Plasmids are principal vehicles of antimicrobial resistance dissemination, moving resistance determinants across bacterial lineages, host species, and environmental reservoirs. Which plasmids warrant concern, and in which host contexts, is harder to establish: current tools describe plasmid content and mobility but do not capture host context or the uncertainty associated with incomplete genomic data. This project aims to develop machine learning models that score plasmid risk from surveillance data of varying quality for use in public health decision-making.

Risk scores will be derived from features such as mobility potential, resistance-gene content, and host range, inferred in the first instance from genome assemblies, with training and evaluation temporally separated so that scores are tested against subsequent dissemination rather than prevalence already recorded. Whether read-based sketches can approximate assembly-based performance is a secondary question, and one that determines how far the approach extends to unassembled or lower-quality data.

A key aim will be to develop models that are interpretable as well as predictive by contextualising inference with biological sequence composition, plasmid backbone relatedness, protein-level features of mobility, virulence, and resistance genes, host metadata, and epidemiology. Validation will proceed from characterised clinical isolates, where phenotype is known, to closed plasmid sequences, where structure is certain, to independent metagenomes, where neither is available. Performance, robustness to fragmented input, and generalisation across bacterial lineages will be assessed at each stage.