Facile, high‐throughput protein purification enabled by high‐efficiency protease elution Authors: Edson N Cárcamo Noriega, Ian S Truebridge, Frank D Teets, James W Bowman, Nathalia Rodriguez, Tessa A Howard, Christopher D Bahl Published: 2026 ABSTRACT AbstractWhile advances in artificial intelligence have made protein design widely accessible, protein production and characterization remain a bottleneck. Protease‐cleavable affinity tags are commonly used to improve yield and purity of recombinant proteins, but tag removal adds labor and complexity. Previous attempts to use a selective protease to elute cleaved fusion protein during affinity chromatography have suffered from poor digestion efficiency that greatly diminishes yield, or protease contamination in the eluate. Here, we describe a novel SUMO protease construct that supports rapid, high‐yield protease elution. This is the keystone of an end‐to‐end DNA‐to‐protein workflow that we optimized for speed, parallelizability, and generalizability. While our workflow is intended for an automated liquid‐handler, it can easily be performed manually, making it broadly accessible. We demonstrate the method by producing and characterizing 96 disparate proteins derived from mesophilic organisms. This provides a benchmark for methods development and a curated dataset useful for machine learning.Keywords: affinity chromatography, high‐throughput, laboratory automation, protein expression, protein purification, recombinant protein High‐throughput protein production is increasingly essential for both protein engineering and basic research such as mutational series or ortholog panels where rapid experimental feedback is critical for probing sequence–structure–function relationships. Despite advances in expression systems and affinity purification strategies, producing high‐quality material for a large number of disparate proteins is still a bottleneck. Current approaches often rely on harsh elution conditions or non‐parallelizable chromatography steps which reduce yields, require time‐consuming and resource‐intensive steps for performing buffer exchange, and limit throughput (Gräslund et al., 2008; Walls & Loughran, 2011). Immobilized metal affinity chromatography (IMAC) is a broadly employed method for purifying poly‐histidine tagged protein following heterologous over‐expression in host organisms such as Escherichia coli (Block et al., 2009; Porath et al., 1975). Washing IMAC resin with a buffer containing a high concentration of imidazole is the standard method for elution. However, there are three major limitations to high‐imidazole elution. First, proteins from the expression host can contaminate the eluate and reduce purity. Imidazole competes for binding to immobilized metal ions and therefore releases His tagged proteins as well as host protein contaminants that are also captured from the cell lysate (Johnson et al., 1996; Soltaninasab et al., 2022). This problem is exacerbated when the heterologous His tagged protein is weakly expressed or too much IMAC resin is used, as unsaturated binding capacity that is not occupied by heterologous His tagged protein is free to capture host protein contaminants. Second, imidazole elution is inefficient when performing IMAC in batch, as is typical when performing protein purifications using multiwell plates. While IMAC under flow allows for non‐equilibrium binding, IMAC in batch is taking place under equilibrium binding conditions. As a result, the His tagged proteins are not fully released in a single elution step, and multiple elution steps are required (Nejadmoghaddam et al., 2011). This reduces overall yield and increases protocol time and reagent consumption. Third, the resulting eluate contains a high concentration of imidazole which often interferes with downstream measurements and applications, including enzymatic activity, biological activity assays, protein–protein interactions, circular dichroism (CD) spectroscopy, etc. Consequently, imidazole is typically removed through buffer exchange or desalting. This adds time and cost, while lowering yield and increasing the complexity of automated workflows (Terpe, 2003). Affinity tags have the potential to interfere with the activity, stability, or solubility of the proteins they are fused to (Terpe, 2003; Waugh, 2005). Thus, the incorporation of a cleavage site for a selective protease is commonly used to remove the tag, but additional purification steps to remove the tag and protease, as well as additional buffer exchange, are required (Kapust et al., 2001; Malakhov et al., 2004; Terpe, 2003; Young et al., 2012). Protease‐mediated elution (or “protease elution” for short) has long been considered an attractive strategy for affinity purification; this is where the target protein is released directly from the affinity resin through site‐specific cleavage that separates the protein from the affinity tag. There are many potential advantages to protease elution, like improved purity and avoiding downstream buffer exchange. Two opposing challenges have stymied this approach to date. First, if the protease used for elution lacks the same affinity tag as the recombinant protein being purified, the protease will contaminate the eluate and require additional steps to remove. Second, if the protease possesses the same affinity tag as the recombinant protein being purified, the protease is captured by the affinity chromatography resin and has restricted access to the substrate; the result is unacceptably low yields. For this reason, protease elution has yet to be widely adopted despite its potential to make protein purification faster and simpler. On the other hand, self‐cleaving affinity purification systems based on inteins have been developed to enable tag removal directly on the affinity resin, thereby eliminating the need for exogenous proteases and additional purification steps (Chong et al., 1997; Wood et al., 1999). However, intein‐mediated cleavage often requires specific buffer conditions, can exhibit slow or incomplete cleavage kinetics, and may suffer from premature cleavage during expression, which can reduce yield and reproducibility (Shah & Muir, 2014). These limitations can complicate their implementation in high‐throughput workflows. Here, we describe a method to enable highly efficient protease elution: insert a long, intrinsically disordered protein linker between the affinity tag and the protease domain (Figure 1). By increasing the distance between the immobilization point and the enzyme active site, we greatly expand substrate accessibility and achieve elution efficiencies that are equivalent to chemical elution or un‐tagged protease elution, all without contaminating the eluate with protease or chemical (i.e., imidazole). Cth SUMO protease engineering and activity. The elution reaction was incubated at room temperature with 500 RPM orbital shaking with 150 nM Cth SUMO protease construct. Eluted sfGFP from triplicate samples was quantified by fluorescence intensity (λex/λem = 488/510 nm), and elution using high‐imidazole buffer is shown as a dashed line. Cartoon representations of the three different Cth SUMO protease constructs are shown for illustration (PDB ID: 6DG4). With this longstanding challenge solved, we then reevaluated the entire protein production process and found opportunities to improve every step. Our overarching philosophy is to ensure broad compatibility with standard liquid handling robotics to maximize reproducibility and throughput, while minimizing time and labor. The result is a complete DNA‐to‐protein workflow that bypasses many time‐consuming and labor‐intensive steps, such as the generation of clonal expression strains and downstream buffer exchange of purified protein. To evaluate the optimized workflow, we assembled a Benchmark Set of 96 proteins derived from mesophilic organisms with maximally different sequences and structures. We avoided proteins from thermophilic organisms or de novo designed proteins because they tend to be very stable and are therefore easier to work with; our goal was to assess the broad applicability of our method. Following purification, we performed a comprehensive suite of biochemical and biophysical characterization experiments. The results confirm that the proteins produced in high‐throughput are of high quality, and they provide a standardized, labeled dataset for others to use for computational analysis and machine learning. RESULTS Design of an immobilized protease construct Affinity tagged protease struggles to cleave substrate that is immobilized via the same affinity tag, and we hypothesized that affinity capture restricts the enzyme from accessing its substrate. Therefore, improving access to the substrate should rescue the protease's activity, and to achieve this, we sought to introduce a long, intrinsically disordered protein between the affinity tag and protease domain. We selected the 26.4 kDa protein CAHS_107838 from Paramacrobiotus richtersi because it is soluble, sufficiently long, is devoid of liabilities such as unpaired cysteine residues, and it was verified to be intrinsically disordered by nuclear magnetic resonance spectroscopy (Boothby et al., 2017). As an added bonus, CAHS_107838 is used by P. richtersi to resist damage caused by desiccation, and we speculated that it may provide additional benefits to improve the stability of a protein fused to it (especially to lyophilization). To develop an efficient protease elution method, we also needed a highly processive, highly selective protease with good solution behavior. We previously identified and characterized the SUMO protease homologue from Chaetomium thermophilum (Cth). We selected this enzyme because it is more thermostable, more soluble, and has a ~4‐fold faster V max than the SUMO protease Ulp1 from Saccharomyces cerevisiae, all while retaining selectivity for substrate proteins tagged with S. cerevisiae SUMO Smt3 (Lau et al., 2018). For the affinity capture method, we selected a 10×His tag for performing IMAC with nickel nitrilotriacetic acid (Ni‐NTA) magnetic beads (Kim et al., 2007). We also sought a test substrate that is stable and easy to quantify in a mixed solution containing multiple other proteins. We selected superfolder GFP (sfGFP) because it is hyperthermostable and can be quantified via fluorescence measurements (Pédelacq et al., 2006). Thus, the test protein has an N‐terminal 10×His tag, followed by SUMO Smt3, and with sfGFP at the C‐terminus. Characterization of immobilized protease activity We expressed and purified three different Cth SUMO protease constructs: with no affinity tag, with an N‐terminal 10×His tag, and an N‐terminal 10×His tag followed by CAHS_107838 (Figure 1). Each of these three protease constructs was able to be expressed and purified from E. coli, and all exhibited favorable solution behavior: for example, no noticeable inclusion bodies following clarification of the cell lysate, no precipitation during or after protein purification, no noticeable viscosity when pipetting the purified protein, etc. Next, we tested the efficiency of each of these protease constructs to digest immobilized 10×His‐SUMO‐sfGFP. We pre‐loaded Ni‐NTA magnetic beads with substrate protein, added 150 nM protease, and incubated at room temperature. The amount of sfGFP in the eluate was quantified by fluorescence intensity at 510 nm, and we compared this to the amount of undigested substrate eluted from washing the beads with 400 mM imidazole (Figure 1). As hypothesized, Cth SUMO protease that lacks an affinity tag is able to access substrate; the reaction reached completion within 30 min, and over 90% of the sfGFP was released into solution. This eluate also contained nearly all of the Cth SUMO protease that was added to the beads (Figure S1, Supporting Information). Conversely, the 10×His tag Cth SUMO protease struggled to access substrate; we found that only ~10% of the sfGFP was eluted after 30 min, and only ~25% after 100 min when we terminated the reaction (Figure 1). While the elution was inefficient, this protease construct was not observed in the eluate by SDS‐PAGE (Figure S1). We observed the best features of these two conditions combined when evaluating the construct that possesses the CAHS_107838 linker between the 10×His tag and Cth SUMO protease domain; sfGFP elution was nearly complete within 30 min (Figure 1), and no contaminating protease was observed in the eluate by SDS‐PAGE (Figure S1). For simplicity, we will henceforth refer to this construct as “AIProtease” (named after the company where the authors are employed). We also tested all three protease constructs at 300 and 600 nM in the elution buffer, and no substantial differences in elution kinetics were observed (Figure S2). However, the highest concentration of 10×His tag proteases also resulted in contamination of the eluate with uncleaved substrate protein as well as cleaved 10×His tag SUMO, presumably by competing for binding to the resin. We did not quantify concentrations of protease lower than 150 nM, but anecdotally, we found that reducing the concentration further required extended incubation times to achieve full elution. We therefore selected 150 nM for the protease concentration going forward, as this balances several factors: it avoids displacing bound substrate protein from Ni‐NTA, and it contains enough protease to rapidly elute captured substrate protein. Because AIProtease uses a cysteine nucleophile, a small amount of reductant should be included in the buffers while preparing this protein for use as a reagent. However, it is not necessary to supplement the elution buffer with reductant; indeed, AIProtease exhibited remarkable activity across a wide range of solution conditions (Table S1). As expected for a cysteine nucleophile protease, AIProtease displayed reduced activity under acidic conditions. This behavior is consistent with the catalytic mechanism which requires a deprotonated thiolate. At lower pH, the equilibrium shifts toward a protonated thiol nucleophile, which reduces catalytic activity. Despite this limitation, AIProtease remained active under a broad range of neutral and mildly basic buffer conditions commonly used for biochemical and structural studies (Table S1). For the linker, CAHS_107838 successfully performed the function that we hypothesized, so we did not make or test other SUMO protease constructs with different intrinsically disordered linkers. We also produced and tested a superTEV (Correnti et al., 2018) construct with the CAHS_107838 linker, but we found this construct to have no detectable protease activity (data not shown). TEV protease recognizes a linear peptide substrate; we suspect that the intrinsically disordered CAHS_107838 linker, for which we did not find evidence of auto‐digestion, may bind competitively and occlude access to the active site. Bottlenecks to facile high‐throughput protein production With high‐efficiency protease elution from affinity chromatography solved, we turned our attention to other points in the protein production process that create challenges for high‐throughput workflows, and we identified opportunities to improve throughput and reduce cost throughout (Figure 2). A high‐throughput protein production platform. The five step workflow begins with DNA assembly and results in purified protein. The process takes 3 days, is able to purify up to 96 proteins in parallel, and each step has been optimized for speed, efficiency, and compatibility with laboratory automation. Images of protein structures (PDB IDs: 6DG4, 2B3P) were generated using Illustrate (Goodsell et al., 2019). Obtaining expression While it is possible to purchase synthetic genes cloned into an expression plasmid from a commercial provider in bulk, this is cost prohibitive for many researchers, especially for large numbers of different constructs. Linear, double‐stranded DNA is much more affordable, but it requires a reliable, low‐cost, automation‐friendly method for inserting synthetic genes into an expression plasmid. Generating expression strains Once an expression plasmid is obtained, it is used to transform bacteria to produce an expression strain. The traditional approach is to generate a clonal strain by plating the culture and selecting a single colony from an agar plate; this strain is often validated by Sanger sequencing to ensure it is carrying the correct gene product. These steps are time, resource, and labor intensive, and automation requires specialized colony picking robots. Protein expression Culture conditions dictate the amount of heterologous protein that can be overexpressed from E. coli. Multiwell plates are essential for any high‐throughput workflow, but can make it challenging to deliver sufficient oxygen to support aerobic metabolism during logarithmic growth. Furthermore, richer media recipes can support more cell growth and enable higher yields from the same cultural volume. Producing cell extracts for protein purification Complete and efficient lysis of bacterial cells is surprisingly challenging. While detergent lysis methods are available, the reagents are costly when scaled to large numbers. Furthermore, it isn't always possible to fully remove detergent from purified protein, and detergent introduces challenges when used with automated liquid handling by increasing the viscosity of the lysate and causing bubbles to form. Construct assembly and transformation We sought to identify a method to assemble a linear DNA fragment into a circular plasmid with sufficient yield and fidelity that the reaction can be used directly for bacterial transformation. We selected Golden Gate assembly because of its simplicity (one‐pot assembly of multiple fragments), scarless junctions, and high efficiency driven by type IIS restriction enzymes and ligation cycling (Engler et al., 2008; Potapov et al., 2018) (Figure 3b). We also evaluated other assembly strategies such as Gibson assembly (Gibson et al., 2009), but anecdotally, we found them to perform less reliably in a high‐throughput context. Assembly and transformation efficiency. (a) Schematic representation of the plasmid architecture, highlighting the construct components: (i) the translational ramp MAKIH, (ii) a 10×His tag, (iii) the SUMO domain Smt3, and (iv) the cloning site. (b) Construct design for Golden Gate assembly with plasmid pAIP4, illustrating the Type IIS restriction enzyme BsaI recognition sites (shown in bold). The 5′ CTGG and 3′ GTTC overhangs at the cleavage site are indicated. The insert handles are designed with matching overhangs that are in‐frame with the plasmid, providing seamless gene integration. The final construct includes N‐terminal Ser‐Gly and C‐terminal Gly‐Ser scars, followed by double‐stop codons. (c) Nanopore sequencing analysis of the transformed population of Escherichia coli following sfGFP insertion into plasmid pAIP4. A total of 1434 mapped reads confirmed the correct insertion of the gene of interest, with a 95% confidence interval (CI) of ±0.04% and no significant mutations or indels (median Q score = 17.4). The analysis detected no aberrant constructs or re‐ligated plasmid. (d) The minimum equivalent of CFUs required to achieve enrichment of a starter culture was determined via serial dilution following bacterial transformation of the Golden Gate assembly reaction. Three biological triplicates were performed with selection on agar and in broth. After 8 h of incubation, the relative expression levels in broth were quantified by measuring the fluorescence intensity of the cultures (λex/λem = 488/510 nm), and colonies were counted on the agar plates. To optimize the gene assembly and transformation workflow, we employed 10×His‐SUMO‐sfGFP as a reporter. We titrated the plasmid:insert molar ratio and reaction time in Golden Gate assembly reactions (Table S2) and found that a 60‐min incubation at 37°C with a 1:3 plasmid:insert molar ratio maximized assembly efficiency while balancing speed and reagent consumption. To reduce contamination from unligated plasmid—which carries an in‐frame gene encoding 10×His‐SUMO (Figure 3a) that would compete during IMAC purification—the assembly protocol includes a post‐ligation incubation at 60°C for 10 min; this inactivates T4 ligase while allowing the BsaI restriction enzyme to linearize any remaining plasmid DNA lacking an insert (Chockalingam et al., 2020). We also sought to bypass the generation of clonal expression strains. In doing so, we must accept that some fraction of the expression culture will not carry a perfect gene insertion. However, incorrect gene insertions will often not result in a soluble protein product, and ones that do need to be sufficiently rare that they are an inconsequential contaminate in the final purified protein. To achieve this, we developed a protocol whereby the DNA assembly reaction is transformed directly into E. coli, which is then cultured in liquid broth. While technically polyclonal, this culture should be equivalent (or nearly so) to a clonal strain. This was assessed by transforming the assembly reaction into DH10B cells, culturing the transformed population with antibiotic selection, isolating the plasmid DNA, and analyzing the sequences using nanopore sequencing. We found that >95% of constructs contained the correct gene insertion with no statistically significant mutations or indels (Figure 3c), and plasmid DNA lacking a gene insert was undetectable. In this study, we used Lemo21(DE3) as the expression strain, which carries the regulatory plasmid pLemo (controlling T7 lysozyme expression) and a chromosomally integrated T7 RNA polymerase system to tune protein expression levels (Schlegel et al., 2012). Expression strains such as Lemo21(DE3) generally exhibit lower transformation efficiencies than cloning strains (Chan et al., 2013). To determine the minimum number of transformants required to reliably establish a liquid culture in microwell plate format, we measured colony‐forming units (CFUs) and growth in non‐inducing “starter” media following transformation with the assembly reaction. Transformation efficiency was quantified by plating serial dilutions and counting CFUs, and liquid cultures were inoculated in parallel. We observed that approximately 188 CFUs are sufficient to reliably seed a starter culture of 1 mL that grows to saturation and supports reproducible protein expression in the 24‐well workflow (Figure 3d). Because we are not validating the gene insertion of every culture, it is not ideal to proceed with protein production from a single transformant, even though this is sufficient to generate a saturated culture through clonal expansion. Therefore, we selected transformation conditions that yielded ~900 CFU to improve tolerance to DNA sequence changes due to synthesis errors or aberrant gene assembly. This efficiency was readily achieved using chemically competent Lemo21(DE3) cells prepared with the Zymo Mix & Go! Competent Cells Kit according to the manufacturer's instructions. In order to minimize reagent consumption, we cloned and transformed 5 μL of the cloning reaction (10 ng/μL of DNA, determined by the plasmid amount) mixed with 50 μL of cells. In contrast, lower transformation efficiencies resulted in extended lag phases and increased variability in population growth (Coleman et al., 2003; Irwin et al., 2010), which can directly impact protein yield in subsequent culture steps. Expression cultures Heterologous protein expression systems are metabolically demanding and can slow or inhibit cell growth (Studier & Moffatt, 1986). Therefore, protein expression is optimally induced when the culture has reached log‐phase growth and the media still contains sufficient nutrients to support high‐yield protein production. Autoinduction media elegantly balances these requirements while avoiding the need to manually monitor cultures and add an inducer (Studier, 2005). In this study, we compared two different autoinduction formulations: ZYM‐5052, the original autoinduction medium, and a custom formulation we developed based on Terrific Broth (Tartoff & Hobbs, 1987) that we call TBM‐5052. We found that TBM‐5052 outperformed ZYM‐5052 in terms of biomass production and overall protein yield, likely because it contains more nutrients (Figure 4a). Optimization of expression culture conditions. (a) Expression of 10×His‐SUMO‐sfGFP was evaluated across different media recipes. Fifty milliliter cultures were grown in baffled flasks at 25°C with shaking at 250 RPM for 16 h in triplicate. TBM‐5052 yielded significantly higher expression levels compared to ZYM‐5052 (p < 0.001). (b) Culture volume and shaking speed were co‐varied to identify the conditions yielding maximal overexpression of heterologous protein using 24‐deep‐well plates, with 10×His‐SUMO‐sfGFP used as a reporter protein. Two milliliters of media with orbital shaking at 800 RPM was used for all subsequent experiments. In both experiments, relative expression levels were quantified by measuring the fluorescence intensity of the cultures (λex/λem = 488/510) and the mean fluorescence of triplicate samples is reported. We chose to use 24‐deep‐well plates for high‐throughput bacterial culture as this format balances parallel capacity while maximizing cultural volume (Saez & Vincentelli, 2014; Vincentelli & Romier, 2013). We evaluated the impact of shaking speed and culture volume on heterologous protein production in TBM‐5052 media with 10×His‐SUMO‐sfGFP as the reporter. We found that faster shaking speeds always resulted in higher protein expression, presumably due to better aeration. However, shaking the culture above 800 revolutions per minute (RPM) in our shaker, which has a 3 mm orbit, caused the culture media to splash onto the gas permeable sealing membrane at the top of the well. While 0.5 mL of media per well resulted in the highest protein yield per mL of culture, 2 mL per well yielded the highest total amount of protein per culture (Figure 4b). We also observed that well shape affected growth and expression, with square wells performing better than round wells (data not shown), presumably because square wells induce more fluid cavitation during shaking. Reproducibility is essential for the success of any high‐throughput workflow. Potential sources of variation include Golden Gate assembly, bacterial transformation, inoculation and growth of starter cultures, inoculation and growth of expression cultures, and autoinduction of heterologous protein expression. To assess consistency, we performed 96 replicates of the workflow starting with linear double‐stranded DNA and terminating with the harvest of the expression cultures. We again used 10×His‐SUMO‐sfGFP as the reporter and evaluated protein expression in whole bacterial cells. Fluorescence measurements yielded a coefficient of variation below 5%, indicating high reproducibility (Figure S3). Protein extraction Many different approaches for homogenizing bacterial cells have been developed (Shehadul Islam et al., 2017), and chemically induced lysis using enzymes and mild detergents is common for high‐throughput workflows because mechanical cell disruption is challenging to parallelize. We first sought to identify a low cost detergent of known composition (i.e., not a proprietary, secret formulation), and n‐Nonyl‐β‐D‐glucopyranoside was selected for evaluation (Blakesley et al., 2002); we compared lysis by this detergent with sonication (Figure 5a). For these experiments, we used cell pellets obtained from 2 mL cultures grown under the optimized conditions described above, corresponding to a wet pellet weight of 0.03 ± 0.005 g. Optimization of protein extraction and capture. (a) Comparison of lysis efficiency using sonication versus detergent‐based lysis with n‐Nonyl‐β‐D‐Glucopyranoside, using 10×His‐SUMO‐sfGFP as a reporter. (b) The temperature of the lysates were measured after 10 cycles of sonication at 50% amplitude with different pulse programs. Samples were kept on ice throughout. (c) The relative extraction efficiency and temperature of lysates during sonication were evaluated in response to pulse cycle number. Sonication was performed at 50% amplitude in cycles of 30 s on and 30 s off with samples kept on ice throughout. (d) Batch binding kinetics of heterologously overexpressed 10×His‐SUMO‐sfGFP in clarified lysates. Complete capture of ~250 μg of protein was achieved within 30 min using 50 μL or more of 25% slurry of Ni‐NTA magnetic beads. All experiments were performed in triplicate. Across all conditions tested, sonication resulted in more efficient cell disruption than detergent lysis, as measured by sfGFP abundance in the clarified lysate, even at relatively high detergent concentrations (up to 2% wt/vol). In addition, detergent‐treated samples exhibited increased viscosity compared to sonicated samples, which reduced clarification efficiency and interfered with accurate pipetting during the transfer of the clarified lysate. Detergent‐containing lysates also tended to generate persistent foam during mixing and pipetting, further complicating handling in automated workflows. Anecdotally, we found that viscosity could be partially reduced by diluting the lysate, but the larger volume complicated clarification and purification steps. We concluded that sonication is the ideal method for cell disruption, but the single‐probe sonicator we used during testing is not amenable to parallel sample processing. To overcome this limitation, we identified a commercially available 24‐tip horn sonicator (Figure 2). Importantly, sonication generates a substantial amount of heat, which increases the temperature of the lysate and can denature proteins if not controlled. Next, we performed a series of experiments where we varied the volume of lysis buffer added to the cell pellet, the sonication pulse duration, and the number of sonication cycles (Figure S4) to identify conditions that maximize soluble protein extraction while controlling the temperature of the lysate. In each case, a 24‐well plate was incubated in an ice‐cold water bath during sonication (Figure S5). We found that resuspending each pellet in 2 mL of lysis buffer containing lysozyme, Benzonase, and an EDTA‐free protease inhibitor cocktail, followed by sonication for 10 cycles of 30 s on and 30 s off at 50% amplitude maximized protein extraction while maintaining lysate temperatures below 15°C (Figure 5b,c). Protein purification Traditional column‐based Ni‐NTA IMAC purification is challenging to parallelize and automate, so we instead chose to use Ni‐NTA magnetic beads paired with a 24‐well magnet plate. Placing the 24‐well plate on the magnet collects the magnetic beads at the bottom of each well and allows liquid to be aspirated without removing the magnetic beads. The three steps: liquid aspiration, liquid dispensation, and bead resuspension can be iterated to capture protein from cell lysate, wash away contaminants, and elute pure protein into solution; this approach is ideally suited for use with automated liquid handlers (Meng et al., 2008). The amount of Ni‐NTA magnetic beads used and the length of time they are incubated with the bacterial cell lysate will impact the capture of His‐tagged protein. Using 10×His‐SUMO‐sfGFP as a reporter, we performed an experiment where we varied the amount of magnetic beads and the duration of capture (Figure 5d). We determined that 12.5 μL of Ni‐NTA magnetic beads was sufficient to capture approximately 250 μg of soluble 10×His‐SUMO‐sfGFP, and 30 min was sufficient to capture nearly all of the heterologously overexpressed protein from the lysate. For our high‐throughput protocol, we chose to use four times this amount of Ni‐NTA magnetic beads (i.e., 50 μL) to ensure we have the capacity to capture up to ~1 mg of protein, as some proteins express more than others, and we wanted to make sure there was sufficient binding capacity available to capture the AIProtease used during elution. Automation To maximize throughput while minimizing human labor, we developed a suite of automated protocols for liquid handling robots for five key steps of our workflow: gene assembly, bacterial transformation, culture inoculation, protein purification, and post‐production sample processing. For clarity, post‐production refers to the process of transferring the eluate from four 24‐well plates containing magnetic Ni‐NTA beads and consolidating into a single 96‐well plate (Figure S6), as well as preparing samples for downstream experiments such as protein concentration assays and SDS‐PAGE. While we focused on the Beckman Coulter Biomek i7, our workflow is compatible with any liquid handler that is capable of moving microwell plates. Regardless of whether the workflow is performed manually or using automation, the following specialized equipment is highly recommended: a high‐speed shaker incubator with a 3 mm orbit, and a 24‐tip horn sonicator. We recommend an automated workstation featuring a dual‐arm configuration with a 1 mL 96‐channel head and Span‐8 pipetting, plus rotational microwell plate grippers to access all deck positions. A grid‐based deck should accommodate all required labware and tips for each protocol: 10 tip‐loading positions, 19 regular positions for microwell plates and magnet plates, trash containers, and wash stations. Additionally, we utilize devices on the deck that can perform specialized tasks such as heating/cooling, shaking, and an automated thermocycler (Figure S7). While we believe this is the ideal setup, our method can readily be adapted to other liquid handling workstations that do not possess all of these features, but doing so will necessitate more human interaction. The platform can process up to 96 proteins in parallel, with the ability to scale depending on the instrument's capacity. Benchmark set To evaluate the generalizability of our method, we selected a set of 95 unique proteins (“Benchmark Set”), plus sfGFP as a positive control. These proteins meet the following criteria: single‐domain, structure deposited in the PDB, disulfide‐free, from mesophilic organisms, and previously expressed in E. coli (Table S3). The Benchmark Set was selected to maximize diversity and be as broadly representative as possible. Because standard root‐mean‐squared deviation (RMSD) algorithms can produce inconsistent results on proteins with very different sequences (Zhang & Skolnick, 2005), we elected to use TM‐align (Zhang & Skolnick, 2005), a sequence‐independent structural alignment algorithm that uses dynamic programming to align on secondary structural content and more accurately quantitate the structural similarity between dissimilar proteins. The benchmark set had a mean TM‐align score of 0.28, 67% of pairs falling under the 0.3 threshold that suggests different folds (Figure 6a). Sixty‐four CATH superfamilies (Knudsen & Wiuf, 2010) are represented among the 79 proteins with an assigned superfamily, giving a Shannon entropy of 4.06 (97.5% of theoretical maximum) and a Simpson's Diversity Index of 0.018. The Benchmark Set of proteins spans a broad range of physicochemical and structural properties. (a) Distribution of pairwise TM‐align scores calculated for all protein pairs in the Benchmark Set. (b) Scatter plot illustrating that the Benchmark Set spans a diverse and uncorrelated range of calculated isoelectric points and the fraction of hydrophobic solvent‐accessible surface area (SASA) from the total SASA. (c) Scatter plot showing calculated molecular weight versus purified protein yield per reaction, indicating no apparent correlation between protein size and production yield. (d) Distribution of melting temperatures (T m) measured by differential scanning fluorimetry (DSF). (e) Overlay of size‐exclusion chromatography (SEC) chromatograms illustrating the distribution of elution profiles and protein dispersity. (f) Overlay of circular dichroism (CD) spectra normalized as mean residue ellipticity (MRE), demonstrating the range of secondary structure signatures within the benchmark proteins. The overlayed traces in panels (e) and (f) are shown with randomly assigned colors to enhance visual differentiation. The Benchmark Set was produced in triplicate using the described workflow and automated protocols, and the entire process from receipt of linear DNA fragments to purified protein was completed in 3 days (Figure 2). We defined a success as: a total yield of >100 μg of purified protein from 2 mL of culture, and >90% purity as evaluated by SDS‐PAGE. By these criteria, we successfully produced 87 out of 96 proteins (>90%) in the Benchmark Set (Figure S8). While purifying such a broad range of proteins, we encountered an unexpected issue with substrate preferences for the Cth SUMO protease. The S. cerevisiae SUMO protease Ulp1 has been reported to tolerate a wide range of P1′ and P2′ residues (excluding proline) without substantially affecting cleavage efficiency (Malakhov et al., 2004; Zhang et al., 2022), and we expect this conclusion will apply to the Cth SUMO protease given the conserved structure and substrate compatibility (Lau et al., 2018). However, we observed that some proteins exhibited poor digestion efficiency with Cth SUMO protease that was not correlated with the amino acid identity adjacent to the cleavage site, but instead may be due to the tertiary structure context of the P1′ and P2′ residues. To overcome this, we included a two amino acid serine‐glycine “scar” at the N‐terminus of the heterologous protein sequence to provide some space between the cleavage site and the protein domain; this rescued protease activity for these substrates (Figure S9). Thus, we included an SG scar at the N‐terminus of all constructs to ensure broad compatibility regardless of the structural context of the P1′ and P2′ residues of the heterologous protein being produced. Next, we performed a series of experiments to characterize the purified proteins, with the goal of evaluating the quality of the material produced by our high‐throughput process. First, we performed differential scanning fluorimetry (DSF) to evaluate folding cooperativity and thermostability. 68/87 (78%) of the proteins displayed a cooperative unfolding transition that allowed us to calculate a melting temperature (T m) (Figure 6d). We performed analytical size‐exclusion chromatography (SEC) to evaluate the dispersity of the purified proteins (Figure 6e). Fifty‐three proteins in the Benchmark Set exhibited monodisperse behavior, eluting as a single, symmetric peak consistent with a properly folded and non‐aggregated species. Only 12 proteins exhibited broad peaks or altered retention times, suggesting partial oligomerization or conformational heterogeneity; this is not surprising given the structural diversity of the Benchmark Set. Importantly, no systematic aggregation or degradation trend was observed. Finally, we measured CD spectra from 190 to 260 nm at 25°C (Figure 6f). We obtained CD spectra consistent with the secondary structure composition observed in the Protein Data Bank for 64/87 (74%) of the Benchmark Set. Spectra for predominantly helical proteins had the characteristic negative peaks centered around 208 and 222 nm (e.g., 2D5J, 5MDT, 5HFQ, 2EYI, 3D1B, and 5H28), while proteins whose structure is predominantly comprised of beta sheets had the characteristic negative peak centered around 218 nm (e.g., 1KQX, 2H3L, 3RZY, 3S60, 5IHW, 5M2Y, and 5NJO). Spectra for proteins with a mixed topology had the characteristic spectra with a broad, flattened negative peak spanning 208 to 222 nm (e.g., 1HKA, 1JL1, 1TP9, 5YPV, 6F2Q and 6WEY). A lack of CD signal was due a high content of regions lacking secondary structure (e.g., 1FL0, 2H7W, 3H6Q, 3HNX, 3KT9, 3TOW, and 4A02), but 13 samples were at a concentration too low for accurate measurement by CD spectroscopy (≤0.1 mg/mL). DISCUSSION We have developed a scalable, high‐throughput protein production platform that integrates a carefully engineered expression plasmid, streamlined cloning and transformation protocols, optimized expression and purification workflows, and an efficient on‐bead protease‐mediated elution strategy. Our goal is to contribute to the democratization of high‐throughput protein science, and to provide a foundation that future efforts can build upon. All expression plasmids reported in this study are available at Addgene under the Uniform Biological Material Transfer Agreement and Industry Material Transfer Agreement https://www.addgene.org/browse/article/28252872/. The complete Benchmark Set of 96 proteins is available as a curated collection from Addgene, where it can serve as a resource for others who are developing or evaluating protein expression and purification methods. All automation scripts for liquid‐handling, detailed protocols for the entire workflow, as well as experimental data generated in this study are fully annotated and available under the permissive CC‐BY‐NC‐SA 4.0 license at https://github.com/aiproteins/high_throughput_protein_production. Future improvements to this platform may be made by increasing the number of different proteins that can be produced in parallel. Here, we primarily utilized 24‐deep‐well plates. The use of 96‐well plates would substantially increase the number of samples that can be processed in parallel, but this will necessitate decreasing the liquid volume of every step for each sample. Achieving this will require simultaneous improvements at many steps in the process, including: further improving the expression yield per culture, higher throughput methods for mechanical cell lysis, and downstream assay miniaturization to utilize less total protein. Further reductions in the cost of DNA synthesis will also be required for substantial advances in protein production throughput to be financially practical to implement for many researchers. ONLINE METHODS Expression plasmid In this study, we use a custom expression plasmid pAIP4 (Addgene, Plasmid # 233686), based on pET29b(+) (Rosenberg et al., 1987) that incorporates modular features optimized for high‐throughput workflows (Figure 3a). The N‐terminal fusion contains three key elements. First, a “translational ramp” sequence, Met‐Ala‐Lys‐Ile‐His (MAKIH), was included to facilitate ribosomal exit of the nascent polypeptide, thereby enhancing protein expression (Verma et al., 2019). Overlapping with the C‐terminal residue of the translational ramp, a 10×His tag enables purification by IMAC. A SUMO tag (Smt3) was incorporated to improve solubility and stability, and to enable digestion by a SUMO‐specific protease (Butt et al., 2005). We purposefully designed two amino acid additions to each terminus of inserted proteins; that is, “scars.” At the N‐terminus, we added SG to ensure broad compatibility with SUMO protease. By including conserved scars at both termini, it is possible to amplify gene inserts with universal primers to support re‐use of gene constructs across multiple expression plasmids without the need for custom primers (Figure S10). The plasmid includes the two most prevalent E. coli stop codons immediately following the C‐terminal SG residues to minimize ribosomal skipping and ensure accurate termination of translation. Between the SUMO tag and stop codons, we inserted a Golden Gate assembly site for use with the BsaI Type IIS restriction enzyme for modular gene insertion (Figure 3b). Optimal 4‐base overhangs (5′: CTGG, 3′: GTTC) maximize ligation fidelity (Potapov et al., 2018). For producing proteins where the introduction of N‐ and C‐terminal scar residues is undesirable and modular gene insertion is not required, we recommend using pAIP53 for scar‐less gene insertion (Addgene, Plasmid # 258140). DNA constructs and assembly Linear, double‐stranded DNA encoding genes of interest were ordered as eBlocks or gBlocks Gene Fragments (Integrated DNA Technologies, Coralville, IA, USA), depending on the length of the DNA fragments and complexity of the sequence. Reverse translation and codon optimization for E. coli was performed using DNAworks (Hoover & Lubkowski, 2002). 5′ and 3′ handles containing BsaI recognition sites were incorporated into the sequences to enable Golden Gate Assembly with plasmid pAIP4 (Figure 2b). Assembly reactions were performed using the NEBridge Golden Gate Assembly Kit (BsaI‐HF v2) (New England Biolabs, Ipswich, MA, USA) according to the manufacturer's instructions. Briefly, 150 ng of plasmid pAIP4 was combined at a 1:3 molar ratio with gene insert to a final reaction volume of 20 μL. Reactions were performed in a high‐throughput format using a 96‐well polymerase chain reaction (PCR) plate and incubated in a thermocycler for 1 h at 37°C with a heated lid, followed by 10 min at 60°C to inactivate the enzymes. Nanopore sequencing Five microliters of the Golden Gate assembly reaction was electroporated into 100 μL of DH10B E. coli cells using a 1 mm gap electroporation cuvette at a voltage of 1.8 kV, with a target time constant from 4 to 5 ms. The transformation was recovered for 1 h in SOC broth at 37°C. The transformation reaction was grown overnight in 1 mL of LB broth supplemented with 50 μg/mL kanamycin. The cell pellet was recovered by centrifugation at 10,000 × g for 5 min, and plasmid DNA was isolated using a Monarch Spin Plasmid Miniprep Kit (New England Biolabs, Ipswich, MA, USA). Three hundred nanograms of DNA was sent for Plasmid‐EZ nanopore sequencing (Genewiz, South Plainfield, NJ, USA). Transformation and starter cultures Chemically competent E. coli cells were prepared using the Mix & Go! Competent Cells Kit (Zymo Research, Irvine, CA, USA), following the manufacturer's instructions, aliquoted in 96‐well PCR plates, and stored at −80°C. For transformation, 5 μL of the Golden Gate assembly reaction was mixed with 50 μL of chemically competent E. coli Lemo21(DE3) cells and incubated on ice or with a 4°C chiller block for 30 min. Next, cells were heat shocked by incubation at 42°C for 30 s, followed by incubation on ice/4°C for 5 min. Cells were transferred to a 96‐deep‐well plate containing 200 μL of SOC broth per well, and incubated for 1 h at 37°C with orbital shaking at 250 RPM for recovery. Selection for transformants was performed concurrently with Starter Culture growth by adding 250 μL of LB broth supplemented with 0.2% (wt/vol) D‐glucose, 100 μg/mL kanamycin and 68 μg/mL chloramphenicol per well to the recovery cell plates, which were subsequently incubated overnight at 37°C with orbital shaking at 800 RPM with a throw diameter of 3 mm (Multitron plate shaker, INFORS HT, Bottmingen, Switzerland). Expression cultures Protein expression cultures were grown in sterile 24‐deep‐well plates with square wells and U‐bottoms (Thomson, Catalog # 931565‐G‐1X). Twenty microliters of saturated starter culture was used to inoculate 2 mL of TBM‐5052 autoinduction broth supplemented with 100 μg/mL kanamycin, 34 μg/mL chloramphenicol, 1 mM L‐Rhamnose, and 0.01% (vol/vol) Antifoam 204. Starter culture inoculation from a 96‐deep‐well plate to four 24‐deep‐well plates was performed according to Figure S11. Expression cultures were incubated at 25°C for 16 to 24 h with orbital shaking at 800 RPM with a throw diameter of 3 mm (Multitron plate shaker, INFORS HT, Bottmingen, Switzerland). Following incubation, the expression cultures were harvested by centrifugation at 3,000 × g for 15 min. The supernatant was discarded, and the cell pellets were frozen at −80°C. Protein extraction and purification Cell pellets from expression cultures in 24‐deep‐well plates were removed from the −80°C freezer and thawed at room temperature; the freeze/thaw cycle facilitates cell lysis. Cell pellets were resuspended in 2 mL of Lysis Buffer, which consists of IMAC Wash Buffer (500 mM NaCl, 20 mM imidazole, 20 mM NaPi pH 7.4) supplemented with 2 mM MgCl2, 0.25 g/L hen egg‐white lysozyme, 25 U/mL TurboNuclease (Accelagen, San Diego, CA, USA), and 1× EDTA‐free Protease Inhibitor Cocktail (Thermo Scientific, Waltham, MA, USA). Cell lysis was performed using a 24‐tip horn sonicator (Qsonica, Newtown, CT, USA; Cat. No. 4579) with a Q500 Sonicator set to 50% amplitude. The pulsing protocol was: 30 s on and 30 s off for a total of 10 cycles. Samples were chilled with an ice‐water bath (Figure S5) to maintain the temperature <15°C during sonication (Figure 5). The lysates were clarified by centrifugation at 3,000 × g for 15 min. The clarified lysates were transferred to a fresh 24‐deep‐well plate containing 200 μL of a 25% slurry of Ni‐NTA magnetic beads pre‐equilibrated in IMAC Wash Buffer. Batch binding of heterologously overexpressed protein from the lysate was performed at room temperature with constant shaking (500 RPM) for at least 30 min. Magnetic beads were washed three times each with 2 mL of IMAC Wash Buffer. During each wash step, the magnetic beads were captured using a 24‐well magnet plate (Alpaqua, Beverly, MA, USA; Cat. No. A000270), and the wash solution was aspirated and discarded. The magnetic beads were resuspended by shaking (500 RPM) during the wash steps. Following the IMAC Wash Buffer washes, the beads were washed with 2 mL of phosphate‐buffered saline (PBS) at pH 7.4 (Teknova, Catalog # P0196) to remove residual imidazole. For protein elution, the beads were resuspended in 500 μL PBS supplemented with 150 nM AIProtease and incubated at room temperature with constant shaking (500 RPM) for approximately 2 h. After incubation, the beads were again captured using a 24‐well magnet plate, and the eluate containing purified, heterologously over‐expressed protein (which no longer contains the 10×His‐SUMO fusion domain) was transferred from four 24‐deep‐well plates to one 96‐deep‐well plate (Figure S11). Protein concentration and purity was evaluated in high‐throughput by Qubit Protein Assay Kit (see below) and by sodium dodecyl‐sulfate polyacrylamide gel electrophoresis (SDS‐PAGE), respectively. Production of Cth SUMO proteases Proteases used in this study were expressed in Lemo21(DE3) cells transformed with pET29b plasmids containing the genes for the respective constructs (Table S4). Starter cultures were prepared in LB broth supplemented with 0.2% (wt/vol) D‐glucose, 50 μg/mL kanamycin, and 34 μg/mL chloramphenicol. Protein expression was carried out in 500 mL of TBM‐5052 broth in Pyrex Fernbach 2.8 L baffled flasks. Cells were harvested after 16 h of expression. The IMAC Wash Buffer and Lysis Buffer are the same as before, with the addition of 1 mM dithiothreitol (DTT) to both. Post‐harvest, cell pellets were subjected to a freeze–thaw cycle and then resuspended in 150 mL of Lysis Buffer. Cells were lysed by sonication using a 0.5 inch tip sonicator at 80% amplitude for 20 cycles of 30 s on and 30 s off. The lysate was clarified by centrifugation at 3,000 × g for 15 min. The clarified lysate was batch bound to 20 mL of Ni‐NTA agarose (pre‐equilibrated with IMAC Wash Buffer) for 1 h with gentle mixing. Next, the Ni‐NTA agarose was packed into a gravity column and washed with 10 column volumes (CVs) of IMAC Wash Buffer. Proteases were then eluted with 1 CV of an IMAC Elution Buffer (500 mM NaCl, 400 mM imidazole, 20 mM NaPi pH 7.4, 1 mM DTT). The eluate was buffer‐exchanged into PBS supplemented with 1 mM DTT using a Sephadex G‐25 HiPrep desalting column (Cytiva, 17508701). For the untagged Cth SUMO protease, the expression construct contained an N‐terminal 10×His tag, followed by a 5×GGS linker, and a cleavage peptide for TEV protease (Dougherty et al., 1989). The purified protein for this construct was incubated with 7×His tag TEV protease (New England Biolabs; Catalog # P8112S) to cleave the 10×His tag from the Cth SUMO protease. Following cleavage, the un‐tagged Cth SUMO protease was separated from the cleaved 10×His tag and 7×His‐tagged TEV protease by reverse IMAC (i.e., the column flow‐through is retained rather than the eluate). The concentration of all three Cth SUMO protease constructs were determined by measuring absorbance at 280 nm, using the calculated extinction coefficients for each: Cth SUMO protease ɛ = 37,720 M−1 cm−1, 10×His tag‐Cth SUMO protease ɛ = 37,720 M−1 cm−1, and AIProtease ɛ = 45,170 M−1 cm−1. Protein concentration determination Protein concentration was determined using a Qubit Protein Assay Kit (Thermo Fisher Scientific, Waltham, MA, USA; Cat. No. Q33212) conducted in a 96‐well plate format. The assay was performed according to the manufacturer's instructions to ensure accuracy and reproducibility. Briefly, 5 μL of each protein sample, diluted to fall within the assay's linear range, was transferred to a 96‐well plate containing 95 μL of the Qubit dye solution and incubated at room temperature for 15 min. Fluorescence measurements were performed using an excitation at 470 nm and emission at 570 nm on a SpectraMax i3x spectrometer (Molecular Devices, San Jose, CA, USA). A standard curve was generated using the Qubit BSA Protein Standards at various concentrations in a range of 0.125–0.500 mg/mL. Throughout this study, sfGFP was utilized as a reporter protein to optimize various steps of the protein production workflow. To assess the relative expression levels of cultures, 10 μL of each 10×His‐SUMO‐sfGFP expression culture was diluted 1:10 in ultrapure water and transferred to a black 96‐well plate. The fluorescence intensity was measured using a SpectraMax i3x spectrometer plate reader, with excitation and emission wavelengths set at 485 and 510 nm, respectively. Circular dichroism spectroscopy CD spectroscopy measurements were performed using a Chirascan Q100 (Applied Photophysics, UK) equipped with a Peltier temperature control unit. Protein samples were adjusted to 0.2 mg/mL in 50 mM NaPi, pH 7.4; samples where the protein concentration was <0.2 mg/mL were measured without dilution. Samples were loaded into a 0.5 mm path‐length quartz cuvette using the instrument's autosampler, and steady‐state spectra were recorded from 190 to 260 nm with 1 nm steps and 1 nm bandwidth at 25°C. Each spectrum represents the average of three scans, and a baseline spectrum was recorded for the buffer and subtracted. Data were converted to mean residue ellipticity for comparison between samples. Differential scanning Fluorimetry DSF was performed using a QuantStudio 3 Real‐Time PCR System (Thermo Fisher Scientific, Waltham, MA, USA). Protein samples were adjusted to 1 μM when possible; lower‐concentration samples were measured as is. Samples were mixed with Sypro Orange dye (Thermo Fisher Scientific, Waltham, MA, USA) at ≥5‐fold molar excess of the protein sample (i.e. 5 μM). Fluorescence was monitored with excitation at 490 nm and emission at 575 nm over a temperature gradient from 25°C to 99°C at 0.05°C/s. Each sample was measured in triplicate. Melting temperatures (T m) were determined from the derivative of the fluorescence intensity using Protein Thermal Shift Software 1.4 (Thermo Fisher Scientific, Waltham, MA, USA). Size‐exclusion chromatography Protein dispersity was analyzed by SEC on a Nexera 40 Series high‐performance liquid chromatography (HPLC) system (Shimadzu, Kyoto, Japan) using a TSKgel G3000SWXL column (7.8 mm × 300 mm, Tosoh Bioscience, Tokyo, Japan) according to the manufacturer's recommendations. The column was equilibrated with 50 mM NaPi pH 7.4, 150 mM NaCl, at a flow rate of 0.5 mL/min and a column temperature of 25°C. Five microliters of purified protein was injected and eluted under isocratic conditions. The elution profile was monitored by UV absorbance at 230 nm. AUTHOR CONTRIBUTIONS Ian S. Truebridge: Conceptualization; investigation; methodology; validation; writing – review and editing. Frank D. Teets III: Conceptualization; investigation; writing – review and editing; writing – original draft; methodology; validation; visualization; software; formal analysis; data curation. Edson N. Cárcamo Noriega: Conceptualization; investigation; writing – original draft; methodology; validation; visualization; software; formal analysis; project administration; data curation; writing – review and editing. Tessa A. Howard: Writing – review and editing; validation; methodology; investigation. Christopher D. Bahl: Conceptualization; investigation; funding acquisition; writing – original draft; writing – review and editing; resources; supervision; project administration. Nathalia Rodriguez: Writing – review and editing; methodology; validation; investigation. James W. Bowman III: Conceptualization; writing – review and editing; project administration; investigation. Supporting information Table S1. Relative activity of AIProtease with varying solution conditions. Table S2. Golden gate assembly reaction optimization. Table S3. Benchmark set layout in a 96‐well plate. Table S4 Amino acid sequences of SUMO proteases used in this work. Figure S1. Comparison of protease construct carryover into the elution fraction. SDS–PAGE analysis of final elution products obtained using three different protease constructs. When the Cth SUMO protease was used without a His tag, contaminating protease is clearly visible in the eluate. Addition of an N‐terminal 10×His tag to the Cth SUMO protease enabled efficient removal of the protease during purification, eliminating it as a contaminant in the final product. Gels were run at different times with the primary purpose of assessing sample purity; therefore, band intensity should not be interpreted as a quantitative comparison of yield. Figure S2. Protease elution kinetics using the three different constructs were tested at 300 and 600 nM. The elution reaction was incubated at room temperature with 500 RPM orbital shaking, and eluted sfGFP was measured as fluorescent intensity (λex/λem = 488/510 nm). Imidazole elution is shown (dashed line) as a reference. All experiments were performed in triplicate. Figure S3. Evaluation of protein expression reproducibility. The platform concatenates multiple individual protocols: DNA assembly, bacterial transformation, starter culture growth, and expression culture growth. Each of these steps presents the opportunity to introduce error and variability that can impact protein expression yield. Using 10×His‐SUMO‐sfGFP expression as a reporter, we evaluated the consistency of protein expression levels. Fluorescence intensity of the culture was measured (λex/λem = 488/510) at 1:10 dilution in ultrapure water. The coefficient variation of the 96 individual expressions across four 24‐well plates of biological replicates is less than 5%. Figure S4. Sonication optimization. (A) Optimal cell suspension volume was determined by evaluating the extraction efficiency of 10×His‐SUMO‐sfGFP using different volumes of lysis buffer. For this experiment, a cell pellet derived from 2 mL of culture was used. Sonication was performed at 50% amplitude using a 24‐tip horn for 10 min continuously without cooling. Fluorescence intensity of the clarified lysate was measured (λex/λem = 488/510) and normalized by multiplying by the volume of the lysis buffer used. (B) The sonication time required to achieve complete lysis was assessed. Cell pellets were resuspended with 2 mL of lysis buffer and sonicated at 50% amplitude using a 24‐tip horn in intervals of 60 s without cooling. After every interval, a sample of the lysate was collected and the temperature of the lysate was measured (which was approximately 10 s of pause in between pulses). Samples from different time points were clarified by centrifugation and the extraction efficiency was determined by measuring sfGFP fluorescence intensity. Maximum extraction was achieved by the fourth cycle, that is, after 4 min of sonication. However, this resulted in a lysate temperature >50°C. Figure S5. Sonication ice bath. To allow the 24‐well plate to be partially submerged in ice‐cold water during sonication, a small hole was made in the upper part of the plate to allow trapped air to escape. This increased the surface contact between the plate and the cooling water, improving heat dissipation. (A) The location of the hole at the upper edge of the plate is indicated with a blank ink mark. (B) The 24‐tip sonicator head helps keep the plate submerged. The slight buoyancy of the plate allows the tips to self‐align with the centers of the wells. As the tips make contact with the bottom of the wells during sonication, the plate is gently pushed downward, automatically adjusting the vertical alignment and ensuring a consistent distance between the tips and the well bottoms. Figure S6. Protein production automated workflow. The platform consists of five protocols: (1) Golden‐Gate assembly, (2) transformation, (3) inoculation, (4) purification and (5) post‐production. Figure S7. Biomek i7 Workstation deck. An ideal deck configuration to execute the workflow possesses multiple features and accessories: 10 Tip Loading positions (TL), 10 standard deck positions (P), two trash containers (TR), two plate chillers (CP), four plate shakers (BS), an Automated Thermocycler (ATC1), a wash station for the 96‐channel head (lavender colored block on the left), and a wash station for the Span‐8 channels (W1). Liquid handling is powered by a 1 mL 96‐channel head and 1 mL Span‐8 channel pipettes. Two accessory grippers allow the labware to be transported with complete coverage of the deck. Figure S8. Production results for the Benchmark Set. (A) SDS‐PAGE of AIProtease‐eluted protein from the Benchmark Set. (B) Mean protein concentration in the eluate from three biological replicates. The automated platform was able to consistently produce 87 out of 96 proteins in the Benchmark Set. Figure S9. Effect of the N‐terminal SG scar on protease cleavage efficiency and protein yield. We observed a substantial decrease in Cth SUMO protease cleavage efficiency with ~10% of the proteins in the Benchmark Set (shown here) without the addition of an N‐terminal di‐peptide “SG” scar. This experiment was performed in triplicate. Figure S10. Option to transfer gene inserts between plasmids. Schematic representation of the cloning process, where the gene of interest is transferred from the pAIP4 plasmid to another plasmid with compatible overhangs. Amplification of the gene of interest can be achieved using primers containing recognition sites for other type IIS restriction enzymes, such as BtgzI, which offers a longer distance between the recognition site (shown in bold) and the cleavage site (indicated in the figure). The DNA insert contains 5′ CTGG and 3′ GTTC overhangs that allow it to be inserted into different plasmids using Golden Gate Assembly. Figure S11. Plate formatting strategies. We recommend two different strategies for transferring samples between 96‐well plates and 24‐well plates. (A) “Quadrant” format for manual pipetting using a 12‐channel multichannel pipette. Pipette tips must be pre‐arranged so that only 6 of 12 pipette channels are loaded, with every other channel skipped. (B) “Corner” format for use with a Span‐8 pipetting system on an automated liquid handler. Because the spacing between channels is not fixed, pipetting columns enables faster processing, and this format conserves labware usage when ≤72 proteins are being produced. ACKNOWLEDGMENTS The authors also extend our gratitude to the entire team at AI Proteins, Inc. for their collaborative spirit, technical assistance, and valuable discussions. Cárcamo Noriega EN, Truebridge IS, Teets FD III, Bowman JW III, Rodriguez N, Howard TA, et al. Facile, high‐throughput protein purification enabled by high‐efficiency protease elution. Protein Science. 2026;35(9):e70746. 10.1002/pro.70746 Review Editor: Aitziber L. Cortajarena Contributor Information Edson N. Cárcamo Noriega, Email: edson.carcamo@aiproteins.com. Christopher D. Bahl, Email: chris.bahl@aiproteins.com. DATA AVAILABILITY STATEMENT The data that support the findings of this study are openly available at https://github.com/aiproteins/high_throughput_protein_production. REFERENCES Associated Data Supplementary Materials Table S1. Relative activity of AIProtease with varying solution conditions. Table S2. Golden gate assembly reaction optimization. Table S3. Benchmark set layout in a 96‐well plate. Table S4 Amino acid sequences of SUMO proteases used in this work. Figure S1. Comparison of protease construct carryover into the elution fraction. SDS–PAGE analysis of final elution products obtained using three different protease constructs. When the Cth SUMO protease was used without a His tag, contaminating protease is clearly visible in the eluate. Addition of an N‐terminal 10×His tag to the Cth SUMO protease enabled efficient removal of the protease during purification, eliminating it as a contaminant in the final product. Gels were run at different times with the primary purpose of assessing sample purity; therefore, band intensity should not be interpreted as a quantitative comparison of yield. Figure S2. Protease elution kinetics using the three different constructs were tested at 300 and 600 nM. The elution reaction was incubated at room temperature with 500 RPM orbital shaking, and eluted sfGFP was measured as fluorescent intensity (λex/λem = 488/510 nm). Imidazole elution is shown (dashed line) as a reference. All experiments were performed in triplicate. Figure S3. Evaluation of protein expression reproducibility. The platform concatenates multiple individual protocols: DNA assembly, bacterial transformation, starter culture growth, and expression culture growth. Each of these steps presents the opportunity to introduce error and variability that can impact protein expression yield. Using 10×His‐SUMO‐sfGFP expression as a reporter, we evaluated the consistency of protein expression levels. Fluorescence intensity of the culture was measured (λex/λem = 488/510) at 1:10 dilution in ultrapure water. The coefficient variation of the 96 individual expressions across four 24‐well plates of biological replicates is less than 5%. Figure S4. Sonication optimization. (A) Optimal cell suspension volume was determined by evaluating the extraction efficiency of 10×His‐SUMO‐sfGFP using different volumes of lysis buffer. For this experiment, a cell pellet derived from 2 mL of culture was used. Sonication was performed at 50% amplitude using a 24‐tip horn for 10 min continuously without cooling. Fluorescence intensity of the clarified lysate was measured (λex/λem = 488/510) and normalized by multiplying by the volume of the lysis buffer used. (B) The sonication time required to achieve complete lysis was assessed. Cell pellets were resuspended with 2 mL of lysis buffer and sonicated at 50% amplitude using a 24‐tip horn in intervals of 60 s without cooling. After every interval, a sample of the lysate was collected and the temperature of the lysate was measured (which was approximately 10 s of pause in between pulses). Samples from different time points were clarified by centrifugation and the extraction efficiency was determined by measuring sfGFP fluorescence intensity. Maximum extraction was achieved by the fourth cycle, that is, after 4 min of sonication. However, this resulted in a lysate temperature >50°C. Figure S5. Sonication ice bath. To allow the 24‐well plate to be partially submerged in ice‐cold water during sonication, a small hole was made in the upper part of the plate to allow trapped air to escape. This increased the surface contact between the plate and the cooling water, improving heat dissipation. (A) The location of the hole at the upper edge of the plate is indicated with a blank ink mark. (B) The 24‐tip sonicator head helps keep the plate submerged. The slight buoyancy of the plate allows the tips to self‐align with the centers of the wells. As the tips make contact with the bottom of the wells during sonication, the plate is gently pushed downward, automatically adjusting the vertical alignment and ensuring a consistent distance between the tips and the well bottoms. Figure S6. Protein production automated workflow. The platform consists of five protocols: (1) Golden‐Gate assembly, (2) transformation, (3) inoculation, (4) purification and (5) post‐production. Figure S7. Biomek i7 Workstation deck. An ideal deck configuration to execute the workflow possesses multiple features and accessories: 10 Tip Loading positions (TL), 10 standard deck positions (P), two trash containers (TR), two plate chillers (CP), four plate shakers (BS), an Automated Thermocycler (ATC1), a wash station for the 96‐channel head (lavender colored block on the left), and a wash station for the Span‐8 channels (W1). Liquid handling is powered by a 1 mL 96‐channel head and 1 mL Span‐8 channel pipettes. Two accessory grippers allow the labware to be transported with complete coverage of the deck. Figure S8. Production results for the Benchmark Set. (A) SDS‐PAGE of AIProtease‐eluted protein from the Benchmark Set. (B) Mean protein concentration in the eluate from three biological replicates. The automated platform was able to consistently produce 87 out of 96 proteins in the Benchmark Set. Figure S9. Effect of the N‐terminal SG scar on protease cleavage efficiency and protein yield. We observed a substantial decrease in Cth SUMO protease cleavage efficiency with ~10% of the proteins in the Benchmark Set (shown here) without the addition of an N‐terminal di‐peptide “SG” scar. This experiment was performed in triplicate. Figure S10. Option to transfer gene inserts between plasmids. Schematic representation of the cloning process, where the gene of interest is transferred from the pAIP4 plasmid to another plasmid with compatible overhangs. Amplification of the gene of interest can be achieved using primers containing recognition sites for other type IIS restriction enzymes, such as BtgzI, which offers a longer distance between the recognition site (shown in bold) and the cleavage site (indicated in the figure). The DNA insert contains 5′ CTGG and 3′ GTTC overhangs that allow it to be inserted into different plasmids using Golden Gate Assembly. Figure S11. Plate formatting strategies. We recommend two different strategies for transferring samples between 96‐well plates and 24‐well plates. (A) “Quadrant” format for manual pipetting using a 12‐channel multichannel pipette. Pipette tips must be pre‐arranged so that only 6 of 12 pipette channels are loaded, with every other channel skipped. (B) “Corner” format for use with a Span‐8 pipetting system on an automated liquid handler. Because the spacing between channels is not fixed, pipetting columns enables faster processing, and this format conserves labware usage when ≤72 proteins are being produced. Data Availability Statement The data that support the findings of this study are openly available at https://github.com/aiproteins/high_throughput_protein_production.