An easy-to-use high-throughput selection system for the discovery of recombinant protein binders from alternative scaffold libraries

Abstract

Selection by phage display is a popular and widely used technique for the discovery of recombinant protein binders from large protein libraries for therapeutic use. The protein library is displayed on the surface of bacteriophages which are amplified using bacteria, preferably Escherichia coli, to enrich binders in several selection rounds. Traditionally, the so-called panning procedure during which the phages are incubated with the target protein, washed and eluted is done manually, limiting the throughput. High-throughput systems with automated panning already in use often require high-priced equipment. Moreover, the bottleneck of the selection process is usually the screening and characterization. Therefore, having a high-throughput panning procedure without a scaled screening platform does not necessarily increase the discovery rate. Here, we present an easy-to-use high-throughput selection system with automated panning using cost-efficient equipment integrated into a workflow with high-throughput sequencing and a tailored screening step using biolayer-interferometry. The workflow has been developed for selections using two recombinant libraries, ADAPT (Albumin-binding domain-derived affinity proteins) and CaRA (Calcium-regulated affinity) and has been evaluated for three new targets. The newly established semi-automated system drastically reduced the hands-on time and increased robustness while the selection outcome, when compared to manual handling, was very similar in deep sequencing analysis and generated binders in the nanomolar affinity range. The developed selection system has shown to be highly versatile and has the potential to be applied to other binding domains for the discovery of new protein binders.

Graphical Abstract

Graphical Abstract.

Introduction

The development of specific protein binders, based on antibodies or alternative scaffolds, has been instrumental for acquiring new knowledge regarding the natural presence and activities of proteins, as well as supplying new therapeutic agents. Binders are routinely selected from large protein libraries using display technologies such as phage, cell, or ribosome display (Smith, 1985; McCafferty et al., 1990; Frenzel et al., 2016; Dreier and Plückthun, 2018; Jaroszewicz et al., 2022).

The phage display system allows for easy scale-up and parallelization, meaning the possibility to select with multiple tracks for numerous targets simultaneously. However, the selection process requires a lot of manual work and handling multiple tracks quickly becomes cumbersome, particularly when selecting in solution with biotinylated target proteins, followed by capture on streptavidin-coated magnetic beads. Panning with the target proteins in solutions is preferred over surface panning as it allows the target molecule to be more accessible and reduces the role of avidity favoring high-affinity interactions (Zimmermann et al., 2020). The workload of this multi-step panning can be tremendously reduced by automation using robotics in the form of a particle processor. Automation and implementation of a plate-based format allows for increased throughput and reproducibility, decreased hands-on time and in addition to that, less consumables needed. However, systems for automated panning already in use require very high-priced equipment, making it unaffordable for many labs (Ch’ng et al., 2019). Furthermore, automation of the selection process alone shifts the bottleneck further to the screening and evaluation part of the workflow. Traditionally, phage display selections are followed by screening of single clones using ELISA combined with Sanger sequencing. While this is a straightforward strategy, it is usually limited to a couple hundred clones, resulting in only a snapshot of the whole selection output. To overcome this limitation, next-generation, also called high-throughput or deep sequencing is used for the identification of large selection outputs. Studies comparing the screening of antibody libraries with high-throughput sequencing and ELISA have shown, that while top-ranking clones were identified by both technologies, deep sequencing enabled the discovery of more specific binders among rare clones (Ljungars et al., 2019; Passariello et al., 2021). Short read high-throughput sequencing such as Illumina is limited in the length of reads that can be produced, which results in heavy and light chain matching being lost in antibody libraries, however, this is not an issue for synthetic libraries with shorter randomized regions that can be covered by a single read.

Two such synthetic libraries have been used to generate binders in the present study, both based on scaffolds constituted of three alpha-helices with a molecular weight between 6 and 8.5 kDa which are conveniently produced in E. coli. The albumin-binding domain-derived affinity protein (ADAPT) library is based on one of the albumin-binding domains of streptococcal protein G. The ABD scaffold has been previously engineered in our group, resulting in high affinity and specific binders toward various targets of interest for purification, diagnostic or therapeutic applications which have reached a first-in-human phase I clinical trial (Alm et al., 2010; Nilvebrant et al., 2011, 2013, 2014; Garousi et al., 2015; Bragina et al., 2021). Most recently, the ADAPT scaffold has been further engineered for simultaneous binding to human serum albumin (HSA) and another target-of-choice (Witting et al., 2021). The second library employed in this study is the calcium-regulated affinity (CaRA) library, which is based on the engineered B domain, called Z, of staphylococcal protein A. Previously, a calcium-binding loop has been inserted in the scaffold between helix 2 and 3, rendering the binding to the Fc of IgG calcium-dependent (Kanje et al., 2018). In a recent study, this scaffold has been developed further for calcium-regulated binding to novel antigens. For this aim, a randomized surface has been introduced in helix 1 and 2 which is similar to that of the Affibody (Jönsson et al., 2022). In this study, we aimed to provide an efficient, high-throughput workflow for the selection of binders from these large synthetic protein libraries. The workflow consists of three parts, (i) a time-efficient automated panning using an inexpensive magnetic-bead handling device with subsequent compatible plate-based amplification, (ii) analyses of a broader repertoire of promising clones through Illumina sequencing, (iii) and finally, rapid detection of desired binding characteristics through BLI-screening of cell supernatant, allowing for small scale production without purification for the assessment of the binding kinetics. To validate and show the versatility of the workflow, we conducted new selections against diverse target proteins, namely an antibody part (IgE Cε3-Cε4 with the CaRA library), a membrane-anchored glycoprotein and a soluble ligand (CEACAM5 and CCL7 with the ADAPT library).

Overall, our results demonstrate the power and versatility of the developed system, which not only significantly speeds up the discovery process when screening large synthetic libraries but also improves the robustness by reducing human error and being able to precisely control the panning. With this approach, we were able to select binders from the CaRA and ADAPT library with a wide range of affinities and characteristics, offering a rich source of diverse binders that have the potential to be further developed as novel therapeutic agents.

Materials and Methods

Production and biotinylation of target proteins

All recombinant target proteins used in this study were produced in CHO cells by the Human Secretome Project (Stockholm, Sweden) as described by Tegel et al., 2020. The protein purity was assessed by SDS-PAGE and analytical size exclusion chromatography (SEC). The target protein CCL7 was further purified by SEC (Table SI). All target proteins were biotinylated by incubation for 30 min at room temperature (RT) with a 6–12× molar excess of EZ-link Sulfo-NHS-LC-Biotin (Thermo Fisher Scientific, Waltham, MA, USA) in PBS (Phosphate-buffered saline) or TBS (Tris-buffered saline) pH 7.4. Excess of non-reacted biotin reagent was removed by performing a buffer exchange using Illustra NAP-5 columns (Cytiva, Stockholm, Sweden). The degree of biotinylation was assessed by performing a binding test with streptavidin-coated magnetic beads (Dynabeads™ M-280 Streptavidin, Thermo Fisher Scientific, Waltham, MA, USA).

Manual panning procedure with amplification in shake flasks

Manual selections and amplification in flasks with the ADAPT library against protein targets TNFα and Angiogenin were performed as described in detail in Witting et al., 2021.

Automated panning during phage display and amplification in 24-well plates

A magnetic bead handling device TANBead Maelstrom™ 8 Autostage (Taiwan Advanced Nanotech, Taoyuan, Taiwan) was used for automated selections in 96-deep well plates. Four rounds of panning were carried out using biotinylated recombinant target protein in solution with capture on streptavidin-coated magnetic beads (Dynabeads™ M-280 Streptavidin). To avoid selecting non-specifically binding phages, the 96-well plate, the magnetic beads and the spin tips were blocked with PBST or TBST (0.1% Tween-20) supplemented with 0.5% gelatin or 3% bovine serum albumin (BSA) beforehand. Additionally, streptavidin-binding phages were depleted by a negative selection step preceding each selection round in which the phage library diluted in PBST or TBST with 0.1% gelatin or 3% BSA was incubated in blocked tubes with 0.5 mg of magnetic beads for 30 min at RT with 150 rpm. The supernatants from the negative selection were used as input for the selection with the biotinylated target proteins. The pre-selected supernatants were mixed with 150, 100, 50, and 25 nM biotinylated target protein, for the four rounds respectively, in the first wells of the 96-well plate. Further, magnetic beads, wash buffer (PBST or TBST) and elution buffer (50 mM glycine, pH 2.0/50 mM or 5 mM EDTA pH 6.0) for the respective selection track were added to the plate. A beforehand programmed selection protocol was started which contained mixing of phage library and target for 2 h during the first round and 1 h for the second to fourth round at 500 rpm, capture with magnetic beads for 10 min at 500 rpm, followed by two, four, eight and twelve washes, respectively, at 1500 rpm for 1 min each, elution for 10 min in elution buffer and finally capture of the magnetic beads. The solution with eluted phages was neutralized with equal parts 10% 1 M Tris-HCl in PBS, pH 8.0, or 1 M CaCl₂ in TBST.

The eluted phages were amplified by infecting XL1 blue E. coli cells in a 10–100× excess compared to eluted phages in a 24-well plate. 2 ml XL1 blue E. coli cell culture in TSB (30 g/l tryptic soy broth) with 10 μg/ml tetracycline was added to each well and the eluates were split over 1–4 wells, depending on the estimated number of phages. Afterward, the cells were allowed to be infected for 15 min without shaking at 37°C, followed by 15 min shaking at 300 rpm, thereafter 2 ml TSB supplemented with 200 μg/ml carbenicillin and 10 μg/ml tetracycline was added. After 30 min incubation at 270 rpm, a 10× excess of M13K07 helper phages, compared to the number of cells at infection, was added. After another 90 min incubation at 200 rpm, the cells were pelleted by centrifugation of the plate at 3500 × g for 20 min and the resulting pellets were resuspended in 4 ml of TSB + YE (30 g/l tryptic soy broth and 5 g/l yeast extract) with 100 μg/ml carbenicillin, 25 μg/ml kanamycin and 0.1 mM IPTG. Finally, the cells were incubated around 16 h overnight at 30°C and after pelleting the cells, the amplified phages were recovered by precipitation using 4% polyethylene glycol in 0.5 M NaCl at 3500 × g for 30 min at 4°C the next day. An additional wash step by resuspending the phage pellets in water followed by a second precipitation was performed before the phages were resuspended in the selection buffer and used as input for the next selection round. Phage titers were calculated from titrations of infected E. coli cells.

Polyclonal phage ELISA

Amplified phage stocks from each selection round were diluted in PBSC (PBS + 0.5% Casein) or TBSC (TBS + 0.5% Casein) and incubated with target proteins (10 μg/ml) or IgG (0.1 μg/ml) coated on 96-well ELISA half-area plate for 1 h at RT with slow shaking. After 3× 10 min washes with PBST, anti-M13-HRP antibodies (HRP/ANTI-M13 Monoclnl CONJUG. 27-9421-0, Cytiva, Stockholm, Sweden) were diluted 1:5000 in PBSC and incubated for 1 h. After 3× 10 min washes with PBS, bound anti-M13-HRP was detected using TMB substrate (Pierce™ TMB Substrate Kit, Thermo Fisher Scientific, Waltham, MA, USA) and absorbances were read at 450 nm.

High-throughput sequencing and data analysis

Phagemid DNA was prepared from XL1 blue E. coli that had been infected with eluates from the respective selection rounds during the amplification using a QIAprep Miniprep kit (Qiagen, Hilden, Germany). The library sequences were amplified by performing a PCR with 50 ng of the phagemids and 5 pmol of an oligo containing the forward adapter sequence and 5 pmol of an oligo containing the reverse adapter sequence together with a unique index sequence specific for each sample. The PCR was run for 15 cycles and the resulting products were extracted from a 2% GTG agarose gel using a QIAquick gel extraction kit (Qiagen). All samples were pooled to a final concentration of 10 nM and sequenced using an Illumina MiSeq v2 instrument. Obtained FASTQ files were analyzed using in-house software (unpublished Karlander et al., 2022). Quality control was performed by depleting sequences with incorrect lengths and bases with a Phred quality score < 30. Following that, sequences were ranked according to their relative frequency.

Subcloning and production of lead candidate binders

A number of high-ranked variants were chosen for each of the target proteins CEACAM5 and CCL7 from the ADAPT library (Witting et al., 2021) and IgE Cε3-Cε4 from the CaRA library (Jönsson et al., 2022). The gene variants from selections against CEACAM5 and CCL7 were recovered by a PCR-based strategy. Sequence-specific oligos which covered helix 1 and 2 and included an N-terminal His₆-sequence as well as restriction sites were used as forward and reverse primers in a PCR reaction with plasmids containing the library as a template. Afterward, restriction cloning was performed to clone the variants into a T7 inducible expression vector containing the non-randomized helix 3 (Witting et al., 2021). Gene variants from selections against IgE Cε3-Cε4 were synthesized by Thermo Fisher Scientific GENEART GmbH (Regensburg, Germany) and also cloned into a T7 inducible expression vector after PCR amplification with primers introducing an N-terminal His₆-sequence and restriction sites. All cloned variants were sequence-verified by Sanger sequencing (Eurofins Genomics, Ebersberg, Germany) and the variants were expressed in BL21(DE3) E. coli.

High-throughput screening using biolayer interferometry

The chosen ADAPT and CaRA variants were expressed in BL21 E. coli in 24-well plates. After overnight production in 4 ml TSB + YE, the cells were pelleted by centrifugation of the plate at 3500 × g for 20 min at 4°C. The resulting cell pellets were lysed in 500 μl lysis buffer containing 0.5 mg/ml lysozyme (from chicken egg white, Sigma-Aldrich, Burlington, Massachusetts, United States), DNase I (Roche, Basel, Switzerland) and 1 mM MgCl₂ in 20 mM Tris, pH 8, for 1 h at RT with slow shaking. The lysed cells were pelleted by centrifugation and the supernatants were used for screening in an Octet 96e red (Pall ForteBio, Fremont, CA, USA). The supernatants were diluted 1:1 in PBST or TBST with 1 mM CaCl₂ and the His-tagged proteins were loaded on Octet® Anti-Penta-HIS (HIS1K) Biosensors (Sartorius AG, Göttingen, Germany) for 600 s with 1000 rpm. After a baseline step with buffer for 300 s, the sensors with loaded proteins were dipped into a well with the respective target diluted to 1000 nM for the association for 300 s. Dissociation in wells with buffer solution was done for 600 s followed by regeneration of the sensors with 10 mM glycine, pH 1.5 alternating between buffer and regeneration solution 3× 20 s each. A blank sensor was used as a reference.

Purification of lead variants

After production in BL21 E. coli as described in the previous section but here in 100 ml culture volume, the cells were lysed by sonication. The ADAPT variants were purified by affinity chromatography using a resin with coupled human serum albumin (HSA) produced in-house and the CaRA variants were purified by immobilized metal-ion affinity chromatography using HisPur™ cobalt resin (Thermo Fisher Scientific, Waltham, MA, USA). The purity and molecular weight were assessed by SDS-PAGE and LC-MS.

Characterization of target binding by surface plasmon resonance (SPR)

Binding analysis of the purified ADAPT and CaRA variants to their targets was conducted using surface plasmon resonance (SPR) Biacore T200 and 8 K systems (Cytiva). All analyses were performed using a CM-5 sensor chip (Cytiva) and all ligands were immobilized using amine coupling. The CaRA variants were directly immobilized on the sensor surface, while the ADAPTs were either captured on HSA (CCL7 variants) or the target was immobilized (CEACAM5). The respective analytes were injected in 1:2 dilutions in TBST with 1 mM CaCl₂ or PBST (0.05% Tween-20) at a flow rate of 30 μl/min at 25°C. The surfaces were regenerated with 10 mM HCl. The binding curves were analyzed and kinetic parameters were determined using the Biacore Insight Evaluation software, according to the respective expected interaction.

Structural evaluation by circular dichroism (CD)

The secondary structure content was evaluated using a Chirascan circular dichroism spectrometer (Applied Photophysics, Surrey, UK). All samples were diluted to 0.2 mg/ml and measured in a cell with an optical path length of 1 mm. The secondary structure content was evaluated by measuring the ellipticity from 260 to 195 nm at 20°C.

Results and Discussion

Design of an affordable easy-to-use high-throughput workflow

A high-throughput selection system with an implemented automated selection procedure was developed for the discovery of recombinant protein binders using phage display. To make the procedure faster and allow for a more thorough screening, a workflow for high-throughput analysis was established as shown in Fig. 1. In the first step, a magnetic bead handling device was used for an automated selection procedure including incubation of biotinylated target proteins of interest with the phage library, several carefully designed washes and finally an elution step retrieving the target-binding phages. To simplify the handling and to increase the throughput, a protocol for the amplification of eluted phages in a 24-well plate format was developed. After four selection rounds, polyclonal phage ELISA was performed to validate the success of the selection before the outputs were analyzed by Illumina MiSeq deep sequencing. Further, the workflow included a high-throughput screening step using BLI where potential binders were evaluated directly from unpurified supernatants.

Fig. 1.

Schematic description of the high-throughput selection system. The first step involves automated selection from the phage-displayed protein library using a magnetic bead handling device and amplification of the outputs of each selection round in plate format. The success of the selection may be studied by polyclonal phage ELISA before the selection outputs are analyzed using deep sequencing to find the most enriched variants over the background. Lastly, potential binders are screened in a high-throughput manner using biolayer interferometry.

Selection outcome of automated panning and amplification in plates is comparable to a manual workflow outcome

To develop and optimize a simplified selection procedure including an automated selection step and amplification in plate, the new semi-automated system and its outcome were compared to a successful manual procedure selecting against the same target proteins. The target proteins TNFα and Angiogenin (ANG) were chosen for the comparison (Witting et al., 2021). Four rounds of phage display selection were performed using the ADAPT library with a size of 5*10⁹ variants (Witting et al., 2021). To ensure high efficiency in all different steps, three diverse experiments were performed: (i) manual selection and amplification in flasks, (ii) automated selection and amplification in flasks (TNFα only (Fig. S1 available at PEDS online)) and (iii) automated selection and amplification in plates. Analysis of the output from each selection round by polyclonal phage ELISA showed an increase in specific target binding after the third and fourth rounds for both the automated and the manual selection against TNFα (Fig. S2). A comparison of phage titers during phage amplification in E. coli shows that similar titers are obtained with amplification in flasks and plates (Fig. 2). Based on the estimation that less than 10% of the phages express a library member (Russel et al., 2004), it is desired to amplify the eluted phages at least 2000-fold in order to ensure 100 copies of each clone. This large excess of phages is most important in early rounds as the number of unique variants decreases throughout the selection process. Fig. 2 illustrates that the desired amplification rate from output to amplified input for the next selection round is highly exceeded with both amplification methods in selections against TNFα and ANG.

Fig. 2.

Comparison of phage titers after flask vs. plate amplification from (a) the TNFα selection and (b) the ANG selection. The phage titer from amplification in flask and in plate are comparable showing that sufficient amplification can be achieved with the optimized protocol in plates. The error bars show the standard deviation from two technical replicates.

The outputs from the second, third and fourth selection rounds were sequenced using Illumina MiSeq and the sequence analysis shows that the gene distributions are shifted as desired and in a similar way in all performed selections (Fig. 3). The gene diversity is high in the beginning and decreases over the course of the selection rounds when certain variants become more enriched over the background thus indicating a successful selection.

Fig. 3.

Gene distribution as percent (0–50%) of all sequenced genes from selections against TNFα and ANG comparing manual vs. automated selection. Each cross represents a single gene variant. The gene diversity decreases from earlier to later selection rounds in both the manual and automated selections in a similar way which indicates successful selections as certain variants are enriched.

A detailed analysis of the amino acid distribution in the 11 randomized positions of the ADAPT library shows a very similar selection outcome in all different selections against both TNFα and ANG, regardless of the degree of automation. After the fourth selection round in the automated process, the gene distribution was analyzed and compared with the manual selection outcome for both TNFα and ANG (Fig. 4a and b). It has to be noted that the manual TNFα selections were performed using a version of the ADAPT library that contains cysteines and prolines in library positions 30, 31 and 34, whereas the automated selections were done with a newer version of the library which did not contain these two amino acids, as can be seen in Fig. 4 (gray and light pink). The outputs from the different selections are very similar, especially in the positions that show a clear enrichment throughout the selections.

Fig. 4.

Amino acid distribution in the randomized library positions of selection output after round 4. (a) Automated selection followed by amplification in plate against TNFα vs. manual selection (ADAPT 2nd gen). (b) Automated selection followed by amplification in plate against ANG vs. manual selection. The positions which are likely to be target-specific due to their clear enrichment, appear to result in a very similar distribution of certain amino acids, independent of the selection method used.

Overall, we found that while the selection outcome was comparable to the successful manual procedure, the automation of the panning greatly improved the robustness and reproducibility of the protocol, irrespective of the selector. Consequently, the speed of mixing and time of the washes during automated panning can be adjusted very precisely, giving more control over the selection stringency.

Selections against new target proteins using the high-throughput selection system confirm the robustness and efficiency

To test and validate that the newly developed selection workflow could be used for finding binders toward various soluble proteins without being limited to certain target requirements, selections against three new target proteins with different characteristics were made using two different protein libraries. The target proteins chosen were CCL7 and CEACAM5 for selection with the ADAPT library (Witting et al., 2021) and IgE Cε3-Cε4 for selection with the CaRA library (Jönsson et al., 2022). We chose these proteins for their relevance as potential therapeutic targets and to demonstrate the applicability of the selection workflow for proteins of different molecular weights and properties (Table SI). Four rounds of selection using the high-throughput system were performed for all targets.

High-throughput sequencing data

The outputs of the selections against CCL7, CEACAM5 and IgE Cε3-Cε4 were sequenced using Illumina MiSeq which was chosen as a cost-efficient alternative to Sanger sequencing due to the number of reads obtained and the possibility of finding enriched binders earlier during the selection process. For all targets except CEACAM5, outputs from rounds 2 to 4 were sequenced, but for CEACAM5 only the output of the second round was sequenced, due to low yields of correct length DNA in the sample preparation of rounds 3 and 4. The gene distributions of the total number of reads shows enrichment of specific variants for the selection against CCL7 and IgE Cε3-Cε4 which increases over the course of the rounds (Fig. S3). The most enriched sequence variants from rounds 3 and 4 for CCL7, round 4 for IgE Cε3-Cε4 and round 2 for CEACAM5 were clustered in dendrograms according to their sequence similarity (Figs S4–S7). This shows a reasonable diversity even in later panning rounds allowing the selection of sequences that might not have been identified without extensive screening. A few variants were selected from the most prevalent clusters for each of the target proteins and they were synthesized, subcloned and produced in E. coli.

High-throughput screening of candidates using BLI enables faster kinetic analysis

A protocol for easy and high-throughput target binding analysis, that makes it possible to screen for desired kinetic characteristics directly from E. coli lysates using biolayer interferometry (BLI), was developed. Based on sequence similarity and enrichment over the background, potential binders were selected from the deep sequencing outputs for further analyses. Since the high-throughput sequencing shows the selection of enriched clones earlier in the selection process, as early as from round 2 for the target CEACAM5, previously needed extensive screening approaches become more redundant. Hence, it could be time- and resource-efficient to downscale the subsequent screening throughput and instead tailor it to include the advantage of receiving kinetic estimates of target binding earlier in the workflow. This would enable earlier selection of lead candidates with desired binding characteristics. Therefore, we developed a general BLI protocol that starts with capturing His-tagged binding domains directly from the lysates onto sensors pre-immobilized with anti-Penta-HIS antibodies that are subsequently moved into wells with their respective target protein in well-defined concentrations. In Fig. 5, binding curves for analyses of three different pairs of monomeric binders and target proteins are shown (Fig. S9). Taking all binding analyses into consideration, all variants except ADAPT_{CCL7_91}, ADAPT_{CEACAM5_22} and CaRA_{IgE Cε3-Cε4_a8} show interaction with their respective target protein, although with different kinetics (Fig. S8a–c). Furthermore, cross-reactive binding to other targets was not observed, indicating specificity of the binders (Fig. S8a–c). No binding or low signals in the BLI analysis correlated with later-on assessed properties such as improper folding (Table SII), further indicating that this screening step is well suited to identify desired candidates early.

Fig. 5.

Screening for target binding using biolayer interferometry. Some of the most promising variants are (a) CaRA_{IgE Cε3-Cε4_1,} (b) ADAPT_{CEACAM_4} and (c) ADAPT_{CCL7_1}. The variants were captured on sensors from lysates by their His-tags for association with their respective target at a concentration of 1000 nM.

The BLI data obtained from screening the lysates of chosen candidates are satisfactory for a first assessment of the kinetic interaction and comparable to the parameters assessed by SPR of the purified variants in a further evaluation (Fig. 6a–c). However, it is expected that the reliability of the kinetic parameters from the lysate screening is not as accurate as for the purified binders since baseline drifts can be an issue. Moreover, this screening step was done with only one concentration of the analyte.

Fig. 6.

(a–c) Binding affinities evaluated by SPR. Variants were a) immobilized on the sensor surface (CaRA_{IgE Cε3-Cε4_1}, b) the target was immobilized (ADAPT_{CEACAM5_4}, or c) captured on HSA (ADAPT_{CCL7_1}), and the respective analytes were injected in at least three concentrations. The assumed binding model fits (a, bivalent binding b, 1:1 binding and c, two-state reaction) are shown in black. Regeneration with 10 mM HCl reduced the surface activity in (b), however, other regeneration solutions such as 50 mM NaOH and 0.05% H₃PO₄ were inferior. (d–f) Secondary structure evaluation by CD confirms alpha-helicity of all the selected proteins.

Deeper characterization of binders against new target proteins by SPR and CD validates the properties discovered using the high-throughput scheme

To further evaluate the accuracy and sensitivity of the screening method based on BLI, selected variants were purified using affinity chromatography and their target binding was analyzed using SPR. The respective target protein or binder was immobilized on the sensor chip surface. Furthermore, the target specificity was tested by injecting the variants over another irrelevant protein. Interestingly, all binders found using the parallel high-throughput format that showed a target affinity in the SPR analysis were also positive in the BLI screening and vice versa, which validates the sensitivity and accuracy of the BLI method and the applicability of the entire process. Hence, the screening method is fast and simultaneously sensitive and therefore very suitable to reliably sort out binders with promising kinetic characteristics as a first characterization step. In total, 22 binders were assessed for the three targets (Table SII). The most promising binders could already be identified in the screening step from unpurified lysates, making it possible to omit purification for the initial characterization. More importantly, it was possible to get a first insight into the binding kinetics, which could later be confirmed by SPR. In Fig. 6a–c, binding curves from SPR for the same binders as shown in Fig. 5a–c are displayed for comparison. Since IgE Cε3-Cε4 is in dimeric form and here used as the analyte, the binding curves were fitted to a bivalent binding model (Fig. 6a and Table SIV). ADAPT_{CEACAM_4} displayed a very high affinity of 0.13 nM to CEACAM5 when fitted to a 1:1 binding model (Fig. 6b). The low sensorgram response may be caused by partly shielding of the epitope due to the amino-linked immobilization and the pronounced glycosylation. The interaction of ADAPT_{CCL7_1} with its target was fitted to a two-state reaction model, due to the captured ADAPT molecule likely undergoing conformational change upon target binding. The calculated K_D of this interaction is 643 nM. Further, the secondary structures of the binding domains were evaluated by CD spectroscopy which confirmed the expected alpha-helical conformation (Fig. 6d–f) of these alternative scaffold proteins. All in all, we found that it was possible to efficiently select and characterize binders with favorable characteristics that display affinities to their targets in the nanomolar range with this approach. These outcomes and characteristics are comparable to binders achieved in previous selection campaigns with these two protein libraries (Witting et al., 2021; Jönsson et al., 2022).

The binders selected in this campaign toward IgE Cε3-Cε4, CEACAM5 and CCL7 will be further evaluated for their potential use as therapeutics in upcoming studies. Especially the selected CEACAM5 binder shows very favorable characteristics with a picomolar affinity. Moreover, other selection campaigns using the described methods have generated binders against a broad range of targets (Table SIII).

While our new selection approach generated many promising binders, we found that besides reducing the overall development time and consumable costs, there was also a dramatic decrease in hands-on time during the phage display selection process itself (Table I). When comparing the time needed for the same number of selection tracks, including washes and elution, it became clear that while the number of targets and washes considerably increases the hands-on time in the manual workflow, this is not the case for the semi-automated workflow. In the semi-automated workflow, the same amount of time is needed for the parallel selection of up to 16 tracks with 3 washes or 8 tracks with 9 washes. Further, in order to elevate the throughput even more, it is possible to simultaneously select binders against 24 targets in the same selection plate if reusing the same wells for the washes. Although it was not evaluated, we speculate that the risk for re-binding of phages would be very low when selecting against different target proteins. Also, by keeping the mixing speed of the spin tips as high as in this developed protocol, rebinding of low-affinity binders would be neglectable. Since the automated selection with the magnetic bead handling device takes around 1–3 h, depending on the selection time and number of washes, it is also possible to start several runs per day.

Table I.

Comparison of hands-on selection time using the manual and semi-automated workflow

	Manual workflow	Semi-automated workflow	Manual workflow	Semi-automated workflow
Number of targets/tracks	8		16
Number of washes	9		3
Hands-on time selection	(5 + 1 × 7) × 9 min	10 min	(5 + 1 × 15) × 3 min	10 min
	110 min		60 min
Hands-on time amplification	(15 + 5 × 7) min	20 min	(15 + 5 × 15) min	25 min
	50 min		90 min
Total hands-on time	160 min	30 min	170 min	35 min

For the manual workflow we calculated 5 min per wash and an additional 1 min for each target or track during the selection and 15 min hands-on time with an additional 5 min for each track during the amplification in flasks due to the separate handling. For the semi-automated workflow, we calculated 10 min for preparing the selection plate and 20–25 min in total for the plate amplification with little time increase per additional target added due to the parallelized batch handling.

Besides the increase in throughput, we find that our selection system comprises additional advantages such as high reproducibility and ease of adjustment of stringency. While reproducibility is a natural advantage of automation, this protocol can also be easily adjusted, for example by regulating the speed of the mixing during washes according to desired stringency. The selection using magnetic beads in solution makes this approach suitable for a broad range of targets such as soluble ligands and membrane proteins. The magnetic bead handling device for automated selection provides a cost-efficient alternative to other high-priced liquid handling systems.

Conclusions

We developed and evaluated a new, semi-automated selection system for recombinant binder proteins using phage-display which has been shown to efficiently enrich selective binders with down to nanomolar affinity to their target. The selection system has been developed and evaluated for two recombinant libraries, ADAPT and CaRA, but we believe that it is equally applicable to other types of binding domains. Further, when comparing the outcome with a traditional, manually performed selection campaign, the outcome is very similar, while reducing hands-on time and cost of goods remarkably. We found that the automation of the panning greatly improves the robustness and reproducibility of the selections, enabling fine-tuning of the panning protocol. Additionally, within the high-throughput pipeline it was possible to characterize new binders regarding affinity characteristics in a highly parallelized manner. While other semi-automated or automated systems require high-priced very high-throughput equipment for panning (Konthur and Walter, 2002; Konthur et al., 2010; Ch’ng et al., 2019), which then shifts the bottleneck to the subsequent screening and testing of candidates, our workflow is designed with integrated steps, going from selection using more affordable equipment to a first assessment of kinetic parameters of the candidates. By utilizing this semi-automated selection system, the path from selection to evaluation of lead candidates is affordable and efficient, while still remaining highly tunable.

Edited by: Dr Feng Ni

Author contributions

Marit Möller (Formal analysis [lead], Investigation [equal], Methodology [equal], Validation [equal], Visualization [equal], Writing—original draft [lead], Writing—review & editing [equal]), Malin Jönsson (Formal analysis [supporting], Investigation [supporting], Methodology [supporting], Writing—original draft [supporting], Writing—review & editing [equal]), Magnus Lundqvist (Investigation [supporting], Software [lead], Validation [supporting], Visualization [supporting], Writing—review & editing), Blenda Hedin (Formal analysis [supporting], Investigation [supporting], Methodology [supporting], Writing—review & editing [supporting]), Louise Larsson (Formal analysis [equal], Investigation [equal], Methodology [equal], Writing—review & editing [supporting]), Emma Larsson (Data curation [supporting], Formal analysis [supporting], Investigation [supporting], Methodology [supporting], Writing—review & editing [supporting]), Johan Rockberg (Methodology [equal], Software [supporting], Writing—review & editing [supporting]), Mathias Uhlén (Funding acquisition [supporting], Writing—review & editing [supporting]), Sarah Lindbo (Conceptualization [equal], Investigation [equal], Methodology [equal], Supervision [equal], Writing—review & editing [supporting]), Hanna Tegel (Investigation [equal], Methodology [equal], Writing—review & editing [supporting]) and Sophia Hober (Conceptualization [equal], Funding acquisition [lead], Investigation [equal], Methodology [equal], Project administration [lead], Resources [lead], Supervision [lead], Validation [equal], Writing—review & editing [equal])

Funding

This work is financially supported by Vinnova (the CellNova center, grant 2017-02105), the Swedish research council (grant 2016-04717 and 2021-04289) and the Knut and Alice Wallenberg Foundation.

Data availability

The authors confirm that the data supporting the findings of this study are available within the article and its supplementary material. The code used for the sequencing analysis is available upon request.