A new-to-nature carboxylation module to improve natural and synthetic CO2 fixation

The capture of CO2 by carboxylases is key to sustainable biocatalysis and a carbon-neutral bio-economy, yet currently limited to few naturally existing enzymes. Here, we developed glycolyl-CoA carboxylase (GCC), a new-to-nature enzyme, by combining rational design, high-throughput microfluidics and microplate screens. During this process, GCC’s catalytic efficiency improved by three orders of magnitude to match the properties of natural CO2-fixing enzymes. We verified our active-site redesign with an atomic-resolution, 1.96-Å cryo-electron microscopy structure and engineered two more enzymes that, together with GCC, form a carboxylation module for the conversion of glycolate (C2) to glycerate (C3). We demonstrate how this module can be interfaced with natural photorespiration, ethylene glycol conversion and synthetic CO2 fixation. Based on stoichiometrical calculations, GCC is predicted to increase the carbon efficiency of all of these processes by up to 150% while reducing their theoretical energy demand, showcasing how expanding the solution space of natural metabolism provides new opportunities for biotechnology and agriculture. Biological CO2 fixation is restricted to few enzymes and pathways, limiting its value in environmental protection and agricultural productivity. Now, a new-to-nature CO2-fixing enzyme allows CO2-dependent assimilation of glycolate in a designed pathway, and its use for different applications is demonstrated.

T he conversion of the C 2 compounds glycolate and glyoxylate into C 3 metabolites plays a central role in many carbon metabolic processes, such as photorespiration and fatty acid assimilation, as well as formato-and methylotrophy. Moreover, glyoxylate is the product of several synthetic CO 2 fixation pathways 1,2 , where it serves as a hub metabolite that connects to central carbon metabolism.
However, there are only very few natural metabolic routes that allow the direct conversion of these C 2 intermediates into C 3 metabolites, and all of them result in the loss of carbon. The glyoxylate cycle 3 and the recently described β-hydroxyaspartate cycle 4 arrive at C 4 compounds, which need to be decarboxylated to generate C 3 metabolites. Similarly, photorespiration and the glycerate pathway convert two glyoxylate molecules into glycerate through the release of CO 2 5,6 . The inevitable loss of CO 2 in all of these pathways strongly limits their carbon efficiency, which is especially apparent for photorespiration. It was estimated that in hot and dry climates, agricultural crop yield is diminished by up to 50% due to photorespiratory carbon losses 7 . Thus, circumventing energy and carbon loss during glycolate assimilation through synthetic pathways is expected to enhance productivity substantially.
Recently, the tartronyl-CoA (TaCo) pathway was proposed as a direct route for the assimilation of glycolate into central carbon metabolism 8 . This hypothetical pathway was designed to fix CO 2 instead of releasing it, and is expected to outperform all naturally evolved glycolate assimilation routes (Fig. 1). However, the pathway remained theoretical until now, as neither tartronyl-CoA nor the individual enzymatic reactions of the TaCo pathway are known to occur in nature. This presented a challenge for its realization, with the engineering of glycolyl-CoA carboxylase (GCC)-the key enzyme of the TaCo pathway-as the biggest obstacle.
Here, we show the successful reconstitution and in vitro implementation of the complete TaCo pathway, applying rational design and high-throughput evolution of enzymes. Furthermore, we interface the TaCo pathway with several biotechnologically and agriculturally relevant processes (in particular, photorespiration, ethylene glycol conversion and synthetic CO 2 fixation) as proof of principle.

Results
Identifying the enzymatic framework for the TaCo pathway. To establish a biosynthetic route to the non-native carboxylation substrate glycolyl-CoA, we investigated two possible options: a coenzyme A (CoA) transfer from another acyl-CoA donor by CoA transferases, as well as the direct ligation of glycolate to CoA by acyl-CoA synthetases. We screened 11 different native and engineered enzymes (Supplementary Table 1). All of the tested enzymes catalysed the activation of glycolate. The best transferase from Clostridium aminobutyricum (AbfT) showed a catalytic efficiency of 120 M −1 s −1 for the activation of glycolate with acetyl-CoA as the CoA donor (Supplementary Table 1).
The best CoA synthetase was an acetyl-CoA synthetase (ACS) homologue from Erythrobacter sp. NAP1 (EryACS1), which showed a catalytic efficiency of 20 M −1 s −1 with glycolate. It is well known that ACSs are post-translationally regulated through lysine acetylation in vivo 9,10 . To suppress post-translational inactivation of EryACS1 during protein production, we changed surface loop residue Leu641 into a proline, which has been reported to prevent acetylation of ACS from Salmonella enterica 11 . Indeed, EryACS1 Leu641Pro showed a twofold higher specific activity, but also a tenfold increased apparent K m for glycolate (Supplementary Table 1). Therefore, we decided to create a lysine acetylase knockout strain for protein expression (Escherichia coli BL21 (DE3) AI ΔpatZ). When we produced EryACS1 in this knockout strain, the catalytic efficiency of the enzyme was increased almost 30-fold to 540 M −1 s −1 (Supplementary Table 1), which is sevenfold higher than the catalytic efficiency of a previously reported engineered E. coli ACS for the activation of glycolate (82 M −1 s −1 ; deacetylated enzyme) 8 . Based on a homology model created with ACS of S. enterica (Protein Data Bank (PDB) ID: 2P2B; 62% sequence identity) 12,13 , we identified Val379 as a target for directed mutagenesis to open up the active site for accommodation of the slightly larger glycolate ( Supplementary  Fig. 1). A substitution of Val379 by alanine (Val379Ala) had previously been reported to enhance the activity of the enzyme with the slightly larger propionate in S. enterica ACS 12 . We tested three different Val379 substitutions (Supplementary Table 1), isolating the variant Val379Ala (glycolyl-CoA synthetase (GCS)), which showed an improved apparent K m for glycolate (13 ± 3 mM) and a catalytic efficiency of 853 M −1 s −1 (Table 1). Based on the favourable kinetic parameters and thermodynamic considerations (that is, the irreversibility of the reaction because of immediate hydrolysis of inorganic pyrophosphate in vivo), as well as the independence from other acyl-CoA pools in vivo compared with transferases, we decided to further rely on the engineered GCS for glycolyl-CoA synthesis (Table 1 and Supplementary Fig. 1).
In the TaCo pathway, the hypothetical product of glycolyl-CoA carboxylation, tartronyl-CoA, is reduced to glycerate via tartronic semialdehyde. We tested different reductases and found that the bifunctional malonyl-CoA reductase from Chloroflexus aurantiacus 14 (CaMCR; tartronyl-CoA reductase (TCR)) was able to convert tartronyl-CoA directly to glycerate in two steps with a k cat of 1.4 s −1 and a very favourable apparent K m of 26 µM for tartronyl-CoA (Table 1 and Supplementary Table 1).

Identification and engineering of GCC.
With both GCS and TCR in hand, we focused our efforts on identifying a suitable enzyme candidate for the key reaction of the TaCo pathway, the carboxylation of glycoyl-CoA to tartronyl-CoA. We screened for promiscuity in four different biotin-dependent propionyl-CoA carboxylases (PCCs), because of the structural similarity between glycolyl-CoA and propionyl-CoA (Supplementary Table 1). Of the tested enzymes in this study, only PCC from Methylorubrum extorquens (MePCC) had minuscule activity with glycolyl-CoA. Co-expression of the cognate biotin ligase gene of M. extorquens resulted in very low but measurable activity (k cat = 0.01 s −1 ; Fig. 2 and Supplementary  Table 1) accompanied by a high ratio of futile ATP hydrolysis compared with tartronyl-CoA formation (~100:1).
Next, we sought to alter the substrate preference of MePCC. Such engineering of biotin-dependent acyl-CoA carboxylases had only been attempted in a few studies [15][16][17] . For structure-guided rational design, we built a structural model of MePCC, for which we obtained a 3.48-Å cryo-electron microscopy (cryo-EM) dataset (Fig. 2, Supplementary Fig. 2 Table 2). We then targeted several residues in the first and second shell of the active site of the carboxyltransferase subunit, which form the substrate binding pocket. The selected residues were assumed to be involved in direct or indirect binding of the natural substrate propionyl-CoA, which differs in the Cα position from glycolyl-CoA by a hydroxyl group (Fig. 2, Supplementary Fig. 3 and Supplementary Table 1). To accommodate this hydroxyl group of glycolyl-CoA in the active site, we created variant GCC M2, in which we introduced a Tyr143His substitution to enable direct hydrogen bonding between enzyme and substrate, and further directed the hydroxyl group towards His143 via a Asp407Ile substitution, which had been reported to enhance substrate promiscuity in a PCC from Streptomyces coelicolor 15 . We introduced a third substitution in the active site's second shell (Leu100Ser), to facilitate the formation of a hydrogen-bonding network and further strengthen the interaction of His143 with glycolyl-CoA. Compared with the wild type, the corresponding triple mutant (GCC M3) exhibited a more than 50-fold increased catalytic efficiency and a more than 15-fold decreased futile ATP hydrolysis (Fig. 2a-c and Supplementary Table 3).

and Supplementary
Next, we applied directed evolution using high-throughput approaches to screen for mutants with further decreased futile ATP hydrolysis. We diversified the carboxyltransferase subunit of GCC M3 using error-prone PCR. To assess whether the libraries contained enzymes with decreased futile ATP hydrolysis, we established a high-throughput microfluidics screen ( Supplementary  Fig. 4). In this screen, the formation of tartronyl-CoA was monitored in a TCR-coupled assay under ATP-limiting conditions within picolitre droplets. We used previously described microfluidic workflows 18 to develop short-and long-term multiplexed endpoint assays with single E. coli cells expressing individual GCC variants . Analysis of the pooled libraries showed that 4.7% of the variants exhibited more favourable ATP hydrolysis ratios compared with GCC M3 (Supplementary Fig. 8). This fraction of positive hits made it feasible to switch to microplate screens, which had the additional benefit that time-dependent kinetics of single variants could be directly obtained. We assessed each library separately to identify and characterize potentially improved enzyme variants. The best variant obtained from two rounds of subsequent random mutagenesis (GCC M5) possessed two additional substitutions (Ile450Val and Trp502Arg; Fig. 2e), a 560-fold increased k cat of 5.6 s −1 and a more than 25-fold lower ATP hydrolysis compared with the MePCC wild type (Table 1 and Fig. 2a-c). Overall, GCC M5 had a catalytic efficiency of 3.6 × 10 4 M −1 s −1 for glycolyl-CoA carboxylation, which is three orders of magnitude higher compared with the wild type (Supplementary Table 3) and similar to the catalytic efficiencies of naturally occurring biotin-dependent acyl-CoA carboxylases ( Supplementary Fig. 9).
To exclude incomplete biotinylation of our GCC variants during protein production, we performed avidin-gel shift assays, which confirmed full biotinylation of all characterized enzymes ( Supplementary  Fig. 10). Furthermore, we assessed the thermostability of all , after the addition of GCS, after the addition of GCC and after the addition of TCR. e-g, Performance of the TaCo pathway for different applications. The data are shown for two independent experiments each. e, Demonstration of the TaCo pathway as synthetic photorespiratory bypass using malate as a read-out (Fig. 4). The malate yield was increased in the presence of the TaCo pathway (inset). f, Ethylene glycol assimilation via the ethylene glycol module and the TaCo pathway. Shown is glycerate formation from ethylene glycol in an optimized experimental set-up. g, The TaCo pathway as an additional carbon-fixing module for the CETCH cycle. Shown is glycerate formation from propionyl-Coa and CO 2 as starting substrates under optimized experimental conditions. enzymes by circular dichroism spectroscopy. All GCC variants exhibited melting temperatures similar to the wild type, indicating that enzyme stability was not affected through our mutations ( Supplementary Fig. 11). Finally, we determined the catalytic efficiency of the enzymes with the natural substrate, propionyl-CoA. The catalytic efficiency varied among the different variants (Supplementary Table 3) and was decreased about twofold in GCC M5 compared with the wild type. Taken together, all of these results   Fig. 12e). The strong discrepancy in resolution and electron density between the β subunit core and the α subunits may hint at extreme flexibility within α subunits. It is therefore possible that, during catalysis, it is not just the biotin carboxyl carrier protein domain that undergoes conformational changes to bridge the distance between the active sites of the biotin carboxylase and the carboxyltransferase, as proposed by the swinging domain model 19,20 . Instead, also, the biotin carboxylase domains themselves may move or bend towards the β subunit core. In both structures, we also observed the biotin cofactor-interestingly, in a position analogous to what was reported for the crystal structure of a chimeric PCC holoenzyme 19 . The biotin is notably not located in a catalytically relevant position, as the nitrogen atom that accepts the carboxyl group is in close coordination contact with a backbone carbonyl oxygen ( Supplementary Fig. 13d). Moreover, although the biotin is inserted into the neighbouring β subunit, it is still about 10 Å away from the site of the carboxyl transfer to the CoA thioester substrate. The fact that this exact position of the biotin was observed in our cryo-EM structures, as well as in the crystal structure of a homologous enzyme from another organism, suggests that this location is a possible parking position for the cofactor. Unfortunately, we were not able to observe the supplied substrate glycolyl-CoA in the active site of the carboxyltransferase β subunits, but only free CoA ( Supplementary Fig. 13b,c), probably due to spontaneous hydrolysis of the CoA thioester bond. Nevertheless, the resolution of the β subunit core of GCC M5 was high enough to model almost 900 ordered solvent water molecules with very low B factors. This allowed us to determine the actual distances between important active-site residues, which indicated that the Leu100Ser substitution did not directly hydrogen bond to His143 as initially assumed (Fig. 2e), but probably provided space for a new hydrophilic interaction between His143 and Asp171 in GCC M5. The Tyr143His substitution that was introduced to accommodate the hydroxyl group of glycolyl-CoA was actually held in a more favourable rotamer conformation by Asp171 through an almost perfect H bond of 2.8 Å ( Fig. 2e and Supplementary Fig. 3). During random mutagenesis, two additional substitutions were introduced: Ile450Val and Trp502Arg. The Ile450Val substitution is located close to the active site in an α helix ( Fig. 2e and Supplementary Fig. 3). Compared with the wild type, the Ile450Val substitution may slightly influence interatomic distances during catalysis. In contrast, the Trp502Arg substitution is far away from the active site, close to the rotational symmetry centre of the β subunit core, where it only affects the position of the loop region between Lys496 and Lys503. Its impact on enzyme activity cannot be simply rationalized by our structures, which represent a single non-catalytic state of the enzyme.
In vitro reconstitution of the TaCo pathway. Having established and engineered all of the enzymes, we next calculated the thermodynamic profile of the TaCo pathway using component contribution 21 . Overall, the max-min driving force (MDF)-representing the minimum thermodynamic driving force via the pathway reactions after optimizing metabolite concentrations within a physiological range 22 -was above 7 kJ mol −1 for the TaCo pathway (Methods and Supplementary Fig. 14). This high MDF predicts that all pathway reactions could work with minimal backward flux (that is, <20% of the total flux) 22 and hence close to the maximum rate 23 . In fact, the tartronyl-CoA module seems to be almost as thermodynamically favourable as the Calvin-Benson-Bassham (CBB) cycle itself (MDF approaching 8 kJ mol −1 ; Supplementary Fig. 14) and more thermodynamically favourable than most central metabolism pathways, including Embden-Meyerhof-Parnas glycolysis (MDF < 2 kJ mol −1 ) 22 . Next, we aimed to confirm the functionality of the full TaCo pathway by sequentially reconstituting its reaction sequence and isotopic labelling with 13 C-bicarbonate ( Fig. 1b-d). The reconstituted TaCo pathway converted glycolate into glycerate at a rate of 27 ± 1 nmol min −1 mg −1 total protein ( Supplementary Fig. 15). To continuously operate and optimize the TaCo pathway, we coupled it to different ATP regeneration modules 2,24,25 . While the use of a polyphosphate kinase-based system was limited due to the precipitation of polyphosphate at concentrations above 20 mM, a phosphocreatine-based system proved three times more effective compared with the polyphosphate system ( Supplementary Fig. 16). With these optimized conditions, we next aimed to test the TaCo pathway for three different potential applications.
The TaCo pathway as photorespiratory bypass. As proof of concept, we first interfaced the TaCo pathway with photorespiration. Natural photorespiration yields 2-phosphoglycolate (2-PG), which is recycled back into the CBB cycle through a complicated reaction sequence of 11 enzymes releasing NH 3 and CO 2 ( Supplementary  Fig. 17a). According to theoretical calculations, replacing natural photorespiration with the TaCo pathway would allow the direct conversion of 2-PG into the CBB cycle intermediate 3-phosphoglycerate (3-PGA) with only five enzymes, while fixing an additional carbon, instead of releasing it, thereby increasing the carbon efficiency from 75% to 150% (Supplementary Table 4). Furthermore, flux balance analysis showed that the combination of the CBB cycle and the TaCo pathway requires 21% less ATP and 29% less reducing equivalents for the net formation of one 3-PGA molecule from three CO 2 molecules, compared with the CBB cycle coupled with natural photorespiration ( Fig. 3 and Supplementary Table 4). The data represent means ± s.d., as determined from n = 18 independent measurements using nonlinear regression. a EryaCS1 (Val379ala). b MePCC (asp407Ile, Tyr143His, Leu100Ser, Ile450Val and Trp502arg). c CaMCR.
Starting from the photorespiration product 2-PG, we tested the TaCo pathway together with 2-PG phosphatase and glycerate kinase (GlxK) using 13 C-labelled bicarbonate. Our synthetic pathway produced (R)-glycerate, which was further converted into labelled phosphoglycerate at a rate of 12.3 nmol min −1 mg −1 , showing that the TaCo pathway can be successfully interfaced with photorespiration ( Supplementary Fig. 18).
We further aimed at testing the TaCo pathway under conditions mimicking 100% RuBisCO oxygenation by the addition of equimolar amounts of 3-PGA and 2-PG. For quantification, we developed a malate read-out module based on isotopic labelling. The read-out module converts 3-PGA into malate, while 2-PG is only converted to malate when the TaCo pathway is active (Fig. 4a). Isotopic labelling with 13 C-bicarbonate allowed us to distinguish the fraction of malate derived from 3-PGA from that derived through the TaCo pathway from 2-PG. 3-PGA is converted into malate via one carboxylation step (that is, the carboxylation of phosphoenolpyruvate (PEP) to oxaloacetate by PEP carboxylase; Fig. 4). This carboxylation introduces one 13 C label, which leads to a single-labelled malate. The conversion of 2-PG into malate only takes place in the presence of the TaCo pathway. This requires two carboxylation reactions (and therefore double incorporation of H 13 CO 3 − ): first, the carboxylation of glycolyl-CoA to tartronyl-CoA by GCC; and second, the carboxylation of PEP to oxaloacetate by PEP carboxylase (Fig. 4). Thus, the malate formed via the TaCo pathway is double labelled while malate formed from 3-PGA is only single labelled. Malate formation increased in the presence of the TaCo pathway by 130 µM, corresponding to a surplus of 33%. 13 C labelling confirmed that the malate surplus was provided by the TaCo pathway (Figs. 1e  and 4). In a control experiment, in which the TaCo pathway was disconnected from the read-out module by omission of GlxK, no additional malate was formed, but 139 µM glycerate accumulated (Fig. 4d).
Notably, GlxK from E. coli only accepts the (R)-stereoisomer of glycerate 31,32 , thus confirming that the physiologically relevant stereoisomer (R)-glycerate is the product of the TaCo pathway. Overall, these results showcase the potential of the TaCo pathway as synthetic, carbon-positive photorespiratory bypass that is still able to drive CO 2 fixation, notably even in a 100% RuBisCO oxygenation reaction (that is, maximum photorespiration in the CBB cycle).
Ethylene glycol conversion via the TaCo pathway. Next, we tested the TaCo pathway in the context of ethylene glycol conversion. Ethylene glycol, a constituent of polyethylene terephthalate and a defrosting agent, is an environmental pollutant that is degraded by aerobic microbes via the glycerate pathway under CO 2 release 33,34 . We conceived of a TaCo pathway-based route for the conversion of ethylene glycol into the central metabolite glycerate that would increase the carbon efficiency of ethylene glycol assimilation from 75% to 150% (Fig. 1a). We combined GCC and TCR of the TaCo pathway with (L)-lactaldehyde dehydrogenase of E. coli (FucO 35 , which oxidizes ethylene glycol into glycolaldehyde) and an aldehyde dehydrogenase of Rhodopseudomonas palustris BisB18 (PduP 8,36 , to convert glycolaldehyde into glycolyl-CoA). A first version of the synthetic pathway produced 77 µM glycerate over 2 h with an initial rate of 0.4 nmol min −1 mg −1 in vitro ( Supplementary Fig. 19). To improve the initial ethylene glycol oxidation, we replaced FucO with Gox0313 from Gluconobacter oxydans 37 (Supplementary  Table 1) and introduced a water-forming NADH oxidase 38 to maintain a high NAD + concentration. Together with the integration of an efficient ATP regeneration system, glycerate production was increased to 485 µM at a fivefold higher rate of 2.1 nmol min −1 mg −1 , demonstrating efficient CO 2 -dependent conversion of the environmental pollutant and plastic component ethylene glycol into a central carbon metabolite through the TaCo pathway ( Fig. 1f and Supplementary Fig. 19). Synthetic CO 2 fixation by the TaCo pathway. Finally, we tested the TaCo pathway in the context of the crotonyl-CoA/ ethylmalonyl-CoA/hydroxybutyryl-CoA (CETCH) cycle, a recently developed synthetic CO 2 fixation pathway 2 . The primary CO 2 fixation product of the CETCH cycle is glyoxylate, which can be converted into glycolate using glyoxylate reductase (Fig. 1a). Extending the CETCH cycle with the TaCo pathway would add another CO 2 fixation step and allow direct formation of the central carbon metabolite glycerate from three CO 2 molecules. Thus, the use of GCC directly increases carbon efficiency of the CETCH cycle by 25% (Supplementary Table 5). For proof of principle, we coupled the CETCH cycle and TaCo pathway and optimized their interplay in several rounds. When we initially combined the 17 enzymes of CETCH (version 5.4) 2 with a semialdehyde reductase for glyoxylate reduction (Gox1801 from G. oxydans) and the TaCo pathway, the TaCo pathway produced glycerate at rates of 0.1 and 0.4 nmol min −1 mg −1 in continuous and stepwise assays, respectively ( Supplementary Fig. 20). Further analysis identified two potential bottlenecks of the overall system. First, the use of formate and formate dehydrogenase to regenerate NADPH led to production of the dead-end metabolite formyl-CoA due to a side reaction of GCS, trapping CoA and consuming ATP. Second, succinyl-CoA reductase of Clostridium kluyveri (CkSucD), which was used in the CETCH cycle, had a side reactivity with glycolyl-CoA of 311 ± 21 nmol min −1 mg −1 . To optimize the system, we switched to glucose dehydrogenase for NADPH regeneration and replaced CkSucD with a homologue from Clostridioides difficile (CdSucD) that specifically converted glycolyl-CoA only at 53 ± 21 nmol min −1 mg −1 . Furthermore, we used higher concentrations of phosphocreatine for enhanced ATP regeneration. Our final set-up produced 331 µM glycerate at an almost 50-fold higher rate of 4.8 nmol min −1 mg −1 TaCo enzymes (Fig. 1g and Supplementary   Fig. 20), which is comparable to CO 2 fixation rates of the CETCH cycle alone (5 nmol min −1 mg −1 ) 2 . These results demonstrate successful combination of the TaCo pathway with the synthetic CETCH cycle into a more carbon-and energy-efficient autotrophic pathway.

Conclusions
We have constructed a new-to-nature metabolic reaction sequence for the CO 2 -dependent assimilation of glycolate and demonstrated its use for different applications. To develop the TaCo pathway, it was necessary to identify and engineer enzymes able to catalyse reactions, which to the best of our knowledge are not known from any natural pathways. A crucial step was the engineering of a new-to-nature CO 2 -fixing enzyme, GCC, which shows great potential to improve photosynthetic yield in natural and synthetic carbon fixation, even under very low CO 2 conditions and with 100% oxygenation of RuBisCO 8 . Our engineered enzymes of the TaCo pathway show turnover numbers (1.4-11.1 s −1 ) that are comparable to naturally evolved enzymes, which show on average a k cat of 10 s −1 (ref. 39 ), while most RuBisCOs notably fall in the range of 1-10 s −1 (ref. 40 ). Nevertheless, to unlock the full potential of the TaCo pathway, future efforts might involve directed evolution to further improve enzyme activities, enhance flux through the pathway and potentially eliminate the unfruitful ATP hydrolysis of GCC.
Overall, our results showcase how hitherto unknown but theoretically feasible enzyme reactions can be developed on the scaffold of naturally existing proteins to extend the solution space of natural metabolism [41][42][43][44][45] . We expect similar future approaches to allow access to novel routes with carbon and energy efficiencies superior to naturally evolved pathways 46 , which may greatly impact current efforts in biotechnology and metabolic engineering.

Methods
Materials. Chemicals were obtained from Sigma-Aldrich, Carl Roth, Santa Cruz Biotechnology and Merck. NaH 13 CO 3 was obtained from Cambridge Isotope Laboratories. Biochemicals and materials for cloning and protein expression were obtained from Thermo Fisher Scientific, New England Biolabs and Macherey-Nagel. Coenzyme A was bought from Roche Diagnostics. Materials and equipment for protein purification were obtained from GE Healthcare, Bio-Rad and Merck Millipore. Pyruvate kinase/lactic dehydrogenase, malic dehydrogenase, glucose-6-phosphate dehydrogenase, glucose dehydrogenase and PEP carboxylase were bought from Sigma-Aldrich.
Library generation of GCC variants. Plasmid libraries of randomly mutagenized GCC were created by megaprimer-based whole-plasmid PCR (MEGAWHOP) 47 .
To generate randomized fragments of the β subunit of GCC M3 (pTE1412) for the first round of mutagenesis, error-prone PCR 48 Table 6), 10% (vol/vol) dimethyl sulfoxide, 50 ng template DNA of pTE1412 and 200-500 µM MnCl 2 in a 50 µl reaction. The randomized fragments were digested with DpnI, purified by agarose gel electrophoresis and used as mega primers for a whole-plasmid PCR (MEGAWHOP), as described elsewhere 47 , or subjected to another error-prone PCR reaction to further increase the mutation rate. The MEGAWHOP reaction (50 µl) contained 1× KOD Hot Start reaction buffer (Novagen), 0.2 mM dNTPs, 1.5 mM MgSO 4 , 500 ng megaprimer, 50 ng template plasmid (GCC M3; pTE1412) and 2.5 U KOD Hot Start DNA polymerase (Novagen). The MEGAWHOP product was purified, digested with DpnI, again purified, and transformed into ElectroMAX DH5α (Thermo Fisher Scientific) to ensure a high number of transformants in the resulting libraries.
To estimate the mutation rate for the different concentrations of MnCl 2 used in the error-prone PCR, the plasmids of six randomly picked clones after MEGAWHOP were purified, sequenced and analysed for nucleotide exchanges. Library 1_1 contained 3.6 mutations per kilobase pair (kbp) (two times 500 µM MnCl 2 in subsequent error-prone PCR); library 1_2 contained 1.1 mutations per kbp (500 µM MnCl 2 ); and library 1_3 contained 0.2 mutations per kbp (200 µM MnCl 2 ).
For the second round of random mutagenesis, plasmid pTE3100 (GCC M4) was used as a template for error-prone PCR as well as for MEGAWHOP, following the procedure described above. Here, we aimed to create libraries with mutation rates of 0.2 and 1.1 per kbp and used 200 and 500 µM MnCl 2 , respectively.
Screen of GCC libraries in microtitre plates. The libraries 1_1, 1_2 and 1_3 were transformed into E. coli BL21_BirA (for the bacterial strains used, see the Supplementary Methods) and colonies were picked into 96-deep-well plates (PlateOne) with lysogeny broth (Miller) containing 100 µg ml −1 ampicillin and 50 µg ml −1 spectinomycin. The plates were incubated over night at 37 °C with subsequent transfer into fresh 96-deep-well plates with lysogeny broth (Miller), 100 µg ml −1 ampicillin, 50 µg ml −1 spectinomycin and 2 µg ml −1 biotin to an OD 600 of 0.1. Protein expression was induced with 0.25 mM isopropyl β-d-1-thiogalactopyranoside at an OD 600 of 0.4-0.6 and the cells were incubated over night at 25 °C. The cells were lysed using CelLytic B (Sigma-Aldrich) and stored in 20% glycerol at −80 °C. The enzyme activity was measured in a plate reader by the coupled enzyme assay with purified CaMCR (TCR). We used small-volume 384-well plates (Greiner Bio-One) with 2 µl cell extract, 100 mM 3-(N-morpholino)propanesulfonic acid (MOPS)/KOH (pH 7.8), 1 mM ATP, 50 mM KHCO 3 , 500 µg ml −1 CaMCR, 1 mM NADPH, 10 mM MgCl 2 and 1 mM glycolyl-CoA in a reaction volume of 10 µl. The absorbance of NADPH was measured at 340 nm and 37 °C for 10 h with intervals of 47 s in a plate reader (Tecan Infinite M Plex).

Sequential reconstruction of the TaCo pathway.
For sequential reconstruction of the TaCo pathway, the enzymes were added one after the other to enable the detection of the intermediates of the reaction. The assay was run at 37 °C and contained 100 mM MOPS/KOH (pH 7.8), 5 mM ATP, 10 mM MgCl 2 , 2 mM coenzyme A, 4 mM NADPH, 50 mM NaH 13 CO 3 , 2 mM glycolate, 200 mM phosphocreatine, 10 U ml −1 creatine phosphokinase (CPK) and 50 µg ml −1 myokinase. Before starting the reaction, a t 0 sample was taken and immediately quenched with 1% HCl. All subsequent samples were treated in the same way.
To start the reaction, 0.5 mg ml −1 GCS was added. After 20 min, the next sample was taken, followed by the addition of 0.6 mg ml −1 GCC M4. After 20 min, the next sample was taken, followed by the addition of 0.7 mg ml −1 TCR. After 20 min, the final sample was taken. Samples were analysed for glycolyl-CoA and tartronyl-CoA by ultraperformance liquid chromatography-high-resolution mass spectrometry (UPLC-hrMS) and for glycerate using the derivatization method via UPLC with tandem mass spectrometry (UPLC-MS/MS) (for detailed UPLC-MS/MS methods, see the Supplementary Methods).
Cryo-EM sample preparation and data collection. For cryo-EM sample preparation, 4.0 µl of the purified complex (1 mg ml −1 ) containing 1 mM β,γ-imidoadenosine 5′-triphosphate, 2 mM MgCl 2 and 4 mM glycolyl-CoA was applied to glow-discharged Quantifoil 2/1 grids, blotted for 3.5 s with force 4 in a Vitrobot Mark IV (Thermo Fisher Scientific) at 100% humidity and 4 °C and then plunge frozen in liquid ethane that had been cooled by liquid nitrogen.
For MePCC, cryo-EM data were collected on an FEI Glacios transmission electron microscope operated at 200 keV using the SerialEM software package 49 . A total of 2,637 video frames were recorded at a calibrated pixel size of 1.18 Å using a magnification of 36.000×. The total dose was 50.2 e − Å −2 , distributed over 40 frames. Micrographs were recorded in a defocus range of −0.5 to −3.5 μm.
All subsequent imaging processing steps were performed in the cryoSPARC software package. The dose-fractionated videos were gain-normalized, aligned and dose-weighted using the patch motion algorithm. Defocus values were estimated using the CTFFIND4 implementation, and 168,488 particles were automatically chosen using a reference-free blob picker. Particle sorting and reference-free two-dimensional classification were performed to remove non-particle candidates and damaged particles. Ab initio model generation using the stochastic gradient descent algorithm was used to prevent any model bias using a completely data-driven starting model. The particles were further classified using heterogeneous refinement into three classes. The best-aligning class was subsequently subjected to three-dimensional refinement yielding 3.67 Å global resolution and a B factor of −99 Å 2 . After per-particle defocus and higher-order contrast transfer function (CTF) correction in the Refinement_New algorithm, the resolution improved to 3.48 Å with a B factor of −86.8 Å 2 .
For GCC M5, cryo-EM data were acquired with an FEI Titan Krios transmission electron microscope using SerialEM software 49 . Video frames were recorded at a nominal magnification of 105,000× (calibrated physical pixel size: 0.8512 Å px −1 ) using a K3 direct electron detector (Gatan) in correlated double sampling mode and a GIF quantum energy filter (Gatan) at a 20-eV slit width. The total electron dose of ~55 e − Å −2 was distributed over 35 frames. Micrographs were recorded in a defocus range of −0.5 to −2.7 μm.
Cryo-EM micrographs were processed on the fly using the Focus software package 50 if they passed the selection criteria (iciness < 1.05; drift 0.4 Å < x < 70 Å; defocus 0.5 µm < x < 5.5 µm; estimated CTF resolution < 5 Å). Micrograph frames were aligned using MotionCor2 (ref. 51 ) and the CTF was determined using GCTF 52 . Using Gautomatch (http://www.mrc-lmb.cam.ac.uk/kzhang/), 5,398,283 particles were picked template free on 20,057 acquired micrographs. Candidate particles were extracted with a pixel box size of 340 using RELION 3.1 (ref. 53 ) and cleaned using reference-free two-dimensional classification. A total of 2,863,953 particles were imported into cryoSPARC 54 , used for ab initio construction of initial models and subjected to multiple rounds of heterogeneous refinement to obtain the best-aligning 2,181,317 particles. Non-uniform refinement resulted in a reconstruction with an estimated resolution of 2.25 Å and a temperature factor of −83.9 Å 2 . Using several rounds of per-particle defocus-estimated and higher-order CTF refinements, the final refinement yielded a global resolution of 1.96 Å and an improved temperature factor of 71.7 Å 2 (gold-standard FSC analysis of two independent half-sets at the 0.143 cut-off). Local resolution estimation revealed a local resolution of 1.9 Å in the β subunit core of the molecule and a flexible amino-terminal α subunit. Visual inspection of the resulting electron density map revealed protein features that were consistent with the determined high resolution (for example, holes in the aromatic ring systems and the zigzag structure of extended sidechains such as Arg and Lys). Additionally, in the density maps, coordinated water molecules were visible ( Supplementary Fig. 12).
Structural modelling and analysis. Homology modelling was performed using SWISS-MODEL 13 . As a template for homology modelling of EryACS1, the structure of an ACS mutant of S. enterica 12 (PDB ID 2P2B; 62% amino acid identity) was used, which had an adenosine-5′-monophosphate-propyl ester (propyl-AMP) bound-an analogue of the reaction intermediate propionyl-AMP. Based on the position of the propyl-AMP in the S. enterica ACS, a glycolyl-AMP was modelled into the active site of the EryACS1 homology model.
For initial homology modelling of the β subunit of MePCC, the structure of the β subunit of PCC from Ruegeria pomeroyi (PDB ID 3N6R; 72% amino acid identity) 19 was used.
Initial cryo-EM map fitting was performed in UCSF Chimera version 1.14 (ref. 55 ) using homology models based on PDB ID 3N6R 19 . Automatic refinement of the structure was done using phenix.real_space_refine of the PHENIX 1.17.1-3660 suite 56 . Manual refinements and water picking were performed with COOT 0.8.9.2 (ref. 57 ).
All structural depictions were created using PyMOL (the PyMOL Molecular Graphics System; version 1.8; Schrödinger). Modelling of propionyl-CoA and glycolyl-CoA into the active sites of the MePCC and GCC, respectively, was based on the position of methylmalonyl-CoA in the structure of a methylmalonyl-CoA carboxytransferase from Propionibacterium freudenreichii 58 (PDB ID 1ON3; 52% amino acid identity). Manual fitting and adjustments of the CoA thioesters reflecting differences in active-site architectures were done with COOT and PyMOL.
Flux balance analysis. To compare the natural and synthetic photorespiration pathways in terms of consumption of ATP, NAD(P)H and ferredoxin, as well as required turns of RuBisCO, a stoichiometric analysis was performed by applying flux balance analysis with COBRApy 59 . Simplified models were constructed including the reactions of CBB, the specific reactions of each photorespiration pathway and cofactor regeneration reactions (for example, ADP + inorganic phosphate → ATP, NAD + → NADH and NADH + NADP + → NAD + + NADPH), as well as key interconversion reactions, such as adenylate kinase. The ratio between the carboxylation and oxygenation reactions catalysed by RuBisCO was set to 3:1 (25% oxygenation) 60 . The stoichiometric requirement of ATP, NADPH, ferredoxin and RuBisCO turns (that is, the sum of carboxylation and oxygenation reactions) was computed for the production of one 3-PGA molecule. The full code and the list of reactions of each model can be found at https://github.com/he-hai/PubSuppl within the 2020_TaCo directory.
Pathway thermodynamics analysis. MDF analysis 22 was applied to evaluate the thermodynamics feasibility of the tartronyl-CoA module. The Python packages equilibrator_api, equilibrator_assets and equilibrator_pathway were used for the analysis. The change in Gibbs energy of the reactions was estimated using the component contribution method 22 . CO 2 was considered as the substrate for the carboxylation reactions as its concentration is pH independent, unlike that of bicarbonate, thus simplifying the calculations. Metabolite concentrations were constrained to the range 1 µM to 10 mM, as descripted previously 22 . The pH was assumed to be 7.0, the ionic strength was assumed to be 0.25 M and -log[Mg 2+ ] (pMg) was assumed to be 3. The scripts and details can be found at https://github. com/he-hai/PubSuppl within the 2020_TaCo directory.
Reporting Summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
The structural data that support the findings of this study have been deposited in the PDB with the accession codes 6YBP (MePCC wild type) and 6YBQ (GCC M5), and in the Electron Microscopy Data Bank with the accession codes EMD-10770 (MePCC wild type) and EMD-10771 (GCC M5). Other publicly available PDB structures used in this study are 2P2B, 3N6R and 1ON3. Other data supporting the findings of this study are available in the paper and Supplementary Information, or from the corresponding author upon reasonable request.

Code availability
For flux balance analysis of different photorespiratory pathways, the full code and list of reactions of each model, as well as the scripts and details of the pathway thermodynamics analysis, can be found at https://github.com/he-hai/PubSuppl within the 2020_TaCo directory.