E. coli strain DH5α (lab stock) was used for all molecular biology. E. coli strain BL21 Star^TM (DE3) (Novagen) was used for expression and purification of sfGFP containing a C-terminal glycosylation tag (Fisher et al., 2011) and polyhistidine tag (sfGFP^{DQNAT−6xHis}), which was used for in vitro glycosylation reactions. E. coli strain BL21 Star^TM (DE3) was also used for expression of the enzyme BirA, which was used for biotinylation of the Campylobacter jejuni OST enzyme PglB (CjPglB), and for preparing crude S30 extract. E. coli strain CLM24 (Feldman et al., 2005) was used for expression and purification of CjPglB while E. coli strain SCM6 (Wacker et al., 2002) was used for preparation of lipid-linked oligosaccharides bearing C. jejuni heptasaccharide glycans (CjLLOs) .

For both cell-free and cell-based expression of sfGFP^{DQNAT−6xHis}, the pJL1-sfGFP^{DQNAT−6xHis} plasmid (Jaroentomeechai et al., 2018) was used. Plasmid pTrc99a-BirA (lab stock) was used for expression of the BirA enzyme. Plasmid pSPI01A-CjPglB encoding CjPglB with a C-terminal AviTag was constructed as follows. First, the CjPglB^10xHis gene was PCR amplified from plasmid pSN18 (Kowarik et al., 2006a) and the resulting PCR product was then ligated between the NdeI and EcoRI restriction sites in plasmid pSPI01A (Ikonomova et al., 2016), a vector containing the AviTag after the EcoRI cut site. All plasmids were confirmed by DNA sequencing at the Biotechnology Resource Center of the Cornell Institute of Biotechnology.

Preparation of lysates containing CjPglB with a C-terminal AviTag was performed according to previously published methods (Guarino and Delisa, 2012; Jaroentomeechai et al., 2018). Briefly, a colony of E. coli CLM24 carrying plasmid pSPI01A-CjPglB was grown overnight in 5 ml of Luria-Bertani (LB) media supplemented with chloramphenicol. Overnight cultures were then subcultured into 1 L of terrific broth (TB; 24 g/L yeast extract, 12 g/L tryptone, 8 ml glycerol, 10% (v/v) 0.72 M K₂HPO₄/0.17 M KH₂PO₄ buffer) supplemented with chloramphenicol. Cells were grown at 37°C until an optical density at 600 nm (OD₆₀₀) of ∼0.6 and then induced with 100 µM isopropyl β-D-1-thiogalactopyranoisde (IPTG) for 18 h at 16°C. Cells were harvested by centrifugation, after which the pellet was resuspended in Buffer 1 (25 mM TrisHCl, 250mM NaCl, pH 8.5) and lysed using a C5 Emulsiflex homogenizer (Avestin). The lysate was centrifuged to remove cellular debris and the supernatant was ultracentrifuged at 120,000 × g for 1 h at 4°C. The resulting pellet was manually resuspended using a Potter-Elvehjem tissue homogenizer into Buffer 2 (25 mM TrisHCl, pH 8.5, 250 mM NaCl, 1% (w/v) n-dodecyl-β-D-maltoside (DDM), and 10% (v/v) glycerol). Once fully resuspended, the solution was rotated at room temperature to facilitate solubilization of the protein and then ultracentrifuged again at 120,000 × g for 1 h at 4°C.

To prepare BirA-containing lysate, BL21 (DE3) cells carrying pTrc99a-BirA were grown overnight and then subcultured into 250 ml of LB media supplemented with kanamycin. Upon reaching an OD₆₀₀ of ∼0.6, cells were induced with 100 µM IPTG for 18 h at 30°C. Cells were harvested, resuspended in Buffer 1, lysed by homogenization, and centrifuged to remove cellular debris. To prepare biotinylated CjPglB (CjPglB-biotin), CjPglB-containing lysate was mixed with BirA-containing lysate and 5 mM biotin, 10 mM MgCl₂, 10 mM ATP, and 1 EDTA-free protease inhibitor cocktail tablet (Thermo Scientific). The mixture was rotated overnight at 4°C to allow time for biotinylation. CjPglB-biotin was then enriched using HisPur Ni-NTA resin (Thermo Scientific) according to manufacturer’s recommendations and the elution fraction was desalted with buffer containing 50 mM HEPES, pH 7.5, 100 mM NaCl, 5% (v/v) glycerol, and 0.05% (w/v) DDM.

To prepare sfGFP^{DQNAT−6xHis}, BL21 (DE3) cells carrying plasmid pJL1-sfGFP^{DQNAT−6xHis} were grown overnight and subcultured in LB media supplemented with kanamycin. Upon reaching an OD₆₀₀ of ∼0.6, cells were induced with 100 µM IPTG for 18 h at 30°C. Cells were collected, resuspended in buffer containing 50 mM NaH₂PO₄, pH 8, and 300 mM NaCl and lysed as above. The sfGFP^{DQNAT−6xHis} was purified using HisPur Ni-NTA resin as above. The final protein was desalted using buffer containing 20 mM HEPES, pH 7.5, 500 mM NaCl, and 1 mM EDTA.

CjLLOs were prepared by organic solvent-based extraction according to a protocol that was adapted from previous methods (Kowarik et al., 2006b; Jaroentomeechai et al., 2018). Briefly, SCM6 cells carrying plasmid pMW07-pglΔB (Ollis et al., 2014) were grown overnight in LB media supplemented with chloramphenicol. Cells were then subcultured into 1 L of TB media, grown at 37°C until reaching an OD₆₀₀ of ∼0.7, then induced with a final concentration of 0.2% (w/v) L-arabinose for 16 h at 30°C. After induction, cells were harvested by centrifugation, the pellet re-suspended in methanol, and the cells dried for 2 days at room temperature. After drying, the cells were collected and subsequently suspended in 12 ml 3:2 (v/v) chloroform:methanol, 20 ml water, and 18 ml 10:10:3 (v/v/v) chloroform:methanol:water. After each step, sonication was used to facilitate extraction of LLOs. After the first two sonication steps, centrifugation was used to separate shorter sugars and water-soluble compounds in the supernatant from the pellet. After the final step, centrifugation was used to pellet the cellular debris and the supernatant was collected and dried at room temperature. After drying, the LLOs were resuspended in buffer containing 10 mM HEPES, pH 7.5, and 0.01% (w/v) DDM and stored at −20°C.

Microfluidic masters were made on silicon wafers according to standard photolithography protocols at the Cornell NanoScale Science and Technology Facility. Briefly, SPR220-7.0 photoresist was spun onto silicon wafers and exposed using an ABM Contact Aligner. Wafers were developed using Microposit MIF 300. Coated wafers were etched to the desired depth using a Unaxis 770 Deep Silicon Etcher, which was confirmed by using a Tencor P10 profilometer. Remaining photoresist was removed via plasma cleaning, and a coating of (1H,1H,2H,2H-perfluorooctyl) trichlorosilane (FOTS) was applied using a MVD-100 to allow for easy removal of polydimethylsiloxane (PDMS). Microfluidic devices were made by pouring degassed PDMS (mixed 1:10 with crosslinker) and curing for 5 h at 60°C. PDMS molds were cleaned with ethanol and MilliQ water, before being dried with nitrogen gas. Final devices were assembled after oxygen plasma cleaning at 700 μm for 25 s and sealed with a Piranha washed (70/30 (v/v) H₂SO₄/H₂O₂ for 10 min) glass coverslip. Devices were placed in a 70°C oven for 10 min to promote bonding of the PDMS to the glass.

S30 crude extracts for CFPS reactions were prepared using a simple sonication-based method (Kwon and Jewett, 2015). Briefly, BL21 (DE3) cells were grown in 1 L of 2xYTPG media (16 g/L tryptone, 10 g/L yeast extract, 5 g/L NaCl, 7 g/L K₂HPO₄, 3 g/L KH₂PO₄, 20 g/L glucose) and harvested upon reaching an OD₆₀₀ of ∼3.0. Cell mass was washed three times in Buffer A (10 mM Tris-acetate, pH 8.2, 14 mM magnesium acetate, 60 mM potassium glutamate and 2 mM dithiothreitol) then resuspended in a ratio of 1 ml of Buffer A to 1 g wet cell mass. The resuspended cells were sonicated with an optimal energy input (reported based on the volume obtained after resuspending cells) and centrifuged at 30,000 × g to obtain S30 extract, and the supernatant stored at −80°C. No run-off reaction was needed for the BL21 (DE3) extract.

CFPS reactions consisted of a mixture of components at a final concentration of 13 ng/μL plasmid DNA, 40% (v/v) S30 crude extract, 1.2 mM adenosine triphosphate (ATP), 0.85 mM guanosine triphosphate (GTP), 0.85 mM uridine triphosphate (UTP), 0.85 mM cytidine triphosphate (CTP), 34 μg/mL L-5-formyl-5, 6, 7, 8-tetrahydrofolic acid (folinic acid); 170 μg/ml of E. coli tRNA mixture, 130 mM potassium glutamate, 10 mM ammonium glutamate, 12 mM magnesium glutamate, 2 mM each of 20 amino acids, 0.33 mM nicotinamide adenine dinucleotide (NAD), 0.27 mM coenzyme-A (CoA), 1.5 mM spermidine, 1 mM putrescine, 4 mM sodium oxalate, 33 mM phosphoenolpyruvate (PEP), 100 μg/mL T7 RNA polymerase.

For CFPS in a microcentrifuge tube, 15-μL reactions were conducted in 1.5-ml microtubes in a 30°C incubator. For CFPS on-chip batch reactions, CFPS reaction mixtures were manually inserted into the microfluidic device using a syringe and incubated at 30°C in a moist environment to prevent evaporation. For CFPS on-chip reactions with continuous flow, two mixtures were prepared—one containing S30 crude extract and T7 RNA polymerase and the other containing the rest of the CFPS components—that when combined contained all components diluted to the final concentrations of a standard CFPS reaction. These mixtures were flown into the microfluidic device using a syringe pump where the reactants had a total residence time in each chip of 30 min.

Silane-PEG5000-biotin (Nanocs, Inc.) was dissolved in 95% (w/w) ethanol in water according to the manufacturer’s recommendations. The solution was manually pushed into the microfluidic devices and left to react for 2 h at room temperature. Devices were flushed with 100 μL of MilliQ water and then PBS at a flowrate of 10 μL/min. A solution of 100 μg/ml NeutrAvidin (Thermo Scientific) was then introduced and left to bind to the surface for 1 h. For loading of purified CjPglB-biotin, the devices were rinsed with PBS and then buffer containing 50 mM HEPES, 100 mM NaCl, 5% (v/v) glycerol, and 0.01% (w/v) DDM at pH 7.5. The purified CjPglB-biotin was then introduced into the device and allowed to bind overnight at 4°C; unbound enzyme was rinsed away before use.

For off-chip in vitro glycosylation (IVG) reactions, mixtures consisted of components at a final concentration of 17 μg/ml of purified sfGFP^{DQNAT−6xHis}, 170 μg/ml solvent-extracted CjLLOs, 10 mM MnCl₂, and 0.1% (w/v) DDM. For microcentrifuge tube reactions, IVG reaction mixtures were supplemented with 170 μg/ml purified CjPglB-biotin to a final volume of 30 μL and reactions were conducted in 1.5 ml microtubes in a 30°C incubator.

For on-chip glycosylation experiments, purified CjPglB-biotin was immobilized on the functionalized surface of the device and the IVG reaction mixture was continuously pushed through the channels using a syringe pump with a total residence time of 30 min per chip. The reaction was heated by placing the microfluidic chip on a hot plate to maintain the internal temperature of the device at 30°C and confirmed by using a thermocouple in a similar arrangement. The product was then collected at the outlet of the device and either saved for analysis or recirculated through the device again to measure the effect of increasing residence times.

The microfluidic device used for protein purification was designed with a series of posts at the end of the channel to entrap chromatography resin in the channel. For each device, we manually introduced Ni-charged profinity resin (Bio-Rad) into the channels before use. CFPS reactions expressing sfGFP^{DQNAT−6xHis} were then introduced into the inlet of the device with a total residence time of 30 min per chip and the outlet was collected and analyzed as the flowthrough fraction. The device was rinsed with PBS containing 10 mM imidazole at a flowrate of 2 μL/min and any unbound protein was collected and analyzed as the wash fraction. Finally, the target protein was eluted from the resin with PBS containing 300 mM imidazole at a flow rate of 2 μL/min. The fluorescence of each fraction was analyzed using a microplate reader to determine the GFP concentration and assayed for purity using standard SDS-PAGE with Coomassie blue staining.

For immunoblot analysis of IVG products and CjPglB-biotin, samples were diluted 3:1 in 4× NuPAGE LDS sample buffer (Invitrogen) supplemented with 10% beta-mercaptoethanol (v/v). IVG products were boiled at 100°C for 10 min while CjPglB-biotin samples were held at 65°C for 5 min. The treated samples were subjected to SDS-polyacrylamide gel electrophoresis on Bolt™ 12% and 4–12% Bis-Tris Plus Protein Gels (Invitrogen). The separated protein samples were then transferred to polyvinylidene difluoride (PVDF) membranes. Following transfer, the membranes containing IVG samples were blocked with 5% (w/v) milk in TBST (TBS, 0.1% (v/v) Tween 20) and then probed with horseradish peroxidase (HRP) conjugated anti-His antibody (1: 5,000) (Abcam, catalog # ab1187) or rabbit polyclonal serum, hR6, that is specific for the C. jejuni heptasaccharide glycan (1:10,000) (kindly provided by Markus Aebi) for 1 h. To detect hR6 serum antibodies, goat anti-rabbit IgG conjugated to HRP (1:5,000) (Abcam, catalog # ab205718) was used as the secondary antibody. The membranes containing CjPglB-biotin samples were blocked with 5% (w/v) bovine serum albumin (BSA) in TBST and then probed with ExtrAvidin-Peroxidase (1:10,000) (MilliporeSigma, catalog #E2886) for 1 h. After washing five times with TBST for 5 min, the membranes were visualized with Clarity ECL substrate (Bio-Rad) using a ChemiDoc^TM MP Imaging System (Bio-Rad).

The design of our microfluidic-based glycoprotein biosynthesis platform integrated three key processes: protein expression, protein glycosylation, and protein purification (Figure 1). In the first module of the device, sfGFP bearing a C-terminal DQNAT glycosylation motif (Fisher et al., 2011) that is optimally recognized by CjPglB (Kowarik et al., 2006a; Gerber et al., 2013) and a hexahistidine tag was expressed using crude S30 extract derived from E. coli, which enabled transcription and translation of the target protein on chip. We chose sfGFP as the acceptor protein so that the protein production and purification processes could be visualized and easily quantified during optimization of the microfluidic system. Next, in the second module, site-specific glycosylation was achieved by subjecting the newly expressed sfGFP^{DQNAT−6xHis} to components derived from a well-characterized bacterial N-linked glycosylation pathway, which occurs natively in the bacterium C. jejuni and has been functionally transferred to E. coli ³⁶ . These components included CjPglB as the glycan conjugating enzyme and CjLLOs comprised of the C. jejuni GalNAc₅(Glc)Bac heptasaccharide linked to undecaprenol-pyrophosphate as the glycan donor. CjPglB and its cognate N-glycan structure were chosen here for proof-of-concept experiments because of the high transfer efficiency that has been observed with these components both in vivo (Ollis et al., 2014; Perregaux et al., 2015) and in vitro (Guarino and Delisa, 2012; Jaroentomeechai et al., 2018). However, in a notable departure from previous works, we sought to site-specifically biotinylate CjPglB and subsequently immobilize it in the device using biotin and streptavidin interactions, thereby enabling reuse of this important membrane protein biocatalyst (Lizak et al., 2011). Lastly, in the third module, the sfGFP^{DQNAT−6xHis} product was selectively enriched using a microfluidic device loaded with affinity resin that facilitated reversible capture of the hexahistidine-tagged glycoprotein product. The modularity of the device was designed to enable optimization of each unit operation and to allow flexible biosynthesis of different glycoproteins by simply interchanging acceptor protein target plasmids, glycosylation enzymes, LLO donors, affinity tags, and chromatography resins.

For the microfluidic device design, we aimed to create a system where protein synthesis, glycosylation, and protein purification could happen continuously in series at a fixed flow rate. Therefore, we fabricated individual chips to serve as building blocks that could be serially connected to increase the residence time of a particular process as needed. To test this design, we used an etched silicon wafer as a mold to fabricate channels in polydimethylsiloxane (PDMS) that was subsequently attached to glass slides. We chose PDMS because it enabled low-cost microfluidic fabrication that was sufficiently robust for device prototyping. Each microfluidic chip involved a serpentine channel design (width = 200 μm, depth = 120 μm, volume = 11 µL) that was inspired by previous work in which a similarly designed microfluidic bioreactor resulted in enhanced CFPS productivity (Timm et al., 2015). Additionally, we hypothesized that long serpentine channels with a high surface area-to-volume ratio would promote efficient glycosylation by allowing sufficiently high levels of CjPglB enzyme to be tethered to the device, thereby increasing the probability of contact with substrates. For the purification module, an immobilized metal affinity chromatography (IMAC) strategy was implemented whereby 25-µm posts were spaced apart from one another at the outlet of the device and the resulting channels were filled with Ni²⁺-charged beads for efficient hexahistidine-tagged protein capture.

As a first test of our design, we measured the on-chip protein titers obtained from the protein synthesis module following two modes of operation—batch and continuous flow—and compared these to the titers produced from one-pot reactions performed in standard microcentrifuge tubes. For these experiments, we generated crude S30 extract from E. coli strain BL21 Star^TM (DE3) using a low-cost, sonication-based method (Kwon and Jewett, 2015) and the resulting extract was primed with plasmid pJL1-sfGFP^{DQNAT−6xHis} to drive the expression of sfGFP^{DQNAT−6xHis}. In a standard 15-µL, one-pot CFPS reaction using a microcentrifuge tube, we produced 11.9 μg/ml of sfGFP^{DQNAT−6xHis} in 2 h (Figure 2A; Supplementary Figure S1). To determine how the microfluidic environment affected sfGFP expression, we next performed batch-mode CFPS reactions in the first module of the microfluidic device. Specifically, the device was quickly filled with the same CFPS reaction mixture and fluorescence evolution was monitored in 30-min increments. When all CFPS components were present, fluorescence emission in the microfluidic channels gradually increased over time (Figure 2B), corresponding to production of 9.4 μg/ml of sfGFP^{DQNAT−6xHis} in 2 hours (Figure 2A). This result confirmed that the microfluidic environment itself had little-to-no effect on batch-mode CFPS productivity. It is also worth noting that surface blocking within the device was sufficient to allow sfGFP^{DQNAT−6xHis} clearance from the channels by simple rinsing.

We next investigated the effect of continuous flow on CFPS-based sfGFP expression. To generate a device that could accommodate a two-hour residence time (and thus be directly comparable to the batch-mode experiments above), we created a multi-chip system by linking individual devices with short pieces of tubing. Two input streams, one containing plasmid, energy, salts, and metabolites and the other containing S30 extract and T7 polymerase, met at the inlet and were mixed via diffusion between the two parallel streams as they moved through the channels (Figure 2C). In a four-chip system, corresponding to a two-hour residence time, we observed increasing fluorescence along the length of the channels from the inlet to the outlet corresponding to production of 38.3 μg/ml of sfGFP^{DQNAT−6xHis} (Figures 2A,D). Fluorescence across the width of the channels was uniform, indicating that the solution was well-mixed. Additionally, when comparing the fluorescence generation in two-, three-, and four-chip systems, corresponding to one, one and a half-, and two-hour residence times, respectively, we observed non-linear protein production with the maximum production rate occurring between one and one and a half hours (Supplementary Figures S2A,B). Importantly, the production rates in equivalent chips were similar, indicating that linking chips in series is a viable method for varying the residence time. Lastly, when comparing the titers of the two modes of on-chip operation relative to the microcentrifuge tube reaction for a two-hour residence time, we observed that sfGFP^{DQNAT−6xHis} produced on-chip in batch mode was statistically similar to off-chip production in a microcentrifuge tube, whereas introducing flow to the system significantly improved production by several-fold compared to both batch operations (Figure 2A). This increase in production has also been observed by others (Timm et al., 2015) and can be attributed to shorter diffusion lengths in the microfluidic channels.

In the protein glycosylation module, we sought to develop an OST tethering strategy that would allow for efficient protein glycosylation as the reaction substrates—the acceptor protein and LLOs—were continuously flown over the immobilized OST enzymes within the device. The advantage of OST tethering is that it enables creation of a local environment with a high concentration of OST enzyme that is reused in continuous operation. Such a reusable configuration is significant because OSTs are integral membrane proteins that are laborious and time consuming to produce in purified form (Schoborg et al., 2018). For surface immobilization of CjPglB, we leveraged avidin-biotin technology because it afforded the opportunity to site-specifically modify the OST with biotin such that enzymatic activity was minimally affected. To this end, an AviTag was genetically fused to the C-terminus of CjPglB, providing a unique site for covalent biotin conjugation by separately prepared BirA enzyme. Biotinylation of CjPglB was confirmed by immunoblot analysis using commercial ExtrAvidin-Peroxidase that specifically detects biotin (Figure 3A). To verify that enzymatic activity of CjPglB-biotin was not diminished by this modification or subsequent tethering onto a solid support, we performed off-chip in vitro glycosylation (IVG) reactions in a microcentrifuge tube using purified sfGFP^{DQNAT−6xHis} as acceptor protein, CjLLOs as glycan donor, and either untethered CjPglB-biotin or CjPglB-biotin that was tethered to commercial streptavidin beads. Immunoblot analysis of the sfGFP^{DQNAT−6xHis} produced in these reactions was performed using an anti-His antibody to detect the protein/glycoprotein and hR6 serum that specifically recognizes the C. jejuni heptasaccharide. These blots revealed 100% conversion of sfGFP^DQNAT to the glycosylated form (g1) in reactions with both untethered and tethered CjPglB-biotin, but only when the microcentrifuge tube for the latter reactions was shaken to keep the beads well suspended in solution (Figure 3B). In the absence of shaking, the beads were observed to sink to the bottom of the microcentrifuge tube so that CjPglB-biotin was not well dispersed within the reaction mixture, thereby reducing glycosylation efficiency as evidenced by the detection of sfGFP^DQNAT in a predominantly aglycosylated form (g0). Importantly, these results confirmed that CjPglB tolerated both site-specific biotinylation and tethering to a solid surface without any measurable loss in enzyme activity.

Encouraged by these results, we went on to investigate a strategy for surface tethering of CjPglB-biotin within the channels of our microfluidic device. To provide an evenly distributed, functionalized surface having low non-specific adsorption of other biomolecules, we modified the surface of our device with a silane-PEG5000-biotin moiety. This molecular weight of PEG has been shown to effectively reduce non-specific binding (Teramura et al., 2016) and to improve surface coverage compared to traditional coupling methods (Baranowska et al., 2015). Here, silane-PEG5000-biotin provided a highly selective binding surface that was observed to promote higher loading capacity compared to non-specific adsorption to non-biotinylated silane-PEG5000 when visualized with fluorescently labeled streptavidin (Supplementary Figure S3A). Comparing the surface coverage of the functionalized PEG brush to that of the non-covalent random adsorption also showed that we had a 30% increase in streptavidin coverage, allowing us to load more enzyme onto the surface of the device. Next, unlabeled NeutrAvidin was immobilized on the silane-PEG5000-biotin surface and was observed to bind fluorescently labeled, free biotin (Supplementary Figure S3B), indicating that unliganded binding sites in surface-bound NeutrAvidin, which has four putative biotin-binding pockets, were available to capture additional biotin groups. Collectively, these experiments confirmed that silane-PEG5000-biotin provided a highly selective, passivating surface that increased binding capacity.

To evaluate this tethering strategy in the context of CjPglB, we coated the channels of our microfluidic device with silane-PEG5000-biotin, followed by the addition of streptavidin and then CjPglB-biotin (Figure 3A). To determine whether immobilization of CjPglB in this manner resulted in a glycosylation-competent device, we first performed on-chip IVG reactions in batch mode without flow. This involved manually pushing IVG reaction components—sfGFP^{DQNAT−6xHis} and CjLLOs—over CjPglB that was tethered in the microfluidic device. The sfGFP^{DQNAT−6xHis} product was collected from the chip and analyzed by immunoblotting, which revealed barely detectable glycosylation that was significantly less efficient than the glycosylation observed for an on-chip, batch-mode control reaction performed concurrently in a microcentrifuge tube (Supplementary Figure S4A). To determine if continuous flow would remedy this issue, we next flowed the IVG reaction components over the device-tethered CjPglB across a series of chips, each with a reaction residence time of 30 min. In parallel, batch reactions in microcentrifuge tubes were conducted at the same time for comparison. For these off-chip reactions, we calculated the maximum amount of enzyme that could theoretically be bound to the microfluidic surface and used that amount in the microcentrifuge-based reactions. It should be noted that this amount is likely higher than what is tethered within the device. The sfGFP^{DQNAT−6xHis} products from these reactions were analyzed by immunoblotting as above, with readily detectable glycosylation occurring in the on-chip, continuous-flow system that was on par in terms of efficiency with the off-chip microcentrifuge reactions (Figure 3C; Supplementary Figure S4B). Interestingly, the addition of flow even appeared to enhance the reaction kinetics, akin to what was observed in the CFPS module.

In the third module of our device, we sought to capture polyhistidine-tagged sfGFP^{DQNAT−6xHis} using an affinity capture strategy. By selectively binding our target protein, unwanted cellular debris, cofactors, and other waste products generated from the upstream reactions can be easily removed by flow-based rinsing. The glycoprotein product can then be recovered by elution with buffer containing a high concentration of imidazole. Using a design based on earlier works (Xiao et al., 2018; Zilberzwige-Tal et al., 2019), we prepared a PDMS microfluidic device with posts at the outlet that could be packed with commercial Ni²⁺-charged beads, thereby enabling on-chip IMAC (Figure 4A). To test this strategy, we attempted to purify sfGFP^{DQNAT−6xHis} in CFPS reaction mixtures that were flowed through the device with the initial exit stream collected as the flowthrough. Next, we switched the inlet stream to buffer for washing the IMAC resin and removing any non-specifically bound proteins. Finally, we eluted the hexahistidine-tagged protein product using imidazole. The loading and elution steps were monitored by fluorescence imaging of the device (Supplementary Figure S5A) while the composition of each purification fraction was analyzed by SDS-PAGE analysis (Figure 4B; Supplementary Figure S5B). Based on multiple trials, we achieved 78 ± 10% purity in the final product (Figure 4B). This range of purities is to be expected because metal-binding proteins and histidine-rich regions in proteins are naturally present in cells which bind to the nickel resin and elute along with the his-tagged protein of interest. While this purity may not be acceptable for human therapeutic purposes, a product produced in our system could potentially be used in animal studies if using detoxified cell-free lysate (Stark et al., 2021). It should be noted that more complicated device configurations may improve the overall capture efficiency; nonetheless, our results are comparable to other microfluidic capture strategies (Murphy et al., 2019). To determine the efficiency of product capture, we measured the fluorescence of each fraction and calculated the percent sfGFP^{DQNAT−6xHis} that was present. While there was some variation in the capture efficiency, we reproducibly captured 45 ± 14% of total produced sfGFP^{DQNAT−6xHis} (Figure 4C). This simple strategy for protein purification provides a convenient way to obtain a purified final protein product using inexpensive reagents and gentle elution conditions. Because of the modularity of our design, other types of resin (e.g., glycan-binding affinity reagents) could be used in place of, or in addition to the set-up shown here depending on the desired separation. Additionally, multiple devices could be connected for larger scale purifications.