Machine learning-driven prediction of Salmonella enterica virulence for enhanced risk assessment in poultry in Mexico

Strain-level heterogeneity in Salmonella enterica pathogenic potential is not routinely incorporated into quantitative microbial risk assessment (QMRA), where homogeneous virulence assumptions are commonly applied. This study evaluated whether supervised machine learning (ML)-based genomic classification could provide a structured approach to integrating strain-level variation into risk modeling. Publicly available S. enterica genomes annotated with clinical metadata were used to train classifiers assigning isolates to clinically derived risk categories. Two genomic feature representations were compared: pangenome-wide gene presence/absence matrices and curated functional gene sets. Ensemble models trained under nested cross-validation achieved strong predictive performance, with pangenome-based models outperforming models restricted to curated functional gene sets. Feature importance analyses indicated that accessory genome variation contributed substantially to class discrimination. The trained classifiers were applied to an independent cohort of poultry-derived isolates to examine out-of-sample prediction patterns. Model-derived classifications were subsequently incorporated into an established QMRA framework representing high-risk chicken consumption scenarios in central Mexico. Integration of virulence categories modified dose-response assumptions and redistributed prevalence across risk strata, producing scenario-dependent differences in estimated probabilities of illness and projected case numbers. The median estimated risk varied according to the genomic stratification approach and scenario. For the ML pangenome-based model, risk estimates ranged from 5.84 × 10^-10 to 1.02 × 10^-3. The ML model based on curated functional genes produced estimates from 5.7× 10^-10 to 1.67 × 10^-3. These were compared with a previous 13-gene model (5.72 × 10^-10 to 1.70 × 10^-3) and a serotype-based approach (1.42 × 10^-9 to 6.00 × 10^-3). These findings indicate that genome-informed stratification can be operationally embedded within QMRA and, importantly, that the way strains are classified directly influences estimated risk levels. Because predicted risks varied substantially depending on whether isolates were categorized as high-, moderate-, or low-virulence, accurate genomic classification is critical for appropriately capturing strain-level heterogeneity can influence predicted public health outcomes within modeled scenarios.