A Kinetic Platform to Determine the Fate of Hydrogen Peroxide in Escherichia coli

Abstract

Hydrogen peroxide (H2O2) is used by phagocytic cells of the innate immune response to kill engulfed bacteria. H2O2 diffuses freely into bacteria, where it can wreak havoc on sensitive biomolecules if it is not rapidly detoxified. Accordingly, bacteria have evolved numerous systems to defend themselves against H2O2, and the importance of these systems to pathogenesis has been substantiated by the many bacteria that require them to establish or sustain infections. The kinetic competition for H2O2 within bacteria is complex, which suggests that quantitative models will improve interpretation and prediction of network behavior. To date, such models have been of limited scope, and this inspired us to construct a quantitative, systems-level model of H2O2 detoxification in Escherichia coli that includes detoxification enzymes, H2O2-dependent transcriptional regulation, enzyme degradation, the Fenton reaction and damage caused by •OH, oxidation of biomolecules by H2O2, and repair processes. After using an iterative computational and experimental procedure to train the model, we leveraged it to predict how H2O2 detoxification would change in response to an environmental perturbation that pathogens encounter within host phagosomes, carbon source deprivation, which leads to translational inhibition and limited availability of NADH. We found that the model accurately predicted that NADH depletion would delay clearance at low H2O2 concentrations and that detoxification at higher concentrations would resemble that of carbon-replete conditions. These results suggest that protein synthesis during bolus H2O2 stress does not affect clearance dynamics and that access to catabolites only matters at low H2O2 concentrations. We anticipate that this model will serve as a computational tool for the quantitative exploration and dissection of oxidative stress in bacteria, and that the model and methods used to develop it will provide important templates for the generation of comparable models for other bacterial species.

Fig 1. H₂O₂ biochemical reaction network.

A. The kinetic model is separated into three compartments: gas, media, and intracellular. It includes spontaneous reactions (black lines, details in S2 Table) and enzymatic reactions (blue lines, details in S3 Table). Metabolite abbreviations can be found in S1 Table. Information regarding enzyme degradation can be found in S2 Table, and enzyme expression is described in S3 Table. For clarity, •OH reaction products with amino acids and protons are not included in the diagram.

Fig 2. Systematic approach to construct a kinetic model of H₂O₂ metabolism.

Uncertain parameters in each of the ten model structures are optimized on wild-type clearance data of 10, 25, 100, and 400 μM H₂O₂, starting from 1,000 random initial parameter sets. Any models within an evidence ratio of 10 (ER≤10) are used to calculate cumulative H₂O₂ clearance by the different detoxification pathways. If the calculated H₂O₂ distributions between the models are inconsistent, simulations are used to suggest experiments that differentiate between the disagreeing models, experiments are executed, and the optimization is performed on all experimental data for the model structures that had at least one parameter set with an ER≤10. Once consistent H₂O₂ distributions are realized, we identify an ensemble of parameter sets that can all describe the data comparably well with an MCMC procedure. We then assess whether H₂O₂ distributions are consistent across the entire ensemble. If the calculated H₂O₂ distributions are consistent, we undertake forward predictions.

Fig 3. Parameter training on wild-type data and analysis of acceptable models.

A-D. Parameters for each of the ten different structures were optimized on wild-type clearance of 10 (A), 25 (B), 100 (C), and 400 (D) μM H₂O₂. Models were ranked using an AIC-based method (Methods), and the 35 models with an ER≤10 were considered viable. Experimental data (solid points) represents at least three biological replicates, with error bars showing the standard error of the mean. Windows represent the maximum and minimum of the 35 acceptable models. Solid lines within the window show the most likely model. E-H. Prediction of the amount of H₂O₂ cleared by the two major detoxification pathways, AHP (orange) and combined catalase activity (black), after boluses of 10 (E), 25 (F), 100 (G), and 400 (H) μM H₂O₂. Each line represents the prediction from a single model. I-L. Prediction for H₂O₂ clearance of 10 (I), 25 (J), 100 (K), and 400 (L) μM H₂O₂ after removal of all catalase activity (ΔkatE ΔkatG). Structure 2 (green) and structure 3 (purple) models predict different clearance dynamics after this perturbation, suggesting that data obtained from this mutant could be used to discriminate between the model structures.

Fig 4. Parameter training on wild-type and ΔkatE ΔkatG data and analysis of acceptable models.

A-D. Parameters for each of the two remaining structures (structures 2 and 3) were optimized simultaneously on clearance of 10 (A), 25 (B), 100 (C), and 400 (D) μM H₂O₂ by wild-type (red) and a ΔkatE ΔkatG mutant (blue). The clearance of 400 μM H₂O₂ by ΔkatE ΔkatG was omitted from the training procedure because significant cell death was observed (S1H Fig). Experimental data (solid points) represents at least three biological replicates, with error bars showing the standard error of the mean. Windows represent the maximum and minimum of the fits from the 965 acceptable models. Solid lines within the window show the most likely model. E-H. Prediction for the amount of H₂O₂ cleared by the two major detoxification pathways AHP (orange) and combined catalase activity (black) after boluses of 10 (E), 25 (F), 100 (G), and 400 (H) μM H₂O₂. Each line represents the prediction from a single model. I-L. Prediction for the amount of H₂O₂ cleared by the individual catalases HPI (pink) and HPII (green) after boluses of 10 (I), 25 (J), 100 (K), and 400 (L) μM H₂O₂. Each line represents the prediction from a single model.

Fig 5. Model training on wild-type, ΔkatE ΔkatG, ΔkatE, and ΔkatG data and analysis of acceptable models.

A-D. Parameters for the one remaining structure were optimized simultaneously on clearance of 10 (A), 25 (B), 100 (C), and 400 (D) μM H₂O₂ by wild-type (red), ΔkatE ΔkatG (blue), ΔkatE (purple), and ΔkatG (green) data. The clearance of 400 μM H₂O₂ by ΔkatE ΔkatG and ΔkatG were omitted from the training procedure because significant cell death was observed (S1H, S1P Fig). Experimental data (solid points) represents at least three biological replicates, with error bars showing the standard error of the mean. Windows represent the maximum and minimum of the fits from the 40 acceptable models. Solid lines within the window show the most likely model. E-H. Prediction for the amount of H₂O₂ cleared by the two major detoxification pathways AHP (orange) and combined catalase activity (black) after boluses of 10 (E), 25 (F), 100 (G), and 400 (H) μM H₂O₂. Each line represents the prediction from a single model. I-L. Prediction for the amount of H₂O₂ cleared by the individual catalases HPI (pink) and HPII (green) after boluses of 10 (I), 25 (J), 100 (K), and 400 (L) μM H₂O₂. Each line represents the prediction from a single model.

Fig 6. Predictions for clearance during carbon deprivation.

A-D. Ensemble predictions for H₂O₂ clearance of 10 (A), 25 (B), 100 (C), and 400 (D) μM H₂O₂ by wild-type in M9 10 mM glucose (red) and M9 lacking glucose (orange). The spontaneous rate of H₂O₂ degradation differs in media with and without glucose, and this parameter was optimized using cell-free controls and adjusted to reflect the different spontaneous degradation during simulations. E-H. Ensemble simulations for H₂O₂ clearance of 10 (E), 25 (F), 100 (G), and 400 (G) μM H₂O₂ by wild-type with only NADH depletion (blue) or translation inhibition (black). Since these were controls for the–glucose predictions, the spontaneous H₂O₂ degradation rate matched that of media lacking glucose. I-L. Experimental measurement of clearance of 10 (I), 25 (J), 100 (K), and 400 (L) μM H₂O₂ by glucose-deprived (orange) and CAM-treated (black) cultures, shown with their predicted profiles. Experimental data (solid points) represents three biological replicates, with error bars showing the standard error of the mean. Results for clearance of 400 μM H₂O₂ by CAM-treated cultures were omitted based on significant cell death (S1X Fig). In all simulations, windows represent the maximum and minimum of the predictions from the 4,000 models in the ensemble. Solid lines within the window show the most likely model.