Results

Results

We hypothesize that adding the Cysteinator to dark fermentative feedstocks will strengthen hydrogen-producing bacteria (HPB) propagation, enhance HPB hydrogenase activity, and weaken hydrogen-consuming bacteria (HCB) propagation. To validate this hypothesis, we conducted the experiments listed on our notebook and modeling pages. With their respective procedures described in the experiments and modeling pages.

The results detailed below serve as a proof-of-concept to inform future iterations of the Cysteinator, and validate that a both the L-cysteine overexpression plasmid and L-cysteine activated kill-switch are functional together in an engineered E. coli K-12 BW25113 strain. The standards for device functionality are as follows:

1. To produce excess L-cysteine extracellularly to a concentration of 300 mg/L of media.
2. L-cysteine concentration triggers device growth arrest in the bioreactor and enhances fermentative bacteria to propagate.

To best characterize device behavior, we first simplified testing into validating the two plasmids separately. The L-cysteine overexpression plasmid was tested through the ninhydrin assay, a quantitative measurement of L-cysteine concentrations and optical density measurements. The L-cysteine activated kill-switch was tested with a bacterial viability, western blot, and optical density measurements.

Wet Lab

Ninhydrin Assay

As our engineering page details, ligation and synthesis of our DNA has inhibited our ability to perform assays on bacteria transformed with correct plasmid constructs. However, we have still performed many iterations of the ninhydrin assay. Through extensive testing we were able to create valuable modifications to the 1967 Gaitonde Protocol [1] that optimized efficiency and minimized degradation of the final Ruhemann's Purple (RP) solution resulting in more accurate L-cysteine concentration readings.

The conversion of the raw qualitative findings of the ninhydrin assay to more useful quantitative data happens by running a calibration curve. A calibration curve takes known concentrations of L-cysteine and their OD560 values to approximate a mathematical equation capable of relating OD560 values to true L-cysteine concentration. To produce the most accurate calibration curves, we questioned four steps in the Gaitonde protocol [1]:

1. If the intermediate ice bathstep is necessary.
   1. Our specific concern was that the ice bath step contributed to degradation due the time between the boiling step and optical density reading.
2. If the adding ethanolstep is necessary.
   1. The ethanol is used to dilute the RP solution, and we surmised that our low optical densities were an effect of this over dilution.
3. If the preparing the ninhydrin reagentfresh every day is necessary.
   1. Since the ninhydrin reagent does not react with the other reagents in the sample until the boiling step, we thought that we would decrease the time burden of this assay by preparing the reagent days before the assay is run.

After our testing we concluded:

1. The ten-minute incubation period does not have a significant impact on the protocol. We observed subtle decreases in OD560 values when compared to the Gaitonde Protocol. However, because the ninhydrin assay is long and time-sensitive, it was determined that the ten-minute incubation period was unnecessary.
2. The exclusion of the addition of ethanol resulted in OD560 readings outside of the linear range of absorbance. This led to inaccurate readings and unreliable results. For this reason, we concluded that the addition of ethanol has a significant impact on the protocol.
3. As the prepared ninhydrin solutions had crystals that fell out of solution, it was determined that preparing the ninhydrin solution fresh every day has a significant impact on the protocol.
4. The best blank for the Ninhydrin assay is all the reagents of a typical sample without HCL. This solution accounts for the change in OD560 readings that occur independent of HCL addition.

Additionally, we decided to blank our spectrophotometer with a solution containing all reagents of a typical sample, excluding HCL (acetic acid, ninhydrin reagent 2, and SM1 media), because HCL acts as an activating reagent. By blanking with this solution instead of solely SM1 media, we accounted for any changes in the OD560 absorbance that may have been introduced by the assay conditions.

The experience of optimizing a time-sensitive ninhydrin assay led us to determine unique techniques that maximize accurate readings. Read more about them on our contributions page.

Figure 1 depicts calibration curve results prior to optimization testing. 5 readings were taken at each concentration with negligible error and variance.

The ninhydrin calibration curve performed before the modifications showed no evidently discernible relationship between L-cysteine concentration and OD560 (see Figure 1). The ninhydrin calibration curve performed after the post-optimization modifications were able to model the expected logarithmic relationship between L-cysteine concentration and OD560 (see Figure 2). These results further validated that our modifications were justified.

Plasmid Assembly

After addressing our assembly challenges, we were able to successfully assemble our L-cysteine overexpression plasmid construct through traditional restriction cloning into the pUC19 vector. Blue/White colony screening yielded positive qualitative and sequence results (see Figure 3).

Our immediate next steps are:

• transforming the L-cysteine overexpression plasmid into E. coli JM109 cells and beginning our assay scheme to determine the amount of L-cysteine export and the speed of cell growth.
• performing a similar blue/white screen of OneShot CcdB Survival Cell colonies transformed by the L-cysteine-induced kill-switch plasmid.

Effects of Modifications on L-cysteine production

The first aim of the flux-balance analysis (FBA) analysis was to understand how the modifications we made in the L-cysteine overexpression plasmid affect overall L-cysteine production/export and biomass growth. A lexicographic optimization was done on the modified and unmodified systems, and those results are detailed in Table 1. Table 1 shows that our circuitry increases L-cysteine export by nearly twofold.

Simulation	Biomass/Growth Rate (h^-1)	L-cysteine Export Flux (mmol/gDW*h)
Unmodified	0.117	7.263
Modified	0.201	14.03
Fold Change	1.71	1.93

Table 1. shows the effects of modifications to the GEM on overall growth rate and L-cysteine export

While performing optimization, the model predicts how fluxes are carried through metabolic networks/pathways. The top 15 reactions carrying the highest flux are illustrated in Figure 3. It is noted that most of these reactions involve using glucose and acetate for biomass growth. Upon further analysis of the fluxes through all pathways, the L-cysteine production pathways depicted in Figure 2 all carried flux. Thus, we can expect the modifications made in genetic circuits used in wet lab to similarly increase L-cysteine production experimentally.

Media Conditions

Another mechanism for adding constraints before performing FBA is through media conditions and limiting the metabolites available for uptake. The initial simulations were run using SM1, LB, and thiosulfate. Because adding thiosulfate is an additional cost to our device design and function, the model was used to understand how thiosulfate and carbon sources like glucose affect biomass growth and L-cysteine production.

From Figure 5B, it is apparent that after thiosulfate uptake reaches around 8 mmol/gDW*h, L-cysteine export does not increase significantly and plateaus. These results are critical for understanding how much thiosulfate to add to the medium to optimize L-cysteine production without added costs, especially for making the device cost-effective for industrial application. Biomass growth similarly does not continue to increase linearly as glucose uptake increases. It is important to note that because FBA predicts fluxes assuming steady state conditions, specific optimal concentrations cannot be determined from this model. However, the results still provide strong insights to consider when preparing the device for industrial application.

Predicting Flux Under Varying Conditions

Furthermore, these conditions are likely to change if our device were to be implemented in a bioreactor. Dark fermentation bioreactors contain wastewater and sludge for biohydrogen production. Through modifying and testing various media conditions, we can predict how our device will behave under these different conditions and use the results further down the line for industrial use. Wastewater used in bioreactors commonly comes in three forms—real wastewater (i.e. from factories), synthetic wastewater, and solid waste. Sugar factory wastewater (i.e. from beet sugar) and synthetic wastewater contain several of the same components like sucrose and acetate [16]. Solid waste is typically composed of fruit/vegetable waste and office paper waste, which contain glucose, fructose, and xylose [17]. The original SM1 conditions had glucose as the main carbon source. To test different bioreactor conditions, two representative medium conditions were developed to represent real/synthetic wastewater and solid waste. The upper bounds for uptake rates were calculated similar to the SM1 medium based on the concentration of each medium component found in literature and the component’s molecular weight. Although we expect to add our device with the SM1 medium to wastewater, the media conditions tested excluded SM1 medium components to understand the effects of the wastewater without other contributing factors. The real/synthetic wastewater conditions incorporated sucrose, acetate, butyrate, lactate, and ammonium in addition to amino acids and trace metals. The solid waste conditions incorporated fructose and xylose in addition to amino acids and trace metals. The uptake bounds for amino acids and trace metals were kept the same from initial simulations.

While testing the real/synthetic wastewater conditions, the main carbohydrate added to the medium was sucrose. Thus, sucrose was initially expected to be used for growth in place of glucose. However, after running optimization and evaluating Figure 5A, it was apparent the model did not follow the initial conjecture. Biomass growth was independent of sucrose uptake. With further evaluation of the iML1515 model, it was found that the model did not contain downstream reactions for hydrolyzing sucrose 6-phosphate to use it as a carbon source. This is because the E.coli K-12 strain and its derivatives lack the csc operon, which is responsible for sucrose catabolism [18]. This operon is found in other strains like E.coli EC3132 and E.coli W. Although the specific strains tested in wet lab are E.coli K-12 derivatives, future directions for the project include testing the genetic circuits in various chassis. Thus, the same type of testing and analysis can be used to evaluate the effect of sucrose on the functionality of the device in different chassis.

Furthermore, biomass growth and L-cysteine export were still possible without sucrose or glucose present. From analyzing the reactions carrying flux in the model, acetate and threonine uptake reactions were found to carry high fluxes. These reactions were tested to see if they had an effect on biomass growth. From the analysis, it was seen that the model was able to utilize threonine and amino acids for growth. These results highlight the flexibility of our device to function under various and nonideal conditions, necessary for a device to be implemented in variable environments.

In addition to real/synthetic wastewater, solid waste was also tested. To determine the effects of different carbon sources like fructose and xylose, glucose was removed from the medium conditions although it is commonly present in solid waste. These simulations yielded slightly higher biomass growth and L-cysteine export flux of 0.222 h^-1 and 14.17 mmol/gDW*h, respectively, compared to initial simulations. With further analysis, it was observed that these results were mainly attributed to the addition of fructose to the medium. E.coli K-12 was able to naturally metabolize this carbon source. These results are shown in Figure 5C-D. Thus, we expect higher yields with the addition of glucose as well. Thiosulfate uptake similarly plateaus and does not contribute to increased L-cysteine export after reaching a certain uptake rate. This signifies that once an optimal concentration of thiosulfate is determined, adding more thiosulfate does not make the system more effective.

Summary

The FBA model, incorporating the iML1515 GEM and enzyme constraints, provided insights into the effects of genetic modifications on the metabolic network of E.coli K-12 BW23115 and its ability to produce and export L-cysteine. The results validated the effect of the genetic modifications made experimentally and predicted a 1.71-fold increase in biomass growth and a 1.93-fold increase in L-cysteine export. After testing media conditions, the model simulations showed that thiosulfate and glucose uptake plateaued after reaching certain rates, revealing that the cost-effective aspect of the device design can be optimized by identifying the ideal concentrations of these medium components. Testing various media conditions representative of dark fermentation bioreactor environments demonstrated the metabolic flexibility of the design to use various carbon sources like fructose and threonine for growth, while identifying limitations such as the K-12 strain’s inability to metabolize sucrose. These findings offer actionable insights for scaling the device and selecting the final industrial chassis.

Molecular Docking Results

Figure 7 illustrates the 3D structure of each phase of oligomerization of CcdR. The binding patterns follow those of the FFRP transcription factor family of proteins. Specifically, in the AsnC family of proteins, the homodimer is held together by interactions between the anti-parallel beta sheets of the C-terminal domain [11]. The same binding pattern is observed in the dimer structure in Figure 7B. The dimer-dimer interface mainly consists of hydrophobic contacts, and each dimer pair forms a side of the binding pocket for the ligand. AsnC proteins also form an octamer with P4 symmetry, which is observed in the predicted octamer structure of CcdR as well.

Overall, all of the predicted structures in AlphaFold have confidence scores consistently higher than 90, indicating high accuracy. Furthermore, the interface predicted template modeling (ipTM) score for the dimer and octamer structures were higher than 0.8 as shown in Table 2. Values higher than 0.8 represent confident high-quality predictions.

Oligomerization Phase	Predicted Kd Value (M)	ipTM Score from AlphaFold
Dimerization	8.4e-18	0.89
Tetramerization	8.9e-09	0.6
Octamerization	7.9e-16	0.88
CcdR Octamer + L-cysteine	7.2e-15	N/A

Table 2. show Kd values predicted for each phase of oligomerization by PRODIGY

The Kd values were directly inputted into the mechanistic model to represent the binding affinity and the likelihood of the formation of the final octamer.

Given the hydrophobic nature of L-cysteine, it is reasonable that the ligand preferentially binds within the hydrophobic dimer–dimer interface.

Mechanistic Model Results

For this section, I would like to recollect the guiding questions that we started with

• How long on average does the killswitch take to activate?
• What is the average internal concentration of cysteine at that time?
• What is the desired promoter of ccdA such that we ultimately reach our target concentration at cell death?

and to answer these questions with the data we have collected. Starting with question 1, which is the main output from the mechanistic model. Due to the stochastic expression of CcdB, output graphs can vary drastically; to remedy this, Figure 10 overlays 1000 trials of our system. The black bars represent the average first passing time and its standard deviation with \(\bar{t}_f=19469.38\pm8937.72\) s. Firstly, the standard deviation is notably quite large, though this is expected as the system is known to and expected to have a certain variance. By definition, we have induced variability into this system, and therefore, this cannot be used as an indicator of model accuracy. Though we can clearly see a clustering of values within that range that begins to thin out significantly beyond. To that end, we find that \(64.5\%\) of our \(t_f\) values fall within the domain \(\{0,\bar{t}_f+\sigma\}\). Meaning we estimate that approximately 2/3 of the population's kill-switch will activate within the first 0.5 hours of a significant increase in cysteine concentration.

Figure 11 displays this information in a slightly more cohesive manner and displays a very interesting result. We see that, due to the central limit theorem, we actually recover the Poissonian statistical distribution that we originally put in. This fits quite well, showing us that the decision to use Poissonian statistics was justifiable. An \(R^2\) of 0.8885 is fairly consistent with a good fit; typically, fits closer to 1 are considered better. We still must ask the question of how accurate our average is to answer our second question. To do so, we take the standard error of the mean for our results, which is given by Eq. 35

\[ SEM\equiv\frac{\sigma}{\sqrt{N}} \] Where \(\sigma\) is the standard deviation and N is the number of outcomes in our sample space, for our system, we find that \(SEM=\pm282.64\), meaning our average is accurate within \(2.9\%\) of the actual value. With that degree of accuracy, we can take our average to be a fairly good approximation of the average of our system, and it can therefore be used to answer question 2. By looking at our dFBA outputs at our average time point, we see that the average internal cysteine concentration at the time of kill-switch activation is 2.144 mM. This is of a reasonable magnitude for intracellular cysteine under an over-expression pathway as found in "High Levels of Intracellular Cysteine Promote Oxidative DNA Damage by Driving the Fenton Reaction" (Park and Imlay, 2003). In this paper, it was found that an intracellular concentration of 1.5 mM is typically reached by a cysteine-overexpressing device. This is indicative that our model may overestimate concentrations in general, which is not unexpected, but is certainly a shortcoming of our model. To find what promoter strength we needed for CcdA we first looked at the available Anderson promoters that we had on hand and their relative strength according to the base promoter. We then used a tool from the Salis lab, specifically from "Automated design of thousands of nonrepetitive parts for engineering stable genetic systems"(Houssain et al., 2020), to calculate the rate constant for the base promoter in units of \(\frac{1}{M*s}\), the equation for which is detailed below \[ k_j=\frac{1}{2}\sum_{i=1}^{2}\frac{\frac{RNA_{i,j'}}{\sum RNA_i}}{\frac{DNA_{i,j'}}{\sum DNA_i}} \] Where i represents the replicate trial, j' represents the j'th concentration amongst all of the listed promoters for both DNA and RNA. These measurements were made for 2 replicates and are then averaged over those 2 replicates to increase the accuracy of the final rate constant. From this Eq. 36, we were able to determine the strength of the base level Anderson promoter and, therefore, the strength of the others that we had on hand.

Promoter	Relative strength to BBa\_J23100	Rate (\(\frac{1}{M \cdot s}\))
BBa\_J23100	1	0.1834
BBa\_J23102	0.86	0.157724
BBa\_J23105	0.24	0.044016
BBa\_J23109	0.04	0.007336

Table 3 shows the relative promoter strengths and transcription rates.

For each of our 3 available promoters, we tested the above values for the expression rate constant of CcdA, \(k_{AMRNA}\), to see how it affects our dynamical system. The results of this are in Figures 12 and 13:

As can be seen in the case of the BBa\_J23102 Figure 12, the level of expression was such that it would always outweigh the expression of CcdB. This would therefore make activation of our kill switch almost impossible, meaning we would have no control over cell death. Additionally, looking at the case of BBa\_J23109, Figure 13, we see that expression of CcdA cannot outweigh that of CcdB. This does not allow sufficient time for cysteine to overexpress and is therefore not viable. With BBa\_J23105, Figure 10, we are in the goldilocks zone, and expression of the CcdA/CcdB systems perfectly outweigh each other, allowing for fine-tuned control of the kill-switch. Ideally, we would prefer a slightly less powerful promoter, as currently, we would reach an external cysteine concentration of about 433 mg/L with the BBa\_J23105 promoter. To reach our desired external concentration of 300 mg/L, we would need an average first passage time of 15945 s. An additional constraint on this is the availability of discretized promoter strengths, necessitating an approximation based on the closest available strength. Based on this analysis, we have found that the best Anderson promoter to optimize our device is BBa\_J23105.

dFBA Results

The dFBA model without integration with the mechanistic model predicted extracellular metabolites, biomass concentrations, and intracellular/extracellular L-cystiene concentrations over time. However, upon integration with the mechanistic model, the transcription factor concentrations were too low to activate the kill switch. We are currently in the process of troubleshooting the error. The predicted intracellular L-cysteine concentrations from the dFBA reach the target concentration levels of around 300 mg/L, which should be sufficient to activate the kill-switch. The low levels of the transcription factor are likely attributed to the assumption that 8 L-cysteine molecules are required to form the complex and the formulation of the Hill-like approximations used in the mechanistic model. We expect to identify a solution and accurately predict the kill-switch activation time.