• Loading metrics

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

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

  • Kristin J. Adolfsen, 
  • Mark P. Brynildsen


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.

Author Summary

Bacterial hydrogen peroxide (H2O2) response networks contain essential virulence factors for a number of pathogens. Without these systems, infecting bacteria fall prey to host immune cells and cannot establish or sustain an infection. The reaction networks and regulatory features involved are complex, which suggests that computational modeling would facilitate quantitative dissection and analysis of these systems. However, current models of H2O2 reaction networks have been of limited scope. Here, we constructed a systems-level H2O2 detoxification model for Escherichia coli, and used it to understand how the network responds to different H2O2 concentrations and insults. We anticipate that this model and comparable ones for other species that are facilitated by its construction will be useful in identifying and understanding methods to sensitize pathogens to immune attack. Such strategies hold great promise for the development of next generation antibiotics, since agents that impair oxidative stress defense systems would focus selective pressure to infection sites, and therefore exhibit slow resistance development and little impact on commensal bacteria.


Reactive oxygen species (ROS) are critical immune antimicrobials used in the first line of defense against infections, where phagocytic cells of the innate immune response use NADPH oxidase to generate an “oxidative burst” of superoxide (O2•) after engulfing pathogens in a phagosome [1, 2]. The O2 can then be dismutated to H2O2 [2], which readily diffuses across the bacterial membrane where it damages sensitive biomolecules, reacts with ferrous iron to produce the highly deleterious •OH [3], or is detoxified by specialized enzymes. The importance of the oxidative burst to immunity is highlighted by the incidence of recurring infections within and shortened life expectancy of patients with defects in NADPH oxidase, a condition known as chronic granulomatous disease (CGD) [4]. In addition, many pathogens including Bacillus anthracis [5], Coxiella burnetti [6], Chlamydia trachomatis (serovars E, K, and L2) [7], Salmonella enterica (serovar Typhimurium) [8], Mycobacterium tuberculosis [9, 10], Staphylococcus aureus [11], Helicobacter pylori [12], Streptococcus pyogenes [13], and Enterococcus faecalis [14] require H2O2 defense systems to establish or sustain infections. Interestingly, beyond its use by immune cells, bacteria also use H2O2 against each other, such as when Streptococcus pneumoniae stimulates prophage induction and cell death in Staphylococcus aureus by generating H2O2 during niche competitions [15].

Accordingly, bacteria have evolved various pathways to detoxify H2O2. While the importance of these H2O2 detoxification systems has been established [16], there are gaps in knowledge regarding the kinetic interplay between them under different conditions. Escherichia coli K-12 encodes one alkyl hydroperoxidase (AHP) and two separate catalases for detoxifying H2O2, which differ in regulation and/or reaction mechanism. AHP and catalase HPI expression are induced by OxyR during oxidative stress, whereas catalase HPII expression is up-regulated in stationary phase and does not increase in the presence of H2O2 [1719]. AHP requires one molecule of NADH per reaction cycle, coupling the rate of detoxification achievable by this enzyme to catabolism, whereas H2O2 is the only substrate in the catalase reaction cycle. AHP has been shown to act as the primary scavenger of endogenously produced H2O2, and is efficient at detoxifying low concentrations of H2O2 (<20 μM), whereas catalase is known to dominate clearance at higher concentrations (>50 μM) [20, 21]. Since the result of H2O2 exposure whether that be bacteriostasis, mutagenesis, cell death, or continued growth depends on a kinetic competition for the molecule, it is important to have a quantitative, systems-level understanding of its biochemical reaction network. Due to the complexity of H2O2 biochemical reaction networks, computational models are necessary for interpretation of H2O2 detoxification data and prediction of system behavior.

As a result of its importance as a signaling molecule, the most complete models of H2O2 metabolism currently available were developed for mammalian systems [2224]. They have included H2O2 elimination by antioxidants (e.g., glutathione and thioredoxin) and enzymes (e.g., catalase, glutathione peroxidase, glutathione reductase, glutaredoxin, and peroxiredoxin), and processing of oxidized protein thiols [2224]. However, these models were specific to mammalian physiology and did not include transcriptional regulation, enzyme degradation, side reactions of H2O2 with sensitive biomolecules such as methionine and pyruvate, or the related reactive oxygen species O2• and •OH and their associated reactivity (e.g., •OH rapidly oxidizes all twenty amino acids and glutathione). Although models equivalent to those of mammalian systems have yet to be described for bacteria, there has been progress in modeling subsystems of the H2O2 response network under H2O2 stress, such as the thioredoxin system in E. coli [25].

Here, we have generated a kinetic model of H2O2 stress in E. coli whose components are depicted in Fig 1. The biochemical reaction network is compartmentalized into media and intracellular spaces, includes spontaneous and enzymatic detoxification of H2O2, transcriptional regulation and inactivation of detoxification enzymes, and reactions of H2O2 and its degradation intermediates (e.g., •OH) with biomolecules (e.g., pyruvate, glutathione, all twenty amino acids). Parameters were informed from literature or trained using an iterative and integrated computational and experimental approach (Fig 2). The design criteria we chose to use to develop the model stipulated that consistent discrimination between clearance contributions by the major detoxification systems (AHP, HPI, and HPII) needed to be achieved. Once the design criteria were met, remaining parametric uncertainty was accounted for with use of a Markov chain Monte Carlo (MCMC) procedure to explore the viable parameter space and assemble an ensemble of models that performed comparably well with the training data [26]. The ensemble was then used to quantitatively investigate the importance of carbon availability and translation to H2O2 detoxification, and its predictions were experimentally confirmed.

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.


Design criteria

Our aim was to construct a systems-level kinetic model of H2O2 detoxification in E. coli that could provide consistent predictions of H2O2 distributions among its different detoxification pathways after exposure to a range of initial H2O2 boluses. To accomplish this goal in the most efficient way possible, we adopted the systematic approach shown in Fig 2. Briefly, we began with a minimal number of experiments, wild-type clearance of different initial H2O2 concentrations. After optimizing uncertain parameters, we selected models based on their relative likelihood, also referred to as their evidence ratio (ER) [2732], discarding models more than ten times less likely than the most-likely model in our set (ER≥10). If the acceptable models did not uniformly attribute H2O2 detoxification to the same pathways, we performed simulations to suggest experiments that could resolve the disagreement. Those experiments were then performed, and data used to arrive at updated parameter estimates. This process was continued until we arrived at a model or set of models that rendered consistent H2O2 distributions. Since some parameters may not have been important to H2O2 clearance under the conditions used here, and therefore, unlikely to be informed by the training procedure, we explored the parameter space using a previously developed MCMC procedure [26] to assemble an ensemble of parameter sets that could all describe the H2O2 clearance data comparably well (ER≤10). In this way, we could ensure that forward predictions were not dependent on ill-defined parameters. We note that this procedure also accounts for cases in which parameter pairings or more complex relationships rather than absolute values are important by varying all parameters simultaneously when walking away from known viable points. Also, before proceeding to forward predictions, we confirmed that all the models within the ensemble still satisfied the design criteria.

Kinetic model of H2O2 metabolism in E. coli

We constructed a compartmentalized reaction network that includes spontaneous and enzymatic reactions present in an E. coli culture under H2O2 stress, transcriptional regulation of AHP and HPI, and degradation/inactivation of the major detoxification enzymes AHP, HPI, and HPII. Uncertainty exists with regard to the dynamics of enzyme degradation/inactivation in the presence of H2O2, as well as the possibility of an H2O2 gradient across the membrane. Specifically, the enzymes could be degraded or inactivated in an H2O2-independent manner, either with a fixed degradation constant [33, 34], or optimized to account for the varying degradation rates of different proteins [35, 36]. Alternatively, the H2O2 detoxification enzymes could be poisoned by their own substrate [3739], following bimolecular [40] or more complex kinetics [37].

In addition to the indeterminacy in degradation/inactivation kinetics, there is evidence supporting [41] and opposing [42] the presence of an H2O2 gradient across the cell membrane. Models accounting for these various possibilities are presented in Table 1, along with their corresponding number of uncertain parameters. The introduction of unknown parameters has the potential to improve agreement between model simulations and experimental data solely by increasing the flexibility of the model. For this reason, we calculated the relative likelihood of models, otherwise known as their evidence ratio (ER) [2732] based on their respective Akaike Information Criterion (AIC), which is a commonly used statistical metric that weighs goodness of fit against model complexity when discriminating between competing models [27, 4345]. Models with a relative likelihood of ten times less than the best model in the set (ER≤10) were considered acceptable, whereas others were discarded.

Table 1. Model types.

Rate equations can be found in S2 and S3 Tables, and a more complete description can be found in the Methods.

Experimentally-driven model refinement and parameter estimation

When optimizing parameters simultaneously on clearance of 10, 25, 100, and 400 μM boluses in wild-type cultures, 35 of the 10,000 models had an ER≤10 and were considered viable models. The windows of simulation results of these 35 models are presented in Fig 3A–3D along with the experimental clearance data they were trained on. None of these models contained a gradient; 30 were structure 2 models, and 5 were structure 3. We note that structures that contain a gradient could be favored over those with no gradient under different conditions (e.g., training on data from a single H2O2 concentration); however, our goal was to arrive at a model that could describe a wide range of bolus concentrations, and the gain in simulation accuracy for gradient models did not justify addition of the extra parameter as determined by the ER for the experimental conditions considered here. When the H2O2 distributions of the acceptable models were analyzed, the utility of AHP and the catalases separated into two distinct groups at all bolus concentrations (Fig 3E–3H). The reaction fluxes through AHP and HPI+HPII can be found in S2A–S2D Fig, and those also separated into two distinct groups. We found that these two groups represented predictions made by the two model structures. At all bolus concentrations, structure 2 models predicted a greater contribution by AHP than did structure 3 models. Indeterminacy in catalase null mutant simulations (Fig 3I–3L) suggested that experiments on a strain lacking both HPI (katG) and HPII (katE) would resolve this discrepancy.

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.

Simultaneous training of models on wild-type and ΔkatE ΔkatG clearance data was able to resolve the uncertainty between structures 2 and 3. This training iteration resulted in 965 models that all had an ER≤10 (Fig 4A–4D), and all acceptable models were structure 3, which suggests that bimolecular H2O2-dependent enzyme degradation is an important feature of the detoxification network. We note that clearance of 400 μM H2O2 by ΔkatE ΔkatG was omitted because significant cell death was observed (S1H Fig), and the models were not designed to simulate cell death and possible lysis. All models predicted similar distributions across the major pathways (Fig 4E–4H), but diverged when we looked more closely at the individual catalase contributions (Fig 4I–4L). Reaction fluxes through the major pathways and individual catalases can be found in S2E–S2H Fig and S3A–S3D Fig, respectively. The different parameter sets predicted a range of clearance profiles after removal of either catalase (S4 Fig), suggesting that data obtained from these mutants would resolve the disagreement between models.

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.

Training uncertain model parameters on wild-type, ΔkatE ΔkatG, ΔkatE, and ΔkatG data resulted in 40 parameters sets from the 1,000 random initializations that were within an ER of 10 (Fig 5A–5D). All of these models agreed regarding how H2O2 distributes across not only the major pathways (Fig 5E–5H), but also the individual catalases (Fig 5I–5L). Reaction fluxes through the major pathways and individual catalases can be found in S2I–S2L Fig and S3E–S3H Fig, respectively. The consistent distributions satisfied our design criteria, so we proceeded with the generation of an ensemble of viable parameter sets with which to make forward predictions.

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.

Exploration of the viable parameter space

The identification of universal “sloppiness” in computational biological models [46], meaning many parameters are poorly constrained after fitting on experimental data, led to the development of a number of methods designed to identify ensembles of parameter sets that could comparably describe the data and be used to assess the robustness of forward predictions [26, 46]. Methods such as “brute force” uniform sampling or Gaussian sampling become impossible with increasingly complex models, so computational biologists have turned to the use of Monte Carlo techniques to explore the viable parameter space efficiently (e.g., HYPERSPACE [26] and SloppyCell [46]). Here, we used a previously developed MCMC method [26] to explore the parameter space, initiating a random walk away from each of the 40 acceptable parameter sets and keeping 100 viable sets with an ER≤10 for each point. This resulted in an ensemble of 4,000 parameter sets that could all capture our experimental observations, and allowed us to assess robustness of our predictions to parametric uncertainty. In addition, before proceeding, we ensured that all models in the ensemble satisfied our design criteria (S5 Fig).

Minimal H2O2 detoxification model

Based on the tight predictions that AHP, HPI, and HPII would dominate clearance (Fig 5), we sought to determine the minimal reaction network required to capture all of our data. To do this, we adopted a previously used two-tiered approach that first deletes reactions in a random order, and then re-optimizes uncertain parameters to determine if adjusted parameters would allow deletion of additional reactions [33]. Beginning with the best model in our ensemble and using this method, we determined that 70 out of our 75 reactions could be removed without increasing the ER beyond a threshold of 10. The essential reactions to the network were the major detoxification enzymes (AHP, HPI, and HPII) and degradation of AHP and HPI. In the case of AHP, a drop in active enzyme could indicate degradation, or alternatively a decrease in available NADH, which is held constant during simulation. On the other hand, H2O2 is the only substrate of catalase, which suggests that the importance of degradation reflects a decrease in concentration of functional enzyme.

Parametric sensitivity analysis

In addition to identifying reactions in the network that are dispensable to capturing the H2O2 clearance data presented in Fig 5, we identified those uncertain parameters that influenced the simulations. Using the optimal parameter set, we individually varied each of the 13 optimized parameters within their bounds. Changes to the Fenton reaction rate constant and Fe2+ and Fe3+ initial concentrations never increased the ER to beyond 10. All other uncertain/trained parameters perturbed simulations to varying degrees, and their impact was quantified and displayed in S6 Fig

Impact of carbon deprivation on H2O2 detoxification

With the model developed and an ensemble of viable parameter sets identified, we sought to assess its predictive capabilities on a physiologically-relevant environmental perturbation. There is growing evidence that microbial killing within macrophages is a combined effect of the toxic environment and a scarcity of nutrients [4749]. We therefore chose to investigate how H2O2 detoxification changes during carbon starvation. In the absence of an exogenous source of energy and carbon, the abundance of reducing equivalents can fall to limiting levels [50] and energetic processes such as translation can be hampered [51]. These effects could impact the stress response network by limiting AHP activity and inhibiting H2O2-dependent induction of AHP and HPI. To determine if carbon starvation in the media used here leads to NADH depletion, we directly measured NADH and NAD+ in M9 media cultures with and without glucose and found that a significant reduction in NADH occurred in carbon-starved cultures (S7 Fig). To see if carbon starvation depresses NADH to levels that inhibit enzyme activities, we measured respiration, which is an NADH-driven process, in M9 media in the presence and absence of glucose and found it to be significantly impaired when glucose was omitted (S8 Fig). In addition, to see if a lack of carbon reduces NADH to levels that impair AHP activity, we monitored clearance of 10 μM H2O2 in a strain with AHP as the lone major detoxification enzyme (ΔkatE ΔkatG), and found that omission of glucose completely inhibited H2O2 clearance in this strain (S9 Fig). In accordance with these results, model predictions indicated that if NADH was not held constant, AHP would drain it from the system in less than a second (S10 Fig). The impact of glucose starvation on induction of AHP and HPI expression was also assessed with the use of GFP reporter plasmids. Omission of glucose completely inhibited H2O2-dependent induction (S11 Fig). Therefore, to simulate the impact of carbon deprivation, NADH concentrations were no longer held constant and protein production was set to zero.

Ensemble predictions for carbon deprivation (- glucose) were made using the complete reaction network and are shown in Fig 6A–6D, along with the carbon-replete control (+ glucose). These predictions were experimentally confirmed and the data are presented in Fig 6I–6L, orange). To quantify how the different elements of glucose deprivation (NADH limitation, inhibition of translation) contributed to the observed phenotypes, we investigated the individual effects of NADH depletion or translation inhibition with simulation controls (Fig 6E–6H). At lower treatment concentrations (10 and 25 μM H2O2), starvation was predicted to slow detoxification as a result of NADH depletion, whereas inhibition of protein synthesis was predicted to have a negligible effect. At 100 μM H2O2, glucose-deprived cultures were predicted to clear H2O2 comparably to glucose-fed cultures, with neither reducing equivalent availability nor enzyme production substantially hindering detoxification. The impact of starvation at 400 μM H2O2 was predicted to be largely mediated by translation. We note that although selective inhibition of NADH production and usage was not feasible due to the wide variety of sources and sinks, targeted inhibition of translation was experimentally tractable with the use chloramphenicol (CAM) (S11 Fig). Experimental confirmation of clearance by CAM-treated cultures is shown in Fig 6I–6K. Unfortunately, CAM-treatment led to cell death at 400 μM H2O2 (S1X Fig), which prevented direct confirmation of the prediction at that concentration. Interestingly, this cell death suggested that translation of some protein other than AHP or HPI is important to survival at 400 μM H2O2, because we demonstrated that carbon deprivation inhibited induction of AHP and HPI at 400 μM H2O2 (S11 Fig), and it is known that carbon deprivation can have promoter-specific effects [51] and CAM stops synthesis of all proteins.

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.


The toxic nature of H2O2 makes it an ideal weapon in inter-species warfare, and it is used as such by the immune system during infection [2] and even by other bacteria in niche competitions [15]. Bacteria have evolved numerous defense systems, which can differ significantly in their substrate requirements, reaction mechanisms, and regulation, and the complexity of these defense networks and the broad reactivity of H2O2 necessitate the use of computational modeling for quantitative interpretation and prediction of H2O2 distributions in cells [52]. Due to its importance as a signaling molecule, models of H2O2 metabolism in mammalian systems have been constructed, and they have included enzymatic detoxification of H2O2 [2224], oxidation of cysteine residues [22], transport of H2O2 across membranes [2224], and oxidation of targets involved in signaling [24]. However, beyond their specificity for mammalian systems, none have accounted for uncertainty in optimized parameters or included synthesis or inactivation of enzymes, side reactions of H2O2, or other reactive oxygen species present in the network. In bacteria, modeling efforts have focused on subsystems affected by H2O2 stress, such as that of the thioredoxin system in E. coli, which included the oxidation of thioredoxin and the reduction of oxidized thioredoxin, methionine sulfoxide, protein disulfides, and 3’-phosphoadenosine-5’-phosphosulfate [25]. These previous efforts inspired us to construct a quantitative, systems-level model of H2O2 stress in E. coli that includes media and cellular compartment-specific species and reactions; H2O2-dependent transcriptional regulation, inactivation, and activity of H2O2 detoxification enzymes; reductases to reduce oxidized species; O2• and •OH and their related reactions (e.g., oxidation of all twenty amino acids by •OH); and reactions of H2O2 with other metabolites such as glutathione and the α-keto acid pyruvate. In addition, we addressed structural uncertainty in the model using an iterative computational and experimental methodology, and assessed parametric uncertainty using an MCMC procedure, which enabled the robustness of model predictions to be assessed. Similar ensemble approaches have become popular methods to account for parametric uncertainty [5362], and several techniques have been developed to efficiently explore parameter spaces [26, 46].

One power of quantitative computational modeling is its ability to predict emergent systems behavior [33, 44, 6365]. For instance, Schaber and colleagues used an ensemble of possible models describing different hypotheses regarding the mechanism of the high osmolarity glycerol (HOG) pathway in yeast to uncover novel features of the pathway [44]. Here, we leveraged our model to gain a quantitative understanding of how carbon deprivation, which bacteria encounter in phagocytes [4749], affects H2O2 detoxification by E. coli. Accounting for the NADH limitation and translational inhibition that occurs with carbon source starvation, our simulations were able to correctly predict H2O2 clearance dynamics. Upon dissection of simulation results, delayed clearance at lower concentrations was attributed to reduced AHP activity from NADH depletion, whereas at higher concentrations carbon-starved cultures resembled carbon-replete cultures because pre-expressed HPI dominated H2O2 detoxification, suggesting that both NADH availability and induction of AHP and HPI synthesis at H2O2 concentrations > 25μM were of minor importance. We note that changes in concentrations of other metabolites, such as ATP, occur in carbon-starved cultures [66], and that they were not explicitly accounted for here because they did not directly act as a substrate in any of the reactions of the model. Rather we anticipate that some of the metabolite perturbations were implicitly accounted for because they contributed to the inhibition of translation and/or depletion of NADH that were included. These data demonstrate that the nutritional status of the environment can have a major impact on bacterial H2O2 defenses, but the extent of that impact depends on the quantitative level of H2O2. Beyond carbon deprivation, it would be interesting to see how other types of starvation (e.g., sulfate, iron) influence H2O2 detoxification, since bacteria are subjected to oxidative stress in various scenarios [1, 15, 67], and we expect that dependencies distinct from those of carbon source starvation could be observed if other types of limitation influence NADH availability and translation differently.

In immune cells, phagocytized bacteria are exposed to ROS with an oxidative “burst” from NADPH oxidase, which then tapers over time [6870]. In this work, we examined detoxification of a burst of H2O2 with bolus treatments, and note that more complex treatment dynamics could be handled by the model developed here. For instance, the model could be adjusted for continuous treatment by adding an H2O2 delivery reaction, which could be achieved experimentally with a fed-batch reactor. Alternatively, H2O2 could be provided through indirect means, such as with exposure to redox-cycling agents like paraquat [71]; though obtaining accurate estimates of H2O2 production from such compounds could be a challenge, because generation would be cell-dependent. Different H2O2 delivery dynamics could have a profound impact on the kinetic competition for H2O2, and as long as H2O2 influx can be accurately accounted for the model developed could prove invaluable for interrogating its distribution.

One area of growth for the platform we developed is adaptation of the model to allow for analysis of lethal H2O2 concentrations. In its current form, the size of the cellular compartment is fixed and enzymatic reactions do not occur in the media compartment. To model lethal concentrations of H2O2, cell lysis has to be accounted for in terms of reduction in the volume and surface area of the cellular compartment and the addition of certain enzymatic activities to the media compartment, such as catalase, which can function when released from cells. In addition, it might be necessary to diversify the cellular compartment if all cells that die do not lyse but contain compromised translational and/or catabolic activities. Despite this added complexity, the ability to accurately simulate lethal damage, such as that involving DNA and the membrane, could provide insight into H2O2-induced death.

Models akin to the one described here will improve understanding of bacterial defenses against host immune responses, and possibly suggest targets for novel anti-virulence therapies [52]. For example, an existing model of nitric oxide (NO•) defenses in E. coli [33] has provided valuable predictions regarding NO• delivery rates that maximize antimicrobial activity [63]. Additionally, it provided a framework that allowed for a model-guided investigation of the underlying mechanism of an NO•-sensitizing mutation, ΔclpP, in E. coli [72]. The success of the NO• model provided inspiration to develop a similar model for H2O2, which is another toxic, diffusible metabolite used by immune cells when fighting infection [1, 2]. We anticipate that the H2O2 model developed here will yield novel quantitative insight into the kinetic competition for H2O2 in E. coli and provide a framework for the mechanistic investigation of perturbations that affect clearance, while illuminating targets to sensitize bacteria to immune attack.

Materials and Methods

Bacterial strains

E. coli K-12 MG1655 was used in all experiments. ΔkatE and ΔkatG mutations were transduced into MG1655 from their respective strains in the Keio collection [73] by the P1 phage method. The ΔahpCF mutation was generously provided by Michael Kohanski and transduced into MG1655 using the P1 phage method. All antibiotic markers were cured out using pCP20 [74]. The known antioxidant pyruvate (25 mM) was included in all LB agar plates to prevent a toxic build-up of H2O2 [75]. Deletions were PCR checked for proper chromosomal integration with a forward primer external to the gene and reverse primer within the kanamycin resistance cassette (kanR) before curing. Internal primers were used to check for gene duplication. In cured strains, external primers before and after the gene were used to check for proper scar size. All PCR primers are listed in S4 Table.

Measurement of H2O2 clearance

Overnight cultures were inoculated from −80°C stocks and grown for 20 hours in 1 mL LB + 30–75 U/mL catalase (bovine liver catalase at 2,000–5,000 units/mg protein: Sigma Aldrich), then used to inoculate 20 mL M9 10 mM glucose medium + 30–75 U/mL catalase to an OD600 of 0.01 in 250 mL baffled flasks. Catalase was added to prevent the possibility of H2O2 accumulation in strains lacking major detoxifying enzymes, and added to wild-type cultures to maintain consistency across strains. The catalase concentration was chosen based on the amount required to maintain growth in a mutant lacking all major detoxification enzymes (ΔkatE ΔkatG ΔahpCF), beyond which increasing the catalase concentration no longer increased growth rate or terminal cell density in the case of overnights. Cultures were grown at 37°C with shaking at 250 rpm for 8 h (OD600 0.3–0.6, 0.15 for the slower growing ΔkatE ΔkatG ΔahpCF). After the 8 hour growth period, 12 mL of culture was removed to a pre-warmed 15 mL Falcon tube and centrifuged at 37°C and 4,000 rpm for 10 min. 10.8 mL of spent media was removed, the cell pellet resuspended, and 1 mL transferred to a warm 1.5 mL microcentrifuge tube. Cells were washed a total of four times to remove all catalase. Washes consisted of spinning down at 14,000 rpm for 2 min, removing 980 μL of media, and resuspending the cell pellet with 980 μL fresh warm media. For samples lacking glucose during challenge with H2O2, glucose was omitted during the final wash step and in the inoculated flask. For CAM treatment assays, all wash steps were performed with 100 μg/mL CAM.

Prior to inoculation with washed cells, 20 mL fresh M9 10 mM glucose media in 250 mL baffled flasks were warmed to 37°C. A bolus of 10, 25, 100, or 400 μM H2O2 was added to the flasks, and the time 0 point was measured, after which flasks were inoculated to an OD600 of 0.01. At desired time points, 200 μL was removed to a 1.5 mL microcentrifuge tube and centrifuged at 15,000 for 3 min. 150 μL of the supernatant was moved to a sterile microcentrifuge tube and stored at 4°C until H2O2 concentration could be measured. Samples were assayed for H2O2 within 2 h of harvesting. H2O2 in the supernatant was measured using the Amplex Red Hydrogen Peroxide/Peroxidase kit (Life Technologies) according to the manufacturer’s instructions after dilution to below 10 μM H2O2. A standard curve spanning 0 to 10 μM H2O2 was used to calculate H2O2 concentrations. A fresh standard curve was produced for each Amplex Red assay to account for increasing background fluorescence over the course of the day due to the sensitivity of Amplex Red to both light and air [76].

To assess whether centrifuging to remove cells was sufficient, we compared values from this method to sterile filtering (0.22 μm pore size) of samples, as well as centrifuging + filtering, for the 30 min point in the 400 μM H2O2 clearance assay (S12 Fig). Centrifuging alone was not significantly different from filtering (p = 0.45) or centrifuging + filtering (p = 0.40) based on a two-sample t-test with unequal variance. To determine if the exogenous catalase that was added to the pre-culture steps affected clearance profiles, we performed identical experiments for wild-type propagated without catalase in all steps (S13 Fig). The presence of catalase in the pre-culture did not affect wild-type clearance of H2O2 at any concentration.

Measurement of colony forming units (CFUs)

To determine whether H2O2 treatment resulted in cell death, we quantified CFUs throughout the clearance assays. After isolating the H2O2-containing supernatant for Amplex Red assays as described above, an additional 30 μL supernatant was removed from centrifuged samples and discarded, to achieve a greater fold-dilution of H2O2 during the first wash step. In the first wash step, 980 μL of PBS was added and the cell pellet was resuspended. Samples were centrifuged again at 15,000 rpm for 3 min, 980 μL of the supernatant was removed, and the cell pellet was resuspended a final time in 80 μL PBS. Plating was performed using the serial dilution method, and samples were plated on LB agar containing 25 mM pyruvate to scavenge any residual H2O2 remaining in the pellet and any endogenously produced H2O2 in scavenging-deficient strains. Plates were incubated for 16 h at 37°C prior to counting colonies.

Oxygen measurement

MG1655 cultures were grown and washed identically to the H2O2 clearance assay. After the final wash, the resuspended cells were used to inoculate 10 mL of pre-warmed M9 with or without glucose in a 50 mL Falcon tube containing a sterile magnetic stirring bar, immersed in a stirred water bath at 37°C, to an OD600 of 0.1. Cells were allowed to consume oxygen for ten minutes before being treated with 5 mM KCN to halt respiration, which consumes the majority of O2 in E. coli cell cultures under these conditions [63]. The percent oxygen saturation was measured at a frequency of one reading per second using the FireStingO2 fiber-optic O2 meter with the OXROB10-CL2 robust oxygen miniprobe (PyroScience, GmbH). Temperature fluctuations were compensated for using the TDIP15 temperature sensor (PyroScience GmbH) and the FireSting Logger Software. The equilibrium oxygen concentration was used to convert the percent saturation to concentration, and was determined by calibrating the probe in ultrapure Milli-Q water at 37°C, which has an oxygen concentration of 210 μM [77], and transferring the probe to air-saturated M9 media. The equilibrium concentration of the media matched that of the ultrapure water.

NAD+/NADH measurement

Overnight cultures were inoculated and grown identically to the H2O2 clearance assay, and used to inoculate 20 mL M9 10 mM glucose medium + 30–75 U/mL catalase to an OD600 of 0.01 in 250 mL baffled flasks. Cultures were grown at 37°C with shaking at 250 rpm to an OD600 of 0.2 (~6.5 h). Four 1 mL aliquots were transferred from the flask to warm, 1.5 mL microcentrifuge tubes and centrifuged at 15,000 rpm for 3 min. The media was removed, and the pellets were resuspended with 1 mL fresh M9 with or without 10 mM glucose and transferred to warm test tubes. The tubes were then incubated at 37°C with shaking at 250 rpm for 60 min. The time 0− point was taken directly from the flask prior to centrifuging and resuspension.

NAD+ and NADH were measured using the EnzyChrom NAD/NADH Assay Kit (BioAssay Systems) following the manufacturer’s protocol, except for a brief sonication step. For each measurement, 400 μL (NAD+) or 800 μL (NADH) of cell culture was transferred from the flask (time 0) or test tube (60 min) to a 1.5 mL microcentrifuge tube. The tubes were centrifuged for 3 min at 15,000 rpm, and 380 μL (NAD+) or 780 μL (NADH) of supernatant was removed and discarded. The cell pellets were resuspended in 100 μL of either NAD or NADH extraction buffer and sonicated for 20 s at room temperature at an amplitude of 10 using a Fisher Scientific Model 50 Sonic Dismembrator. The extracts were heated at 60°C for 5 min, before adding 20 μL of assay buffer and 100 μL of the opposite extraction buffer. Samples were then vortexed briefly and centrifuged for 5 min at 14,000 rpm. The supernatants were used for the NAD/NADH assay following manufacturer’s instructions. An NAD+ standard curve from 0–2 μM was generated each day and used to convert absorbance to concentration. The standard curve underwent extraction protocols identical to cell samples, including sonication and heating. Since NAD+ and NADH produce identical standard curves, and NADH is more unstable, only NAD+ was provided in the kit and was used to convert both NAD+ and NADH absorbance to concentration.

Reporter assays

MG1655 was transformed with pUA66 PahpC-gfp, pUA66 PkatG-gfp, and pUA66 PkatE-gfp, which were all obtained from a pre-existing library [78]. Overnight cultures were inoculated from −80°C stocks and grown for 20 h in 1 mL LB + 30–75 U/mL catalase + 30 μg/ml kanamycin for plasmid retention, then used to inoculate 20 mL M9 10 mM glucose medium + 30–75 U/mL catalase + 30 μg/ml kanamycin to an OD600 of 0.01 in 250 mL baffled flasks. Cultures were grown, washed, and treated with H2O2 identically to the protocol described in the Amplex Red assay protocol. Washed cells were fixed before inoculation to the H2O2-containing flasks to provide a time 0− sample for each condition. Final points were sampled and fixed when ~90% of the H2O2 had been cleared by wild-type in 10 mM glucose M9 media: 30 min for the 10 μM H2O2 flask, 40 min for the 25 μM H2O2 flask, 1 h 15 min for the 100 μM H2O2 flask, and 2 h for the 400 μM H2O2 samples. Fixing involved removing 1 mL culture to a microcentrifuge tube and centrifuging at 15,000 rpm for 3 min, removing 980 μL supernatant, and resuspending with 480 μL 4% paraformaldehyde (PFA). After 25 min at room temperature, the samples were again centrifuged at 15,000 rpm for 3 min, 480 μL of the PFA was removed, and the pellet was resuspended with 980 μL 1X PBS. Samples were stored at 4°C until analysis by flow cytometry on an LSR II flow cytometer (BD Biosciences, San Jose, CA), where green fluorescence was measured on a per cell basis. Fluorescence was measured using 488 nm excitation and a 525/20 bandpass filter, and data were acquired using FACSDiVa software (BD Biosciences, San Jose, CA).

Model framework

The modeling framework used in this work largely followed that used by Robinson and Brynildsen [33]. It is composed of a system of ordinary differential equations that are numerically integrated to provide predicted species concentration over time. We begin with a mole balance for all species in the model: (1) where N is an s x 1 vector representing the total amount of given species in moles, S is the s x r stoichiometric matrix, and v is an r x 1 vector representing reaction rates in moles per time. Here, s indicates the number of species in the model, and r indicates the number of reactions in the model. Since most reaction rates are calculated on the basis of concentration per time, the rate vector was converted into these units in the following manner. (2) where Vrxn is an r x r diagonal matrix representing the volumes of the compartments in which the reactions are taking place: Vcell for intracellular reactions, Vmedia for reactions taking place only in the media, and Vtotal for exchange reactions. Due to experiments also being performed on a concentration basis, N was converted into units of concentration by performing the following operation. (3) where Vspec is a diagonal s x s matrix of species compartment volumes: Vcell for intracellular species, Vmedia for species in the media compartment, and Vtotal for species that freely diffuse across the cell membrane (e.g., O2, H2O2 in non-gradient models). The left-hand side of the equation is equivalent to dC/dt when the volume does not vary appreciably over the course of the experiment. To avoid having a culture-volume-specific model, we transformed the volume dependencies into volume fractions by multiplying and dividing by Vtotal (Vcell/Vtotal, Vmedia/Vtotal, Vtotal/Vtotal) and rearranging the equation with the use of the commutative property of scalar multiplication. (4) (5) (6) where Fspec is an s x s diagonal matrix of the volume fractions for species and Frxn is an r x r diagonal matrix of the volume fractions for reactions. By making this adjustment, we avoid the need of requiring total culture volume as an input into the model, and simplify the input to optical density (OD600) that can readily be converted to volume fractions.

Initial species concentrations

Most initial species concentrations were obtained from literature (S1 Table)[21, 25, 7990]. The equilibrium concentration of oxygen was determined by calibrating a FireStingO2 fiber-optic O2 meter with the OXROB10-CL2 robust oxygen miniprobe (PyroScience, GmbH) in ultrapure Milli-Q water at 37°C and transferring it to air-saturated M9 10 mM glucose medium at 37°C. We calculated the concentration in our media by comparing it to the known value in deionized water [77], which is 210 μM. The value in our media was equivalent. The initial H2O2 concentration was set to the initial average value of the data when optimizing parameters (e.g., 25.89 μM instead of 25 μM for wild-type) to avoid penalizing model fit for experimental error. When making forward predictions, the concentration was set to the anticipated initial concentration (e.g., exactly 25 μM).

Initial species that were trained on experimental data included AHP, HPI, HPII, Fe2+, and Fe3+. While experimental measurements on AHP [21], HPI [82], and HPII [82] are available, their concentrations vary with environment and growth phase, as shown in our reporter assays (S11 Fig). AHP, HPI, and HPII are the major H2O2 detoxification systems in E. coli. We therefore allowed flexibility in their initial concentrations, constraining them to be within the range of 0–20 μM. Because individual concentrations of Fe2+ and Fe3+ are unresolved, but experimentally measured to have a combined concentration of 10 μM, we allowed their initial concentrations to vary from 0 to 10 μM.

HPI and HPII reaction mechanism

Catalase activity follows a ping-pong mechanism, reacting with one H2O2 to form a reactive intermediate, followed by reaction with a second H2O2 molecule to return the enzyme to its original form [91, 92]. Given that the substrate in the first and second reaction is the same, the rate equation simplifies to a Michaelis-Menten type structure. While inhibition of catalase activity by H2O2 becomes apparent at high concentrations (e.g., greater than 100 mM H2O2 for E. coli HPII [92]), it is assumed negligible at the concentrations used in our experiments, and Michaelis-Menten kinetics are appropriate. Rate equations and constants can be found in S3 Table (Reactions 69 and 70). While HPI has the ability to utilize other reducing agents at low H2O2 concentrations, this activity is significantly slower than its catalase activity (about 1% the kcat of its catalase activity [21]), so the peroxidase activity was assumed to negligibly contribute to H2O2 detoxification in this study.

AHP reaction mechanism

The AHP reaction cycle begins when the peroxidatic cysteine of AhpC reacts with H2O2 to form a sulfenic acid, which resolves to form a disulfide bond with another cysteine residue. The active AhpC is regenerated by its reductase partner AhpF, which uses NADH as an electron donor [38]. We modeled this cycle using ping-pong reaction kinetics, with H2O2 and NADH as the substrates, a structure which has been used previously [38]. Kinetic parameters were not available for E. coli, but a protein BLAST search [93] revealed 98% protein sequence identity for AhpC and 95% identity for AhpF between E. coli MG1655 and S. Typhi, so available parameters for S. Typhi were used [38, 94]. Additional information and rate constants can be found in S3 Table (Reaction 71).

Expression of detoxification enzymes

In this work, the main experimental variable was the concentration of H2O2, and therefore, we opted for simplicity and only H2O2-dependent regulation of gene expression was considered. The expression of catalase HPI and AHP increase in response to H2O2 [17, 95], whereas the expression of catalase HPII is not dependent on H2O2 [17, 18]. These dependencies were confirmed using transcriptional reporters for ahpC, katE, and katG, and measuring fluorescence on a per cell basis using flow cytometry (S11 Fig). Following previous dynamic models, gene expression was modeled using a Hill equation with a coefficient of n = 1 [33, 34], except for HPII which had initial concentration but was not expressed further. In addition, transcription was assumed to be limiting in the production of active enzyme, as assumed previously [33, 34], and the bioavailability of ferroheme b, which is an essential cofactor of HPI, was assumed to not be rate limiting. HPI and AHP are expressed according to Reactions 67 and 68, respectively, in S3 Table. The maximum expression rate and Hill equation constants KAHP-exp,H2O2 and KHPI-exp,H2O2 are informed during parameter optimization. The bounds on the maximum expression rates are based on the highest and lowest maximum expression rates found in the work of Kotte and colleagues [34], and have been used previously when optimizing unknown expression rates [33]. Bounds on KAHP-exp,H2O2 and KHPI-exp,H2O2 were approximated by the work of Kotte and colleagues [34], which varied from approximately 2 nM to 1 mM. Here, we allowed variation from 0 to 1 mM. We note that while the parsimonious treatment of H2O2-dependent expression was sufficient in this work, exploration of new environments could necessitate a more comprehensive modeling of transcriptional regulation, since expression of the detoxification enzymes in this work are known to depend on numerous regulators (e.g., OxyR, RpoS, FIS, Fur) [96].

Modeling of enzyme degradation

In previous studies, enzymes have typically been modeled as undergoing first order degradation with a universal constant [33, 34]. However, rates can vary greatly among different proteins [35, 36], and some evidence suggests that H2O2 detoxification enzymes, such as catalase and alkyl hydroperoxidase, can be poisoned by their own substrate [3739]. Inactivation of catalase has been described using bimolecular kinetics [40], but it has also been suggested that poisoning should be of the same general form as the enzyme’s reaction kinetics [37]. The AhpC component of alkyl hydroperoxidase is reduced by AhpF after oxidation by H2O2. If there is not enough NADH present or AhpC reacts with more H2O2 before encountering AhpF, the cysteine sulfenic acid formed by the first oxidation can be further oxidized to sulfinic or sulfonic acid, rendering it inactive [38, 97]. Whether AhpC inactivation is significant at the concentrations used in this work was uncertain [39]. We therefore allowed it to be degraded in a first order or bimolecular manner. Due to the indeterminacy of how these enzymes are degraded, we included all of these possible degradation scenarios. Rate equations can be found in S2 Table. We note that since AHP requires a co-substrate, NADH, and that we assume that NADH is constant (unless otherwise noted), if AHP inactivation were found to be important it could reflect degradation of AHP or reduced availability of NADH (Reaction 71).

Parameters or bounds on parameters for enzyme degradation/deactivation rate equations varied with the method of degradation. For constant degradation with a constrained constant, we used the "general" protein degradation rate reported by Kotte and colleagues [34]. When optimized, the constant degradation rate had bounds set by the longest [35] and shortest [36] protein half-life we found in literature. For both bimolecular and the more complex inactivation, bounds were based loosely on rate constants found for Aspergillus niger and bovine catalase [37, 40]. Based on the gross difference between the organisms tested in those studies (fungus and mammal) and our own (bacteria), as well as the orders of magnitude difference between the Aspergillus niger and bovine catalase rates in both the bimolecular and more complex kinetics studies, these parameters were constrained to two orders of magnitude lower than the lowest reported value, and two orders of magnitude higher than the highest reported value.

Modeling of H2O2 gradient across the membrane

While H2O2 rapidly diffuses across bacterial membranes at a rate similar to that of water, there is evidence for and against the existence of a gradient across the membrane [41, 42]. For example, an Ahp−Kat+ strain cocultured with Ahp−Kat− in the presence of a low H2O2 concentration can outcompete its scavenging deficient neighbors and multiply under the stress, whereas the deficient strain only grows after the catalase proficient strain has cleared the environmental H2O2 [41]. On the other hand, dilute suspensions of catalase proficient strains are readily killed by high concentrations of H2O2 similarly to scavenging deficient strains, while high-density catalase proficient strains can not only survive challenge with H2O2 but also protect deficient neighbors [42]. There have been strides in the ability to measure H2O2 intracellularly, with the introduction of genetically-encoded indicators (HyPer) [98] and the ability to use Amplex Red intracellularly with expression of a mutated ascorbate peroxidase [99]. However, the dependence of HyPer fluorescence on reductase activity, the impact of HyPer on cellular scavenging capacity [100], and the difficulties associated with converting measurements from either method to absolute H2O2 concentrations led us to use measurements of external H2O2 and statistical metrics (AIC) to assess the suitability of modeling the system with or without a gradient. In one set of models, the intracellular and extracellular H2O2 concentrations were equal. In the other set, we allowed for a gradient by modeling transport across the membrane as a convective mass transport process, with the effective mass transport coefficient being an additional parameter for optimization. The lower bound on the effective mass transfer coefficient was set as the permeability coefficient of H2O2 across E. coli cell membranes in unstirred culture [41], adjusted for cell area and the cell density for our system. The upper bound was set two orders of magnitude higher than the permeability, to account for increased mass transfer in our chaotic shake flask system.

Spontaneous degradation of H2O2

The rate of spontaneous degradation of H2O2 into H2O and O2 was determined using the MATLAB function lsqcurvefit after monitoring H2O2 concentration over time in cell-free controls for each media condition (M9 10 mM glucose with and without CAM and M9 lacking glucose). Samples were collected at time 0, 20 min, 40 min, and 60 min for 10 and 25 μM H2O2-containing flasks; time 0, 1 h, 2 h, and 3 h for 100 μM H2O2-containing flasks; and 0 h, 1 h, 2 h, 3 h, and 4 h for 400 μM H2O2-containing flasks. All control data for each media condition was fit simultaneously (e.g., 10, 25, 100, and 400 μM H2O2 in M9 10 mM glucose) while accounting for experimental error by weighting points in the sum squared of residuals (SSR) calculation by the inverse of the variance for that data point [43, 101103]. The optimized rate constants were 0.0324 h-1 for M9 10 mM glucose, 0.0331 h-1 for M9 10 mM glucose with 100 μg/mL CAM, and 0 h-1 for M9 lacking glucose. The spontaneous degradation rate was assumed to be equivalent in the intracellular and media compartments.

Other reactions

Other reactive oxygen species in the reaction network include O2-• and •OH. The model includes endogenous production of O2-• (Reaction 4) and its dismutation to H2O2 (Reactions 3, 74–75), as well as its reactions with other molecules (Reactions 1, 25, 26, 55) and production by other reactions in the network (Reaction 49). Additionally, the Fenton reaction produces •OH (Reaction 24), which can react with amino acids (Reactions 5–23) and other compounds in the network (Reactions 1, 27, 36, 37, 48). The rate constant for the Fenton reaction were variable in literature, so bounds were set as the slowest and fastest reported rates [21].

Other reactions include glutathione oxidation, reduction, and reaction with other molecules (Reactions 29–32, 42–55), and oxidation of methionine (Reaction 2) and pyruvate (Reaction 57) by H2O2. All spontaneous reactions, rate equations, and rate constant can be found in S2 Table [21, 3437, 3941, 81, 85, 88, 91, 92, 104111]. Enzymatic rate equations and rate constants can be found in S3 Table [38, 91, 92, 94, 111117], and include methionine sulfoxide reductase (Reaction 64), thiol peroxidases (Reaction 65–66), thioredoxin reductase (Reaction 72), and glutathione reductase (Reaction 73).

Parameter optimization

Parameters were optimized by minimizing the SSR using the built-in MATLAB function lsqcurvefit. Because experimental error varied for each time point, we weighted each data point’s contribution to the SSR by the inverse of the variance of that point [43, 101103]. The initial concentration of H2O2 in the model was changed to match the experimental data before optimizing. Due to the nonlinear nature of the optimization, each model structure was initialized with random parameter sets within the defined bounds a total of 1,000 times. The progression of minimum SSR found by each of these iterations is shown in S14 Fig. A plateau suggests that additional iterations would not lead to substantial improvement in the fit. The slowest converging optimization reached a plateau after 639 optimizations.

The number of parameters optimized varied according to model structure, as presented in Table 1. Parameters related to HPI and AHP expression (4 parameters total) were optimized in all structures. The Fenton reaction rate constant varied in literature (1 parameter), and intracellular Fe2+ and Fe3+ concentrations (2 parameters) were unresolved. Additionally, initial concentrations of the major detoxifying enzymes (3 parameters) were unknown and can vary with growth environment and stage of cell growth. Beyond these 10 parameters, all structures that did not have constant enzyme degradation with a universal degradation constant added 3 parameters, and including an H2O2 gradient added 1 parameter (a convective mass transfer coefficient).

Model ranking and selection

The introduction of additional parameters, such as enzyme degradation rate constants and mass transport coefficients, has the potential to improve fit solely by increasing the flexibility of the model. To account for the utility of additional parameters, we ranked models based on their evidence ratios (ER), or likeliness relative to the most likely model in the set. For each model, we calculated its Akaike Information Criterion corrected for small sample size (AICc) [43]: (7) where n represents the sample size and K is the number of estimable parameters. Here, n is the number of data points used in the fitting procedure, and K is the number of model parameters plus 1 because regression estimates SSR and parameter values [43]. We account for unequal variances within the data by using weighted least squares, where each point is weighted by the inverse of its variance.

The weight of evidence for a given model in a set of M models is given by the following [43]: (8) where Δi = AICi-min(AIC). With this, an ER can be calculated, which represents the relative likelihood of a model compared to the best model in the set [27]: (9)

A larger ER indicates a more unlikely model. In this work, models with an ER greater than 10 were discarded, a cutoff that has been used previously to discard models during model selection [29].

Generation of the ensemble

To account for parametric uncertainty when making predictions, we generated an ensemble of parameter sets that all predicted the data within an ER≤10. We initially attempted to use the software HYPERSPACE [26], which is a three-step process that provides a uniform sampling of the viable parameter space. However, the calculation of our cost function was computationally expensive, which lengthened the time required per iteration, and the method had not converged within 100,000 iterations. Therefore, we utilized the pre-existing Markov chain Monte Carlo function within the HYPERSPACE software, but started it from all 40 viable parameter sets that met our design criteria. We allowed approximately 200 random steps away from each viable point, and randomly selected 100 of those parameters sets with an ER≤10 from each random walk. This process generated 4,000 parameter sets that had an ER≤10.

Identification of the minimal H2O2 model

We identified the minimal model that was able to capture our data (ER≤10) using a previously developed two tiered approach [33]. In the first tier, reactions were removed from the best model in a random order and the ER was calculated. If the deletion of a reaction increased the ER above its threshold of 10, it was returned to the model and the process continued through the remaining reactions. This random deletion process was repeated 100 times. In the second tier, parameters were re-optimized after the deletion of each of the remaining reactions. If the optimization produced a model that returned below an ER of 10, the parameters were changed and the process continued.

Supporting Information

S1 Fig. CFUs/mL during clearance assays.

A-D. CFUs/mL during wild-type clearance of 10 (A), 25 (B), 100 (C), and 400 (D) μM H2O2 in M9 10 mM glucose media. E-H. CFUs/mL during ΔkatE ΔkatG clearance of 10 (E), 25 (F), 100 (G), and 400 (H) μM H2O2 in M9 10 mM glucose media. I-L. CFUs/mL during ΔkatE clearance of 10 (I), 25 (J), 100 (K), and 400 (L) μM H2O2 in M9 10 mM glucose media. M-P. CFUs/mL during ΔkatG clearance of 10 (M), 25 (N), 100 (O), and 400 (P) μM H2O2 in M9 10 mM glucose media. Q-T. CFUs/mL during wild-type clearance of 10 (Q), 25 (R), 100 (S), and 400 (T) μM H2O2 in M9 media lacking glucose. U-X. CFUs/mL during wild-type clearance of 10 (U), 25 (V), 100 (W), and 400 (X) μM H2O2 in M9 10 mM glucose media with 100 μg/mL CAM. Y. CFUs/mL during ΔkatE ΔkatG clearance of 10 μM H2O2 in M9 media lacking glucose. Z. CFUs/mL during ΔkatE ΔkatG clearance of 10 μM H2O2 in M9 10 mM glucose media with 100 μg/mL CAM. AA. CFUs/mL during ΔahpCF ΔkatE ΔkatG clearance of 10 μM H2O2 in M9 10 mM glucose media. Experiments were performed with three biological replicates. Error bars show the standard error of the mean. Asterisks indicate significant (p<0.05) CFU loss from the initial value based on a two-tailed t-test with unequal variance performed on log-transformed values.


S2 Fig. Reaction flux through AHP and HPI+HPII.

Reaction flux through the two major detoxification systems AHP vs. HPI+HPII are shown as a function of time. A-D. Reaction fluxes for the 35 acceptable models after fitting on wild-type data (Fig 3). E-H. Reaction fluxes for the 965 acceptable models after fitting simultaneously on wild-type and ΔkatE ΔkatG data (Fig 4). I-L. Reaction fluxes for the 40 acceptable models after fitting on wild-type, ΔkatE ΔkatG, ΔkatE, and ΔkatG data (Fig 5). Each line represents the prediction from a single model.


S3 Fig. Reaction flux through HPI and HPII.

Reaction flux through the two catalases HPI and HPII are shown as a function of time. A-D. Reaction fluxes for the 965 acceptable models after fitting simultaneously on wild-type and ΔkatE ΔkatG data (Fig 4). E-H. Reaction fluxes for the 40 acceptable models after fitting on wild-type, ΔkatE ΔkatG, ΔkatE, and ΔkatG data (Fig 5). Each line represents the prediction from a single model.


S4 Fig. Prediction for H2O2 clearance by ΔkatE and ΔkatG.

Predicted clearance of 10 (A), 25 (B), 100 (C), and 400 (D) μM H2O2 by ΔkatE, and 10 (E), 25 (F), 100 (G), and 400 (H) μM H2O2 by ΔkatG in M9 10 mM glucose media. Each line represents the prediction from one of the 965 acceptable models trained on wild-type and ΔkatE ΔkatG H2O2 clearance in M9 10 mM glucose media (Fig 4). Wide distributions on clearance dynamics suggest that these single mutants could be used to discriminate between models.


S5 Fig. Ensemble consistency.

To ensure that none of the models in our ensemble violated the design criteria, we checked the consistency of predictions for H2O2 distribution across the detoxification pathways for the 4,000 model set. A-D. Prediction for the amount of H2O2 cleared by the two major detoxification pathways AHP (orange) and combined catalase activity (black) after boluses of 10 (A), 25 (B), 100 (C), and 400 (D) 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 (E), 25 (F), 100 (G), and 400 (H) μM H2O2. Each line represents the prediction from a single model.


S6 Fig. Parameter sensitivity analysis.

Beginning from the best parameter set in our ensemble, parameters were varied between their bounds. Parameters that increased the ER to beyond our threshold of 10 are shown in the figure. The Fenton reaction rate constant and Fe2+ and Fe3+ initial concentrations did not substantially affect the ER.


S7 Fig. [NAD+] and [NADH] dependence on glucose availability.

Exponentially growing cells were transferred to fresh M9 10 mM glucose or M9 lacking carbon. Time 0- points were measured before resuspension in fresh media. Data represents the average of four biological replicates, and error bars show the standard error of the mean. Cells have a significantly lower NADH level after 60 minutes in carbon-free media (p = 0.035), as determined by a two-tailed t-test with unequal variance. A higher cell density (OD600 = 0.2) than that used in the H2O2 clearance assays was necessary to exceed the limit of detection of the kit (BioAssay Systems EnzyChromTM NAD/NADH Assay Kit).


S8 Fig. Dependence of respiration on glucose availability.

Exponentially growing cells were washed, resuspended in media with or without glucose, and used to inoculate M9 media +/- glucose to an OD600 of 0.1. A higher density than that used for H2O2 clearance assays was necessary to observe a measureable drop in O2. Cells were allowed to consume oxygen for ten minutes before being treated with 5 mM KCN to inhibit respiration. The solid lines show the average of three biological replicates, and windows represent the standard error of the mean. We found that cultures in M9 media with 10 mM glucose efficiently consumed oxygen via respiration, whereas glucose deprived cultures consumed very little.


S9 Fig. AHP activity is removed in glucose-deprived cultures treated with 10 μM H2O2.

To explore whether omitting glucose from the media effectively eliminated AHP activity in a regime where it dominates (10 μM H2O2), we compared H2O2 clearance in ΔkatE ΔkatG in M9 minimal media with glucose (blue), without glucose (orange), and with glucose and 100 μg/mL CAM (black) to ΔkatE ΔkatG ΔahpCF (red) and cell-free controls (green). Removal of the two other major detoxification systems leaves AHP, which requires one NADH for every reaction cycle, as the only major detoxification system. When glucose is omitted from the media, H2O2 clearance is eliminated, as evidenced by ΔkatE ΔkatG in M9 minimal media without glucose never differing significantly from ΔahpCF ΔkatE ΔkatG or a cell-free control based on a two-tailed t-test with unequal variance. These results were not due solely to inhibition of translation, another effect of glucose starvation (S11 Fig), based on clearance abilities of CAM-treated, ΔkatE ΔkatG cultures.


S10 Fig. Predicted NADH concentration when its levels are not maintained.

Under conditions of glucose deprivation, AHP drains NADH in less than a second after exposure to 10 (A), 25 (B), 100 (C), and 400 (D) μM H2O2. Windows show the maximum and minimum predictions of the 4,000 models in the ensemble. The solid line indicates the prediction made by the most likely model. Note that the x-axis here is in seconds, not hours. We note that in all other simulations, the NADH concentration was held constant (S1 Table).


S11 Fig. Synthesis of GFP under control of katE, katG, and ahpC promoters.

Wild-type cells transformed with pUA66 PkatE-gfp (A-D), pUA66 PkatG-gfp (F-I), or pUA66 PahpC-gfp (J-M) were exposed to H2O2 in M9 10 mM glucose (blue), M9 lacking glucose (red), or M9 10 mM glucose + 100 μg/mL CAM (green) media. We confirmed that the lack of expression of GFP from pUA66 PkatE-gfp after exposure to H2O2 was not due to a defect in the vector by including a 16 h overnight control in M9 10 mM glucose. As expected, katE expression increases in stationary phase as determined by an increase in fluorescence after 16 h (0 μM H2O2 panel). In all cases, an empty vector control (black) was included to account for auto-fluorescence. Two biological replicates were analyzed on different days for each experiment. A representative replicate is shown here. Solid lines indicate time 0- distributions. For the PkatE-gfp control, the time 0- line in the 0 μM H2O2 panel represents inoculation after an 8 h growth period to mid-exponential phase. Dashed lines show distributions after ~90% of the H2O2 has been cleared by wild-type in M9 10 mM glucose (see Methods), or 16 h for the PkatE-gfp control (0 μM H2O2 panel).


S12 Fig. H2O2 measurement after centrifuging and/or sterile filtering to remove cells.

For the 30 min point in the 400 μM H2O2 clearance assay, three aliquots of sample were removed. One was centrifuged for 3 min at 15,000 rpm and the supernatant was removed identically to our protocol, one was left on the bench during the 3 min spin and then sterile filtered with an 0.22 μM syringe filter, and one was centrifuged and then sterile filtered. The experiment was performed in triplicate. Error bars show the standard error of the mean. Measurements from centrifuging alone were not significantly different from filtering (p = 0.45) or centrifuging + filtering (p = 0.40) based on a two-sample t-test with unequal variance.


S13 Fig. Clearance assays without exogenous catalase.

To determine whether exogenous catalase in the overnight or flask growth prior to the assay affected H2O2 clearance profiles for wild-type E. coli, we performed clearance assays in M9 media with 10 mM glucose after omitting the catalase in all pre-processing steps. A-D. H2O2 concentration did not differ significantly at any time point. E-H. CFU loss was not observed under either condition.


S14 Fig. Minimum sum of squared residuals (SSR) with increasing optimizations performed.

Each parameter set optimization was randomly initialized 1,000 times. Presented here is the minimum SSR that had been found at each of the 1,000 iterations (model index) for the following optimizations performed: model structure 1 (A), 2 (B), 3 (C), 4 (D), 5 (E), 6 (F), 7 (G), 8 (H), 9 (I), 10 (J) trained on wild-type data; model structure 2 (K) and 3 (L) trained on wild-type and ΔkatE ΔkatG data; model structure 3 (M) on wild-type, ΔkatE ΔkatG, ΔkatE, and ΔkatG data. This data suggested that more than 1,000 initializations would provide very little return for additional computational time invested.


S1 Table. Model species.

All metabolites and enzymes are listed with their initial concentrations, references, and relevant notes. Bounds are listed for uncertain parameters.


S2 Table. Spontaneous reactions.

All spontaneous reactions are provided with their rate constants, references, and relevant notes. Bounds are listed for uncertain parameters.


S3 Table. Enzymatic reactions.

All enzymatic reactions are provided with their rate equations, constants, and references. Bounds are listed for uncertain parameters, which are indicated by an asterisk.



We thank the National BioResource Project (National Institute of Genetics, Japan) for its support of the distribution of the Keio Collection, and Michael Kohanski for providing the ΔahpCF mutation. We would like to acknowledge Elliot Horlick, who did preliminary work on the project, as well as Sophia Leonard, who helped to develop the Amplex Red assay used in this work. We thank Christina J. DeCoste, John Grady, and the Flow Cytometry Resource Facility for assistance with flow cytometry experiments. In addition, we thank Professors Joshua Rabinowitz, Stanislav Shvartsman, and José Avalos for their suggestions on this project, and Jonathan Robinson for his help and suggestions throughout the project.

Author Contributions

Conceived and designed the experiments: KJA MPB. Performed the experiments: KJA. Analyzed the data: KJA MPB. Wrote the paper: KJA MPB.


  1. 1. Flannagan RS, Cosio G, Grinstein S. Antimicrobial mechanisms of phagocytes and bacterial evasion strategies. Nature reviews Microbiology. 2009;7(5):355–66. Epub 2009/04/17. pmid:19369951
  2. 2. Diacovich L, Gorvel JP. Bacterial manipulation of innate immunity to promote infection. Nature reviews Microbiology. 2010;8(2):117–28. Epub 2010/01/16. pmid:20075926
  3. 3. Imlay JA. Pathways of oxidative damage. Annu Rev Microbiol. 2003;57:395–418. pmid:14527285
  4. 4. Fang FC. Antimicrobial reactive oxygen and nitrogen species: concepts and controversies. Nature reviews Microbiology. 2004;2(10):820–32. Epub 2004/09/21. pmid:15378046
  5. 5. Crawford MA, Aylott CV, Bourdeau RW, Bokoch GM. Bacillus anthracis toxins inhibit human neutrophil NADPH oxidase activity. J Immunol. 2006;176(12):7557–65. Epub 2006/06/06. pmid:16751402
  6. 6. Siemsen DW, Kirpotina LN, Jutila MA, Quinn MT. Inhibition of the human neutrophil NADPH oxidase by Coxiella burnetii. Microbes Infect. 2009;11(6–7):671–9. Epub 2009/04/22. pmid:19379824
  7. 7. Tauber AI, Pavlotsky N, Lin JS, Rice PA. Inhibition of human neutrophil NADPH oxidase by Chlamydia serovars E, K, and L2. Infect Immun. 1989;57(4):1108–12. Epub 1989/04/01. pmid:2538397
  8. 8. Hebrard M, Viala JPM, Meresse S, Barras F, Aussel L. Redundant Hydrogen Peroxide Scavengers Contribute to Salmonella Virulence and Oxidative Stress Resistance. J Bacteriol. 2009;191(14):4605–14. pmid:19447905
  9. 9. Li Z, Kelley C, Collins F, Rouse D, Morris S. Expression of katG in Mycobacterium tuberculosis is associated with its growth and persistence in mice and guinea pigs. J Infect Dis. 1998;177(4):1030–5. Epub 1998/04/16. pmid:9534978
  10. 10. Manca C, Paul S, Barry CE, Freedman VH, Kaplan G. Mycobacterium tuberculosis catalase and peroxidase activities and resistance to oxidative killing in human monocytes in vitro. Infect Immun. 1999;67(1):74–9. pmid:9864198
  11. 11. Cosgrove K, Coutts G, Jonsson IM, Tarkowski A, Kokai-Kun JF, Mond JJ, et al. Catalase (KatA) and alkyl hydroperoxide reductase (AhpC) have compensatory roles in peroxide stress resistance and are required for survival, persistence, and nasal colonization in Staphylococcus aureus. J Bacteriol. 2007;189(3):1025–35. Epub 2006/11/23. pmid:17114262
  12. 12. Harris AG, Wilson JE, Danon SJ, Dixon MF, Donegan K, Hazell SL. Catalase (KatA) and KatA-associated protein (KapA) are essential to persistent colonization in the Helicobacter pylori SS1 mouse model. Microbiology-Sgm. 2003;149:665–72.
  13. 13. Brenot A, King KY, Janowiak B, Griffith O, Caparon MG. Contribution of glutathione peroxidase to the virulence of Streptococcus pyogenes. Infect Immun. 2004;72(1):408–13. Epub 2003/12/23. pmid:14688122
  14. 14. La Carbona S, Sauvageot N, Giard JC, Benachour A, Posteraro B, Auffray Y, et al. Comparative study of the physiological roles of three peroxidases (NADH peroxidase, Alkyl hydroperoxide reductase and Thiol peroxidase) in oxidative stress response, survival inside macrophages and virulence of Enterococcus faecalis. Mol Microbiol. 2007;66(5):1148–63. Epub 2007/11/01. pmid:17971082
  15. 15. Selva L, Viana D, Regev-Yochay G, Trzcinski K, Corpa JM, Lasa I, et al. Killing niche competitors by remote-control bacteriophage induction. Proc Natl Acad Sci U S A. 2009;106(4):1234–8. Epub 2009/01/15. pmid:19141630
  16. 16. Bogdan C, Rollinghoff M, Diefenbach A. Reactive oxygen and reactive nitrogen intermediates in innate and specific immunity. Curr Opin Immunol. 2000;12(1):64–76. Epub 2000/02/19. pmid:10679404
  17. 17. Loewen PC, Switala J, Triggs-Raine BL. Catalases HPI and HPII in Escherichia coli are induced independently. Arch Biochem Biophys. 1985;243(1):144–9. Epub 1985/11/15. pmid:3904630
  18. 18. Schellhorn HE, Hassan HM. Transcriptional regulation of katE in Escherichia coli K-12. J Bacteriol. 1988;170(9):4286–92. Epub 1988/09/01. pmid:3045091
  19. 19. Imlay JA. The molecular mechanisms and physiological consequences of oxidative stress: lessons from a model bacterium. Nature reviews Microbiology. 2013;11(7):443–54. Epub 2013/05/29. pmid:23712352
  20. 20. Seaver LC, Imlay JA. Alkyl hydroperoxide reductase is the primary scavenger of endogenous hydrogen peroxide in Escherichia coli. J Bacteriol. 2001;183(24):7173–81. Epub 2001/11/22. pmid:11717276
  21. 21. Mishra S, Imlay J. Why do bacteria use so many enzymes to scavenge hydrogen peroxide? Arch Biochem Biophys. 2012;525(2):145–60. pmid:22609271
  22. 22. Adimora NJ, Jones DP, Kemp ML. A model of redox kinetics implicates the thiol proteome in cellular hydrogen peroxide responses. Antioxid Redox Signal. 2010;13(6):731–43. Epub 2010/02/04. pmid:20121341
  23. 23. Makino N, Sasaki K, Hashida K, Sakakura Y. A metabolic model describing the H2O2 elimination by mammalian cells including H2O2 permeation through cytoplasmic and peroxisomal membranes: comparison with experimental data. Bba-Gen Subjects. 2004;1673(3):149–59.
  24. 24. Brito PM, Antunes F. Estimation of kinetic parameters related to biochemical interactions between hydrogen peroxide and signal transduction proteins. Front Chem. 2014;2:82. Epub 2014/10/18. pmid:25325054
  25. 25. Pillay CS, Hofmeyr JHS, Rohwer JM. The logic of kinetic regulation in the thioredoxin system. BMC Syst Biol. 2011;5.
  26. 26. Zamora-Sillero E, Hafner M, Ibig A, Stelling J, Wagner A. Efficient characterization of high-dimensional parameter spaces for systems biology. BMC Syst Biol. 2011;5:142. Epub 2011/09/17. pmid:21920040
  27. 27. Snipes M, Taylor DC. Model selection and Akaike Information Criteria: An example from wine ratings and prices. Wine Economics and Policy. 2014;3(1):3–9.
  28. 28. Wagenmakers EJ, Farrell S. AIC model selection using Akaike weights. Psychon B Rev. 2004;11(1):192–6.
  29. 29. Uriarte M, Condit R, Canham CD, Hubbell SP. A spatially explicit model of sapling growth in a tropical forest: does the identity of neighbours matter? Journal of Ecology. 2004;92(2):348–60.
  30. 30. Johnson JB, Omland KS. Model selection in ecology and evolution. Trends in ecology & evolution. 2004;19(2):101–8. Epub 2006/05/17.
  31. 31. Mazerolle MJ. Improving data analysis in herpetology: using Akaike's Information Criterion (AIC) to assess the strength of biological hypotheses. Amphibia-Reptilia. 2006;27(2):169–80.
  32. 32. Pond SLK, Frost SDW. A genetic algorithm approach to detecting lineage-specific variation in selection pressure (vol 22, pg 478, 2005). Mol Biol Evol. 2005;22(4):1157–.
  33. 33. Robinson JL, Brynildsen MP. A kinetic platform to determine the fate of nitric oxide in Escherichia coli. PLoS computational biology. 2013;9(5):e1003049. Epub 2013/05/10. pmid:23658508
  34. 34. Kotte O, Zaugg JB, Heinemann M. Bacterial adaptation through distributed sensing of metabolic fluxes. Mol Syst Biol. 2010;6:355. Epub 2010/03/10. pmid:20212527
  35. 35. Mosteller RD, Goldstein RV, Nishimoto KR. Metabolism of individual proteins in exponentially growing Escherichia coli. J Biol Chem. 1980;255(6):2524–32. Epub 1980/03/25. pmid:6987224
  36. 36. Mizusawa S, Gottesman S. Protein degradation in Escherichia coli: the lon gene controls the stability of sulA protein. Proc Natl Acad Sci U S A. 1983;80(2):358–62. Epub 1983/01/01. pmid:6300834
  37. 37. Vasudevan PT, Weiland RH. Deactivation of catalase by hydrogen peroxide. Biotechnol Bioeng. 1990;36(8):783–9. Epub 1990/10/20. pmid:18597274
  38. 38. Parsonage D, Karplus PA, Poole LB. Substrate specificity and redox potential of AhpC, a bacterial peroxiredoxin. Proc Natl Acad Sci U S A. 2008;105(24):8209–14. Epub 2008/01/01. pmid:18165315
  39. 39. Nelson KJ, Parsonage D, Karplus PA, Poole LB. Evaluating peroxiredoxin sensitivity toward inactivation by peroxide substrates. Methods Enzymol. 2013;527:21–40. Epub 2013/07/09. pmid:23830624
  40. 40. DeLuca DC, Dennis R, Smith WG. Inactivation of an animal and a fungal catalase by hydrogen peroxide. Arch Biochem Biophys. 1995;320(1):129–34. Epub 1995/06/20. pmid:7793971
  41. 41. Seaver LC, Imlay JA. Hydrogen peroxide fluxes and compartmentalization inside growing Escherichia coli. J Bacteriol. 2001;183(24):7182–9. Epub 2001/11/22. pmid:11717277
  42. 42. Ma M, Eaton JW. Multicellular oxidant defense in unicellular organisms. Proc Natl Acad Sci U S A. 1992;89(17):7924–8. Epub 1992/09/01. pmid:1518815
  43. 43. Turkheimer FE, Hinz R, Cunningham VJ. On the undecidability among kinetic models: from model selection to model averaging. J Cereb Blood Flow Metab. 2003;23(4):490–8. Epub 2003/04/08. pmid:12679726
  44. 44. Schaber J, Baltanas R, Bush A, Klipp E, Colman-Lerner A. Modelling reveals novel roles of two parallel signalling pathways and homeostatic feedbacks in yeast. Mol Syst Biol. 2012;8:622. Epub 2012/11/15. pmid:23149687
  45. 45. Link H, Kochanowski K, Sauer U. Systematic identification of allosteric protein-metabolite interactions that control enzyme activity in vivo. Nat Biotechnol. 2013;31(4):357–61. Epub 2013/03/05. pmid:23455438
  46. 46. Gutenkunst RN, Waterfall JJ, Casey FP, Brown KS, Myers CR, Sethna JP. Universally sloppy parameter sensitivities in systems biology models. PLoS computational biology. 2007;3(10):1871–78. Epub 2007/10/10. pmid:17922568
  47. 47. Appelberg R. Macrophage nutriprive antimicrobial mechanisms. J Leukoc Biol. 2006;79(6):1117–28. Epub 2006/04/11. pmid:16603587
  48. 48. Lorenz MC, Fink GR. Life and death in a macrophage: role of the glyoxylate cycle in virulence. Eukaryotic cell. 2002;1(5):657–62. Epub 2002/11/29. pmid:12455685
  49. 49. Weiss G, Schaible UE. Macrophage defense mechanisms against intracellular bacteria. Immunol Rev. 2015;264(1):182–203. Epub 2015/02/24. pmid:25703560
  50. 50. Xu YF, Amador-Noguez D, Reaves ML, Feng XJ, Rabinowitz JD. Ultrasensitive regulation of anapleurosis via allosteric activation of PEP carboxylase. Nature chemical biology. 2012;8(6):562–8. Epub 2012/04/24. pmid:22522319
  51. 51. Groat RG, Matin A. Synthesis of Unique Proteins at the Onset of Carbon Starvation in Escherichia-Coli. J Ind Microbiol. 1986;1(2):69–73.
  52. 52. Robinson JL, Adolfsen KJ, Brynildsen MP. Deciphering nitric oxide stress in bacteria with quantitative modeling. Curr Opin Microbiol. 2014;19:16–24. Epub 2014/07/02. pmid:24983704
  53. 53. Wang L, Hatzimanikatis V. Metabolic engineering under uncertainty. I: framework development. Metabolic engineering. 2006;8(2):133–41. Epub 2006/01/18. pmid:16414298
  54. 54. Wang L, Hatzimanikatis V. Metabolic engineering under uncertainty—II: analysis of yeast metabolism. Metabolic engineering. 2006;8(2):142–59. Epub 2006/01/18. pmid:16413809
  55. 55. Tran LM, Rizk ML, Liao JC. Ensemble modeling of metabolic networks. Biophys J. 2008;95(12):5606–17. Epub 2008/09/30. pmid:18820235
  56. 56. Rizk ML, Liao JC. Ensemble modeling for aromatic production in Escherichia coli. PloS one. 2009;4(9):e6903. Epub 2009/09/05. pmid:19730732
  57. 57. Miskovic L, Hatzimanikatis V. Modeling of uncertainties in biochemical reactions. Biotechnol Bioeng. 2011;108(2):413–23. Epub 2010/09/11. pmid:20830674
  58. 58. Jia G, Stephanopoulos G, Gunawan R. Ensemble kinetic modeling of metabolic networks from dynamic metabolic profiles. Metabolites. 2012;2(4):891–912. Epub 2012/01/01. pmid:24957767
  59. 59. Robinson JL, Brynildsen MP. Ensemble modeling enables quantitative exploration of bacterial nitric oxide stress networks. In: de Brujen FJ, editor. Environmental Control of Gene Expression and Adaptation in Bacteria: Wiley-Blackwell; (In press).
  60. 60. Brynildsen MP, Winkler JA, Spina CS, MacDonald IC, Collins JJ. Potentiating antibacterial activity by predictably enhancing endogenous microbial ROS production. Nat Biotechnol. 2013;31(2):160–5. Epub 2013/01/08. pmid:23292609
  61. 61. Adolfsen KJ, Brynildsen MP. Futile cycling increases sensitivity toward oxidative stress in Escherichia coli. Metabolic engineering. 2015;29:26–35. Epub 2015/03/04. pmid:25732623
  62. 62. Khodayari A, Zomorrodi AR, Liao JC, Maranas CD. A kinetic model of Escherichia coli core metabolism satisfying multiple sets of mutant flux data. Metabolic engineering. 2014;25:50–62. Epub 2014/06/15. pmid:24928774
  63. 63. Robinson JL, Miller RV, Brynildsen MP. Model-driven identification of dosing regimens that maximize the antimicrobial activity of nitric oxide. Metabolic Engineering Communications. 2014;1:12–8.
  64. 64. Tymoshenko S, Oppenheim RD, Agren R, Nielsen J, Soldati-Favre D, Hatzimanikatis V. Metabolic Needs and Capabilities of Toxoplasma gondii through Combined Computational and Experimental Analysis. PLoS computational biology. 2015;11(5):e1004261. Epub 2015/05/23. pmid:26001086
  65. 65. Chen ZH, Hills A, Batz U, Amtmann A, Lew VL, Blatt MR. Systems dynamic modeling of the stomatal guard cell predicts emergent behaviors in transport, signaling, and volume control. Plant Physiol. 2012;159(3):1235–51. Epub 2012/05/29. pmid:22635112
  66. 66. Chapman AG, Fall L, Atkinson DE. Adenylate energy charge in Escherichia coli during growth and starvation. J Bacteriol. 1971;108(3):1072–86. Epub 1971/12/01. pmid:4333317
  67. 67. Boles BR, Singh PK. Endogenous oxidative stress produces diversity and adaptability in biofilm communities. Proc Natl Acad Sci U S A. 2008;105(34):12503–8. Epub 2008/08/23. pmid:18719125
  68. 68. Bagaitkar J, Matute JD, Austin A, Arias AA, Dinauer MC. Activation of neutrophil respiratory burst by fungal particles requires phosphatidylinositol 3-phosphate binding to p40(phox) in humans but not in mice. Blood. 2012;120(16):3385–7. Epub 2012/10/23. pmid:23086626
  69. 69. Anderson KE, Boyle KB, Davidson K, Chessa TA, Kulkarni S, Jarvis GE, et al. CD18-dependent activation of the neutrophil NADPH oxidase during phagocytosis of Escherichia coli or Staphylococcus aureus is regulated by class III but not class I or II PI3Ks. Blood. 2008;112(13):5202–11. Epub 2008/08/30. pmid:18755982
  70. 70. Noubade R, Wong K, Ota N, Rutz S, Eidenschenk C, Valdez PA, et al. NRROS negatively regulates reactive oxygen species during host defence and autoimmunity. Nature. 2014;509(7499):235–9. Epub 2014/04/18. pmid:24739962
  71. 71. Greenberg JT, Monach P, Chou JH, Josephy PD, Demple B. Positive control of a global antioxidant defense regulon activated by superoxide-generating agents in Escherichia coli. Proc Natl Acad Sci U S A. 1990;87(16):6181–5. Epub 1990/08/01. pmid:1696718
  72. 72. Robinson JL, Brynildsen MP. An ensemble-guided approach identifies ClpP as a major regulator of transcript levels in nitric oxide-stressed Escherichia coli. Metabolic engineering. 2015;31:22–34. Epub 2015/06/27. pmid:26112956
  73. 73. Baba T, Ara T, Hasegawa M, Takai Y, Okumura Y, Baba M, et al. Construction of Escherichia coli K-12 in-frame, single-gene knockout mutants: the Keio collection. Mol Syst Biol. 2006;2:2006 0008. Epub 2006/06/02.
  74. 74. Datsenko KA, Wanner BL. One-step inactivation of chromosomal genes in Escherichia coli K-12 using PCR products. Proc Natl Acad Sci U S A. 2000;97(12):6640–5. Epub 2000/06/01. pmid:10829079
  75. 75. Nath KA, Ngo EO, Hebbel RP, Croatt AJ, Zhou B, Nutter LM. alpha-Ketoacids scavenge H2O2 in vitro and in vivo and reduce menadione-induced DNA injury and cytotoxicity. Am J Physiol. 1995;268(1 Pt 1):C227–36. Epub 1995/01/01. pmid:7840152
  76. 76. Zhao B, Summers FA, Mason RP. Photooxidation of Amplex Red to resorufin: implications of exposing the Amplex Red assay to light. Free Radic Biol Med. 2012;53(5):1080–7. Epub 2012/07/07. pmid:22765927
  77. 77. Montgomery HAC, Thom NS, Cockburn A. Determination of Dissolved Oxygen by Winkler Method + Solubility of Oxygen in Pure Water + Sea Water. Journal of Applied Chemistry. 1964;14(7):280–+.
  78. 78. Zaslaver A, Bren A, Ronen M, Itzkovitz S, Kikoin I, Shavit S, et al. A comprehensive library of fluorescent transcriptional reporters for Escherichia coli. Nature methods. 2006;3(8):623–8. Epub 2006/07/25. pmid:16862137
  79. 79. Han L, Enfors SO, Haggstrom L. Changes in intracellular metabolite pools, and acetate formation in Escherichia coli are associated with a cell-density-dependent metabolic switch. Biotechnol Lett. 2002;24(6):483–8.
  80. 80. Sundararaj S, Guo A, Habibi-Nazhad B, Rouani M, Stothard P, Ellison M, et al. The CyberCell Database (CCDB): a comprehensive, self-updating, relational database to coordinate and facilitate in silico modeling of Escherichia coli. Nucleic Acids Res. 2004;32:D293–D5. pmid:14681416
  81. 81. Lim CH, Dedon PC, Deen WM. Kinetic analysis of intracellular concentrations of reactive nitrogen species. Chem Res Toxicol. 2008;21(11):2134–47. Epub 2008/10/03. pmid:18828639
  82. 82. Taniguchi Y, Choi PJ, Li GW, Chen H, Babu M, Hearn J, et al. Quantifying E. coli proteome and transcriptome with single-molecule sensitivity in single cells. Science. 2010;329(5991):533–8. Epub 2010/07/31. pmid:20671182
  83. 83. Keyer K, Imlay JA. Superoxide accelerates DNA damage by elevating free-iron levels. Proc Natl Acad Sci U S A. 1996;93(24):13635–40. Epub 1996/11/26. pmid:8942986
  84. 84. Bennett BD, Kimball EH, Gao M, Osterhout R, Van Dien SJ, Rabinowitz JD. Absolute metabolite concentrations and implied enzyme active site occupancy in Escherichia coli. Nature chemical biology. 2009;5(8):593–9. Epub 2009/06/30. pmid:19561621
  85. 85. Kirsch M, Lehnig M, Korth HG, Sustmann R, de Groot H. Inhibition of peroxynitrite-induced nitration of tyrosine by glutathione in the presence of carbon dioxide through both radical repair and peroxynitrate formation. Chemistry. 2001;7(15):3313–20. Epub 2001/09/04. pmid:11531117
  86. 86. Albe KR, Butler MH, Wright BE. Cellular Concentrations of Enzymes and Their Substrates. J Theor Biol. 1990;143(2):163–95. pmid:2200929
  87. 87. Park S, Imlay JA. High levels of intracellular cysteine promote oxidative DNA damage by driving the fenton reaction. J Bacteriol. 2003;185(6):1942–50. Epub 2003/03/06. pmid:12618458
  88. 88. Imlay JA, Fridovich I. Assay of metabolic superoxide production in Escherichia coli. J Biol Chem. 1991;266(11):6957–65. Epub 1991/04/15. pmid:1849898
  89. 89. Nettleton CJ, Bull C, Baldwin TO, Fee JA. Isolation of the Escherichia coli iron superoxide dismutase gene: evidence that intracellular superoxide concentration does not regulate oxygen-dependent synthesis of the manganese superoxide dismutase. Proc Natl Acad Sci U S A. 1984;81(15):4970–3. Epub 1984/08/01. pmid:6087358
  90. 90. Lillig CH, Prior A, Schwenn JD, Aslund F, Ritz D, Vlamis-Gardikas A, et al. New thioredoxins and glutaredoxins as electron donors of 3'-phosphoadenylylsulfate reductase. J Biol Chem. 1999;274(12):7695–8. Epub 1999/03/13. pmid:10075658
  91. 91. Singh R, Wiseman B, Deemagarn T, Jha V, Switala J, Loewen PC. Comparative study of catalase-peroxidases (KatGs). Arch Biochem Biophys. 2008;471(2):207–14. Epub 2008/01/08. pmid:18178143
  92. 92. Obinger C, Maj M, Nicholls P, Loewen P. Activity, peroxide compound formation, and heme d synthesis in Escherichia coli HPII catalase. Arch Biochem Biophys. 1997;342(1):58–67. Epub 1997/06/01. pmid:9185614
  93. 93. Altschul SF, Madden TL, Schaffer AA, Zhang J, Zhang Z, Miller W, et al. Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res. 1997;25(17):3389–402. Epub 1997/09/01. pmid:9254694
  94. 94. Poole LB, Ellis HR. Flavin-dependent alkyl hydroperoxide reductase from Salmonella typhimurium. 1. Purification and enzymatic activities of overexpressed AhpF and AhpC proteins. Biochemistry (Mosc). 1996;35(1):56–64. Epub 1996/01/09.
  95. 95. Tartaglia LA, Storz G, Ames BN. Identification and molecular analysis of oxyR-regulated promoters important for the bacterial adaptation to oxidative stress. J Mol Biol. 1989;210(4):709–19. Epub 1989/12/20. pmid:2693740
  96. 96. Keseler IM, Collado-Vides J, Santos-Zavaleta A, Peralta-Gil M, Gama-Castro S, Muniz-Rascado L, et al. EcoCyc: a comprehensive database of Escherichia coli biology. Nucleic Acids Res. 2011;39(Database issue):D583–90. Epub 2010/11/26. pmid:21097882
  97. 97. Ellis HR, Poole LB. Roles for the two cysteine residues of AhpC in catalysis of peroxide reduction by alkyl hydroperoxide reductase from Salmonella typhimurium. Biochemistry (Mosc). 1997;36(43):13349–56. Epub 1997/10/28.
  98. 98. Belousov VV, Fradkov AF, Lukyanov KA, Staroverov DB, Shakhbazov KS, Terskikh AV, et al. Genetically encoded fluorescent indicator for intracellular hydrogen peroxide. Nature methods. 2006;3(4):281–6. Epub 2006/03/24. pmid:16554833
  99. 99. Dwyer DJ, Belenky PA, Yang JH, MacDonald IC, Martell JD, Takahashi N, et al. Antibiotics induce redox-related physiological alterations as part of their lethality. Proc Natl Acad Sci U S A. 2014;111(20):E2100–9. Epub 2014/05/08. pmid:24803433
  100. 100. Lim JB, Barker KA, Huang BK, Sikes HD. In-depth characterization of the fluorescent signal of HyPer, a probe for hydrogen peroxide, in bacteria exposed to external oxidative stress. J Microbiol Methods. 2014;106:33–9. Epub 2014/08/12. pmid:25109249
  101. 101. Leighty RW, Antoniewicz MR. Dynamic metabolic flux analysis (DMFA): a framework for determining fluxes at metabolic non-steady state. Metabolic engineering. 2011;13(6):745–55. Epub 2011/10/18. pmid:22001431
  102. 102. Antoniewicz MR, Kelleher JK, Stephanopoulos G. Determination of confidence intervals of metabolic fluxes estimated from stable isotope measurements. Metabolic engineering. 2006;8(4):324–37. Epub 2006/04/25. pmid:16631402
  103. 103. Leighty RW, Antoniewicz MR. Parallel labeling experiments with [U-13C]glucose validate E. coli metabolic network model for 13C metabolic flux analysis. Metabolic engineering. 2012;14(5):533–41. Epub 2012/07/10. pmid:22771935
  104. 104. Davies MJ. The oxidative environment and protein damage. Biochim Biophys Acta. 2005;1703(2):93–109. Epub 2005/02/01. pmid:15680218
  105. 105. Aliaga ME, Carrasco-Pozo C, Lopez-Alarcon C, Olea-Azar C, Speisky H. Superoxide-dependent reduction of free Fe3+ and release of Fe2+ from ferritin by the physiologically-occurring Cu(I)-glutathione complex. Bioorg Med Chem. 2011;19(1):534–41. pmid:21115254
  106. 106. Samanta G, Clifford D. Preservation of Arsenic Species: AWWA Research Foundation; 2006.
  107. 107. Kern DM. The hydration of carbon dioxide. J Chem Educ. 1960;37:14–23.
  108. 108. Nikitovic D, Holmgren A. S-nitrosoglutathione is cleaved by the thioredoxin system with liberation of glutathione and redox regulating nitric oxide. J Biol Chem. 1996;271(32):19180–5. Epub 1996/08/09. pmid:8702596
  109. 109. Desagher S, Glowinski J, Premont J. Pyruvate protects neurons against hydrogen peroxide-induced toxicity. J Neurosci. 1997;17(23):9060–7. Epub 1997/12/31. pmid:9364052
  110. 110. Gupta A, Rao G. A study of oxygen transfer in shake flasks using a non-invasive oxygen sensor. Biotechnol Bioeng. 2003;84(3):351–8. Epub 2003/09/12. pmid:12968289
  111. 111. Gray B, Carmichael AJ. Kinetics of superoxide scavenging by dismutase enzymes and manganese mimics determined by electron spin resonance. Biochem J. 1992;281 (Pt 3):795–802. Epub 1992/02/01. pmid:1311175
  112. 112. Boschi-Muller S, Azza S, Branlant G. E. coli methionine sulfoxide reductase with a truncated N terminus or C terminus, or both, retains the ability to reduce methionine sulfoxide. Protein Sci. 2001;10(11):2272–9. Epub 2001/10/18. pmid:11604533
  113. 113. Baker LM, Poole LB. Catalytic mechanism of thiol peroxidase from Escherichia coli. Sulfenic acid formation and overoxidation of essential CYS61. J Biol Chem. 2003;278(11):9203–11. Epub 2003/01/07. pmid:12514184
  114. 114. Reeves SA, Parsonage D, Nelson KJ, Poole LB. Kinetic and thermodynamic features reveal that Escherichia coli BCP is an unusually versatile peroxiredoxin. Biochemistry (Mosc). 2011;50(41):8970–81. Epub 2011/09/14.
  115. 115. Mulrooney SB. Application of a single-plasmid vector for mutagenesis and high-level expression of thioredoxin reductase and its use to examine flavin cofactor incorporation. Protein Expr Purif. 1997;9(3):372–8. Epub 1997/04/01. pmid:9126609
  116. 116. Scrutton NS, Deonarain MP, Berry A, Perham RN. Cooperativity induced by a single mutation at the subunit interface of a dimeric enzyme: glutathione reductase. Science. 1992;258(5085):1140–3. Epub 1992/11/13. pmid:1439821
  117. 117. Carlberg I, Mannervik B. Purification and characterization of the flavoenzyme glutathione reductase from rat liver. J Biol Chem. 1975;250(14):5475–80. Epub 1975/07/25. pmid:237922
  118. 118. Volkmer B, Heinemann M. Condition-dependent cell volume and concentration of Escherichia coli to facilitate data conversion for systems biology modeling. PloS one. 2011;6(7):e23126. Epub 2011/08/11.
  119. 119. Rosen H, Klebanoff SJ, Wang Y, Brot N, Heinecke JW, Fu X. Methionine oxidation contributes to bacterial killing by the myeloperoxidase system of neutrophils. Proc Natl Acad Sci U S A. 2009;106(44):18686 –91. Epub 2009/10/17.