Understanding Iron and Oxidative Stress Response in Escherichia coli Using Multi-phenotype and Ensemble Models

Developing effective antimicrobial strategies requires a predictive understanding of bacterial responses to multiple stress conditions which often result in multiple phenotypes. A microbe's survival and proliferation depend on its ability to manage concurrent, dynamically varying stressors within its microenvironment. However, time-resolved predictive models that capture multi-phenotype responses are lacking, and single-phenotype models often fail to accurately replicate a microbe's reaction to mixed stress conditions. In this work, we develop a mechanistic in-silico model of multi-stress response in Escherichia coli K12 and use it to characterize phenotype dynamics in iron-limited and hydrogen peroxide containing environments. Specifically, we replicate the iron and oxidative stress response networks in E. coli using a system of ordinary differential equations and applied a multi-phenotype parameterization scheme that leverages multi-measure empirical data, augmented metric-based sensitivity analysis, sequential parameter estimation, and ensemble modeling. Our approach resulted in robust models with a 93% accuracy when compared to experimental datasets across 20 stress-response categories, outperforming traditional single-phenotype approaches (80-87%). Analysis of posterior parameter distributions revealed that multi-phenotype optimization eliminates heavy-tailed distributions characteristic of poorly constrained fits and shifts parameter posteriors from boundary-concentrated to centrally localized forms, indicating improved identifiability. Simulation outcomes confirmed key features of E. coli's iron metabolism, showing that moderate peroxide stress in an iron-rich environment creates significant adaptation challenges, leading to a bacteriostatic phenotype. The model provides insights into biochemical mechanisms important to E. coli's temporal response to varying iron availability, with implications for ecological dynamics and pathogenesis. Our parameterization approach highlights the effectiveness of a combination of optimization methods and ensemble modeling in developing predictive models that are robust across multiple phenotypes. Results demonstrate that data structure, specifically the integration of multiple phenotypes and response outputs, proves to be as critical if not more critical than data volume for achieving well-constrained parameter estimates and robust predictions across experimental conditions.