Predictive Modeling of Probiotic-Pathogen Interactions under Threshold-Triggered Antimicrobial Peptide Control

Combining probiotics and antimicrobial peptides (AMPs) may enable targeted pathogen suppression while preserving beneficial bacteria, but the nonselective antimicrobial activity of AMPs can suppress both populations. We developed an experimentally informed synthetic ordinary differential equation (ODE) benchmark and an inverse-design workflow to allow threshold-triggered AMP control. Preliminary growth and AMP-response measurements from Lactococcus lactis and Escherichia coli were used to define plausible simulation ranges rather than to fit a complete multi-strain model. Through 352,000 closed-loop simulations spanning 2,000 biological-profile groups and 176 candidate controllers per group, we implemented and evaluated Tree-Adaptive Regression (TAR), a target-wise stack of single-tree experts trained with group-aware out-of-fold predictions to predict five strain-specific triggering thresholds. Additionally, AMP release amplitude $$U_{\max }$$ was selected separately by post-prediction ODE reinsertion. Among the tree-based models in the final benchmark, TAR had the highest mean target-wise $$R^2$$ and the lowest original-scale RMSE and showed the highest mean ODE-back reference-fidelity $$R^2$$ among the evaluated tree-based models. However, post-prediction $$U_{\max }$$ selection exposed trade-offs among pathogen suppression, probiotic preservation, and AMP exposure. No jointly feasible candidate was identified in the post-prediction $$U_{\max }$$ evaluation, so the selected amplitudes represented minimum-penalty trade-offs rather than universally feasible controllers. The framework is a reproducible computational testbed for prioritizing controller candidates. Its predictions are simulation-derived and require prospective experimental validation.