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

Area of Science:

• Genomics and Bioinformatics
• Molecular Biology
• Epidemiology

Background:

• Plasmids are key vectors for horizontal antimicrobial resistance (AMR) gene transfer.
• Current risk analyses often fail to integrate plasmid gene content, mobility, and network topology.
• Explainable machine learning offers a potential approach to stratify plasmid dissemination risk.

Purpose of the Study:

• To evaluate if machine learning can stratify plasmid dissemination risk by integrating gene content, mobility, and network topology.
• To rigorously test the success and failure points of machine learning models in predicting plasmid risk.
• To assess the utility of a composite PlasmidRisk score derived from multiple features.

Main Methods:

• Integrated data from 72,556 plasmids (PLSDB) with AMRFinderPlus gene records and CARD v3 ontology.
• Utilized MOBsuite for plasmid typing, constructed a co-resistance network, and derived a PlasmidRisk score.
• Employed five-fold cross-validation and external validation using WHO/ECDC carbapenemase designations; adjusted for length, host, and phylum.

Main Results:

• Antimicrobial resistance genes were present in 41.0% of plasmids across 85 drug classes.
• Machine learning models showed high internal consistency (AUC > 0.999) but modest external validation (AUC = 0.607).
• The model performed well for metallo-beta-lactamases (e.g., blaNDM, blaVIM, blaIMP) but poorly for serine-carbapenemases (e.g., blaKPC, blaOXA-48).

Conclusions:

• PlasmidRiskNet serves as a useful pre-screening tool for metallo-beta-lactamase-bearing plasmids.
• The model's effectiveness is limited for compact serine-carbapenemase backbones, necessitating replicon typing.
• Confounder-adjusted evaluation is crucial for determining the true surveillance value of such predictive models.