Design and engineering of E. coli metabolic sensor strains with a wide sensitivity range for glycerate

Microbial biosensors are used to detect the presence of compounds provided externally or produced internally. The latter case is commonly constrained by the need to screen a large library of enzyme or pathway variants to identify those that can efficiently generate the desired compound. To address this limitation, we suggest the use of metabolic sensor strains which can grow only if the relevant compound is present and thus replace screening with direct selection. We used a computational platform to design metabolic sensor strains with varying dependencies on a specific compound. Our method systematically explores combinations of gene deletions and identifies how the growth requirement for a compound changes with the media composition. We demonstrate this approach by constructing a set of E. coli glycerate sensor strains. In each of these strains a different set of enzymes is disrupted such that central metabolism is effectively dissected into multiple segments, each requiring a dedicated carbon source. We find an almost perfect match between the predicted and experimental dependence on glycerate and show that the strains can be used to accurately detect glycerate concentrations across two orders of magnitude. Apart from demonstrating the potential application of metabolic sensor strains, our work reveals key phenomena in central metabolism, including spontaneous degradation of central metabolites and the importance of metabolic sinks for balancing small metabolic networks.


Introduction
Microbial biosensors are gaining prominence as valuable tools for detecting specific environmental components such as toxic pollutants (Paitan et al., 2004;Trang et al., 2005), explosives (de las Heras et al., 2008), and pathogens (Saeidi et al., 2011).Rather than relying on external cues, biosensors can be used to detect internally produced compounds, thus assisting in optimizing the activity of biosynthetic routes.By engineering molecular biosensors for pathway intermediates or products -e.g., transcription factors or riboswitches -it is possible to screen thousands of different strains in short time (e.g., by flow cytometry), and identify variants that support high pathway flux (Jaffrey, 2018;Liu et al., 2015;Williams et al., 2016;Zhang et al., 2015).However, such high throughput screening can still be challenging, especially if the tested library is very large.
In some cases, it is possible to use dedicated gene-deletion strains, auxotrophic for a compound, in order to couple growth to the presence of a compound (He et al., 2018;Mainguet et al., 2013;Meyer et al., 2018;Yishai et al., 2018;Yu and Liao, 2018).Such strains can replace screening techniques by direct selection to identify the few biosynthetic variants whose activity is high enough to produce the relevant compound at sufficiently high amounts, thus enabling growth.However, in most cases, it is difficult to fine-tune the level of the selection strength.Hence, selection-based testing is usually a binary process, with growth observed above a threshold level of the relevant compound.
One solution to this sensitivity challenge is to construct a series of gene-deletion strains, each displaying a different dependency on the relevant compound to support growth.The growth phenotype of the different strains can be used to accurately estimate the availability of the compound in question.Such 'metabolic sensor strains' can be used to detect an externally available chemical or to provide quantitative information on the activity of the biosynthetic pathway producing this compound.
Here, we use a computational procedure based on flux coupling (Antonovsky et al., 2016;Jensen et al., 2019) for the design of metabolic sensor strains with varying dependencies on a given compound, which can be further modulated by controlling the composition of the growth medium.As an application of this approach, we chose to design biosensors for glycerate -the expected metabolic product of recently suggested synthetic photorespiration bypass routes which do not release CO 2 while assimilating Rubisco's oxygenation product, 2-phosphoglycolate, into the Calvin Cycle (Trudeau et al., 2018).These pathways are expected to enhance the carbon fixation rate of C3 plants under all relevant physiological conditions (Trudeau et al., 2018).The glycerate sensor strains will thus provide a platform to test and characterize the activity of these synthetic pathways.
Following the results of our model, we generated six gene-deletion strains, in which central metabolism is divided into several segments.In each of these strains, glycerate serves as a carbon source for a different fraction of cellular biomass.We show that these strains are sensitive to both exogenous supply and endogenous production of glycerate.We systematically characterize the dependence of each strain on the concentration of glycerate and demonstrate that the different strains span two orders of magnitude sensitivity towards this compound.Importantly, we demonstrate an almost perfect correlation between the measured growth dependencies on glycerate to those predicted by the computational platform.Our study therefore demonstrates the applicability of metabolic sensor strains to provide an easy readout -growth yield -under a wide range of compound concentrations.

Computational model for the design of metabolic sensor strains
Glycerate is assimilated in E. coli via the activity of glycerate 2kinase (Zelcbuch et al., 2015), generating the glycolytic metabolite 2phosphoglycerate.To generate glycerate sensor strains, we aimed at strategic gene deletions to isolate this metabolite from other segments of central metabolism.By doing so, we would be able to control which essential cellular building block is derived from glycerate and which will be synthesized from other carbon sources, thus determining cellular demand for glycerate.
To systematically identify potential glycerate sensor strains we used a dedicated computational tool.We started with the core metabolic model of E. coli (Orth et al., 2009) and, as described in the Methods, we modified it for our needs.In particular, we introduced the biosynthetic pathways for serine, glycine, and one-carbon moieties, as these are derived from 3-phoshoglycerate, which is directly adjacent to the entry point of glycerate to central metabolism.We aimed to systematically explore how different combinations of gene deletions and supplemented carbon sources change the dependency of E. coli growth on the availability of glycerate.We were especially interested in gene deletion sets which lead to a wide range of growth dependencies on glycerate as a function of the supporting carbon sources (which are to be given in large excess).The sensitivity of a strain carrying such deletions towards glycerate should be easily tuned by controlling the composition of the medium.
We considered combinations of glycerate with three supporting carbon sources, each entering a different segment of central metabolism -glycerol for upper metabolism, succinate for lower metabolism, and glycine for middle metabolism.We further considered all possible combinations of 1-4 meaningful gene deletions in central metabolism (Methods), as marked in yellow in Fig. 1.For each combination of carbon sources and gene deletions, we applied Flux Balance Analysis (FBA) to calculate the maximal biomass yield, assuming that glycerate is the only limiting carbon source; that is, the concentration of all other carbon sources is substantially higher than that of glycerate (Methods).We discarded non-viable combinations (i.e. that could not grow even with supplemented carbon sources) and combinations in which glycerate is not required to sustain growth.For the rest, we calculated the Glycerate:Biomass Ratio (GBR): the number of glycerate units required to produce one unit of biomass (as defined by the biomass function).Throughout this study, we report the GBR in units of millimoles glycerate per gram of cell dry weight.The GBR enabled us to compare the glycerate requirement associated with different combinations of carbon sources and gene deletions.
Out of the ~28,000 knockout combinations tested by our algorithm, 562 had a positive GBR values in at least one condition (Supplementary Table 1).We ranked these strains by their smallest GBR across the conditions, that is, the condition where they are most sensitive to glycerate.The top 50 gene deletion combinations are shown in Fig. 2. Some of these suggested strains, e.g., Δpgk Δeno ΔglyA, were sensitive to glycerate only under a single combination of carbon sources.Other strains, e.g., Δpgk Δpps Δppck, display different dependencies on glycerate for different combination of carbon sources, as indicated by different GBR values.The strain selection flexibility, as calculated for each strain and shown at the right-hand side of Fig. 2, corresponds to the ratio between the highest and lowest GBR values across the different combinations of carbon sources.As mentioned above, those strains with higher flexibility might be preferable as their sensitivity towards glycerate can be easily controlled by modulating the composition of the medium.

Overview of the chosen 6 metabolic sensor strains
Out of the many gene deletion possibilities, we chose to implement six, which together span a high variability of GBR values across different combinations of carbon sources.These are marked by green coloring in Fig. 2. The 'E' strains (Fig. 3A), were deleted in enolase (Δeno), effectively separating and completely isolating upper metabolism (upstream part of glycolysis, the pentose phosphate pathway, and metabolic routes derived from them) from lower metabolism (downstream part of glycolysis, the TCA cycle, and metabolic routes derived from them) (Zelcbuch et al., 2015).In these strains, succinate must be added as a carbon source for energy production and for the biosynthesis of cellular building blocks that are derived from lower metabolisme.g., glutamate, aspartate, alanine, fatty acids -which together contribute ~72% of the carbons within the cell (Neidhardt et al., 1990).
Strains E1, E2, and E3 differ from each other by the gene deletions that are added on top of enolase.The E1 strain harbors only the enolase deletion, such that glycerate is responsible for the biosynthesis of all cellular building blocks derived from upper metabolism -e.g., serine, glycine, adenine, histidine -together providing ~28% of cellular carbons (Neidhardt et al., 1990).The E2 stain is further deleted in 3phosphoglycerate kinase (Δpgk) (Wellner et al., 2013), such that central metabolism is effectively separated into three segments: (i) upper metabolism, mainly phosphosugar metabolism; (ii) 'middle metabolism', representing phosphoglycerate and its downstream metabolites serine and glycine (from which other essential cellular components are derived, e.g., one-carbons, purines, cysteine); and (iii) lower metabolism, consisting mainly of pyruvate metabolism and the TCA cycle.This strain requires the addition of succinate and glycerol as carbon sources for lower and upper metabolism, respectively.In this case, glycerate serves as a carbon source only for 'middle metabolism', providing ~10% of the cellular carbons (Neidhardt et al., 1990).The E3 strain is further deleted in serine hydroxymethyltransferase (ΔglyA) (Yishai et al., 2017).This strain requires that addition of succinate, glycerol and glycine, where glycerate is required only for the biosynthesis of serine and its direct derivatives -e.g., cysteine -together contributing only ~3% of cellular carbons (Neidhardt et al., 1990).As expected, we found that GBR E1 (4.1 mmol/gCDW) was higher than GBR E2 (1.5 mmol/gCDW), which was higher than GBR E3 (0.46 mmol/gCDW), reflecting a decreased dependency on glycerate for growth.
The 'P' strains are similar to the 'E' strains, but instead of enolase deletion they are deleted in phosphoenolpyruvate (PEP) synthetase (ΔppsA) and PEP carboxykinase (ΔpckA) (Fig. 3B).In these strains, flux from lower metabolism to upper metabolism is blocked -i.e., PEP cannot be regenerated from any lower metabolism intermediate -but flux from upper metabolism to lower metabolism is possible -i.e., PEP can be metabolized to pyruvate (via pyruvate kinase) or oxaloacetate (via PEP carboxylase).Hence, the addition of succinate to these strains is not mandatory as glycerate can feed lower metabolism; yet, if succinate is added, it would not be able to feed middle and upper metabolism.Another difference between the 'E' strains and the 'P' strains is that, in the latter, PEP biosynthesis is strictly dependent on glycerate.The P1, P2, and P3 strains corresponds to the same set of additional deletions as within the E1, E2, and E3 strains, respectively (Fig. 3).
For the 'P' strains, when succinate was available, we found that GBR P1 (4.7 mmol/gCDW) > GBR P2 (2 mmol/gCDW) > GBR P3 (1 mmol/gCDW).The higher GBR of the 'P' strains, as compared to the 'E' strains, is attributed to the fact that, in the former strains, glycerate is also needed for the biosynthesis of PEP, thus increasing biomass dependence on glycerate.When succinate was not provided, such that glycerate was needed also for lower metabolism, we got, as predicted, very high GBR -GBR P1 (27 mmol/gCDW) > GBR P2 (13 mmol/ gCDW) > GBR P3 (12 mmol/gCDW) -reflecting the very high dependence of these strains on glycerate.As expected, unlike the 'E' strains, the 'P' strains are associated with high selection flexibility, as the addition or omission of succinate dramatically changes their GBR.
Overall, the high variability of the calculated GBR values -spanning almost two orders of magnitudes -indicates that our designed strains should be suitable to serve as metabolic sensors with different sensitivities towards the desired compound.

The metabolic sensor strains detecting exogenous glycerate
After completing the construction of the different sensor strains (Methods) we tested whether they indeed grow only on the expected combination of carbon sources.Fig. 4A-F shows the sets of carbon sources that supported the growth of each strain.Within a time frame of ~100 h the observed growth profiles matched those expected.Importantly, all strains, except E2 supplemented with glycine, required glycerate for growth in all conditions (Fig. 4).The glycerate-independent growth of the E2 strain -when glycine was supplementedis to be expected from the strain design as glycine can be converted to serine (Fig. 3A).The E2 strain is thus intended to be used without the addition of glycine, such that both serine and glycine biosynthesis are Fig. 1.An overview of the structure of central metabolism of E. coli.We divided central metabolism into different generalized pathways as indicated by the colors of the arrows.Reactions that were considered for deletion by our software are marked in yellow frame.Potential carbon sources -i.e., glycerate, succinate, glycerol, and glycine -are shown in bold; their entry point to central metabolism is marked by bold arrows.Metabolites which are used to produce biomass building blocks -that is, that are consumed in the biomass reaction -are circled in green.(For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)dependent on glycerate.Indeed, when glycine is omitted, growth of the E2 strain becomes dependent on glycerate (Fig. 4E).While the addition of glycine is not necessary to obtain growth of the E1, P1, and P2 strains, it was found to boost their growth (Fig. 4A, B, D), presumably as it reduces the metabolic burden of glycerate metabolism which no longer needs to supply glycine and serine.The minute growth of the E2 and E3 strains when glycerol is omitted (brown curves) can probably be attributed to the existence of a glycogen reservoir than can feed upper metabolism but only to a limited extent (Wilson et al., 2010).
Longer cultivation resulted in an unexpected outcome, where the E3 strain (but not the other strains) was able to grow without the addition of glycerate, albeit after a long delay and at a very low growth rate (Fig. 4G).We emphasize that the growth experiments were performed in triplicates, which showed identical growth curves (and hence are presented as a single line).Moreover, upon a repeated cultivation of the strain an essentially identical growth profile emerged (two red lines in Fig. 4G, each corresponds to a triplicate experiment performed independently).Hence, the growth of the E3 strain without glycerate represents adaptation rather than mutation as in the latter case we would expect the different replicates to have diverging growth curves.As the E3 strain is expected to show the smallest dependence on glycerate -only serine biosynthesis should depend on glycerate (Fig. 3A)we reasoned that a metabolic leakage of low flux, which does not affect the other strains, short circuit the selection for glycerate in this strain.
To identify the source of the metabolic leakage, we cultivated the strain in the presence of either 13 C-glycine (and non-labeled glycerol and succinate) or 13 C 3 -glycerol (and non-labeled glycine and succinate).We found serine to be labeled thrice only when 13 C 3 -glycerol was added, indicating a metabolic leakage from upper metabolism (Fig. 5).We wondered whether this metabolic leakage depends on the canonical serine biosynthesis route (starting from 3-phosphoglycerate) or rather corresponds to a completely new route.Further deletion of 3-phosphoglycerate dehydrogenase (ΔserA) abolished growth without the addition of serine, confirming that the biosynthesis of serine still depends on the canonical route (purple curve in Fig. 4G).We hypothesized that the source of the metabolic leakage might be the spontaneous dephosphorylation of 1,3-bisphosphoglycerate, which is produced by the reversible activity of glyceraldehyde 3-phosphate dehydrogenase.

Fig. 2.
Glycerate:Biomass Ratio (GBR) of different combinations of reaction deletions and carbon sources.Shown are the 50 strains with the highest sensitivity towards glycerate -that is, lowest GBR -for one of the relevant combinations of carbon sources.Empty cells represent combinations of reaction deletions and carbon sources which cannot grow or for which growth is not dependent on glycerate; in both cases, GBR cannot be defined.The strain selection flexibility, as calculated for each strain, corresponds to the ratio between the highest and lowest GBR avlues across the different combinations of carbon sources.Strains and combinations of carbon sources that were chosen for in vivo implementation are shown in green and with a green 'V' sign, respectively.(For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)Indeed, deletion of this enzyme (ΔgapA) abolished the glycerate-independent growth (yellow curve in Fig. 4G).For further experiments, we therefore used an updated version of strain E3 in which gapA is deleted and is denoted strain E3*.
The spontaneous dephosphorylation of 1,3-bisphosphoglycerate to 3-phosphoglycerate can be attributed to the general reactivity and instability of phosphoanhydrides.Yet, the very low cellular concentrations of 1,3-bisphosphoglycerate (which is practically undetectable (Teusink et al., 2000)) results in very low production rate of 3-phosphoglycerate, barely sufficient to support serine biosynthesis in strain E3 and too low to affect the growth of the other selection strains which have a GBR ≥ 1 mmol/gCDW.

The metabolic sensor strains detecting in vivo production of glycerate
After confirming the response of the sensor strains towards exogenously added glycerate, we wanted to assess their response to glycerate produced endogenously.Towards this aim, we fed the strains with saccharic acid, which is known to be endogenously metabolized to pyruvate and tartronate semialdehyde, where the latter is reduced to glycerate (Monterrubio et al., 2000).As displayed in Fig. 6A-C, all the 'P' strains were able to grow with saccharic acid as a sole source of glycerate.However, of the 'E' strains (Fig. 6D-F), only strain E1 grew with saccharic acid, while strains E2 and E3* -that require much less glycerate for growth -were unable to grow.Our interpretation of these results is that consumption of saccharic acid generates very high levels of glycerate and phosphoglycerate.The 'P' strains can dispose of excess phosphoglycerate via lower metabolism.Yet, in the 'E' strains, lower metabolism is completely isolated from upper metabolism, and hence excess phosphoglycerate cannot be easily disposed of.This problem is less acute in strain E1 which requires a considerable amount of glycerate for growth and which can further dispose of phosphoglycerate via upper metabolism, and especially via the oxidative pentose phosphate cycle.However, strains E2 and E3* require low amounts of glycerate.Moreover, glycerate metabolism in these strains is completely isolated from both upper metabolism and lower metabolism, such that excess phosphoglycerate can easily accumulate and inhibit growth.Supporting this reasoning, we find that, when glycerate is added to the medium, the intracellular concentration of phosphoglycerate in strains E2 and E3* is 3.9 ± 0.9 fold higher than in a WT strain (p-value < 10 −7 ).This illustrates that high sensitivity of a sensor strain -that is, growth on a very low concentration of a sensed compound -can come with a price, where high production can severely inhibit growth.Providing a metabolic sink for the accumulated intermediates, as exists in the 'P' strains, could therefore provide an important safety valve, making the sensor strain more robust.
As further validation of the capability of the metabolic sensor strains to detect internally produced glycerate, we overexpressed the enzyme glyoxylate carboligase in all strains that were further deleted in the corresponding native gcl gene.The expression of this enzyme should thus enable the self-condensation of glyoxylate to generate tartronate semialdehyde, which E. coli can natively reduce to glycerate (Ornston and Ornston, 1969).As shown in Fig. 7, the sensor strains were indeed not able to grow with glyoxylate as a glycerate source, due to the deletion of the native gcl gene; but upon overexpression of the enzyme, growth with glyoxylate was restored.In this case, the rate of glyoxylate conversion to glycerate is low enough such that strains E2 and E3* can consume the latter compound while avoiding the deleterious accumulation of its downstream metabolites.

Quantitative dependence of the metabolic sensor strains on glycerate
Next, we aimed to quantitatively characterize the growth of each strain on varying concentrations of added glycerate.We tested each strain with a gradient of 11 glycerate concentrations, each concentration being 1.5-fold higher than the previous.For the 'P' strains, we checked growth both with and without further addition of succinate.For strains that were expected to be highly dependent on glycerate (i.e., with a high GBR) -for example, strain P1 in which succinate is omitted -we tested a gradient of relatively high glycerate concentrations (e.g., 520 μM to 30 mM).Conversely, for strains in which glycerate provides only a small fraction of biomass building blocks (i.e., with a low GBR) -Fig.3. The metabolic sensor strains chosen for implementation.In each strain, glycerate serves as the precursor for a different set of cellular building block and thus contribute a different fraction of the cellular carbons.(A) The 'E' strains in which upper and lower metabolism are completely isolated from each other by the deletion of enolase (Δeno).Succinate serves as the carbon source of lower metabolism.Depending on further gene deletions -of 3-phosphoglycerate kinase (Δpgk) and serine hydroxymethyltransferase (ΔglyA) -addition of glycerol and glycine are required to support growth.(B) The 'P' strains, deleted in PEP synthetase (ΔppsA) and PEP carboxykinase (ΔpckA), such that flux from lower metabolism to upper metabolism is blocked but not vice versa.Addition of succinate is not mandatory as glycerate can feed lower metabolism.As before, depending on further gene deletions -of 3-phosphoglycerate kinase and serine hydroxymethyltransferase -addition of glycerol and glycine are required to support growth.
Fig. 8 shows that for all strains (with the exception of strain E2, as discussed below) the gradient of glycerate concentrations resulted in a wide range of maximal OD 600 .Similar to the ~1.5 orders of magnitude spanned by the 11 glycerate concentrations tested, so did the resulting OD 600 span ~1.5 orders of magnitude.The diauxic growth observed with some of the strains and carbon source combinations (e.g., Fig. 8B) is difficult to explain, but might be related to the presence of cellular pools (e.g., glycogen feeding upper metabolism) that are initially used to support high growth rate and once depleted (at relatively low OD) are replaced with exogenously provided carbon sources.
Strain E2 displayed a different growth behavior, where the final OD was very similar regardless of the concentration of glycerate (Fig. 8H).Moreover, the growth rate of this strain was very low, with ~250 h required for the strain supplemented with 69 μM glycerate to reach maximal OD.We interpret these results to indicate a slow metabolic leakage of threonine degradation towards glycine (Fraser and Newman, 1975), which can then be further metabolized to serine.This low flux seems to need the initial growth priming provided by the low levels of glycerate and hence was not observed when we tested the strain in the absence of glycerate (Fig. 4E).Importantly, degradation of threonine does not enable growth of the other strains as they either cannot metabolize glycine to serine or are dependent on the metabolism of glycerate to cellular building blocks other than glycine and serine.Following this finding, we excluded strain E2 from our list of suitable glycerate biosensors.

Correlation between predicted and measured strain sensitivity to glycerate
Fig. 9 depicts the maximum OD 600 of each strain as a function of the initial glycerate concentration provided in the growth media (both axes are in logarithmic scale).As expected, the measured values for each strain are approximately arranged along a line with a slope of 1, indicating a linear dependency of growth on the concentration of glycerate (strain E2 deviates from this relationship, as discussed above).The samples associated with strains in which glycerate provides a high fraction of cellular carbons lie to the right-hand side of the figure.Conversely, the samples on the left-hand side of the figure correspond to strains in which glycerate provides only a few cellular building blocks.We define the Glycerate to OD Ratio (GODR) of each strain as the ratio between the glycerate concentration and the maximum OD 600 within a range of concentrations where this ratio is approximately constant (see Methods section).Strains P1, P2, and P3 in which succinate is not added use glycerate to produce most of the cellular building blocks and thus have a high GODR > 20 mM/OD.With the addition of succinate, we observe that GODR P1 (6.3 mM/OD) > GODR P2 (2.3 mM/OD) > GODR P3 (0.76 mM/OD), as expected by the decreasing fraction of cellular carbons that are derived from glycerate.Similarity, we observe that GODR E1 (3.2 mM/OD) > GODR E3 (0.45 mM/OD).The GODR cannot be defined for strain E2 as it cannot serve as a true glycerate biosensor as described above.Overall, our set of sensor strains display a high sensitivity range towards glycerate, with GODR spanning almost two orders of magnitude -from GODR E3 = 0.45 mM/OD to GODR P1 (no succinate) = 29 mM/OD.
Next, we checked whether the computationally calculated GBR indeed predicts the experimentally derived GODR.We found that the two parameters correlate almost perfectly (R 2 = 0.96, Fig. 10A).This supports the validity of the glycerate sensitivity predictions of our computational model and further indicates that the cells utilize their different carbon sources in a nearly optimal manner, without loss of limiting resources.
Further analysis indicated that the predicted GBR values for strains P2 and P3 with succinate disrupt the correlation between GBR and GODR, as they seem to be too low.Hence, we took a deeper look into the predicted metabolic fluxes associated with these strains and found that the model achieves higher yields (and therefore lower GBR values) by replacing the TCA cycle with an oxidative, cyclic flux via the oxidative pentose phosphate pathway.In this predicted mode of growth, all cellular reducing power and energy are derived from glycerol oxidation via this oxidative pentose phosphate cycle (OPPC), thus preventing glycerate oxidation by the TCA cycle and maximizing the utilization of this feedstock for biomass formation.However, replacing the TCA cycle with the OPPC is not supported by any experimental evidence and hence is unrealistic.Therefore, to make the model more realistic, we removed glucose 6-phosphate dehydrogenase, effectively abolishing the OPPC.This resulted in higher GBR values for strains P2 and P3 with succinate -GBR P2 = 23.7 mmol/gCDW and GODR P3 = 22.3 mmol/gCDW -as expected by the additional use of glycerate for generation of reducing power and energy.Notably, the two newly derived values improved the correlation between GBR and GODR substantially, and especially reduced the RMSE by > 50%, from 0.39 to 0.19 (Fig. 10B).This confirms that the OPPC is unlikely to take place and that glycerate oxidation is used to energize the cell.Fig. 4. Systematic testing of the metabolic sensor strains.Each sensor strain was tested with multiple combinations of carbon sources to confirm that growth was possible when an essential ingredient is missing.The growth patterns observed in (A)-(F) match the expectation form Fig. 2, confirming tight selection for the presence of glycerate.Minute growth without glycerol in (C) and (F) can be attributed to the existence of a limited internal glycogen reservoir.(G) Unexpectedly, after a long period, strain E3 was able to grow without glycerate, indicating a metabolic leakage that short-circuited the selection for this compound (red curves).Upon further deletion of serA or gapA this growth was abolished as shown in the purple and yellow lines.All experiments were performed in triplicates that showed identical growth profile ( ± 5%) and hence are represented by a single curve.Where relevant, succinate, glycerol, and glycerate were added at 20 mM, while glycine was added at 4 mM.(For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)Fig. 5. Serine biosynthesis in the E3 strain results from metabolic leakage from upper metabolism.Unlike within a WT strain, feeding with completely labeled glycerol-13 C 3 (and unlabeled succinate and glycine) results in almost fully labeled serine; serine is completely unlabeled when feeding with either glycine-1-13 C or glycine-2-13 C (and unlabeled glycerol and succinate).Where relevant, succinate, glycerol, and glycerate were added at 20 mM, while glycine was added at 4 mM.

Discussion
The construction of the metabolic sensor strains required us to disrupt multiple enzymes in central metabolism.While disruption of key glycolytic enzymes was demonstrated before, e.g., (Irani and Maitra, 1977;Wellner et al., 2013;Zelcbuch et al., 2015), our strains are unique as they harbor multiple such deletions, effectively dissecting central metabolism into multiple disconnected segments.Each segment represents a rather independent metabolic network which requires a dedicated carbon source as a feedstock.Multiple feedstocks can be fed to each metabolic segment, e.g., acetate, pyruvate, and succinate can feed lower metabolism, while glycerol, ribose and xylose can feed upper metabolism.We used glycerol and succinate as carbon sources for upper and lower metabolism, respectively, as these supported the highest growth rate.Notably, as previously reported (Irani and Maitra, 1977;Wellner et al., 2013)), glucose cannot be used as an upper metabolism feedstock as it completely inhibits the growth of strains in which middle glycolysis is disrupted, probably due to catabolite repression.
In strain E3, the metabolic segment that depends on glycerate as a carbon source is especially small, consisting mainly of serine biosynthesis and downstream utilization.This low-demand segment enabled us to observe a metabolic flux which is too small to have a physiological effect within a larger metabolic network.Specifically, we identified that 1,3-bisphosphoglycerate, the product of glyceraldehyde 3-phosphate dehydrogenase, undergoes (most likely) spontaneous dephosphorylation, thus generating 3-phosphoglycerate even if 3-phosphoglycerate kinase is deleted.The rate of the spontaneous Fig. 6.Endogenous metabolism of saccharic acid provides glycerate for the growth of the sensor strains.(A-D) Saccharic acid can replace glycerate when added to the growth medium of strains P1, P2, P3, and E1 (the strains were further supplemented with the required additional carbon source, as indicated in Fig. 2).(E-F) Saccharic acid did not support the growth of strains E2 and E3*, probably due to the lack of metabolic sink for phosphoglycerate.See main text for discussion.All experiments were performed in triplicates that showed identical growth profile ( ± 5%) and hence are represented by a single curve.Fig. 7. Metabolism of glyoxylate to glycerate via an overexpressed enzyme enables growth of the sensor strains.Gcl (glyoxylate carboligase) was expressed in each sensor strain, further deleted in the native gcl gene.Upon overexpression, glyoxylate could replace glycerate and support growth on the strains (the strains were further supplemented with the required additional carbon source, as indicated in Fig. 2).All experiments were performed in triplicates that showed identical growth profile ( ± 5%) and hence are represented by a single curve.Where relevant, succinate, glycerol, and glyoxylate were added at 20 mM, while glycine was added at 4 mM.dephosphorylation is rather low -as evident by the fact that it supported the glycerate-independent growth of only strain E3 -presumably since the concentration of this compound within the cell is very low, < 1 μM (Teusink et al., 2000).Our findings help explaining why 1,3-bisphosphoglycerate is maintained at such a low level, i.e., in order to minimize its wasteful, non-catalytic degradation.More generally, the approach we used in this study, dissecting central metabolism into small segments, can be used to identify low fluxes -spontaneous and enzymatic alike -which cannot be easily identified within a WT strain, but which could shed light on central cellular phenomena.
We designed and implemented a method based on Flux Balance Analysis to calculate the maximal biomass yield on limiting amounts of glycerate (and saturating amounts of the other carbon sources).This enables us to assess the sensitivity of different strains towards glycerate, as expressed by the GBR.We found that our computational estimations correspond almost perfectly to the experimentally derived GODR.
The nearly perfect match between the computational estimation and experimental results points to several important aspects.First, it Fig. 8. Maximal OD 600 of the glycerate sensor strains is determined by the concentration of added glycerate.For each strain, a gradient of 11 concentrations of glycerate was used, each concentration being 1.5-fold higher than the previous.In all cases, with the exception of strain E2 (H), the gradient of glycerate concentrations spanned a broad range of maximal OD 600 values.Strain E2 (H) grew very slowly and reached a very similar final OD regardless of the glycerate concentration, indicating the existence of a slow but persistent metabolic leakage that short-circuited the selection for glycerate.
suggests that the cell is able to optimize the utilization of the limiting carbon source even after undergoing severe disruption of its endogenous metabolism.This is especially true for the 'P' strains growing with succinate, which seem not to waste glycerate on lower metabolism, even though pyruvate kinase and PEP carboxylase are not deleted and could potentially drain the glycerate pool for its use for serine (and glycine) biosynthesis.Second, the observed correlation indicates that using constraint-based metabolic models to quantify the growth yield on multiple carbon sources, only one of which is limiting, is a valid approach to assess the sensitivity of different strains towards a compound.This paves the way for future use of this approach to design metabolic sensors for different metabolites.
Still, the model did fail in one notable case, where the oxidative activity of the TCA was predicted to be replaced by the OPPC, in order to avoid glycerate 'loss' towards energy production.Upon removal of the OPPC, the correlation between GBR and GODR became even better and the RMSE decreased substantially.This serves as a clear reminder that the fluxes predicted by Flux Balance Analysis should always be carefully checked according to available biochemical knowledge and that the underlining metabolic model must be modified to avoid Fig. 9. Maximal OD 600 as function of glycerate concentrations.Strain E2 is shown only for completeness, as it cannot serve as a true glycerate sensor strain (see main text).We define the Glycerate:OD Ratio (GODR) as the scaling factor between glycerate concentration and the resulting maximal OD.GODR was calculated as described in the methods, and the dashed lines correspond to their values.Our sensor strains display a high sensitivity range towards glycerate concentration, as the GODR values span almost two orders of magnitudes, from 0.45 (strain E3*) to 29 (strain P1 with no succinate).Our metabolic sensor strains span two orders of magnitude in their sensitivity towards glycerate.This makes them suitable to detect the exogenous supply or internal production of glycerate at various scales, hence providing a useful tool to test and optimize photorespiration bypass routes that generate this compound (Trudeau et al., 2018).Notably, the most sensitive strain can be used to detect concentrations as low as few tens of μM.Yet, high sensitivity comes with a price, as demonstrated in Fig. 6.That is, the growth of strains which require only a small amount of a compound might be strongly inhibited when it is present at a high concentration.In the case of strain E2 and E3*, this toxic effect probably stems from the accumulation of phosphoglycerate which cannot be effectively metabolized.This serves to emphasize two points.First, low sensitivity strains, such as strain E1 (as compared to strains E2 and E3*), play an important role, as they can be used to detect the desired compound under conditions at which the high sensitivity strains fail to grow.Second, providing a metabolic sink for small networks -as is the case in the 'P' strains but not 'E' strains -is important to enable balanced growth also at unexpected conditions (e.g., high concentration of a compound that is expected to be present at a low concentration).
The 'P' strains hold another advantage over the 'E' strains.Unlike the latter, whose feedstock requirement is rather strict, the carbon sources of the former can be modulated, i.e., supplementation with succinate is not mandatory but, when added, it dramatically contributes to growth.This flexibility would become useful when slowly evolving higher activity of a glycerate biosynthesis route.First, when production rate is still slow, succinate will be supplemented, thus maintaining a low selection pressure for glycerate.Once the activity of the biosynthesis route increases, succinate would be omitted, thus dramatically increasing the selection for faster glycerate production.The capability to tune the selection pressure by controlling the medium composition can thus facilitate continuous evolution of a pathway across several orders of magnitudes of activity.
Just as with genetic biosensors, which might not be fully specific to a single compound, metabolic sensor strains might grow in the presence of compounds other than the ones they are intended to detect.For example, the E3 strain can grow when serine replaces glycerate in the medium, which is to be expected as within this strain glycerate is used solely for the biosynthesis of serine and compounds that are derived from it.Similarly, the E2 and E3* strains can grow when glycerate is replaced with either glycine or serine.While completely eliminating such false detection is impossible, it can be minimized by using sensor strains that depend on the desired compound for the biosynthesis of multiple cellular building blocks which cannot be easily interconverted.For example, the P2 and P3 strains, which depend on glycerate also for the biosynthesis of PEP, are more specific to glycerate than the corresponding E2 and E3* strains, as the presence of glycine and serine cannot support their growth.Using strains with higher specificity and lower probability of false detection would be especially important when using a medium that contains numerous different contaminants.
Overall, our study demonstrates the applicability of metabolic sensor strains to detect a compound over a wide range of concentrations.Such strains could take a prime place alongside genetic biosensors in detecting exogenous as well as endogenous compounds.In the latter case, metabolic sensor strains have an inherent advantage as each single cell is equipped with its own selection mechanism for the enzyme/pathway activity rather than relying on outside screening.This way, the only limitation on the scale of the library is the number of cells that can be cultivate simultaneously, rather than the throughput of the measuring instrument (such as a flow cytometer or plate reader), increasing the capacity by many orders of magnitude.

Strains and gene deletions
Strains used in this study are listed in Table 1.Genomic gene deletions were performed by using two types of E. coli strains.E. coli strain MG1655 (F -λ -ilvG -rfb-50 rph-1) was used as the base strain for the 'P′ auxotroph strains and E. coli strain SIJ488 strain (Jensen et al., 2015) was used as the base strain for the 'E′ auxotroph strains.SIJ488 strain differs from MG1655 by carrying inducible genes that encode for a recombinase (pRed/ET) and a flippase that allow fast turnover for multiple deletions.Most of genomic gene deletions were performed using the Red/ET method (Zhang et al., 2000), recombining the selectable kanamycin resistance at the desired genetic locus (Quick & Easy E. coli Gene Deletion Kit, Gene Bridge, Heidelberg, Germany).The genes glyA, ppsA, and pckA were deleted by P1 phage transduction (Thomason et al., 2007) using the glyA, ppsA, and pckA knock out strains from the Keio collection (Table 1) as a donor (Baba et al., 2006).(We minimize the use of the P1 phage transduction method where possible as it transfers ~100 kb DNA from the donor strain to the host strain and thus may introduce undesired mutations.) For the recombinant gene deletion approach, kanamycin resistance cassettes were generated via PCR -'KO' primers with 50 bp homologous arms are listed in Supplementary Table 1 -using the FRT-PGK-gb2-neo-FRT (Km) cassette.Cells were inoculated in MY medium (see below); upon reaching OD 0.4-0.5, the pRed/ET recombinase gene was induced by addition of 15 mM L-arabinose.After 45-60 min incubation at 37 °C, cells were harvested and washed three times with ice cold 10% glycerol (11,300 g, 30 s, 2 °C).~300 ng of Km cassette PCR-product was transformed via electroporation (1 mm cuvette, 1.8 kV, 25 μF, 200 Ω).After selection on kanamycin, gene deletions were confirmed via PCR using 'KO-Ver' primers (Supplementary Table 1).In order to perform a sequential gene deletion, Km cassette was removed by transformation with a plasmid that encodes for flippase in 'P' strains (no needed for 'E' strains as the gene encoding for filippase plasmid was integrated in genome).50 mM L-rhamnose was added to induce flippase gene expression, in exponentially growing 4 mL MY culture at OD 0.5; induction time was ≥3 h at 30 °C.The successful removal of antibiotic resistance cassette was screened for kanamycin sensitivity and confirmed by PCR (using 'KO-Ver' primers).
For the overexpression of glyoxylate carboligase (gcl), the corresponding gene was amplified from E. coli genomic DNA by PCR (PrimeSTAR® Max DNA Polymerase, TaKaRa) using primer pairs of gcl_F and gcl_R (Supplementary Table 1).The amplicon was first cloned into pJET1.2/blunt(CloneJET PCR Cloning Kit, ThermoFisher Scientific), and the correct insert was cleaved by using restriction enzymes Mph1103I and XbaI to clone in pNivC vector.The correct insertion was confirmed by using primer pairs of pNiv_F and pNiv_R (Supplementary Table 1).This construct was cleaved with EcoRI and PstI to clone it in expression vector pZ-ASS, resulting in pZ-ASS-gcl.The plasmid harboring the pZ-ASS-gcl was transformed in each glycerate sensor strains for further growth experiments.

Media and growth conditions
For generation of the gene deletion strains we used the MY medium: M9 medium with trace elements ( 50 Growth experiments were performed in M9 medium with trace elements, supplemented with the appropriate carbon sources (succinate, glycerol, and glycerate were added at 20 mM while glycine was added at 4 mM).Overnight cultures for growth experiments were incubated in 4 mL in M9 medium with trace elements, supplemented with all relevant carbon sources to ensure growth.Pre-cultures of the strains overexpressing gcl were supplemented with glycerate to ensure the growth, yet glycerate was replaced with glyoxylate when performing plate experiments.Cultures were harvested by centrifugation (11000 rpm, 30 s, 4 °C) and washed three times in M9 minimal medium to eliminate any residual carbon sources.Growth experiments were inoculated to a starting OD 600 of 0.005 and carried out in 96-well microtiter plates (Nunclon Delta Surface, Thermo Scientific) at 37 °C.Each well contained 150 μL of culture and 50 μL mineral oil (Sigma-Aldrich, Germany) to avoid evaporation.A plate reader (Infinite M200 pro, Tecan) was used for incubation, shaking and OD 600 measurements (controlled by Tecan I-control v1.11.1.0).The cultivation program was run as follows; three cycles of 4 shaking phases, 1 min of each: linear shaking, orbital shaking at amplitude of 3 mm, linear shaking, and orbital shaking at amplitude of 2 mm.After each round of shaking (~12.5 min), absorbance (OD 600 nm) was measured in each well.Raw data from the plate reader were calibrated to cuvette values according to ODcuvette = ODplate/0.23.Growth curves were plotted in MATLAB and represent averages of triplicate measurements; in all cases, variability between triplicate measurements was less than 5%.

Isotopic-labeling experiments
In order to 13 C isotope tracing of proteinogenic amino acids cell were grown in M9 containing with the relevant carbon sources.Glycerol-13 C 3 , glycine-1-13 C, glycine-2-13 C (bought from Sigma-Aldrich, Germany) were used as indicated in the main text.After reaching stationary phase, a volume equivalent to 1 mL of OD 600 = 0.5 was harvested and washed in H 2 O by centrifugation.Hydrolysis of proteins was carried out with 6 M HCl, at 95 °C for 24 h (You et al., 2012).HCl was removed over night by incubation at 95 °C under an air stream.Samples were then resuspended in 1 mL H 2 O, centrifuged (5 min, 16,000g) to remove any insoluble compounds, and supernatants used for further analysis.Proteinogenic amino acids were analyzed by UPLC-ESI-MS described previously (Giavalisco et al., 2011) with a Waters Acquity UPLC system (Waters) using a HSS T3 C 18 reversed phase column (100 mm × 2.1 mm, 1.8 μm; Waters).The mobile phase was 0.1% formic acid in H 2 O (A) and 0.1% formic acid in acetonitrile (B).The flow rate was 0.4 mL/min with a gradient of 0-1 min-99% A; 1-5 min -linear gradient from 99% A to 82%; 5-6 min -linear gradient from 82% A to 1% A; 6-8 min -kept at 1% A; 8-8.5 min -linear gradient to 99% A; 8.5-11 min -re-equilibrate.Mass spectra were acquired using an Exactive mass spectrometer (Thermo Scientific) in positive ionization mode, with a scan range of 50.0-300.0m/z.The spectra were recorded during the first 5 min of the LC gradients.Data analysis was performed using Xcalibur (Thermo Scientific).Amino acid standards (Sigma-Aldrich, Germany) were analyzed for determination of the retention times under the same conditions.

Determination of metabolite concentrations
To assess the intracellular concentration of phosphoglycerate, we cultivated the strains in 500 mL shake flasks until reaching mid-exponential phase (OD 600 of 1.0).Culture densities were measured shortly before sampling and the values were later used for normalization.Each culture was sampled three times by filtering cells from 1 mL using MF-Millipore Membrane Filters (round, 0.45 μm, HLPV) and washing (on the filter) with 1 mL of media.The filters with the cells were then immediately placed in cold extraction solution -2:2:1 ratio of acetonitrile:methanol:water at -20 °C.A fully 13 C-labeled internal standard prepared from E. coli was added to each sample.After ≥24 h of incubation at -20 °C, the supernatants were collected while cell debris were separated by centrifugation.The supernatants were dried at 120 μbars (SpeedVac) at room temperature, and stored at -20 °C.
To assess the extracellular concentration of metabolites, we cultivated the strains in culture tubes (Greiner, 14 mL) for ~60 h, until all have reached stationary phase.Samples of 1 mL were taken from each culture and centrifuged for 10 min at 4000 g.The supernatants were stored at -20 °C for further analysis.
The dried cell extracts were re-suspended in 100 μL of MilliQ water, while the extracellular samples were diluted 1:50.All samples, including standards, controls, and calibration samples, were distributed in a sealed 96-well microtiter plate, and injected into an Agilent QTRAP LC-MS/MS with electrospray ionization (Yuan et al., 2012).The results were integrated and analyzed using in-house software described before (Buescher et al., 2010).

Computational platform for designing metabolic sensor strains
In order to design sensor strains, we used a standard constrainedbased framework generally known as Flux Balance Analysis.More specifically, we implement an algorithm that quantifies the strength of coupling between a certain reaction to the biomass rate (Antonovsky et al., 2016), which we term GBR.Testing many combinations of knockouts and carbon sources, ranked their potential to perform as glycerate biosensors based on their GBR values.A related method (Tepper and Shlomi, 2011) and a more recent version called OptAux (Lloyd et al., 2019) similarly identify combinations of gene knockouts that couple a target compound to the growth of the cell (i.e.make it auxotrophic to the target compound) at a predefined growth rate.These previous methods are based on bi-level MILP (Mixed Integer Linear Programming), and are more scalable to large networks such as the genome-scale metabolic network of E. coli.The first of this studies (Tepper and Shlomi, 2011) suggests using integer-cuts (as described in (Pharkya et al., 2004)) to generate alternative optimal knockout combinations for the same auxotrophy target.Although our exhaustive enumeration of all knockout combinations is less efficient than these MILP-based methods, the small scale of the network considered in this

Fig. 10 .
Fig.10.The GBR values calculated by our model predict the experimentally derived GODR almost perfectly.(A) R 2 = 0.96 and RMSE = 0.39.(B) Upon removal of the unrealistic oxidative pentose phosphate cycle (OPPC), which resulted in increased GBR values for strains P2 and P3 with succinate, the prediction improved even further, such that R 2 = 0.99 and RMSE = 0.19.

Table 1
Strains and the essential carbon sources used in this study.