Anticipating the Next Recorded Bacterial Hosts of Plasmid-Borne Antimicrobial Resistance Genes: A Temporal Graph Analysis of Plasmid and Surveillance Data

Background/Objectives: Plasmid-mediated dissemination of antimicrobial resistance genes complicates the control of bacterial infections. Identifying bacterial hosts likely to harbour these genes could guide genomic surveillance before such associations enter public databases. We investigated whether existing plasmid and surveillance data could anticipate which resistance gene–host associations would next be recorded in those databases, and which genomic features inform that prediction. Methods: We integrated 72,556 complete plasmid records and 2,116,241 surveillance isolates, encompassing 458 resistance gene alleles, 574 replicon backbones and 222 bacterial species. A temporal graph model, MobiGraph, was trained on associations recorded before a fixed cutoff and evaluated on subsequently observed associations, using candidate hosts matched for sequencing effort and phylogenetic distance. Ranking within a candidate pool was the primary outcome; threshold-free summaries are reported as secondary. Results: The observed host ranked among the ten highest-priority candidates in 42.4% of test cases; the advantage over gradient boosting on mechanistic features was confined to the ranking metrics, the two being indistinguishable on area under the precision–recall curve. Plasmid incompatibility-group compatibility, previous gene–host associations and shared relaxase classes were prominent predictive features. Across five resistance determinants, including mcr-1, blaNDM and tet(X4), 43–62% of subsequent associations were ranked within the top ten. Among recovered events, the median interval from the prediction cutoff to first recorded observation was 22.5 months. In a separate evaluation of a fixed 500-entry watch list, 65 entries were subsequently recorded as gene–host pairs over 25 months, 13.0% against an 8.2% background rate (95% CI 10.3–16.2%); because the surveillance database does not resolve the replicon, this confirmation does not extend to the predicted backbone. Conclusions: Existing plasmid and genomic surveillance records contain information that can help anticipate which resistance gene–host associations will next be recorded. Whether a newly recorded association reflects recent transfer or earlier under-sampling cannot be resolved from these data. These findings support prioritising bacterial hosts for targeted surveillance and experimental investigation of plasmid-mediated resistance dissemination.