread through a fundamentally different strategy: small, metabolically inexpensive IncN and IncL/M replicons that transfer at high conjugation frequencies precisely because they minimize host burden [35]. Their danger is encoded in transfer kinetics, not gene density. No gene-count-weighted network score can capture that, because the features that make these plasmids dangerous are invisible to content-based analysis. The practical implication for surveillance is direct: replicon typing to identify IncN and IncL/M backbones should be used as a complementary detection layer alongside PlasmidRisk for this carbapenemase subset. The low discriminative AUC for blaOXA-23 (0.313) reflects a distinct biological reality: this gene integrates chromosomally in the great majority of Acinetobacter baumannii clinical isolates [3], and database records attributing blaOXA-23 to plasmid sequences in PLSDB may therefore substantially misrepresent its true genomic context. Surveillance frameworks for OXA-23-type resistance must incorporate whole-genome sequence-based chromosomal localisation rather than plasmid-centric scores. The convergence of all three classifiers above AUC 0.999 in cross-validation is best interpreted as a demonstration of internal feature-label consistency rather than generalizability. Because the high-risk tier labels were derived as the top decile of a composite score itself constructed from the classifier input features, near-perfect internal AUC is mathematically expected and does not constitute evidence that the model will identify novel high-risk plasmids outside the training distribution. This circular label-feature dependency is a genuine structural limitation of the present ML framework. The appropriate remedy, which future work should address, is to define risk labels from an entirely feature-independent source (for example, confirmed clinical outbreak plasmids from genomic surveillance registries) and then validate the score against those labels. Within the present design, the external WHO/ECDC validation (AUC 0.607) is the only truly independent performance estimate. The lower F1 for logistic regression (0.849 versus 0.953 for Random Forest) indicates nonlinear decision boundaries in the feature space that linear models cannot capture. The feature importance hierarchy confirms that gene count and drug class diversity together account for most of the predictive signal, but network degree contributed an additional 18.4% of Gini importance, meaning that network position adds real predictive value beyond what the gene content alone provides. Betweenness centrality contributed a smaller but non-zero 6.2%, suggesting that the bridge plasmid concept has measurable predictive relevance even if it is not the dominant factor. These class-specific findings reinforce a theme running through our broader computational program, in which phytochemical and multi-target screening pipelines are only as trustworthy as the validation design that supports them [[39], [40], [41]]. For PlasmidRiskNet, that principle translates directly into privileging the external carbapenemase benchmark and the confounder-adjusted analyses over the near-perfect but circular internal metrics. Several limitations bound these conclusions. First, all outputs derive from a single repository (PLSDB 2025) and one primary annotation route (AMRFinderPlus with CARD ontology and MOBsuite typing); prevalence estimates, network structure, and the MBL-versus-serine dichotomy could partly reflect curation or tool-specific thresholds, and Supplementary File S2 includes an orthogonal-annotation sensitivity routine to gauge this in future work. Second, the internal machine-learning AUCs are not an independent generalization estimate because the labels were derived from the scored features; the external carbapenemase AUC of 0.607 is the appropriate guide to real-world performance. Third, mobility classes are in silico MOBsuite predictions that we did not validate against experimental conjugation or orthogonal replicon typing, so misclassification could bias mobility-linked comparisons. Fourth, the co-resistance network is a projection of shared gene symbols without null-model or FDR-controlled edge validation in the present version; degree is partly confounded by gene burden, and Supplementary File S2 supplies a degree-preserving null model with prevalence-normalized (Jaccard) similarity and Benjamini-Hochberg edge filtering for a bias-controlled rebuild. Fifth, PLSDB is skewed toward Gram-negative clinical isolates from well-resourced sequencing programs, so host and geographic representation are uneven; the power-law fit is moderate (R2 = 0.58), so the network is heterogeneous rather than scale-free in any strict sense, and circular topology contributed negligibly (Gini importance 0.0003) and can be dropped. Incorporating replicon typing, integron content, fitness-cost estimates, and prospective outbreak-confirmed labels would most strengthen future versions. Conclusions PlasmidRiskNet demonstrates that combining network topology with genomic content and mobility features into a composite score, validated by machine learning and tested against WHO and ECDC clinical priority standards, produces a useful and interpretable risk stratification tool for the plasmid mobilome. The framework works well for the carbapenemase families that tend to inhabit large, gene-dense, conjugative plasmids, namely blaNDM, blaVIM, and blaIMP, and its comparative weakness for blaKPC and blaOXA-48 tells us something important about the limits of content-based scoring and the need for complementary replicon-typing data in surveillance. With 72,556 plasmids analyzed and 29,758 risk scores computed from real data, the framework scales to current database sizes and is designed to update automatically as PLSDB expands. Code is provided in full as Supplementary File S1. Data availability statement This study generated no new sequence data. All primary data are publicly available from PLSDB 2025 (https://ccb-microbe.cs.uni-saarland.de/plsdb2025) and CARD v3 (https://card.mcmaster.ca). Derived per-plasmid risk scores and validation tables are provided as Supplementary Tables S1–S4. The complete reproducible pipeline (Supplementary File S1) and the revision analyses, including confounder-adjusted models, phylum-normalized enrichment, the network null model, and the feature-independent ablation (Supplementary File S2), accompany this article. Ethics This study used publicly available sequence data only. No human participants, clinical specimens, or animal experiments were involved, and ethical approval was not required. Declaration of generative AI and AI-assisted technologies No generative AI used. Funding This research received no specific funding from public, commercial, or not-for-profit funding agencies. CRediT authorship contribution statement Oluwaseun E. Agboola: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Supervision, Validation, Writing – original draft. Samuel S. Agboola: Data curation, Methodology, Project administration, Supervision, Validation, Visualization, Writing – review & editing. Oluwaseun Ruth Olasehinde: Data curation, Investigation, Methodology, Software, Validation, Visualization, Writing – review & editing. Temitope C. Aribigbola: Investigation, Software, Validation, Writing – review & editing. Anuoluwapo Bukola Shaleye: Investigation, Validation, Visualization, Writing – review & editing. I. Adefolake Alebiosu: Investigation, Validation, Visualization, Writing – review & editing. Babatunji E. Oyinloye: Supervision, Visualization, Writing – review & editing. Idowu O. Omotuyi: Investigation, Supervision, Validation, Writing – review & editing. Abel Kolawole Oyebamiji: Investigation, Supervision, Visualization, Writing – review & editing. Olusola Tolulope Fakoyede: Investigation, Validation, Writing – review & editing. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.