De novo biosynthesis of rubusoside and rebaudiosides in engineered yeasts

High-sugar diet causes health problems, many of which can be addressed with the use of sugar substitutes. Rubusoside and rebaudiosides are interesting molecules, considered the next generation of sugar substitutes due to their low-calorie, superior sweetness and organoleptic properties. However, their low abundance in nature makes the traditional plant extraction process neither economical nor environmental-friendly. Here we engineer baker’s yeast Saccharomyces cerevisiae as a chassis for the de novo production of rubusoside and rebaudiosides. In this process, we identify multiple issues that limit the production, including rate-liming steps, product stress on cellular fitness and unbalanced metabolic networks. We carry out a systematic engineering strategy to solve these issues, which produces rubusoside and rebaudiosides at titers of 1368.6 mg/L and 132.7 mg/L, respectively. The rubusoside chassis strain here constructed paves the way towards a sustainable, large-scale fermentation-based manufacturing of diverse rebaudiosides.

C onsidering the prevalence of non-communicable diseases (NCDs), World Health Organization published a global action plan for the prevention and control of NCDs 2013-2020 1 . In the plan, the limitation of excessive calorie intake and the reduction in the content of free and added sugars in food and beverages have been proposed. It is known that a highcaloric/sugar diet increases the risks of NCDs, including diabetes, hypertension, and kidney disease. The sweet taste is the basic taste that animals like the most 2 , but traditional sweeteners are carbohydrates with high-calorie content. Low-calorie sweeteners can be used as sugar substitutes, which can help reduce sugar consumption. In the past, because of safety concerns, artificial sweeteners like saccharine and aspartame, were questioned by consumers and never fully replaced sucrose. More recently, safe and natural sweeteners, with high sweetness and low calories such as steviol glycosides (SGs) have been proposed and commercialized 3 .
As a member of the SGs, rubusoside is about 114-fold sweeter than sucrose 4 . In addition to as a sweetener, Zhang et al. reported that rubusoside was a natural solubilizer of many anti-cancer drugs (such as botulinic acid 5 ). The derivatives of rubusoside, rebaudioside A (Reb A), rebaudioside D (Reb D), and rebaudioside M (Reb M) have even higher sweetness intensities (200-350 times more than sucrose) 6 . Reb D and Reb M are especially interesting because of their better taste with no bitter aftertaste. Traditionally, rubusoside and rebaudiosides have been extracted from Rubus suavissimus (Chinese sweet leaf tea) and Stevia rebaudiana, respectively. However, their large-scale production is limited by the long growth cycle of these plants, their low titer, and their complex extraction process, which makes the production neither economical nor environmental-friendly 7 . Alternatively, synthetic biology-based microbial production, presents shorter process cycles, high efficiency, and simpler extraction processes 8,9 . Recently, synthetic biology approaches to produce rubusoside and rebaudiosides have been developed. However, since the initial enzymatic production of rubusoside and rebaudiosides 3, [10][11][12][13] have been mainly restricted by high substrate cost, microbial de novo biosynthesis is now preferred. Wang et al. built a Reb A producer strain in Escherichia coli 14 , proving the concept feasible. However, synthesis efficiency was low, partially because some of the enzymes in the pathway were poorly expressed in prokaryotic cells, such as the cytochrome P450s. Compared with prokaryotic counterparts, the Generally Recognized as Safe (GRAS) chassis organism Saccharomyces cerevisiae is considered a superior host for expressing plant-derived P450s 15 . Olsson et al. reported that Reb D and Reb M can be synthesized in S. cerevisiae. They found that the UDP-dependent glycosyltransferases (UGTs) UGT76G1 is an enzyme with poor specificity and they engineer it to increase the accumulation of Reb D or Reb M, and reduce the synthesis of the unwanted side-products 16 . However, the rebaudiosides production was still low. According to previous reports, there are three main factors that limit the efficient production of rubusoside and rebaudiosides in yeast. The first one is the presence of rate-limiting steps in the biosynthesis pathway that must be overcome, to avoid that the precursor of diterpenes in the MVA pathway is insufficient, and that plant-derived P450s and UGTs exhibit low catalytic efficiencies 2,3,15 . The second factor is related to the need to improve the tolerance of yeast to rubusoside and rebaudiosides, as terpene accumulation often leads to cytotoxicity or growing pressure 17 . The third factor is related to the complexity of the metabolic network, for which it is important to identify and manipulate genetic targets to optimize it and maximize production 18 . Therefore, to develop more efficient rubusoside and rebaudiosides-producing chassis, it is necessary to address the challenges described above.
Here, we establish a S. cerevisiae chassis for an efficient de novo biosynthesis of rubusoside and rebaudiosides. We remodel the complex metabolic networks by a modular engineering approach, which enables the production of rubusoside and rebaudiosides at titers of 1368.6 and 132.7 mg/L in 15-L bioreactors, respectively. We engineer the substrate channeling system to improve the catalytic efficiency of the P450s. We mine and reinforce the active efflux system and stress-responding regulator to promote the secretion and synthesis of these sweeteners. In addition, we use in silico prediction tools based on genome-scale metabolic models to redistribute and optimize the metabolic networks. We expect that the engineered yeast strain here developed serves as a starting step towards the sustainable, large-scale production of rubusoside and its derivatives.

Results
Construction of de novo rubusoside biosynthetic pathway. In order to build a yeast chassis with the capacity to produce high levels of rubusoside, we divided its complex metabolic pathway into engineering modules (Fig. 1a). In plant, ent-kaurene is synthesized from geranylgeranyl pyrophosphate (GGPP) by copalyl diphosphate synthases and copalyl diphosphate synthases. Interestingly, a kaurene synthase (KS) from Gibberella fujikuroi 19 can directly generate ent-kaurene from GGPP. Therefore, to avoid intermediates loss caused by multi-step reactions, we firstly inserted this KS into the genome of S. cerevisiae CEN.PK2-1C, represents Module A (terpene synthesis module). In the resultant strain SGN01, ent-kaurene was detected, although at very low titers ( Fig. 1b and Supplementary Fig. 1). To enhance its production, the two well-known limiting enzymes in the MVA pathway, tHMG1 (a truncated hydroxymethylglutaryl-CoA reductase 20 ) and IDI1 (isopentenyl diphosphate delta-isomerase), were overexpressed 21 , resulting in strain SGN02. In addition, the precursor (farnesyl diphosphate, FPP) can be diverted towards the production of other metabolites such as ergosterol 22 , which is necessary for cell growth. GGPP is synthesized from IPP and DMAPP via GPP and FPP in S. cerevisiae (Fig. 1a). A mutant farnesyl pyrophosphate synthase (FPS F112A ) was engineered to synthesize GGPP from IPP (isopentenyl diphosphate) and DMAPP (dimethylallyl diphosphate) directly 19 , which can reduce the competition for FPP. Therefore, FPS F112A was introduced into SGN02 strain in order to increase the GGPP pool, creating the strain SGN03. Compared with SGN01, the area of the peak corresponding to ent-kaurene in the strains SGN02 and SGN03 increased to 7.0-and 33.9-times, respectively (Fig. 1b).
For the rubusoside synthesis module (Module C), two UDPglycosyltransferases (UGTs) UGT74G1 and UGT85C2 from S. rebaudiana were integrated into the strain SGN05, generating the strain SGN06. was discovered in the strain SGN06 (Fig. 1c, Fig. 1d), which corresponded to the rubusoside standard ( Supplementary Fig. 4), indicating that de novo biosynthesis pathway of rubusoside and its producer chassis strain were successfully built. The rubusoside titer was 4.5 mg/L in SGN06, and most of it was found outside of the cell (Fig. 1e). All these products of the heterologous pathway were not detected in the original strain S. cerevisiae CEN.PK2-1C (Fig. 1b, c).
Identification and elimination of rate-limiting steps. In strain SGN05, steviol titer was only 5.3 mg/L, and the titer ratio of steviol to its precursor EKA was approximately 1: 8 ( Supplementary Fig. 5), indicating that synthesizing steviol from EKA is a key rate-limiting step in Module B (Fig. 2a). It is difficult to determine whether KO limits the metabolic flux from entkaurene to EKA, due to the lack of the commercial ent-kaurene standard. Therefore, to overcome the rate-limiting issues, all the enzymes in Module B were investigated, including the two P450s (KO and KAH) and their reductase (CPR1). First, the individual genes were integrated into SGN05. Compared with SGN05, steviol titer increased by 91.3% in CYP-01 (overexpressing KO),  37.1% in CYP-02 (overexpressing KAH), and 29.8% in CYP-03 (overexpressing CPR1) (Fig. 2b). Previous studies showed that plant P450s naturally anchored to the endoplasmic reticulum (ER) 25 . We found that KAH and CPR1 contain a transmembrane domain (TMD) ( Supplementary Fig. 6, Supplementary Fig. 7), and anchor to the ER in yeast (Fig. 2c); and the KAH and KO are colocalized to ER ( Supplementary Fig. 8). Then, to investigate whether the cytoplasmic expression of KAH and CPR1 could improve steviol synthesis, we truncated the TMD of KAH (named trKAH) and CPR1 (named trCPR1) (Fig. 2c, Supplementary  Fig. 9). Surprisingly, steviol titer increased by 231.2% in CYP-05, when the TMD of CPR1 was truncated (Fig. 2b), while steviol could not be synthesized at all when the TMD of KAH was truncated (CYP-04, Fig. 2b).
Next, to test whether the rubusoside titers change after releasing the rate-limiting step in Module B, the UGTs in Module C were integrated into the genome of CYP-11, generating the strain SGN07. In SGN07, the rubusoside titer increased to 7.2fold from 4.5 mg/L (SGN06) to 32.2 mg/L (Fig. 2e). Furthermore, to avoid a potential imbalance between ER protein synthesis load and its folding capacity, which has been known to affect the overexpression of P450s 25 , we overexpressed the ER size regulator INO2 by replacing its endogenous promoter (INO2p) with a stronger one (PGK1 promoter, PGK1p), which generated the SGN08 strain. Rubusoside titer increased to 67.7 mg/L (SGN08, Fig. 2e).
Improving the yeast adaptation to rubusoside by in silico prediction and experimental validation. Active efflux is a common method of adaptation to harsh environments in fungi, which is also used to export secondary metabolites and hostderived antimicrobial compounds 29 . Because most of the rubusoside was found outside the yeast cell (Fig. 1e), we speculated that an active efflux system may exist in the yeast plasma membrane (PM) to export rubusoside, which we named here Module E. We firstly screened for a rubusoside efflux pump analyzing known exporters in yeast PM, including those from the ATP-binding cassette (ABC) transporter family, the Multidrug and toxic compound extrusion (MATE) protein family, the Major Facilitator Superfamily (MFS) transporter family, and some other possibly related transporters. Their protein structures were taken from the Alpha Fold Protein Structure Database (https:// alphafold.ebi.ac.uk/), and docked with rubusoside. According to our results ( Supplementary Fig. 11), the affinity of rubusoside with the ABC transporters was higher than with other exporters, and rubusoside could be pulled into most of the ABC transporters channel. Thus, we speculated that ABC transporters may play vital roles in rubusoside secretion. In order to prove that, we used five common inhibitors to destroy the ABC transporters function in yeast, including reserpine, Carbonyl Cyanide m-Chlorophenylhydrazone (CCCP), PAβN, tariquidar, and dexamethasone (DMS), and we found that CCCP and reserpine can weaken the secretion of rubusoside ( Supplementary Fig. 12, Supplementary Fig. 13), indicating that ABC transporters participate in the export of rubusoside.
was increased by 34.0% to 90.7 mg/L in SGN09 strain (YOR1), 129.8% to 155.6 mg/L in SGN10 (PDR11), and 10.1% to 74.5 mg/ L in SGN11 strain (PDR12) (Fig. 3b). Moreover, biomass was higher than that of strain SGN08 (Fig. 3c). In addition, the irregular cell morphology in SGN08 strain, which was caused by the inefficient export of rubusoside, returned to oval (Fig. 3d, Supplementary Fig. 16). Overall, the results indicate that PDR11 is the major efflux pump to mediate rubusoside export.
To further improve rubusoside production, the efflux-pump PDR11 and the stress-response factor MSN4 were simultaneously overexpressed in SGN08, generating the strain SGN13. However, we found that the rubusoside titer was 207.0 mg/L (SGN13, Fig. 3b), which was almost the same as in SGN12 (205.5 mg/L). Based on the RT-qPCR results, PDR11 was upregulated by MSN4 ( Supplementary Fig. 17). This suggests that transport is not limiting production and rubusoside titer can be potentially further improved by increasing the metabolic fluxes towards the rubusoside synthesis pathway.
In silico prediction of engineering targets by genome-scale metabolic models. To maximize the metabolic flux towards rubusoside synthesis, we combined the genome-scale metabolic model (GSMM) with the in silico prediction tool OptKnock, and determined engineering gene targets. Based on the GSMM yeast 8.4.0 36 , we first expanded the model by adding the reactions in Module A-Module C. Then, it was used for in silico simulations by constraint-based flux balance analysis (FBA) 37,38 and knockout targets were identified by OptKnock 39 . As a result, five knockout targets were predicted (Fig. 4a, Fig. 4b); GAL7, ABZ2, ALT1, ALT2, and ARO8. These five genes were individually deleted in SGN13, and the resulting strains were named SGN14 (ΔGAL7), SGN15 (ΔABZ2), SGN16 (ΔALT1), SGN17 (ΔALT2), and SGN18 (ΔARO8). Figure 4c shows that the rubusoside titer increased by 19.4% (247.2 mg/L, SGN14) when GAL7 was knocked out. As shown in Fig. 4b, blocking GAL7 may favor the conversion of glucose-1-phosphate (Glc-1P) into UDP-glucose synthesis, suggesting that UDP-glucose concentration may be a rate-limiting factor for rubusoside production in SGN13. In the other knockouts, rubusoside titer was not markedly improved (Fig. 4c).
Biosynthesis of rebaudiosides using the rubusoside producing chassis. To explore the potential of yeast to produce rebaudiosides, we designed the rebaudioside synthesis module (Module F, Fig. 5a) to be used in the SGN08 strain. In plants, the complex metabolic networks are connected by two UGTs, UGT91D2 and UGT76G1 16 . It has been reported that substrate specificity of UGT91D2 and UGT76G1 is poor, and it is hard to convert rubusoside into rebaudiosides 41 . Previous works have reported higher catalytic activities of the mutants UGT91D2-e0e (V286A and L211M) and UGT91D2-e0w (V253I and T464A) 14 , and higher product specificity of mutant UGT76G1-MUT (T284S and I203V) 16 .
To enhance the expression level and control of UGT76G1-MUT, we replaced the GAL10p with a stronger and inducible promoter DDI2p 43 , generating the strain M23. Here, we controlled UGT76G1-MUT by inducing its expression at different times, 0, 6, 12, 24, and 48 h. We found that rebaudiosides titers were 15.8 mg/L (M23-0 h), 12.9 mg/L (M23-6 h), 11.5 mg/L (M23-12 h), 2.9 mg/L (M23-24 h), and 1.2 mg/L (M23-48 h) (Fig. 5c). Therefore, in the following experiments, the expression of UGT76G1-MUT was induced from the beginning of the fermentation. In addition, to detect whether other types of SGs are synthesized in our rebaudiosides-producing chassis, Progenesis QI v2.4 software was used to analyze the products in the strain M23. According to the identified results, in addition to stevioside, Reb A, Reb D, and Reb M (Fig. 5b), the M23 strain was also producing, 1,2-bioside, Reb B, and Reb E (Fig. 5d). Surprisingly, Reb N, a minor SGs, whose metabolic pathway has not been yet reported, was also identified (Fig. 5d,  Supplementary Data 1). Then, to further improve rebaudiosides production, the stress-responsive regulator MSN4 was overexpressed in the M23 strain. The rebaudiosides titers in the shake-flask increased to 35.2 mg/L (M24 strain) (Fig. 5c), in which the titers of Reb A, Reb D, and Reb M were 6.2, 11.4, and 17.6 mg/L, respectively. Finally, the rebaudiosides titer was raised to 132.7 mg/L in the M24 strain using a 15-L bioreactor and a fed-batch fermentation (Fig. 5e), in which the titers of Reb A, Reb D, and Reb M were 21.5, 44.2, and 67.0 mg/L, respectively.

Discussion
Rubusoside and rebaudiosides are important sugar substitutes with human health-related benefits 44 , which are widely used in the food, pharmaceutical, and beverage industries 6,45 . Their production using microbial cell factories would enable a faster, cheaper, and more sustainable process. Here, we proved that the S cerevisiae CEN.PK2-1C (without the ability to produce SGs) can be engineered to produce rubusoside and rebaudiosides from available carbon sources, such as glucose and ethanol. We demonstrated that using a modular engineering approach it is possible to produce rubusoside at a gram-per-liter level, and rebaudiosides at more than 100 mg/L, with the content of Reb D and Reb M much higher than that of Reb A.
S. cerevisiae has generally been considered a superior host for expressing plant-derived P450s; however, in cases with low catalytic activity or expression level, their efficiency still needs to be improved 15 . Although the catalytic activity of KAH has been improved by its co-expression with CPR1 24 , the steviol titer was still limited by the transformation of EKA into steviol (Supplementary Fig. 5). Here, we present that a substrate channeling system is an excellent strategy to enhance P450s efficiency. By this approach, the steviol titer in the strain CYP-11 reached 40.6 mg/ L, with a ratio of steviol to EKA increased from 1:8 (SGN05) to 1:1.2 ( Supplementary Fig. 24). In addition, the difference in steviol titer between CYP-11 and the original strain was significant (P < 0.001, Fig. 2b). Besides, we found that besides KAH, KO is also a rate-liming enzyme in the P450s module, because after adding KO copies steviol titer was enhanced in CYP-01, and the ent-kaurene accumulation in CYP-11 strain was decreased by 55.7% than SGN05 strain (Supplementary Fig. 25). At the same time, we demonstrated that TMD is essential for the catalytic activity of KAH, but not for CPR1. After truncating the TMD, the highest expression level of trCPR1 was 3.0-fold higher than CPR1 (Supplementary Fig. 26). These strategies for engineering P450s may be applied to improve the production of other natural products in microbes. In addition, they can also be used to solve common problems in the construction of other natural product chassis cells, such as the low substrate transport efficiency caused by the spatial distance of key enzymes, and the low synthesis efficiency caused by compartmentalization.
Another common metabolic engineering challenge is the identification of non-obvious gene targets to optimize the metabolic network and maximize production 48 . Traditionally, the target genes need to be blindly screened from the metabolic networks 49 , but this strategy is time-consuming. GSMMs have become important tools for studying metabolism 50 , which can provide potential metabolic targets and the best metabolic engineering strategy based on model simulations. OptKnock is an algorithm that predicts knockout genes to overproduce the target products under the premise of maximizing the cell biomass 39 . With the help of OptKnock, GAL7 was identified to be involved in a competing reaction sharing the precursor Glc-1-P of UDPglucose synthesis pathway. This indicated that the rubusoside synthesis was limited by the pool of UDP-glucose in the strain SGN13. Insufficient UDP-glucose has been identified as a key rete-liming step in glycosides production in the synthesis of daidzein 15 . Here, we demonstrate that the combination of boosting the UDP-glucose pool and disrupting GAL7, strengthened UDP-glucose synthesis and increased product formation. These results proved the importance of the glycosyl group donor UDP-glucose on glycosides production, and provided an efficient strategy for overcoming the rate-limiting step caused by insufficient UDP-glucose in the biosynthesis of natural glycosides.
Further studies can be performed to improve UGTs through protein engineering. In particular, for Reb M producing strain M24, and while the Reb M titer was higher than the other strains, rubusoside was still accumulated (158.3 mg/L, Supplementary  Fig. 28). Therefore, we plan to focus future experiments on how to improve the transformation rate from rubusoside to Reb M, and enhance the specificities and purities of rebaudiosides using synthetic biology or protein engineering strategies. In this work, the GAL system, in which the transcriptional repressor Gal80 inhibits the activity of the transcriptional activator Gal4, was used to de-couple cell growth with product synthesis. After deleting Gal80, theoretically, GAL promoters can be only activated when glucose is depleted 51 . However, the modified GAL system only controlled by glucose has certain leakage of expression (Supplementary Fig. 14), which does not strictly de-couple the production and cell growth. Therefore, if a biosensor that responds to OD or the key precursors in rubusoside and rebaudiosides synthesis pathway can be successfully designed, it may be feasible to strictly de-couple the yeast growth and rubusoside and rebaudiosides production in the future.
In conclusion, although the biosynthetic pathways of rubusoside and rebaudiosides are complex, our study demonstrates that it is feasible to produce high amounts of rubusoside and rebaudiosides. Finally, the engineered strains and strategies described here have the potential to assist the engineering of complex biosynthetic pathways to produce other natural products.
Genetic manipulation for the construction of strains and plasmids. The heterologous gene sequences used in this study are listed in Table 1 in Supplementary Data 2. The plasmids used in this work are listed in Table 2  were synthesized by GenScript. All the plasmids were constructed in E. coli JM109 (recA1, endA1, thi, gyrA96, supE44, hsdR17Δ (lac-proAB)/F' [traD36, proAB+, lacІq, lacZΔM15]). S. cerevisiae CEN.PK2-1C was employed as the host for rubusoside and rebaudiosides-producing strains construction. S. cerevisiae CEN.PK2-1C was collected from multiple sources over the years, and stored in our lab.
The DNA manipulation was by the in vivo genetic rapid construction method 52 . All the target fragments were obtained by polymerase chain reaction (PCR) using the correct plasmid as the templates, 40-50 bp overlaps of each adjacent fragment were used to implement homologous recombination in yeast. The integration locus of chromosomal in this study was selected from the study of Reider Apel et al. 53 . The transformants with auxotroph markers were selected by the Synthetic Dropout (SD) agar plates (6.7 g/L yeast nitrogen base without amino acid, 20 g/L glucose, 50 mg/L amino acid/ uracil 200 mg/L, and 20 g/L agar). Appropriate amino acids or uracil were added to the culture medium when required (tyrosine 50 mg/L, histidine 50 mg/L, leucine 50 mg/L, or uracil 200 mg/L). After 3 days of incubation at 30°C, the positive colonies were selected by PCR. All the primers used in this study were synthesized by GENEWIZ (Suzhou, China), and listed in Table 3 in Supplementary Data 2. All the strains used in this study are listed in Table 4 in Supplementary Data 2.
Yeast cultivation and identification assays for production. For fermentation, the selected positive colonies were cultivated at 30°C, 220 rpm in 50 mL culture tubes containing 5 mL SD liquid medium. After 16-18 h incubation, cultures were transferred into 250 mL shake-flask with 25 mL YPD medium (20 g/L tryptone, 10 g/L yeast extract, and 20 g/L glucose), and cultivated at 30°C, 220 rpm for 96 h (EKA and steviol production strains), 108 h (rubusoside production strains), or 144 h (rebaudiosides production strains).
The ent-kaurene, EKA, and steviol were extracted from the cell cultures using a mixture extractant (2-propanol: n-hexane = 1: 2) in a 4: 1 ratio (organic: liquid, v/ v). The mixture was vigorously shaken with a vortex mixer for 10 min and then centrifuged at 5000 × g for 5 min to separate the two phases. The supernatant was transferred to sample vials for measurements. For extracting the rubusoside and its derivatives SGs in intracellular, the cell precipitate was broken by a high-pressure homogenizer (UH-06, Union), then the mixture was centrifuged at 5000 × g for 5 min to separate the supernatant. The fermentation supernatant was used to analyze the concentration of rubusoside and SGs extracellular.
The ent-kaurene were detected by GC-MS on GC-MS-QP2010Ultra equipped with an HP-5MS column (30 m × 0.25 mm × 0.25 μm, Shimadzu). The GC-MS operational condition was as follows: initial temperature 80°C for 1 min, followed by a temperature ramp of 15°C/min to 245°C, a ramp of 5°C/min to 300°C; the flow rate of helium was 1.2 mL/min.
Image analysis and fluorescence detection. The green fluorescent protein (GFP) was severally fused to the C-terminus of the genes KAH, CPR1, and trCPR1 by a fusion protein linker (GGGGS 3 ), and the expression cassette was inserted into pY16-TEF1p plasmid through using Gibson Assembly® Cloning Kit (NEB). The red fluorescence protein (mCherry) was fused with the residues 355-479 of the protein SEC12, which anchored to the ER in yeast 54 , and the expression cassette was inserted into pY14-TEF1p plasmid through Gibson assembly. To confirm the subcellular localization of KAH and CPR1 in yeast, the object proteins and the ER marker SEC12 (the C-terminal domain of SEC12, which includes a single TMD specifying ER localization) were individually labeled by GFP and mCherry.
Bimolecular Fluorescence Complementation (BiFC) was used to perform the spatial distance of KAH with CPR1 and trCPR1. The complete yellow fluorescent protein (YFP) was divided into two non-fluorescent fragments: nYFP (the N-terminal 1-185 amino acid residues) and cYFP (C-terminal 186-250 amino acid residues) 28 . nYFP and cYFP were fused with KAH with CPR1 or trCPR1 fragments, respectively. Then two fusion fragments containing nYFP and cYFP were fused by a linker, and the expression cassette was inserted into pY16-TEF1p plasmid through Gibson assembly. When their space is close enough, the yellow fluorescence will be emitted by 488 nm laser line.
The fluorescent protein GFP, mCherry, and YFP were excited by 488, 595, and 510 nm laser lines of an argon-ion laser line of a He-Ne laser, respectively. The fluorescence proteins were observed by Leica microscope (Leica, Mannheim, Germany) fitted with a ×60 oil-immersion objective. Fluorescence emissions were detected with spectral detector sets BP 500-525 for GFP, BP 600-625 for mCherry, and BP 520-555 for YFP. Image analysis was carried out on the Leica LAS X software package and the Image J 1.53k software (National Institutes of Health, Bethesda, MA).
Field emission scanning electron microscopy (FESEM). After fermentation 72 h in YPD, the fresh samples were centrifuged to remove the supernatant and were stored at 4°C. First, samples were fixated with 5% glutaraldehyde then those were rinsed with 0.1 M phosphate buffer solution five times. Next, the samples were dehydrated by ethanol with a gradient concentration of 30, 50, 70, 90, and 100%. Following, after the critical point drying (CPD-300, LEICA) and ion sputtering (ACE-600, LEICA), the samples can be observed by cold field emission scanning electron microscope observation (SU8220, HITACHI) Flux balance analysis. The model used in this study with some modification is based on the model yeast 8.4.0 (https://github.com/SysBioChalmers/yeast-GEM/ releases). The yeast 8.4.0 model was used and deletions of the YEL021W, YCL018W, YDR007W, and YOR202W genes were simulated by setting their corresponding reactions flux bounds to zero. The amino acid uptake rate of the model was set to be 0.1 mmol/g DCW h −1 . The model was manually curated to be functional for rubusoside production. To represent the Rubusoside biosynthesis pathway, the following reactions were added to the model: 'geranylgeranyl diphosphate → entkaurene + diphosphate', 'ent-kaurene + NADPH + O 2 → NADP + + H 2 O + entkaurenoic acid', 'ent-kaurenoic acid + NADPH + O 2 → NADP + + H 2 O + steviol', 'steviol + UDP-D-glucose → H + + UDP + 13-SMG', 'steviol + UDP-D-glucose → H + + UDP + 19-SMG', '13-SMG + H + + UDP-D-glucose→ formaldehyde + UDP + Rubusoside' and '19-SMG + H + + UDP-D-glucose→ formaldehyde + UDP + Rubusoside'. The reconstructed model was subsequently used for in silico simulations. The in silico simulations were performed by FBA, which is a constraint-based flux analysis method 37,38 . For analysis of rubusoside production, the knockout algorithms OptKnock 39 were used to identify knockout targets. The glucose uptake rate of the model was set to be 20 mmol/g DCW h −1 . The synthesis reaction of rubusoside, r_4638, was set as an objective function to identify the potential targets. The in silico simulations were performed using COBRA Toolbox v3.0 55 with Gurobi Optimizer (Gurobi Optimization Inc., Houston TX) in Matlab 2019b (The Mathworks Inc., Natick, MA, USA).
Fed-batch fermentation in 15-L bioreactor. Seed culture was carried out in 500 mL shake-flask containing 100 mL of YPD medium at 30°C with shacking at 220 rpm for 24 h. The seed cultures were inoculated (10%, v/v) into a 15-L bioreactor (T&J-Dtype; T&J Bio-engineering Co., Ltd) with a 12 L medium. The fermentation medium contained 20 g/L yeast extract, 40 g/L tryptone, and 25 g/L glucose. The agitation speed was adjusted from 300 to 800 rpm with a constant air input flow rate of 1.5 vvm, and NH 4 OH was automatically added to maintain the pH at 5.5. Feeding was conducted using a two-stage strategy. In the first stage, the feeding solution consisting of 500 g/L glucose, 9 g/L KH 2 PO 4 , 2.5 g/L MgSO 4 , 3.5 g/ L K 2 SO4, 0.28 g/L Na 2 SO 4 , 10 mL/L trace metal solution, and 12 mL/L vitamin solution 21 was used to sustain fast cell growth. In the second stage, when the strain entered the late-logarithmic growth phase, the feeding solution was changed to 400 g/L ethanol to support rubusoside and its derivatives accumulation. In the feeding stage, the feeding speed was automatically controlled by dissolved oxygenstat (DO-stat), in which the glucose and ethanol concentration was controlled at no more than 5 g/L. The fermentation temperature was 30°C.
Statistical analysis. All experiments were independently carried out at least three times, and the results were expressed as mean ± standard deviation (SD). All the statistical evaluation (p-value) was performed by a two-sided t-test in Excel (Microsoft Office 365), p > 0.05 is presented by no significance (n.s.).