The research linked patients, food and non-food workers, retail pork and chicken, livestock and farm environments, allowing resistance determinants to be tracked across human, animal and food sources. Most resistance features were mobile: 88% of the 421 traits occurred primarily on plasmids or insertion sequences. Forty traits appeared in all seven source types, including food workers, non-food workers, symptomatic patients, pork, chicken, poultry and farm environments. The most frequently shared traits included genes associated with aminoglycoside and cephalosporin resistance, illustrating that resistance determinants were not confined to a single part of the food chain.

Food workers emerged as a potential hotspot rather than a proven transmission source. They carried the greatest number of AMR features, including 134 traits found in no other sampled group. Among these were determinants associated with last-resort antibiotics, including blaCTX-M-64, blaTEM-30, blaNDM-5 and the colistin-resistance determinant mcr-1.1_IS1A; 93.2% of the food workers’ unique features were mobile. After food-worker isolates were down-sampled to match the smaller food-isolate group, they still had 7 percentage points more high-risk traits, with a 95% confidence interval of 3.4 to 11.3 percentage points. A direct comparison with non-food workers from the same occupational screening programme showed a smaller, statistically nonsignificant difference, so the result should be interpreted as an epidemiological association rather than proof that food work caused the enrichment.

The researchers developed a five-by-five risk framework that combined a trait’s observed spread across hosts and Salmonella serovars with its clinical relevance, multidrug-resistance status and mobility. Risk scores ranged from 1 to 25, with 56.5% of the 421 traits assigned to the two highest categories. This approach also incorporated antibiotic-use importance, including the World Health Organization’s AWARE classification, and distinguished resistance genes according to whether they occurred in the chromosome, insertion sequences, plasmids or multiple locations. The framework was designed to evaluate resistance gene–mobile-element combinations rather than treating each gene as an isolated marker.

Bayesian mixed-effects models then tested whether genomic traits predicted observed resistance or susceptibility to 37 antibiotics while accounting for lineage, province, year and source type. Eighteen antibiotic-specific models met the prespecified area-under-the-curve threshold of 0.7, and 339 unique genomic traits were selected across those models. Of these, 205 were high-risk and 98% were mobile in at least one isolate. In a temporal validation using 2021–2022 isolates withheld from model fitting, the mean area under the curve was 0.94 ± 0.04 across the applicable antibiotic models. Performance generally persisted in an independent Guangdong dataset, although reduced precision in some models reflected shifts in baseline resistance prevalence and showed that local recalibration would be required.

Among the 205 high-risk traits selected by the models, 165 had an absolute Pearson correlation greater than 0.5 with at least one of 96 contextual indicators. The strongest recurring associations involved socioeconomic conditions, health-system measures and antimicrobial consumption, while climate indicators were also linked to particular traits. Of the 205 high-risk traits, 101 could be forecast, and 84 were projected to increase in association with indicators that were themselves increasing or stationary. Among those 84, 71% conferred multidrug resistance, 98% had been detected in more than one province and 98% occurred in at least two Salmonella serogroups. These are national time-series associations, however, and do not establish that any indicator directly caused resistance to rise.

The projections differed sharply even among genes affecting the same antibiotic class. The third-generation cephalosporin-resistance genes blaCTX-M-64 and blaTEM-1 were forecast to increase by as much as 20% in prevalence by 2050, whereas blaCTX-M-65 was projected to remain below 5%. For tetracycline resistance, tet(A) and tet(M) were expected to increase while tet(B) was forecast to decline. In the fluoroquinolone models, prevalence of the chromosomal gyrA 87 mutation was projected to decrease, while acquired plasmid-associated determinants such as qnrB17 and qnrS1 were forecast to rise. The latter pattern suggests a possible mechanistic shift in fluoroquinolone resistance, but the study did not directly test the causes of that shift or establish the fitness consequences of individual determinants in China’s food chain.

The researchers used regression models to relate trait prevalence to indicator trends and used Monte Carlo simulations to estimate uncertainty through 2050. Their sensitivity analysis retained 95% of forecastable AMR traits and preserved their projected directions when redundant indicators were removed. Principal-component analysis did not show that the selected traits were concentrated around a single dominant dimension of indicator variation, and 78.2% of forecastable traits were associated with at least two indicator categories. These checks support the robustness of the broad analytical pipeline, although the 25-year horizon is long relative to the maximum 17 years of overlapping data used for calibration, and the forecasts should be updated as observations accumulate.

Important limitations constrain interpretation. Food-worker and non-food-worker isolates came from Guangxi province, whereas other source groups were sampled in different provinces and periods, leaving possible geographic and temporal confounding despite mixed-effects adjustment. Repeated occupational screening could also introduce dependence because anonymized records prevented individual tracking, although positive individuals were treated until testing became negative. National-level contextual indicators masked substantial provincial differences, short-read sequencing limited the accuracy of plasmid reconstruction, and ecological correlations cannot establish causality. Prospective sampling of every source category within the same provinces and time windows would be needed to determine whether food workers truly function as reservoirs. Even with those caveats, the study shows how combining genomic surveillance with contextual data can generate gene-specific hypotheses about where resistance may emerge and which determinants require closer monitoring.

Subject of Research: Genomic surveillance and forecasting of antimicrobial resistance in multidrug-resistant Salmonella across China’s food chain

Article Title: High-risk antimicrobial resistance genomic traits in multidrug-resistant Salmonella across China’s food chain: projections to 2050

Article References: Baker, M., Lu, X., Luo, C., Wu, G., Maciel-Guerra, A., Pinpimai, K., Luo, M., Wang, M., Zhang, Y., Meng, W., Zhu, B., Kan, B., & Dottorini, T. (2026). High-risk antimicrobial resistance genomic traits in multidrug-resistant Salmonella across China’s food chain: projections to 2050. Nature Communications, 17(1), Article 10744. https://doi.org/10.1038/s41467-026-78346-7

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78346-7

Keywords: Antimicrobial resistance, Salmonella enterica, Whole-genome sequencing, Food-chain surveillance, One Health, Mobile genetic elements, Resistance forecasting, Food workers, China, Fluoroquinolone resistance, High-risk, antimicrobial