In vitro transcription-based biosensing of glycolate for prototyping of a complex enzyme cascade

Abstract

In vitro metabolic systems allow the reconstitution of natural and new-to-nature pathways outside of their cellular context and are of increasing interest in bottom-up synthetic biology, cell-free manufacturing, and metabolic engineering. Yet, the analysis of the activity of such in vitro networks is very often restricted by time- and cost-intensive methods. To overcome these limitations, we sought to develop an in vitro transcription (IVT)-based biosensing workflow that is compatible with the complex conditions of in vitro metabolism, such as the crotonyl-CoA/ethylmalonyl-CoA/hydroxybutyryl-CoA (CETCH) cycle, a 27-component in vitro metabolic system that converts CO₂ into glycolate. As proof of concept, we constructed a novel glycolate sensor module that is based on the transcriptional repressor GlcR from Paracoccus denitrificans and established an IVT biosensing workflow that allows us to quantify glycolate from CETCH samples in the micromolar to millimolar range. We investigate the influence of 13 (shared) cofactors between the two in vitro systems to show that Mg²⁺, adenosine triphosphate , and other phosphorylated metabolites are critical for robust signal output. Our optimized IVT biosensor correlates well with liquid chromatography–mass spectrometry–based glycolate quantification of CETCH samples, with one or multiple components varying (linear correlation 0.94–0.98), but notably at ∼10-fold lowered cost and ∼10 times faster turnover time. Our results demonstrate the potential and challenges of IVT-based systems to quantify and prototype the activity of complex reaction cascades and in vitro metabolic networks.

Graphical Abstract

Introduction

Synthetic biochemistry aims to reconstruct biological functions outside of a living cell (“cell-free”). Prominent examples are efforts to reconstitute natural (or new-to-nature) pathways from purified enzymes in vitro [1–9]. Such approaches not only allow us to study the fundamental design principles and function of metabolic networks [2–4] but also bear great application potential. For example, recent work demonstrated the cell-free conversion of the greenhouse gas carbon dioxide (CO₂) into polyketides, terpenes, and antibiotic precursors [5, 6] or the valorization of glucose and other low-cost precursors into monoterpenes and cannabinoids [7–9].

Compared to in vivo systems, cell-free metabolic networks are highly flexible in their composition, can be precisely modified and customized, and allow biochemical reactions to take place under “non-physiological” conditions [10]. Through cell-free systems, a rapid optimization of reaction compositions is possible without the need for molecular cloning, which minimizes time and cost. Consequently, lysate-based cell-free systems have been increasingly used to prototype pathways for the optimal combination and stoichiometry of individual enzymes and components [11–15]. In several cases, these optimized in vitro pathways could also be successfully implemented in vivo. Altogether, this showcases the capabilities of cell-free systems as a broad tool for in vitro and in vivo metabolic engineering purposes.

To optimize complex biological systems with minimal experimental effort, Pandi et al. recently reported a versatile workflow Machine-learning guided Experimental Trials for Improvement of Systems (METIS) that combines laboratory automation with active learning to explore the combinatorial space in iterative design–build–test cycles for (local) optima [16]. METIS successfully helped to improve several biological systems [3, 4, 17, 18], including an in vitro CO₂ fixation cycle of 27 different variables (crotonyl-CoA/ethylmalonyl-CoA/hydroxybutyryl-CoA (CETCH) cycle) [2, 19]. The CETCH cycle converts CO₂ into glycolate and could be improved by >10-fold through METIS. Although active learning–guided workflows are able to drastically minimize the number of samples screened, the screening phase still heavily relies on the use of costly and time-intensive instrumental analytics. In the case of the CETCH cycle, >3000 samples were analyzed by liquid chromatography–mass spectrometry (LC–MS), which requires 12 min per sample for glycolate quantification at a cost of ∼US$7 (Supplementary Note S1). We therefore set out to explore a low-cost and well-scalable in vitro transcription (IVT)-based biosensing method that would allow us to quantify activity of the in vitro metabolic system in the reaction mixture and thus increase throughput.

To establish such an IVT biosensing method, we turned our attention to a system called RNA Output Sensors Activated by Ligand Induction (ROSALIND), which was recently developed to detect pollutants in water samples [20]. The ROSALIND system consists of a linear DNA template encoding the sequence of an RNA aptamer [“Three-way Junction dimeric Broccoli” (3WJdB)] [21] (Fig. 1a). The 3WJdB aptamer is expressed under the control of a T7 promoter and an operator sequence that is repressed by an allosteric transcription factor (aTF). Only in the presence of its specific effector, the aTF releases the operator sequence to allow 3WJdB expression, which in turn results in a green fluorescent readout by stabilizing the fluorogenic dye DFHBI-1T in its fluorescent state.

Figure 1.

Characterization of a glycolate-responsive IVT-based biosensor module. (a) The ROSALIND system is based on the controlled expression of the 3WJdB RNA aptamer and the correlating fluorescence signal of the 3WJdB:DFHBI-1T complex. Expression is regulated by a transcriptional repressor that binds to an operator sequence downstream of a T7 promoter. The repression is lifted in a dose-responsive manner by binding an effector molecule to the aTF. This system allows faster and cheaper sample measurement in microtiter plates than analysis by LC–MS, but its precision is yet unknown. Time estimates refer to glycolate quantification. (b, c) Dose–response of the GlcR sensor module to glycolate in the concentration range of 100 nM to 100 mM glycolate (8 h time point, further time points shown in Supplementary Fig. S5e). The sensor shows an operational range from 10 µM to 20 mM glycolate with an excellent correlation between 16 µM and 8 mM glycolate and a 10-fold dynamic range. Note that IVT biosensing responses are highly time sensitive with better correlation at early time points and better sensitivity at late time points (Supplementary Fig. S5; Table S8). (d) Promiscuity assay of GlcR shows a low-level dose response to dl-lactate and dl-glycerate (4 h time point, further data shown in Supplementary Fig. S6). Raw fluorescence data are standardized to MEF (µM fluorescein). Data are the mean of 3 technical replicates ± SD. IVT output without the effector molecule and without MGlcR is shown as horizontal dotted lines.

Here, we demonstrate a ROSALIND-based biosensing workflow for the rapid prototyping and screening of complex in vitro metabolic systems, using the CETCH cycle as the proof of principle. We developed a glycolate-responsive sensor module to read out the glycolate-forming activity of the CETCH cycle and investigated the inhibitory effects of CETCH cycle components on the IVT system. We identified the availability of free Mg²⁺ as a critical factor for establishing highly robust and quantitative sensing of glycolate production. Notably, the IVT-based biosensing workflow reduces screening costs by an order of magnitude and reduces the analysis time of large sample sets from several days to ∼8 h. This work not only demonstrated that IVT-based reporter systems can be coupled to complex in vitro metabolic systems but also identified critical components and bottlenecks in setting up robust IVT-based screens under such complex and challenging conditions. Our work paves the way for the development of similar IVT-based reporter systems and further guides ongoing efforts to integrate in vitro metabolic networks and IVT–translation systems [22] toward constructing a synthetic cell [23, 24].

Materials and methods

Reagents

Unless stated differently, chemicals were purchased from Merck KGaA (Darmstadt, Germany) and Carl Roth GmbH (Karlsruhe, Germany). Commercial enzymes and bioreagents were purchased from New England Biolabs (Frankfurt am Main, Germany).

Strains and growth media

For molecular cloning, Escherichia coli NEB^® turbo was grown in lysogeny broth supplemented with an appropriate antibiotic (100 µg/ml ampicillin or 34 µg/ml chloramphenicol). For protein production, either E. coli M15 [T7 RNA polymerase (RNAP)] or E. coli BL21-AI [MBP-GlcR (MGlcR)] was grown in terrific broth (TB) supplemented with 100 µg/ml ampicillin or 34 µg/ml chloramphenicol, respectively. All strains used are listed in Supplementary Table S1.

Assembly of plasmids and preparation of linear DNA templates

Oligonucleotides were purchased from Merck KGaA. Synthetic dsDNA was purchased from Twist Bioscience (South San Francisco, CA, USA). Sanger sequencing was performed by MicroSynth (Göttingen, Germany).

Plasmids were generated by Golden Gate Assembly using the modular cloning system proposed by Stukenberg et al. [25]. A mixture of 0.4 nM vector DNA and 4–8 nM insert DNA was assembled using 0.5 U/µl Esp3I or 1 U/µl BsaI-HFv2 and 40 U/µl T4 ligase in 1× T4 ligase buffer. Reactions were cycled 15 times for 1.5 min at 37°C and 3 min at 16°C. Enzymes were heat inactivated for 5 min at 50°C and 10 min at 80°C. The Golden Gate product was transformed into chemically competent E. coli NEB turbo cells, and individual clones were verified by Sanger sequencing using oligonucleotides oSB0021 and oSB0022.

All linear DNA templates were prepared by polymerase chain reaction (PCR) amplification from the respective plasmids using oligonucleotides oSB0021 and oSB0022 and Q5 DNA polymerase, following the vendor’s instructions. All amplified DNA fragments were purified using the NucleoSpin Gel and PCR Clean-up kit (Macherey-Nagel), according to the vendor’s instructions. DNA concentrations were calculated from absorbance measurements at 260 nm (A₂₆₀) using a NanoDrop2000 spectrophotometer (Thermo Scientific, Waltham, MA, USA). To increase the throughput of screening various T7 promoter–operator sequences, we developed a workflow (inspired by previous work [26, 27]), which is described in detail in Supplementary Note S2 and the Supplementary Methods. All plasmids, linear DNA templates, and oligonucleotides used are listed in Supplementary Tables S2, S3, and S4, respectively.

Protein production and purification

MGlcR was produced and purified as previously described by Schada von Borzyskowski et al. [28]. T7 RNAP was produced in an E. coli M15 strain harboring plasmid pQE30-T7 RNAP [29] (strain sAP94). First, a preculture was inoculated in TB, supplemented with 100 µg/ml ampicillin. Cells were grown to high density overnight at 37°C. The preculture was used the following day to inoculate a production culture in TB medium supplemented with 100 µg/ml ampicillin and antifoam reagent. The culture was grown in a baffled flask at 37°C until an optical density (OD₆₀₀) of 0.7 was reached. The culture was then cooled down to room temperature for 30 min, before inducing the culture with 0.5 mM isopropyl β-D-1-thiogalactopyranoside. Cells were grown overnight at 20°C. Cells were harvested at 4000 × g for 20 min at 12°C, and cell pellets were resuspended in twice their volume of Buffer A (50 mM 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid (HEPES) pH 7.5, 500 mM KCl) with 5 mM MgCl₂ and DNase I (Roche, Basel, Switzerland). Cells were lysed by sonication using a SonoplusGM200 (BANDELIN electronic GmbH & Co. KG, Berlin, Germany) equipped with a KE76 tip at 50% amplitude for 3× 1 min of 1-s on/off pulses. Lysates were cleared by centrifugation at 100 000 × g for 1 h at 8°C, and the supernatant was then filtered through 0.45 µm filters (Sarstedt, Nümbrecht, Germany). For affinity purification, an Äkta Start FPLC system (formerly GE Healthcare, now Cytiva, Marlborough, MA, USA) was used with two stacked 1-ml Ni-NTA columns (HiTrap HP, Cytiva). The cleared lysate was loaded onto the columns, which were equilibrated with Buffer A. The column was washed with Buffer A + 75 mM imidazole and eluted with Buffer A + 500 mM imidazole. The eluate was desalted using two stacked 5-ml HiTrap desalting columns (Sephadex G-25 resin, Cytiva) and protein elution buffer (25 mM Tris-HCl pH 7.4, 100 mM NaCl). Protein concentration was calculated from absorbance at 280 nm (A₂₈₀) on a NanoDrop2000, and respective extinction coefficients were calculated using ProtParam (https://web.expasy.org/protparam/). Purified T7 RNAP was aliquoted, flash-frozen in liquid nitrogen, and stored at −70°C. See Supplementary Table S9 for an overview of which batch of T7 RNAP was used in which experiment.

IVT assays

IVT reactions were typically set up, unless stated differently, by adding the following components at their final concentration: IVT buffer [40 mM HEPES (pH 7.8), 8 mM MgCl₂, 10 mM dithiothreitol, 20 mM NaCl, and 2 mM spermidine], 0.2 mM DFHBI-1T and 11.4 mM nucleoside triphosphates (rNTPs, Thermo Fisher Scientific), 0.015 U/µl thermostable inorganic pyrophosphatase (New England Biolabs), and 15 nM DNA template. MGlcR (750 nM) was added to the GlcR module (see Supplementary Fig. S12 for batch-to-batch variations). To ensure stability of rNTPs, stocks at 80 mM rNTPs are buffered in a 200 mM Tris base. The sample volume of an IVT reaction was 20 µl and prepared in replicates of n = 3. The reaction mixture (57.8 µl = 3.4 × 17 µl) was equilibrated at room temperature (RT) for 30 min, before adding the respective analyte in a 1:10 dilution (6.8 µl = 3.4 × 2 µl, e.g. effector molecule, individual CETCH components at an indicated concentration, or CETCH cycle sample) and >10 U/µl T7 RNAP (3.4 µl = 3.4 × 1 µl, see Supplementary Fig. S12 for batch-to-batch variations; activity reference is NEB T7 RNAP, #M0251S). The reaction mixtures were mixed by pipetting, and 3 × 20 µl was immediately transferred into a 384-well, black, optically clear, flat-bottom, nonbinding microtiter plate (Greiner Bio-One, Kremsmünster, Austria; catalog no.: 781906). Plates were centrifuged for 30 s in a small table-top plate centrifuge (VWR, Radnor, PA, USA) before measurement. Reactions were characterized in triplicates on a plate reader (Infinite M200, Tecan, Männedorf, Switzerland) at 37°C, with 30 s of shaking before each fluorescence read at an excitation wavelength of 472 nm and an emission wavelength of 507 nm. Bottom reads of the plate allow for more precise measurements compared to top reads. To convert arbitrary fluorescence measurements to micromolar equivalents of fluorescein (MEFs), serial dilutions of a 12.5-µM stock of fluorescein standard (Invitrogen, catalog no.: F36915) were prepared in dH₂O and measured alongside each IVT assay. To convert arbitrary fluorescence units in MEF, (i) fluorescein fluorescence (in arbitrary units) was linearly regressed with fluorescein concentrations (in µM) and (ii) arbitrary fluorescence units were then divided by the slope of the linear fit. See the study by Jung et al. [20] for a detailed description of MEF standardization.

IVT assays with E. coli lysate were prepared from lysate of E. coli BL21 Star as previously described [30]. Twenty units of RNase inhibitor (NEB, #M0314S) was added to the initial titration of lysate in IVT. Lysate samples were filtered through a molecular weight cut-off (MWCO) membrane using 3- and 10-kDa Amicon filters [Merck Millipore, catalog no.: UFC500308 (3 kDa), UFC501008 (10 kDa)] for 30 min at 14 000 × g and 4°C.

CETCH cycle assays

The production and purification of enzymes were performed as previously described by Sundaram et al. [5]. To test whether stopping reactions by removing enzymes through MWCO filtration yields the same glycolate concentration as stopping reactions by protein precipitation with formic acid, we ran a single CETCH cycle assay (Day 7, Condition 15 [16]) in an 80-µl volume (1.5-ml microcentrifuge tube; started with 100 µM propionyl-coenzyme A (CoA) substrate; 300 rpm shaking in a thermoshaker for 3 h at 30°C; see concentrations in Supplementary Table S6). Two 9-µl samples were quenched with 1 µl of 50% formic acid, and two 25-µl samples were filtered through a 10-kDa MWCO plate (PALL AcroPrep Advance 96-well filter plate; 350 µl, Omega 10K MWCO, catalog no.: 8034) by centrifugation (15 min, 2272 × g, 20°C). Two microliters of the samples was diluted in 18 µl of ddH₂O and used for quantification via LC–MS (method previously described by Pandi et al. [16]).

To generate different glycolate concentrations under constant buffer and cofactor conditions, we ran CETCH cycle assays in which only the methylsuccinyl-CoA oxidase (Mco) concentration was titrated. Six reaction mixtures of Condition 15 (see Supplementary Table S6) were prepared in a 125-µl volume with different concentrations of Mco: 2×, 1×, 0.5×, 0.25×, 0.1×, and no Mco (1× = 26 µM). Reaction mixtures were prepared in 1.5-ml microcentrifuge tubes, started with 100 µM propionyl-CoA, and shaken for 3 h at 30°C and 300 rpm in a thermoshaker. Samples were filtered, and glycolate was quantified as described earlier. Filtered samples were stored at −20°C.

To prepare CETCH cycle samples with varying buffer and cofactor conditions, samples were prepared in a 150-µl volume (1.5-ml microcentrifuge tube; started with the 100 µM propionyl-CoA substrate; 500-rpm shaking in a thermoshaker for 4 h at 30°C). Concentrations of individual CETCH cycle components were varied in the following ranges: HEPES (75–200 mM, pH 7.4–7.8), MgCl₂ (2.5–17.5 mM), creatine phosphate (CP) (5–60 mM), sodium bicarbonate (2.5–60 mM), sodium formate (10–60 mM), CoA (0–5 mM), coenzyme B₁₂ (0–0.1 mM), adenosine triphosphate (ATP) (3–5 mM), dihydronicotinamide-adenine dinucleotide phosphate (NADPH) (2.5–10 mM), propionyl-CoA oxidase (0.10–9.57 μM), crotonyl-CoA carboxylase/reductase (0.62–2.78 μM), epimerase (0.74–6.70 μM), methylmalonyl-CoA mutase (0.61–2.89 μM), succinyl-CoA reductase (3.49–13.08 μM), succinic semialdehyde reductase (0.55–4.97 μM), 4-hydroxybutyryl-CoA synthetase (0.53–12.28 μM), 4-hydroxybutyryl-CoA dehydratase (0.73–3.64 μM), ethylmalonyl-CoA mutase (0.86–2.88 μM), Mco (26.01–46.54 μM), mesaconyl-CoA hydratase (0.28–2.84 μM), malyl-CoA/citramalyl-CoA lyase (2.79–14.73 μM), catalase (KatE, 2.46–8.21 μM), formate dehydrogenase (7.28–40.77 μM), creatine kinase (0.78–3.14 μM), carbonic anhydrase (0.02–0.13 μM), and glyoxylate/succinic semialdehyde reductase (3.31–5.25 μM). For a detailed overview of the involved enzymes, see the study by Sundaram et al. [5] and refer to Supplementary Table S6 for details. All assays were started with 0.1 mM propionyl-CoA. After 4 h, samples were filtered through 10-kDa MWCO spin filters (Amicon Ultra 0.5 ml, Merck Millipore, catalog no.: UFC501008), by centrifuging at 14 000 × g and 4°C for 20 min. Glycolate from filtrates was quantified as described earlier (with the minor difference that 10 µM internal ¹³C-glycolate standard was used), and samples were stored at −20°C.

Overview of the offline IVT biosensing workflow (all details are described earlier)

1. Prepare and run CETCH samples.

2. Filter CETCH samples through a 10-kDa MWCO membrane (plate or spin column-based).

3. Prepare dilution series NIST-traceable fluorescein standard and transfer to a 384-well plate.

4. Prepare ROSALIND reaction mixture on ice:

   1. omit the CETCH sample and T7 RNAP

   2. prepare in a 3.4× scale to prepare three replicates per sample.

5. Aliquot 57.8 µl in PCR tube strips, and equilibrate at RT for 30 min to ensure good repression by GlcR.

6. Add 6.8 µl of CETCH samples to respective wells → 1:10 dilution of the CETCH sample in the IVT sensor.

7. Add 3.4 µl of T7 RNAP to each well.

8. Mix by pipetting up and down a volume of 40 µL.

9. Transfer 20 µl in triplicates in a 384-well plate, centrifuge the plate, and start plate reader measurement.

10. Linearly regress fluorescein standard data to calculate MEF values from arbitrary units for data analysis.

Statistical analysis

All data presented are the mean of n = 3 technical replicates ± SD, unless otherwise stated in the respective caption. Pearson’s correlations were calculated using GraphPad Prism 10.1.0 based on the average values of replicates. We chose Pearson’s correlation over Spearman correlation to calculate the linear, rather than monotonic, relationship between LC–MS-quantified samples and sensor output. We used Student’s t-test (n = 2) or one-way analysis of variance test (n > 2) with a significance threshold of P > .05 to statistically compare the datasets.

Websites

ProtParam was used to calculate the molecular weight and theoretical extinction coefficients of all proteins used in this study (https://web.expasy.org/protparam/) [31].

Results

Establishing GlcR as a glycolate-responsive IVT-based biosensor module

To develop IVT-based reporter systems for in vitro metabolic networks, we chose the CETCH cycle as the model system, because of its pioneering role in synthetic CO₂ fixation, its biological complexity (involving a total of 17 enzymes and 10 nonenzyme components, such as cofactors), and its recent use in a METIS-assisted optimization workflow in which 1000 different combinations (3000 samples) had been already tested [16].

The end product of the CETCH cycle is glycolate, which is produced from CO₂. We therefore set out to construct a glycolate-responsive sensor module from a ROSALIND DNA template and a suitable aTF. A glycolate-responsive transcription factor, GlcC from E. coli, was previously described. However, this protein acts as a transcriptional activator [32–34], which made it incompatible with the T7 promoter–based IVT system of ROSALIND, which strictly relies on transcriptional repression. Thus, we turned our attention to another aTF from Paracoccus denitrificans that was recently reported to regulate glycolate assimilation in the β-hydroxyaspartate cycle [35]. This GntR family transcriptional repressor, named GlcR, is encoded by pden4400, binds the intergenic sequence pden4399-4400, and was shown to unbind in the presence of glycolate [28], which made the protein an interesting candidate for our envisioned IVT-based biosensor.

We purified GlcR as a fusion protein with N-terminal maltose-binding protein (MGlcR) and confirmed its binding and unbinding from the intergenic sequence between pden4400 and pden4397-4399 in the absence and presence of glycolate, respectively, by electrophoretic mobility shift assays (EMSA) (Supplementary Fig. S1a). To identify fragments carrying a putative operator site (glcO, Supplementary Fig. S1b), we next split the 150 base pair (bp)–long intergenic sequence into six fragments, with each fragment composed of ∼60 bp in length and an ∼30-bp overlap with neighboring fragments. The sixth fragment also encoded the first 51 bp of pden4399. EMSA showed that MGlcR bound to four of the six fragments (fragment nos 2–5), and in particular to fragment #3, which was bound by MGclR ∼2- to 4-fold stronger as fragment nos 2, 4, and 5, indicating that GlcR has multiple operator sites.

We then focused on the putative operator site in fragment no. 3. However, further splitting of fragment no. 3 into 20- and 30-bp-long fragments completely abolished MGlcR binding, suggesting that the operator site spans >30 bp (Supplementary Fig. S1c). We next removed base pairs in 4-bp steps from the 5ʹ end of fragment no. 3 and prepared eight ROSALIND templates encoding putative glcO sequences between 60- and 32-bp length (named according to their length, i.e. the 60-bp-long sequence was named glcO₆₀) as part of a P_T7-glcO-3WJdB expression cassette (Supplementary Note S2). When tested for 3WJdB expression in the presence or absence of MGlcR, all eight constructs showed repression between 3-fold (glcO₄₈ and glcO₅₂) and up to 8-fold and 16-fold in the case of glcO₆₀ and glcO₃₆, respectively (Supplementary Fig. S2a).

We continued with the two best sensor constructs, glcO₃₆ and glcO₆₀, and titrated MGlcR over a constant DNA template concentration (25 nM). At 1.25 µM MGlcR (i.e. 50× aTF:DNA template ratio) glcO₃₆ showed an ∼80-fold repression, while glcO₆₀ required 5 µM MGlcR (i.e. 200× aTF:DNA template ratio) to reach a similar level of repression (Supplementary Fig. S2b). At these concentrations, both constructs showed an ∼3-fold derepression with 10 mM glycolate (Supplementary Fig. S2c). We selected glcO₃₆ as the final construct and sought to further increase the sensitivity of the system by reducing the total aTF concentration. It has recently been shown that removing excess aTF molecules, which act as effector chelators, improves the sensitivity of ROSALIND sensor modules [20]. When lowering the DNA template to 15 or 5 nM DNA [and keeping the aTF:DNA ratio at 50× (i.e. 0.75 µM and 0.25 µM MGlcR)], glycolate-induced derepression increased by 6- and 10-fold, respectively, while the signal was only reduced by 8 and 50%, respectively, resulting in a good balance between glycolate sensitivity and total output signal (Supplementary Fig. S2d). We tested the influence of the T7 RNAP preparation on the signal (Supplementary Fig. S3) and confirmed that the sensor was functional in HEPES buffer between pH 7.2 and 7.8, the buffer conditions of the CETCH cycle (Supplementary Fig. S4). As standard conditions for all subsequent IVT-based biosensing experiments, we chose HEPES buffer pH 7.8 with 15 nM glcO₃₆, 750 µM MGlcR, and in-house T7 RNAP, which we refer to as the GlcR sensor module.

The GlcR sensor module is operational over three orders of magnitude

To determine the operational range of the GlcR sensor module, we tested the response of the sensor to glycolate concentrations over six orders of magnitude (from 100 to 100 mM), which defined a limit of detection at 10 µM glycolate and an upper limit at 20 mM (Fig. 1b; Supplementary Fig. S5a and e; Table S8). We observed inhibition at glycolate concentrations of >20 mM. The GlcR sensor module showed an excellent linear response between 16 µM and 8 mM (Pearson’s r = 1.0) and a dynamic range of 10.3-fold after 4 h of incubation (Fig. 1c; Supplementary Fig. S5c and d).

We also tested the specificity of the GlcR sensor module with seven structurally and context-related small organic acids: glyoxylate, acetate, dl-lactate, dl-glycerate, glycine, 2-phosphoglycolate, and 3-phospho-d-glycerate. Notably, the sensor was highly specific for glycolate and showed no response with other C2 acids, including glyoxylate and the amino acid glycine. However, we observed a low dose-responsive signal with dl-lactate and dl-glycerate (1.5- and 1.7-fold to 10 mM effector, respectively; in comparison, glycolate: 9.4-fold), indicating some promiscuity of GlcR with C3 alpha-hydroxy acids (Fig. 1d; Supplementary Fig. S6). However, since these C3 acids are not part of the CETCH cycle, we concluded that the GlcR sensor module could be used for the envisioned IVT-based biosensing system.

Nonenzyme components of CETCH influence online IVT analysis

To assess whether online (i.e. direct) biosensing of CETCH samples would be possible with our IVT biosensor, we next tested the compatibility of IVT with components of the CETCH cycle. The CETCH cycle consists of 27 components—17 enzymes and 10 nonenzyme components, some of which are shared between CETCH and the IVT (e.g. MgCl₂ and ATP). Overall, these nonenzyme components include substrates (bicarbonate and propionyl-CoA), cofactors (ATP, MgCl₂, NADPH, CoA, and coenzyme B₁₂), metabolites for energy supply (formate and creatine phosphate), and a buffer reagent (HEPES) (Fig. 2a). In addition to these 10 nonenzyme components of the cycle, all CETCH enzymes are kept in 20% glycerol, and Pco and Mco are stored additionally with flavin adenine dinucleotide (FAD). We also tested β-mercaptoethanol (β-ME) because it is widely used to store enzymes.

Figure 2.

Individual influence of CETCH cycle components on IVT in the absence of GlcR. (a) Reaction sequence of the CETCH cycle to convert CO₂ into glycolate [2, 19]. Ten nonenzymatic components are actively involved in the cycle, and three additional components are required to maintain enzyme activity during storage. (b) Titration of ATP, coenzyme B₁₂, MgCl₂, and NADPH concentrations in IVT reactions shows a dose-dependent influence of each component on IVT output after 4 h. Detailed data for nine additional components are shown in Supplementary Fig. S7. Raw fluorescence data are standardized to MEF (µM fluorescein). Data are the mean of n = 3 technical replicates ± SD. (c) Heat map describing the influence of all 13 nonenzyme components of the CETCH cycle on IVT output [as shown in (b) and Supplementary Fig. S7]. Data are normalized to IVT output in the absence of the respective component to indicate inhibition, enhancement, and no effect of the screened component at the indicated concentration.

We investigated the individual influence of these 13 different nonenzyme components on our basic IVT system (without the GlcR sensor module, each component titrated separately). For the 10 nonenzyme components of the CETCH cycle, we tested concentrations in our IVT were previously used during METIS-assisted optimization of the cycle by Pandi et al. [16]. For β-ME, FAD, and glycerol, we sampled a wider range of concentrations (Supplementary Table S5). Notably, 8 of the 13 nonenzyme components inhibited the IVT reaction, with ATP, NADPH, and coenzyme B₁₂ showing strong inhibition of up to 74–98% at high concentrations (Fig. 2b and c; Supplementary Fig. S7). ATP inhibition was partially due to an increased competition between ATP and the other three rNTPs (Supplementary Fig. S8). However, the majority of ATP inhibition seemed to be caused by chelation of free magnesium ions, which are essential for T7 RNAP activity, similar to the inhibition of phi29 DNA polymerase by rNTPs [36]. We further speculated that NADPH, CP, CoA, and FAD followed a similar inhibitory mechanism. High concentrations of glycerol were inhibited by 30%. CoA, FAD, and β-ME showed minor inhibition between 12% and 15%, while propionyl-CoA showed no effect. In contrast, four nonenzyme components, bicarbonate, MgCl₂, formate, and HEPES increased IVT output by up to 15% (Fig. 2b and c; Supplementary Fig. S7). While additional Mg²⁺ probably directly increased T7 RNAP activity, the mechanisms behind the slightly increased IVT output of the other three components remained unknown.

Overall, the complex (and partially adverse) effects of the different nonenzyme components on IVT showed that online measurements cannot be simply used for IVT-based biosensing under complex varying conditions of the CETCH cycle. We thus decided to work with an offline workflow, in which samples are quenched and diluted 1:10 before analysis to minimize the effects of the nonenzyme components on our IVT biosensor.

Establishing offline IVT sensing of CETCH samples with one enzyme component varied

We next developed an offline biosensing workflow, in which CETCH cycle variants are run first, and their output is analyzed using our IVT biosensor in a subsequent step (Fig. 3a). To quench CETCH reactions, we separated small molecules from enzymes by filtration through a 10-kDa molecular weight cutoff (MWCO) membrane before analysis of the filtrate (Supplementary Fig. S9) in a 1:10 dilution.

Figure 3.

Glycolate sensing from CETCH cycle samples with a single component, the concentration of enzyme Mco, varied. (a) Schematic of the experimental setup. (b) LC–MS quantification of glycolate from six CETCH cycle samples with the titrated Mco concentration measured in technical triplicates. (c) Time course of glycolate measurement using the GlcR sensor module. Time points shown in (f) are indicated as vertical dotted lines. (d) Time course of IVT measurement in the absence of GlcR showing no differences in inhibition by CETCH cycle samples. (e, f) Correlation between GlcR sensor module output and LC–MS quantification [as shown in (b)] over time. The quality of the correlation is time sensitive and worsens as soon as the first IVT reactions plateau. Data of 4-h and 8-h time points are exemplarily shown in (f). Raw fluorescence data are standardized to MEF (µM fluorescein). Data are the mean of n = 3 technical replicates ± SD (c, d, f).

As a proof of concept, we measured glycolate production from quenched CETCH samples, in which only one enzyme was varied. To that end, we titrated Mco, a critical enzyme known to limit the productivity of the CETCH cycle [16]. We ran six CETCH cycle reactions (Day 7, Condition 15 in the study by Pandi et al. [16], Supplementary Table S6), with different Mco concentrations (0–52 µM). These six conditions yielded different glycolate concentrations (Fig. 3b), which our offline IVT biosensor was able to quantify with high correlations. The time course showed excellent correlation at 4 h (r = 0.97) and between 7 and 9 h (r > 0.98), after which the IVT reactions started to plateau and other factors became limiting (Fig. 3c, e, and f). Overall, these results demonstrated that our offline IVT-based biosensor workflow is able to screen the productivity of CETCH cycle variants with varying enzyme concentrations.

Optimizing offline sensing of CETCH samples with multiple components varied

We next tested whether our offline IVT biosensor workflow was able to quantify glycolate concentrations from CETCH samples of highly diverse composition (enzymes and nonenzyme components varied, Fig. 4a). We prepared six CETCH samples with known productivity (Day 7 in the study by Pandi et al. [16], as specified in Supplementary Table S6) and compared their IVT readout with LC–MS-based quantification of glycolate (Fig. 4b). Overall, the GlcR module and the LC–MS-based method showed r = 0.77 after 4 h (Fig. 4c and d; Supplementary Fig. S10), with Sample 2 underestimating and Sample 3 overestimating the actual glycolate concentrations, respectively.

Figure 4.

Glycolate sensing from CETCH cycle samples with varied concentrations of nonenzyme components and enzymes. (a) Schematic of the experimental setup. (b) LC–MS quantification of glycolate from six CETCH cycle samples of different compositions (detailed compositions listed in Supplementary Table S6), measured in technical triplicates. (c, i) Time course of glycolate measurement with the GlcR module without and with the addition of 20 mM MgCl₂, respectively. (d, h) Correlation between GlcR module output [4-h time point, indicated as dashed lines in (c, i)] and LC–MS quantification [as shown in (b)]. See Supplementary Fig. S10b for correlation coefficients of hourly time points. (e) Correlation between free Mg²⁺ and IVT output in the absence of GlcR. Free Mg²⁺ is approximated as the concentration difference of added MgCl₂ and Mg²⁺ binding to estimate the change in free Mg²⁺ upon addition of the CETCH sample to the GlcR sensor module. (f) Titration from 0 to 30 mM MgCl₂ showed a dose-dependent IVT output in the presence and absence of GlcR. In the presence of 750 nM GlcR, the leakiness of the repressed module increased linearly, whereas in the absence of GlcR, the IVT output was bell-shaped. (g) Effect of additional 20 mM MgCl₂ on the constitutive repressed and derepressed state of the GlcR sensor module. Elevated MgCl₂ concentrations increased the leakiness of the module but did not affect the response to 1 mM glycolate. Data shown are 4-h time points, normalized to data with only MGlcR added. (c–i) Raw fluorescence data are standardized to MEF (µM fluorescein). Data are the mean of n = 3 technical replicates ± SD.

Strikingly, in both samples, the concentration of (free) MgCl₂ seemed to be the critical factor. Sample 3 contained the highest MgCl₂ concentration (17.5 mM) and low concentrations of Mg²⁺-binding cofactors (6.95 mM ATP, NADPH, and CoA). In contrast, Sample 2 contained the lowest MgCl₂ concentration (2.5 mM) and high amounts of Mg²⁺-binding cofactors (11 mM ATP, NADPH, and CoA). Because Mg²⁺ is the cofactor of T7 RNAP and its availability is essential for IVT (see earlier), we speculated that Mg²⁺ availability was strongly affecting the readout in these samples. This was supported by the fact that when we measured the effect of the six CETCH samples on the IVT system without the GlcR module, IVT output correlated well with the approximated concentration of free Mg²⁺ (Fig. 4e, r = 0.93).

We therefore decided to increase the overall Mg²⁺ concentration in our biosensor system to minimize the effect of CETCH cycle samples on Mg²⁺ availability. We examined the Mg²⁺ dependence of the IVT system and the GlcR module in the range of 0–30 mM MgCl₂ (Fig. 4f). Between 15 and 25 mM MgCl₂, the derepressed IVT system showed a broad plateau, while the Mg²⁺ effect on the repression of the system by GlcR was relatively small, indicating that this MgCl₂ concentration range was useful for robust sensing.

Indeed, when increasing the MgCl₂ concentrations from 8 mM to 20 mM in our IVT-based system, this significantly improved the correlation between GlcR module output and LC–MS to r = 0.94 (Fig. 4h; Supplementary Fig. S10), albeit at some increase in total output signal (Fig. 4i), caused by a higher baseline expression of the system (Fig. 4g). Overall, however, this setup established our GlcR IVT biosensor as a reliable glycolate quantification system that worked robustly across different conditions. This was further confirmed by probing E. coli lysate spiked with glycolate (Supplementary Fig. S11), demonstrating the possibility of using our IVT-based biosensing also in bacterial lysates that have become an important platform for pathway prototyping, recently [11, 12, 14, 37].

Discussion

Here, we explored the potential of IVT-based biosensors in cell-free manufacturing and in particular synthetic biochemistry. As a proof of concept, we tested the sensing of glycolate production from the CETCH cycle, a synthetic CO₂ fixation cycle, which was only possible by LC–MS analysis, thus far. Our experiments demonstrate that IVT-based biosensing of complex samples is feasible, with excellent and robust correlations.

Key to establish our biosensor was the finding that IVT-based sensing is highly sensitive to (shared) components that are commonly used in in vitro systems, including ATP, NADPH, and other nucleotide-based and phosphorylated cofactors, which strongly inhibit IVT activity. These results are in line with recent studies that investigated the inhibition of reconstituted transcription, translation, and DNA replication systems [36]. In our study, we extend these findings by providing a fitness landscape of T7 RNAP-based transcription in the presence of common metabolic cofactors. These data are not only relevant for efforts to establish other IVT-based biosensors, but might also be helpful for efforts of integrating metabolic and IVT–translation systems toward constructing a synthetic cell [22–24], and to prototype genetic (RNA) circuits for cell-free biosensing and biocomputing under complex conditions and within artificial compartments [38–42].

On a more practical note, our work shows that IVT biosensing offers several advantages over classical metabolic quantification methods, because it allows us to improve throughput and cost efficiency and does not rely on expensive analytical instrumentation. For example, during METIS-assisted optimization of the CETCH cycle, Pandi et al. screened glycolate production in a 384-well format with LC–MS [16]. At a sample analysis time of ∼12 min, complete data analysis takes ∼80 h. In contrast, analysis with the GlcR module in the 384-well format is finished within 4–8 h, which is at least 10 times faster. In addition, the cost per sample for LC–MS is ∼US$7 (2020–2 average, without instrument purchase included), while IVT-based biosensing reagent costs are ∼US$0.5 for a 20-µl reaction mixture, which makes our system a more cost- and resource-efficient alternative to LC–MS measurements. Increased throughput at lower cost, however, comes at the expense of some accuracy, as various components can interfere with the IVT measurement, which makes LC–MS the unmatched tool for highly precise quantifications (see Supplementary Note S1 for a detailed comparison of IVT-based biosensing and LC–MS and Supplementary Table S7 for IVT reagent costs).

We note that the applicability of our approach depends on the availability of suitable aTF–operator pairs for sensing the metabolite of interest. Today, only a few of aTFs are sufficiently characterized and collected in curated databases such as GroovDB [43]. However, systems biology approaches have proven to be powerful in identifying the role of aTFs in vivo [44–48] and are complemented recently through computational approaches that were able to successfully predict new aTF–operator pairs [49–52] and enzymes converting nondetectable metabolites into detectable ones [53, 54]. At the same time, efforts to engineer aTFs for new effector specificities and enhanced properties are increasing [48, 55–57], which will hopefully increase the repertoire and availability of aTFs to construct new IVT biosensors in the future.

Author contributions

Sebastian Barthel, Luca Brenker, Tobias J. Erb (Conceptualization), Sebastian Barthel, Luca Brenker (Methodology), Sebastian Barthel, Luca Brenker, Christoph Diehl, Nitin Bohra, Simone Giaveri (Investigation), Sebastian Barthel (Visualization), Sebastian Barthel, Tobias J. Erb (Writing—original draft); Sebastian Barthel, Tobias J. Erb (Writing—review & editing), Tobias J. Erb (Funding acquisition), Nitin Bohra (Resources), and Sebastian Barthel, Tobias J. Erb (Supervision).

Supplementary data

Supplementary data is available at SYNBIO Online.

Conflict of interest:

None declared.

Funding

This project was funded by Max Planck Society general funding. S.G. is grateful to the European Molecular Biology Organization postdoctoral fellowship (ALTF 162-2022). N.B. conducted his research within the Max Planck School Matter to Life supported by the German Federal Ministry of Education and Research (BMBF) in collaboration with the Max Planck Society.

Data availability

Data underlying this article are available in FigShare at dx.doi.org/10.6084/m9.figshare.25869418. Plasmids were deposited with Addgene (pTE5400: #221189, pTE5417: #221191, pTE5418: #221192, pTE5419: #221193, and pTE5423: #221194).