Factors associated with plasmid antibiotic resistance gene carriage revealed using large-scale multivariable analysis

Abstract

Plasmids are major vectors of bacterial antibiotic resistance, but understanding of factors associated with plasmid antibiotic resistance gene (ARG) carriage is limited. We curated > 14,000 publicly available plasmid genomes and associated metadata. Duplicate and replicate plasmids were excluded; where possible, sample metadata was validated externally (BacDive database). Using Generalised Additive Models (GAMs) we assessed the influence of 12 biotic/abiotic factors (e.g. plasmid genetic factors, isolation source, collection date) on ARG carriage, modelled as a binary outcome. Separate GAMs were built for 10 major ARG types. Multivariable analysis indicated that plasmid ARG carriage patterns across time (collection years), isolation sources (human/livestock) and host bacterial taxa were consistent with antibiotic selection pressure as a driver of plasmid-mediated antibiotic resistance. Only 0.42% livestock plasmids carried carbapenem resistance (compared with 12% human plasmids); conversely, tetracycline resistance was enriched in livestock vs human plasmids, reflecting known prescribing practices. Interpreting results using a timeline of ARG type acquisition (determined by literature review) yielded additional novel insights. More recently acquired ARG types (e.g. colistin and carbapenem) showed increases in plasmid carriage during the date range analysed (1994-2019), potentially reflecting recent onset of selection pressure; they also co-occurred less commonly with ARGs of other types, and virulence genes. Overall, this suggests that following acquisition, plasmid ARGs tend to accumulate under antibiotic selection pressure and co-associate with other adaptive genes (other ARG types, virulence genes), potentially re-enforcing plasmid ARG carriage through co-selection.

Figure 1

Co-occurrence of antibiotic resistance gene (ARG) types is determined from their presence/absence in the dataset of 14,143 plasmids, and visualised using heatmaps. ARG types are ordered by the inferred timeline of known plasmid-mediated resistance acquisition (see Table 2) from earliest (aminoglycoside, sulphonamide) to most recent (colistin). Counts along the diagonal indicate total plasmids carrying a given ARG type. Counts in the upper-left triangle indicate pairwise ARG type intersections i.e. the number of plasmids where a given pair of ARG types co-occur. Heatmaps are coloured by similarity metrics ((a) Jaccard index, (b) overlap coefficient) indicating the degree of co-occurrence between ARG types (red = more co-occurrence; light blue = less co-occurrence). Heatmaps were generated using custom R code available in a GitHub repository (PlasmidARGCarriage v1.0). (https://github.com/AlexOrlek/PlasmidARGCarriage/blob/v1.0/exploratory_analysis.R).

Figure 2

Multivariable-adjusted association between presumed proxies of antibiotic selection pressure and the log-odds of antibiotic resistance gene (ARG) carriage. The effect of (a) collection date; (b) isolation source; (c) host taxonomy on the log-odds of ARG carriage (y-axis), is shown across 10 ARG types. ARG types are ordered by the date plasmid-mediated resistance was first reported (see Table 2) from earliest (aminoglycoside, sulphonamide) to most recent (colistin). For categorical factors (isolation source, host taxonomy), log-odds ratios indicate the effect of a given factor level, relative to reference factor level, and error bars show 95% confidence intervals. For collection date, smooth curves indicate the predicted effect of sample collection date. An effective degrees of freedom (edf) value > 1 indicates a non-linear smooth. Shaded areas around smooths show 95% confidence intervals. No Firmicutes plasmids encoded known sulphonamide, ESBL, quinolone, or colistin resistance genes, and only 2 Firmicutes plasmids encoded known carbapenem resistance genes; therefore, corresponding odds ratios are not shown.

Figure 3

The multivariable-adjusted association between plasmid genetic traits and the log-odds of antibiotic resistance gene (ARG) carriage. The effect of (a) number of other ARG types; (b) virulence gene presence; (c) biocide/metal resistance gene presence; (d) integron presence, on the log-odds of ARG carriage (y-axis), is shown across 10 ARG types. ARG types are ordered by the date plasmid-mediated resistance was first reported (see Table 2) from earliest (aminoglycoside, sulphonamide) to most recent (colistin). For categorical factors (biocide/metal resistance gene presence, integron presence, virulence gene presence), log-odds ratios indicate the effect of a given factor level, relative to reference factor level, and error bars show 95% confidence intervals. For number of other ARG types, smooth curves indicate the predicted effect of ≥ 1 other resistance gene types (relative to 0 other ARG types). An effective degrees of freedom (edf) value > 1 indicates a non-linear smooth. Shaded areas around smooths show 95% confidence intervals.