Predictive modeling of Salmonella inactivation in a hybrid alternative protein matrix: A comparative evaluation of mechanistic and machine learning approaches

Abstract

• ML ensembles surpassed mechanistic baselines in predicting Salmonella inactivation. • Random Forest performed best with an R 2 of 0.91 for Salmonella inactivation. • SHAP analysis revealed complex interactions between OEO, pH, and temperature. • Validation in commercial hybrid matrices revealed a significant protective effect Ensuring microbial safety remains a key issue in the alternative protein industry. This study examined the inactivation kinetics of Salmonella , a common foodborne pathogen, in a model protein matrix. A Doehlert response surface design experiment was used to evaluate the effects of oregano essential oil (OEO) concentrations (0–0.3% v/w), pH levels (4.7–6.8), sodium chloride concentrations (0.9–3.8% w/w) and temperatures (8–20°C). Primary mechanistic modelling indicated that the biphasic model was the best fit for 73.9% of individual experiments, suggesting the presence of stress-resistant microbial populations. For global predictive models, machine learning algorithms outperformed global mechanistic models. The Random Forest (RF) model achieved the best performance with R² = 0.91, RMSE = 0.52, B f = 1.09 and A f = 2.24, followed by XGBoost. Validation conducted using a commercial hybrid protein substrate demonstrated the protective effect of the matrix. However, the RF model demonstrated strong relative performance (R² = 0.75; RMSE = 1.04) compared to the one-step global Weibull model. Chemical profiling using GC-MS and sensory profiling using an electronic tongue revealed the effects of the matrices on volatile antimicrobial agents. These findings establish a data-driven approach for safety-by-design in the alternative protein industry.