PlasmidRiskNet: An explainable machine-learning framework for antimicrobial resistance plasmid risk stratification

ABSTRACT Background Plasmids 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. Methods From 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. Results AMR 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 bla NDM , bla VIM , and bla IMP (0.72 to 0.73) but at or below chance for bla KPC and bla OXA-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. Conclusions PlasmidRiskNet offers a useful pre-screening layer for MBL-bearing plasmids but not for the compact serine-carbapenemase backbones ( bla KPC , bla OXA-48 ), which require replicon typing. Honest external and confounder-adjusted evaluation, not internal metrics, defines its class-specific surveillance value.