Work

Whole-genome sequencing-driven source tracing of foodborne diseases: from cgMLST/SNP typing toward an integrated framework of pan-genomics and machine learning

Year: 2026

Type: review

Abstract: Core-genome multilocus sequence typing (cgMLST) and core-genome single-nucleotide polymorphism (cgSNP) analysis have made whole-genome sequencing (WGS) the standard tool for foodborne disease surveillance. Fixed-threshold clustering is increasingly unreliable for highly clonal lineages such as Salmonella Enteritidis ST11 and Escherichia coli ST131, whose core genomes differ by only a handful of alleles even between unrelated strains. The central question is whether surveillance should keep simple, easy-to-interpret allele thresholds or move to probabilistic machine-learning models that also use accessory-genome variation. Accessory genes and mobile genetic elements, including prophages, plasmids, and genomic islands, change on timescales that match outbreak transmission and carry niche-adaptation signal absent from the core genome. Gradient-boosting and deep-learning models can combine genomic, epidemiological, and spatiotemporal data to report source probabilities instead of binary cluster assignments. A U.S. CDC random forest trained on 18,661 Salmonella isolates achieved 81% out-of-bag accuracy when every prediction was counted and 91% when limited to high-confidence isolates whose largest predicted source probability was above 0.50. Feature-attribution analysis ranked plasmid-borne and cell-surface genes among the strongest predictors, raising host-adaptation hypotheses that still require experimental testing. The review ends with the threshold-vs.-probability debate, barriers to data sharing and federated-learning alternatives, and the gap between retrospective accuracy and prospective validation in routine surveillance.

Source: Frontiers in Public Health

DOI: 10.3389/fpubh.2026.1968969