Forecasting next-year change in antimicrobial resistance genes in Klebsiella pneumoniae and Acinetobacter baumannii from genomic surveillance data
Background: Surveillance tells laboratories which antimicrobial resistance genes are circulating now, but not which will change next. We built a model that forecasts, for each resistance gene and year, whether the following year brings a previously unseen allele (sequence variant), called emergence, or a change in the most common allele, called turnover. We tested it separately in Klebsiella pneumoniae and Acinetobacter baumannii. Results: We used genomes annotated with AMRFinderPlus: 2,385 gene-years across 331 genes for K.pneumoniae and 1,059 across 216 for A. baumannii. To remove the effect of sequencing effort, each gene-year was sampled to 10 records and the analysis repeated 25 times. Models were trained on earlier years and tested on later ones, and on genes held out of training. On later years every model ranked genes by risk better than simple baselines in both organisms, and results for held-out genes were close. In K.pneumoniae, turnover was predicted better than emergence (best area under the ROC curve, AUROC, 0.909 versus 0.819). In A. baumannii the order reversed (emergence 0.789, turnover 0.722). The two events also relied on different features. Emergence depended mainly on how many alleles a gene had already accumulated, in both organisms. Turnover in K. pneumoniae depended mainly on how dominant the leading allele already was, and a stability index added a small gain. In A. baumannii it was spread across many weaker features. There, 56% of changes in the leading allele involved only alleles already seen, against 23% in K. pneumoniae, which may make them harder to predict. Conclusions: Changes in resistance genes can be forecast a year ahead from routine genomic surveillance data, including for genes the model has not seen. How well it works, and what it depends on, differs between organisms, so models should be built and validated for each organism, not assumed to transfer. Such models could help laboratories decide which resistance genes to monitor most closely. Keywords: antimicrobial resistance; genomic surveillance; forecasting; machine learning; allele dynamics; Klebsiella pneumoniae; Acinetobacter baumannii