Using extension‐based mRNA display to design antibody‐like proteinogenic peptides for human PD‐L1

Abstract

Many peptide drugs rely on nonproteinogenic amino acids and chemical modifications for improved activity and proteolytic stability. However, these features also make drug production expensive and challenging to scale. Here, we engineered small, linear, proteinogenic peptides that bind human programmed death‐ligand 1 (hPD‐L1) with high affinity and stability using mRNA display affinity maturation. The resulting peptides, SPAM2 and SPAM3, have antibody‐like affinities for hPD‐L1 (dissociation constants between ~250 and 300 pM) and are selective for hPD‐L1. Both SPAM2 and SPAM3 compete with hPD‐L1 ligands known to interact with the programmed cell death protein 1 site and are stable in human serum. SPAM3 bound human glioma D423 cells with high affinity in flow cytometry experiments comparable to that of a clinical therapeutic antibody. These results support the use of affinity maturation selections to dramatically enhance the biophysical properties of linear, proteinogenic peptides for translational applications.

1. INTRODUCTION

Cancer cells commonly exploit immunosuppressive pathways to evade immune destruction. For example, the interaction between programmed cell death protein 1 (PD‐1) and human programmed death‐ligand 1 (hPD‐L1) mediates T‐cell growth, function, and survival (Bardhan et al., 2016; Freeman et al., ²⁰⁰⁰). Many monoclonal antibodies target this interaction for therapeutic and diagnostic applications (Doki et al., 2022; Larson et al., ²⁰²²; Melisi et al., ²⁰²¹; Saxena et al., ²⁰²¹; Seto et al., ²⁰²²). While some patients have achieved complete remission with antibody‐based immune checkpoint therapy; many patients receive no benefit and some experience severe immune‐related adverse events (Li et al., 2025).

Although anti‐PD‐L1 antibodies remain clinical workhorses in immune checkpoint therapy, they are challenging to adapt for imaging and radiotherapeutic applications. In part, this is due to their slow systemic clearance rates, which are partially driven by their high molecular weights (Rhoden & Wittrup, 2012; Thurber & Wittrup, ²⁰¹²). For imaging, this results in long delays between radiotracer injection and image acquisition. In the case of targeted systemic radiotherapy, it results in sustained exposure to ionizing radiation with concomitant damage to healthy tissues.

As lower molecular weight alternatives to antibodies, peptides have been developed to target hPD‐L1. WL12, a macrocyclic peptide that potently inhibits the binding of therapeutic antibodies to hPD‐L1 (IC₅₀ = 22 nM), is currently undergoing development as a clinical Positron Emission Tomography (PET) radiotracer (Zhou, Jiang, et al., 2022). Recently, Bristol Myers Squibb reported the discovery of BMS‐986238, which is undergoing investigation as an orally administered therapeutic (Scola, 2025). However, both peptides incorporate nonproteinogenic amino acids and chemical modifications. Although the peptides described above are highly potent and metabolically stable, nonproteinogenic amino acids and chemical functionalities increase the monetary costs and synthetic challenges associated with scaling up production for commercial use. Moreover, these compounds cannot be readily incorporated into other biological constructs, making them unsuitable as targeting vectors for viruses and vaccines or as biological tags for reagent localization.

A fully proteinogenic peptide with high metabolic stability and antibody‐like hPD‐L1 affinity would provide an outstanding targeting vector for next‐generation theragnostic development. This peptide could also be genetically encoded to impart hPD‐L1 affinity into current and next‐generation biologics. One such proteinogenic peptide, TPP‐1, was developed using bacterial surface display and bound hPD‐L1 with K _D = 95 nM (Li et al., 2018). However, TPP‐1 suffered from poor serum stability and required PEGylation for pre‐clinical PET studies (Hu et al., 2019). To our knowledge, there are no proteinogenic PD‐L1‐targeted peptides with antibody‐like affinity and sufficient serum stability for in vivo applications.

Previously, we developed SPAM (MPIFLDHILNKFLILHYA), a short, linear, proteinogenic peptide with mid‐nanomolar affinity for hPD‐L1 (Kamalinia et al., 2020). Although this peptide showed acceptable serum stability (ex vivo), its affinity (K _D = 67 ± 21 nM) is likely insufficient for further pre‐clinical development. In this work, we sought to improve the SPAM scaffold and identify peptides that bind hPD‐L1 with sub‐nanomolar affinity. Previous studies have validated mRNA display as a platform for affinity maturation of proteinogenic peptides and proteins (Cetin et al., 2017; Kamalinia et al., ²⁰²⁰; Nichols et al., ²⁰¹⁸). Our results indicate that modest extensions of the core SPAM motif can impart considerable affinity enhancements without compromising target selectivity. Indeed, the resulting sequences demonstrate mid‐picomolar target affinity, acceptable serum stability, exquisite selectivity for hPD‐L1, and robust activity in living cells.

2. MATERIALS AND METHODS

2.1. Recombinant protein expression and purification

Recombinant protein produced in mammalian cells (extracellular domains of hPD‐L1, mouse PD‐L1 [mPD‐L1], and human PD‐L2 [hPD‐L2]) for mRNA display and binding was purified as previously described (Grindel et al., 2020, 2022; Kamalinia et al., ²⁰²⁰). Briefly, constructs (GeneART/Thermo Fisher Scientific) were designed with a C‐terminal poly‐histidine tag, a BirA biotinylation site, and a thrombin cleavage site in a pcDNA3.4 backbone. A CMV‐driven promoter and Gaussia princeps signal peptide were included to induce high expression and extracellular secretion of the recombinant proteins. Constructs were transiently transfected into HEK293F cells using an ExpiFectamine Transfection kit (Thermo Fisher Scientific A14524) according to the manufacturer's directions. Conditioned media was dialyzed, and proteins were purified by Ni‐nitriloacetic acid chromatography. Following buffer exchange, the protein was biotinylated by recombinant biotin ligase (BirA) overnight at 4°C and exchanged into storage buffer: 50 mM 4‐(2‐hydroxyethyl) piperazine‐1‐ethanesulfonic acid ‐NaOH (HEPES‐NaOH), pH 8.0, 250 mM NaCl, 5% (v/v) glycerol, and 0.01% (w/v) sodium azide.

2.2. mRNA display

A DNA library encoding 22‐mer peptides was constructed for mRNA display based on mutational analysis data for SPAM (Kamalinia et al., 2020). The open reading frame had the design 5′‐ATG NNS NNS NNS NNS 213 313 143 342 113 112 443 422 141 342 313 413 231 NNS NNS NNS NNS GGC AGC GGA TCC AGT GGA GGC‐3′, where N = A, C, T, or G, S = C or G, 1 = 65.9% A, 9.8% G, 14.6% C, 9.8% T, 2 = 17.6% A, 52.9% G, 17.6% C, 11.8% T, 3 = 14.6% A, 9.8% G, 65.9% C, 9.8% T, and 4 = 17.6% A, 11.8% G, 17.6% C, 52.9% T. This enabled the translation of a library of peptides with the format MXXXXDHILNKFWILHYAXXXXGSSSSGS, where X represents a random amino acid, GSSSSGS is a constant C‐terminal spacer, and the underlined positions were doped with 70% wildtype nucleotide probability and 10% probability for each of the other three nucleotides, when adjusted for the different nucleotide coupling rates during chemical synthesis. Peptide‐mRNA fusions were prepared as previously described (Kamalinia et al., 2020). Additional details regarding library construction and peptide‐mRNA fusion preparation are included in the Supporting Information.

Biotinylated hPD‐L1 was immobilized on neutravidin agarose beads (Thermo Fisher 29201) in selection buffer (20 mM HEPES‐KOH, pH 7.5, 150 mM NaCl, 0.2% (v/v) Tween‐20, 1 mg mL⁻¹ BSA, 50 μg mL⁻¹ tRNA) at 4°C for 1 h. The concentration of immobilized hPD‐L1 in 1 mL selection buffer varied by selection round: 300 nM for Round 1, 100 nM for Round 2, 50 nM for Rounds 3 and 4, and 25 nM for Rounds 5, 6, and 7. The beads were then washed with and incubated in blocking buffer (selection buffer + 0.02 mM biotin) at 4°C for 2 h to saturate any remaining unbound neutravidin sites with biotin.

After reverse transcription, the cDNA–peptide fusion library was added to the immobilized hPD‐L1 and incubated for 1 h in 1 mL blocking buffer. This binding occurred at 4°C for Rounds 1, 2, 6, and 7 and at 22°C for Rounds 3, 4, and 5. The beads were washed seven times with blocking buffer to remove nonfunctional sequences. Rounds 6 and 7 included an additional off‐rate‐based selection step after the initial washing step. The off‐rate selection was performed by adding non‐biotinylated hPD‐L1 to the washed beads at a final concentration of 2500 nM in 1 mL of selection buffer followed by incubation at 4°C. After 15 min for Round 6 and 15 h for Round 7, the beads were separated from solution using a Spin‐X centrifuge filter tube (Corning CLS8161). For all rounds, the washed beads were added to a PCR reaction and amplified to regenerate the DNA library for the next selection round.

2.3. Radioligand binding assays

Radiolabeled peptide‐mRNA fusions were prepared as above but with the addition of 2 μL ³⁵S‐methionine (~1–2 μCi, Perkin Elmer) per 25 μL translation. After translation, fusion formation, and dT purification, 1% of the fusion volume was measured in a liquid scintillation counter (Beckman Coulter LS6500) to estimate the total counts. hPD‐L1 was immobilized onto neutravidin agarose beads as above with the amount required for each assay. 80,000–100,000 counts of radiolabeled fusions were added to the hPD‐L1 immobilized beads and incubated for 1 h at the temperature required for each assay. The supernatant was removed, and the beads were washed three times with blocking buffer. The supernatant, washes, and resuspended beads were counted in a liquid scintillation counter. The fraction of fusions bound was calculated as the ratio of counts of the resuspended beads (the bound fraction) over the sum of counts of the supernatant, washes, and resuspended beads (total counts).

For competition binding assays, the desired competitors (synthetic SPAM, synthetic SPAM2, synthetic SPAM3, atezolizumab [SelleckChem A2004], and WL12 [CPC Scientific CANC‐022A]) were dissolved in dimethyl sulfoxide (DMSO) or water (for atezolizumab) to create stocks at 1000X the concentrations required for each assay. 10 nM hPD‐L1 was immobilized onto beads as above and 1 μL of each competitor stock was added and incubated at 4°C for 1 h. The radiolabeled fusions were then added and incubated at 4°C for 1 h. The beads were washed and counted as described above to determine the fraction of fusions bound.

2.4. Biolayer interferometry analysis

The binding kinetics of the SPAM peptides with hPD‐L1 were measured by biolayer interferometry (BLI) using an Octet RH16 (Sartorius). Biotinylated hPD‐L1 (1 μg mL⁻¹) was immobilized onto a streptavidin‐coated biosensor (Sartorius 18‐5019) to a thickness of 1.4 nm, and biosensors without PD‐L1 were used for reference. Dilutions of SPAM (333, 111, 37, 12.4, and 4.12 nM), SPAM2 (6.25, 1.56, 0.391, 0.098, and 0.024 nM), SPAM3 (6.25, 1.56, 0.391, 0.098, and 0.024 nM), and SPAM4 (6.25, 1.56, 0.391, 0.098, and 0.024 nM) were prepared in assay buffer (20 mM HEPES‐KOH, pH 7.5, 150 mM NaCl, 0.05% [v/v] Tween‐20, 5% [v/v] glycerol, 1% [w/v] BSA, 0.1% [v/v] DMSO). The peptides were allowed to associate with the biosensors for 180 s and then allowed to dissociate in assay buffer for 900 s. The sensorgrams were normalized to no‐peptide values and globally fit to a 1:1 Langmuir binding model using Octet Analysis Studio 13.0 software. Each measurement was carried out in duplicate, and the mean values and fitting errors of the kinetic parameters were calculated.

2.5. Flow cytometry

The peptide reagents were tested for binding targets exogenously expressed in Chinese hamster ovary (CHO) cells. Test cells were either untransfected (parental) or stably transfected with hPD‐L1, hPD‐L2, or mPD‐L2 according to previous methods (Grindel et al., 2022; Kamalinia et al., ²⁰²⁰). All cells were cultured in Dulbecco's Modified Eagle Medium (Corning, 10‐013‐CV) supplemented with 10% (v/v) heat‐inactivated fetal bovine serum (Sigma‐Aldrich, F4135) and 1× penicillin–streptomycin (Sigma‐Aldrich, P4333). For flow cytometry, cells were briefly washed with Mg⁺²‐ and Ca⁺²‐free phosphate‐buffered saline (PBS) and dissociated into nonadherent single cells with trypsin (0.25% (w/v) trypsin, 0.2% (w/v) EDTA; Sigma‐Aldrich, T4049). Trypsin was deactivated with full media, cells were centrifuged at 400g for 2 min, and the pellet was resuspended in serum‐free (SF) media. Cells were divided into v‐bottom plates (Corning, 3894) (200,000 cells per well), centrifuged, and resuspended in a dilution series of a fluorescein‐conjugated peptide reagent diluted in cold 1:1 PBS:SF media. Cells were incubated without agitation with the reagents for 1 h at 4°C in the dark. This protocol was followed for glioma D423 cells and breast cancer MDA‐MB‐231 cells and MDA‐MB‐468 cells. For antibody competitive binding, durvalumab or trastuzumab control dilutions were included in the mixture with 500 nM fluorescein‐labeled peptide. After binding, cells were centrifuged at 400g and washed three times with cold PBS, resuspended, and immediately acquired for live single‐cell fluorescent intensity in a Cytek® Northern Lights™ flow cytometer (488‐nm laser excitation). The gain was adjusted so that the highest observed fluorescence intensity was below saturation intensity and background signal was well above zero. Parental CHO cells were used to normalize nonspecific uptake/background. Dissociation constants were determined by GraphPad Prism using the one‐site normalized specific binding model. For competitive binding, nonlinear regression models using one‐site Fit Ki were used to generate a dissociation constant for the antibody.

For peptide competition with fluorescently labeled durvalumab (R&D Systems, Biosimilar, FAB10355G), cells were processed as above but were instead incubated with a dilution series of unlabeled SPAM2, SPAM3, SPAM4, WL12, or control % (v/v) matched DMSO for 45 min at 4°C. The fluorescent antibody or IgG control antibody (Biolegend, 40010, clone MOPC‐21) for background was added at a final concentration of 0.35 nM following incubation. Cells were incubated for an additional 40 min at 4°C in the dark and processed as above for flow cytometry. Values were subtracted for IgG control and normalized to DMSO controls per peptide concentration.

2.6. Peptide solubility and aggregation

Lyophilized peptides were reconstituted in phosphate buffer (PB, 15 mM at pH 6.5, or 50 mM at pH 4.5) with small amounts of 10%–20% (v/v) MeOH or 2,2,2‐trifluoroethanol (TFE) as organic cosolvents. Samples were vortexed; then sonicated for 5 min for maximum solubilization. Sample concentrations were determined by UV‐absorbance using a JASCO V‐670 spectrophotometer based on the combined molar absorptivity at 280 nm of Trp and Tyr residues present in the peptide sequence (ε ₂₈₀ = 5500 M⁻¹ cm⁻¹ per Trp, ε ₂₈₀ = 1490 M⁻¹ cm⁻¹ per Tyr). At higher concentrations of peptide, UV‐absorbances were prone to instability; thus, a constant absorbance over the course of 20 min was confirmed for each measurement to ensure a lack of aggregation. This instability was not observed for lower concentrations and higher percentages of organic solvents. In addition, high concentrations (70–100 μM) and temperatures (40–50°C) were found to trigger irreversible aggregation, as evident in the far‐UV circular dichroism (CD) spectra. Observing this onset of aggregation is strong evidence that SPAM1‐3 are in a monomeric state at physiological temperature (37°C) and at concentrations up to 40 μM.

2.7. Far‐UV CD measurements and denaturation studies

Samples were prepared in a mixture of PB (15 mM at pH 6.5 or 50 mM at pH 4.5) and methanol (PB/MeOH 80:20 vol/vol) in a concentration range between 17 and 70 μM, determined by UV absorbance measurements. In brief, raw CD spectra at varying temperatures (0–90°C) were recorded as previously described (Greenfield, 2006) on a JASCO J‐810 spectropolarimeter with a temperature controller module JASCO PFD‐425S in a quartz cuvette (0.1 cm). The instrument was calibrated using (+)‐10‐camphorsulfonic acid provided by JASCO. CD spectra were recorded in mdeg from 190 to 270 nm, and CD spectra of the blank (solvent only) were collected for subtraction. A temperature ramp from 0 to 90°C with a rate of 45°C per hour under nitrogen flow rate (2 L min⁻¹) was used to obtain CD spectra (averaged of eight scans) at set temperatures representing the unfolding transitions for each peptide. Raw spectra were smoothed, and the baseline was set to 0 between 260 and 270 nm using SpectraGryph 1.2 (Menges, 2022). (http://www.effemm2.de/spectragryph/). The ellipticity scale was then converted from mdeg into molar ellipticity (θ) (deg cm² dmol⁻¹) by dividing with the corresponding number of amide bonds (17 for SPAM and 21 for SPAM2 and SPAM3), the concentration, and the pathlength. To determine the thermal stability of the peptides, a melting curve was generated from CD spectra by plotting temperature versus molar ellipticity at 220 nm for SPAM2 and 222 nm for SPAM3. Maxima obtained from the first derivative plots of these CD‐melts were used to identify the melting temperatures of 40 ± 5°C for both SPAM2 and SPAM3.

3. RESULTS AND DISCUSSION

3.1. Maturation selection and identification of SPAM2

We sought to build upon our previous development efforts to identify linear, proteinogenic peptides with high affinity for hPD‐L1. Our design strategy rested upon using extension libraries with a “doped” core SPAM motif flanked by fully randomized tetrapeptide sequences at the N‐ and C‐termini (Figure 1). Previous mutational analysis of SPAM showed residues 6–18 (DHILNKFLILHYA) represented a core motif for hPD‐L1 binding (Kamalinia et al., 2020). Based on mutation analysis and high‐throughput sequencing data of SPAM, this core was expected to conserve the affinity, binding site, selectivity, and stability of the parental SPAM peptide. Therefore, we designed an mRNA display library where the core region of SPAM was doped at a wildtype amino acid frequency of ~50% (Figure 1). This mutational frequency is sufficient to allow a search for mutations in the core segment that might confer higher affinities, yet low enough that the core was still highly similar to SPAM. We also included four fully random amino acids at the N‐ and C‐terminus to extend the PD‐L1 contact surface and improve peptide affinity.

FIGURE 1.

Affinity maturation of the SPAM peptide. An mRNA display library of 22‐mer peptides was designed based on our previous SPAM peptide (top). The core motif of SPAM was doped at ~50% wildtype amino acid frequency (red) and flanked by eight fully randomized amino acids (green) to provide N‐ and C‐terminal extensions.

Using this library, we performed several rounds of selection using mRNA display to select sequences that bind hPD‐L1 (Figure S1). The average binding of each selection pool was measured using radioligand binding assays to monitor selection progress (Figure S2). Significant target‐specific binding emerged in Round 2, and stringent selection pressures were applied in subsequent rounds to further enrich high‐affinity sequences. To increase the stringency of selection, the concentration of immobilized hPD‐L1 was reduced from 300 to 25 nM, and the binding temperature was increased from 4 to 22°C. In addition, in Rounds 6 and 7, an off‐rate selection was performed by adding 100‐fold excess non‐immobilized hPD‐L1 as a specific competitor, enabling sequences with slower off‐rates to be selected (Boder & Wittrup, 1998).

We performed next‐generation sequencing on the final DNA pool, and several sequences with the highest frequencies were evaluated using radioligand binding assays (Figure S3). All tested peptides showed binding to immobilized hPD‐L1 and little binding to beads without target. Of these peptides, SPAM2 (MWSRSDHNLNKFWILHYSANPS) demonstrated the highest binding. Residues 6–18 of SPAM2 are 77% identical to the core motif from SPAM, and this high similarity suggests that SPAM2 likely retains the same selectivity and binding site as SPAM. Further, the conservation of these core residues validates the choice to use the core region from SPAM as a starting point to select for peptides with enhanced affinity.

3.2. Mutational analysis of SPAM2 and identification of SPAM3

We then performed alanine scanning mutagenesis on SPAM2 to evaluate its structure–activity relationship and identify potential positions or substitutions that might further optimize the sequence. Alanine scanning is a well‐established technique to determine the contribution of individual amino acids in a protein–protein interaction (Clackson & Wells, 1995). All positions of SPAM2 were individually mutated to alanine, and the resulting sequences were evaluated in radioligand binding assays (Figure 2a). The D6A, H7A, N10A, F12A, I14A, and H16A mutations all abolished binding, indicating that these positions are vital for interaction with hPD‐L1. These residues were all randomized in the starting mRNA display library, and their reversion to the wildtype residue in SPAM2 also supports the hypothesis that these wildtype residues are critical for binding hPD‐L1. On the other hand, the N8A, S18A, P21A, and S22A mutants showed no statistical difference in binding relative to the wildtype SPAM2 sequence, suggesting that these positions do not contribute significantly to the binding of hPD‐L1. Many of these less critical positions are located at the C‐terminus, suggesting that SPAM2 could be truncated to a shorter, yet still active peptide.

FIGURE 2.

Mutational analysis of SPAM2. Radiolabeled peptide‐mRNA fusions of SPAM2 mutants were incubated with immobilized hPD‐L1, and the fraction of bound fusions was measured by scintillation counting. (a) SPAM2 mutants with individual alanine mutations and (b) SPAM2 truncations were incubated with 5 nM immobilized hPD‐L1 at 4°C for 1 h. (c) The binding of SPAM2 and SPAM3 was measured at the indicated binding temperatures and hPD‐L1 concentrations. Each measurement was carried out in triplicate, and error bars indicate the standard deviation from the mean (SD). Student's t‐test was used to compare each mutant with SPAM2: *p < 0.05; **p < 0.01; ***p < 0.005; ****p < 0.001 or #p < 0.001. The amino acid sequences of each mutant are listed in Table S1. hPD‐L1, human programmed death‐ligand 1.

We tested several truncations of SPAM2 to determine if a shorter peptide might retain binding of the parental peptide. We tested SPAM2 with a deletion of residues 2–5 (no N), a deletion of residues 19–22 (no C), or a deletion of residues 2–5 and 19–22 (no NC). These deletion constructs were evaluated using radioligand binding assays to determine the relative contributions of each terminus and to further define the boundaries of the core motif region (positions 6–18) (Figure 2b). All truncated sequences showed significantly reduced binding relative to the SPAM2 wildtype sequence. Truncation of the N‐terminus showed a greater decrease in binding than truncation of the C‐terminus, and truncation of both termini almost completely abrogated binding. The larger loss of binding upon truncation of the N‐terminus as compared with the C‐terminus suggests that the N‐terminal residues of SPAM2 contribute significantly to hPD‐L1 binding. On the other hand, in line with the alanine scanning experiments above, the C‐terminal residues contributed less strongly to SPAM2 binding. These results are corroborated by next generation sequencing data, which showed high sequence conservation at the N‐terminus and only modest conservation at the C‐terminus (Figure S4).

Taken together, alanine scanning and truncation data suggest that SPAM2 cannot be truncated without significant loss of affinity for hPD‐L1 and that the core motif alone (residues 6–18) is not sufficient to recapitulate SPAM2 binding. The importance of these residues to hPD‐L1 binding supports our strategy of using an extension selection to add flanking residues to a core peptide region to increase binding affinity. Further, the alanine scanning data suggest that the identity of the amino acids in the N‐ and C‐terminal extensions of SPAM2 is less important, yet they still contribute to hPD‐L1 binding, suggesting that these regions may interact with hPD‐L1 via main chain or backbone contacts or through intramolecular stabilization of the core SPAM2 motif.

While SPAM2 demonstrated the highest binding of all the sequences we tested, the theoretical diversity of the library (20²¹ = ~10²⁷) far exceeds the maximum practical throughput of mRNA display selections (~10¹² to 10¹⁴). It is thus likely that there are other 22‐mer sequences with even higher hPD‐L1 affinity. Previously, we showed the frequency of a given sequence present in late‐round mRNA display selection pools broadly correlates with functionality (Olson et al., 2012). Additionally, Olson et al. (2014) demonstrated that epistasis or nonadditive synergy between mutations is rare in directed evolution systems such as mRNA display. Therefore, we hypothesized that an even higher affinity sequence could be constructed from next generation sequencing data by using the most frequent amino acid at each position (Figure S4) even if that sequence did not appear in next generation sequencing. We thus constructed a “synthetic winner” sequence (MYSSTDHILNKFLILHYALNPT), which we designated as SPAM3. SPAM3 contains eight mutations compared with SPAM2, including at positions 8, 18, and 22, which did not contribute significantly to hPD‐L1 binding in the alanine scanning experiments. SPAM3 demonstrated higher binding than SPAM2 at every tested condition (Figure 2c) in radioligand binding assays with variable hPD‐L1 concentration and binding temperature. These data suggest SPAM3 has higher affinity and supports a frequency‐based approach for identifying and designing new sequences based on high‐throughput sequencing data. Scrambled versions of SPAM2 and SPAM3 showed essentially no hPD‐L1 binding (Figure S5), indicating that the binding of SPAM2 and SPAM3 is sequence‐dependent. SPAM2 and SPAM3 are nonhomologous to other known hPD‐L1 peptide ligands (Table S2).

3.3. SPAM2 and SPAM3 are stable in human serum

Serum hydrolases can degrade linear, proteinogenic peptides and reduce their in vivo effectiveness. To improve proteolytic stability, many peptide drugs include nonproteinogenic amino acids, macrocyclic structure, and other chemical modifications (McGregor, 2008). SPAM2 and SPAM3 are linear peptides comprised solely of proteinogenic amino acids and, as such, would be expected to be rapidly degraded by proteases in human serum. To study this, SPAM2 and SPAM3 were incubated in 25% (v/v) human serum at 37°C (ex vivo). Aliquots at 0, 15, 30, 120, and 300 min were analyzed by liquid chromatography‐mass spectrometry and analytical HPLC to quantify the peak area of each intact peptide (Figure S6). SPAM3 showed only modest time‐dependent degradation, with ~70% of the parent compound remaining intact after 300 min of incubation, while SPAM2 showed no appreciable degradation even after 300 min of incubation. In contrast, GiBP, a linear peptide previously studied in our lab, was shown to be unstable in serum and was rapidly degraded over the course of the experiment (Howell et al., 2014). This experiment suggests that both SPAM2 and SPAM3 are largely stable to degradation in serum. In both cases, on the time scale of a typical PET imaging experiment (<120 min), the majority of each peptide would likely remain intact, suggesting that these peptides have sufficient stability for short half‐life nuclear imaging applications.

3.4. SPAM3 binds hPD‐L1 with antibody‐like affinity at the PD‐1 interface

We determined the kinetic binding parameters of SPAM, SPAM2, and SPAM3 with hPD‐L1 using bio‐layer interferometry (BLI) (Figures 3 and S7) and surface plasmon resonance (SPR) (Figure S8). These parameters are listed in Table 1. The equilibrium dissociation constant for SPAM (K _D = 38.2 ± 1.0 nM by BLI and 36 ± 29 nM by SPR) agrees with our previously published value from fluorescence polarization experiments (Kamalinia et al., 2020). More impressively, SPAM2 (K _D = 303 ± 7 pM) and SPAM3 (K _D = 249 ± 3 pM) bind hPD‐L1 with ~100× higher affinity than SPAM (K _D = 38 ± 1 nM), indicating the extension selection was successful at improving the affinity of the core SPAM motif. The kinetic data also indicate SPAM2 and SPAM3 have both increased association and decreased dissociation rates compared with SPAM, demonstrating the improvements made by the mRNA display selection were not limited solely to reductions in the off‐rate. Additionally, we measured the affinity of SPAM4, a dimeric form of SPAM3, which is discussed further below (Figure S9).

FIGURE 3.

Biolayer interferometry of SPAM peptides with hPD‐L1. The binding kinetics of the interactions between hPD‐L1 and (a) SPAM, (b) SPAM2, (c) SPAM3, and (d) SPAM4 were measured using biolayer interferometry. The indicated peptide concentrations were allowed to associate onto biosensors coated in biotinylated hPD‐L1 for 180 s and then allowed to dissociate for 900 s. The sensorgrams were globally fit to a 1:1 (Langmuir) binding model; the fits are plotted alongside the measured data. hPD‐L1, human programmed death‐ligand 1.

TABLE 1.

Binding kinetics of the SPAM peptides with hPD‐L1.

Peptide	k on (104 M−1 s−1)	k off (10−3 s−1)	K D (pM)
SPAM	141 ± 3	49.0 ± 0.9	38,200 ± 1000
SPAM2	878 ± 21	2.64 ± 0.02	303 ± 7
SPAM3	578 ± 8	1.46 ± 0.01	249 ± 3
SPAM4	223 ± 11	0.219 ± 0.004	99 ± 1

Note: Mean on‐rates (k _on), off‐rates (k _off), and dissociation constants (K _D) are shown along with fitting errors.

Abbreviation: hPD‐L1, human programmed death‐ligand 1.

The affinity of SPAM3 is remarkable when compared to natural ligands and approved therapeutics targeting hPD‐L1. SPAM3 binds hPD‐L1 with more than 10,000‐fold higher affinity than the natural ligand PD‐1 (K _D = 8.2 ± 0.11 μM), highlighting its potential for hPD‐L1 binding in an active tumor‐immune microenvironment (Cheng et al., 2013). Additionally, SPAM3 binds with higher affinity than the single‐chain fragment variable (scFv) of atezolizumab (K _D = 1750 pM) and durvalumab (K _D = 667 pM) and with slightly lower affinity than the scFv of avelumab (K _D = 46.7 pM), making its binding comparable to that of therapeutic antibodies (Tan et al., 2018).

CD spectroscopy is routinely used to study the secondary structures of linear peptides and conformational changes induced by solvent, pH, or temperature. Far‐UV CD spectra of SPAM peptides were recorded in aqueous buffers at pH 4.5 or 6.5 with the addition of methanol (for solubility) or a more structuring cosolvent TFE (Figure 4). For SPAM, a drastic pH‐dependent change of folding could be observed with a β‐pleated sheet preference at pH 4.5 marked by a large negative band at 214 ± 2 nm to a more helical structure at pH 6.5 shown by the negative minima at about 206 and 221 ± 2 nm (Figure 4a). In addition, the thermal denaturation of SPAM recorded at pH 6.5 between 0 and 90°C (Figure 4d) revealed not only the unfolding of the β‐sheet but also a change of state leading to an irreversible aggregation above 50°C. In contrast, SPAM2 was found to have a more stable and well‐behaved antiparallel β‐sheet fold in all solvent and pH conditions tested (positive maxima at 197 ± 2 nm and minima at 214 ± 2 nm) (Figure 4b). In this case, the thermal denaturation study revealed a much smoother transition with a β‐sheet fraying leading to a random coil state (200 ± 2 nm) above 50°C (Figure 4e). Whereas at low pH SPAM3 presents a strong α‐helix character (208/221 ± 2 nm) especially in the presence of NaF (Figure 4c), the peptide fold turns into a mixture of helical and β‐sheet structures at pH 6.5. Unlike SPAM2, SPAM3's antiparallel β‐sheet fold is certainly in equilibrium with some helical structures leading to a complex thermal denaturation mechanism that is likely occurring along with significant aggregation (Figure 4f). Yet the CD‐melts transitions of SPAM2 and SPAM3 at ~40 ± 5°C indicate that these peptides are more thermostable than SPAM. Under all conditions tested, the low solubility of SPAM4 made it challenging to reach a workable window of concentrations for CD spectroscopy.

FIGURE 4.

Circular dichroism (CD) of SPAM peptides. The CD spectra of (a) SPAM, (b) SPAM2, and (c) SPAM3 recorded in (1) 50 mM phosphate buffer (PB) with 20% MeOH (solid green line, pH 4.5; solid blue line, pH 6.5), (2) 50 mM PB with 150 mM NaF in 20% MeOH (dotted green line, pH = 6.5; dotted blue line, pH = 4.50), and (3) 50 mM PB, pH = 6.5 in 20% TFE (black line). Representative thermal denaturations of (d) SPAM, (e) SPAM2, and (f) SPAM3 in 50 mM PB (pH 6.5) and methanol (1:4 vol/vol) with 150 mM NaF (only e, f), between 0 and 90°C showing that SPAM rapidly aggregates whereas the β‐sheet and helical structures of SPAM2 and SPAM3 unfold progressively upon warming. The denatured SPAM peptides did not refold upon cooling. TFE, 2,2,2‐trifluoroethanol.

These experimental results suggest that SPAM2 adopts a predominantly antiparallel β‐sheet fold in solution. If the SPAM peptides bind hPD‐L1 as β‐sheets, then the increased propensity of SPAM2 and SPAM3 for this conformation could contribute to their increased affinity for hPD‐L1 relative to SPAM. To investigate this hypothesis, we used molecular docking and molecular dynamics simulation to model the binding interaction between SPAM2 and hPD‐L1. The simulation indicated SPAM2 binds hPD‐L1 at the PD‐1 binding site and adopts a β‐hairpin conformation (Figures S10 and S11). We corroborated this model using competition radioligand binding assays to assess the binding sites of SPAM2 and SPAM3 on hPD‐L1. We measured the binding of radiolabeled SPAM2 and SPAM3 fusions to resin‐immobilized hPD‐L1 in the presence of atezolizumab (a therapeutic anti‐PD‐L1 antibody), WL12 (a macrocyclic peptide), SPAM, SPAM2, SPAM3, or a DMSO‐only control (Figure S12). The presence of atezolizumab or WL12, which both occlude the PD‐1 binding site (Chatterjee et al., 2017; Zhang et al., ²⁰¹⁷), almost completely eliminated binding of SPAM2 and SPAM3 to hPD‐L1, indicating that SPAM2 and SPAM3 bind at the PD‐1 site. These experiments provide a preliminary hypothesis for the increased affinity of SPAM2 and SPAM3.

3.5. SPAM2 and SPAM3 specifically bind hPD‐L1 on the surface of living cells

With clear indications that SPAM2 and SPAM3 bind hPD‐L1 with picomolar affinity in vitro, we next sought to determine their affinity for hPD‐L1 in a cellular context by testing their binding to hPD‐L1 expressed on the surface of living cells. Flow cytometry demonstrated robust binding of fluorescein‐labeled SPAM2, SPAM3, and SPAM4 (discussed below) to hPD‐L1 on the surface of CHO cells (Figure 5). In flow cytometry experiments, SPAM2‐fluorescein bound to CHO cells with stably expressing hPD‐L1 (Figure 5a) yet showed no significant binding to CHO cells expressing mouse PD‐L1 (mPD‐L1) or the closely related human PD‐L2 (hPD‐L2). This indicates that selectivity for hPD‐L1 is maintained in the affinity‐matured sequences despite hPD‐L1 being 77% and 34% sequence identical to mPD‐L1 and hPD‐L2, respectively (Latchman et al., 2001; Lazar‐Molnar et al., ²⁰⁰⁸; Lin et al., ²⁰⁰⁸). SPAM2‐fluorescein also bound to endogenous hPD‐L1 in glioblastoma (D423) and triple‐negative breast cancer cell lines (MDA‐MB‐231) in a concentration‐dependent manner. Meanwhile, hPD‐L1‐negative MDA‐MB‐486 cells showed no concentration‐dependent labeling with SPAM2‐fluorescein. We note that the enhanced maximum signal observed in CHO hPD‐L1 relative to MDA‐MB‐231 and D423 reflects the significantly higher hPD‐L1 expression in the engineered CHO cell lines relative to the cancer cell lines. The higher signal in CHO‐hPD‐L1 cells in comparison to D423 and MDA‐MB‐231 was confirmed by flow cytometry with a biosimilar version of the durvalumab anti‐PD‐L1 antibody (Figure S13) and western blot analysis of total hPD‐L1 expression (Figure S14).

FIGURE 5.

Binding of SPAM2 and its variants with hPD‐L1 on the surface of live cells. (a) SPAM2‐fluorescein was assessed for concentration‐dependent binding in multiple cell lines. (b) SPAM2 binding was compared directly to SPAM3 and SPAM4 in CHO hPD‐L1 expressing cells. Values were normalized to those observed with parental CHO cells. K _D values with 95% confidence interval (CI) values were determined using specific binding models in GraphPad software. CHO, Chinese hamster ovary; hPD‐L1, human programmed death‐ligand 1.

The native IgG forms of therapeutic anti‐hPD‐L1 antibodies likely benefit from avidity effects and bind hPD‐L1 with higher apparent affinities than their corresponding scFvs. With the hope of leveraging bivalent avidity, we developed dimeric forms of SPAM2 and SPAM3 by synthesizing two monomers of each peptide linked via a central bis‐Fmoc lysine at the C‐terminus. We used polyethylene glycol (PEG) spacers with either 5 or 10 repeats to test several different spacers that we hypothesized would be necessary to impart sufficient distance between each peptide arm to facilitate bivalent interactions with hPD‐L1.

We triaged these dimeric peptide variants against CHO‐hPD‐L1 and D423 using flow cytometry to assess concentration‐dependent binding (Figure S15). Of all the constructs, the SPAM3 dimer with PEG₁₀ spacers showed the highest hPD‐L1 affinity in both cell lines and was designated SPAM4 (Figure S9). In general, bivalent forms of SPAM3 performed better than their monovalent forms, but surprisingly SPAM2 dimers bound worse than their monomeric counterparts. Among all bivalent peptides, the PEG₁₀ spacer was universally superior to the PEG₅ spacer, suggesting increased affinity from bivalency requires appropriate geometry and distance. More importantly, SPAM4 showed a greater improvement over SPAM3 in D423 cells (~20×) than in CHO‐hPD‐L1 cells (~5×) while having similar affinity for hPD‐L1 across both cell types. BLI analysis indicated a 99 pM affinity for SPAM4 (Figure 3 and Table 1).

To quantitatively determine the relative binding of these peptides, we determined the binding of SPAM2, SPAM3, and SPAM4 to hPD‐L1 on the surface of CHO hPD‐L1 cells using flow cytometry (Figure 5b). SPAM2 (K _D = 15.3 nM) is the weakest binder, agreeing with the in vitro binding data above. SPAM4 showed the strongest binding (K _D = 3.47 nM), followed by SPAM3 (K _D = 5.15 nM); though this difference is not statistically significant (overlapping 95% confidence intervals). The K _Ds of SPAM3 and SPAM4 are very close to that of durvalumab in the CHO‐hPD‐L1 cell line (Figure S13A; 2.5 nM), further demonstrating their antibody‐like affinities. The scrambled versions of SPAM2, SPAM3, and SPAM4 showed no significant binding to CHO‐hPD‐L1 cells (Figure S15D), confirming that the peptide: protein interaction is sequence‐dependent. Further, SPAM3 and SPAM4 showed no binding to CHO‐mPD‐L1 and CHO‐hPD‐L2 cells at the highest concentration tested (1200 nM; Figure S15E), demonstrating conservation of target selectivity.

With confirmation of high affinity binding across multiple cell lines, we sought to use cell‐based flow cytometry inhibition experiments to confirm that the binding site of the SPAM peptides overlaps with PD‐1 as determined by the in vitro antibody displacement radioligand competition assays above (Figure S12). CHO hPD‐L1 (Figure 6a), D423 (Figure 6b), and MDA‐MB‐231 (Figure 6c) cells were incubated with increasing concentrations of SPAM2, SPAM3, SPAM4, WL12 (positive control), or DMSO (as a background control), followed by the addition of fluorescein‐labeled biosimilar durvalumab mAb. Durvalumab and WL12 both inhibit hPD‐1/PD‐L1 engagement by binding to the PD‐1 interface (Lesniak et al., 2019; Tan et al., ²⁰¹⁸). Thus, if the peptides displace durvalumab, then they likely bind to the PD‐1 interface in vivo. SPAM2, SPAM3, and SPAM4, along with the control WL12 peptide, competed with fluorescent durvalumab in a concentration‐dependent manner in all tested cell lines. This displacement indicates at least a partial overlap of the binding surface with durvalumab and, therefore, the PD‐1 binding interface. These data further validate the in vitro competition data showing that the SPAM peptides compete with atezolizumab, which also binds the PD‐1 site (Figure S12). Interestingly, IC₅₀ values for SPAM2, SPAM3, SPAM4, and the WL12‐positive control showed significant variance across cell lines. The SPAM peptides showed similar IC₅₀ values as WL12 in CHO‐hPD‐L1 cells but showed higher IC₅₀ values in D423 and MDA‐MB‐231 cell lines (Figure 6d). The binding of the biosimilar durvalumab antibody also varied across cell lines (Figure S13A–C) with CHO‐hPD‐L1 cells showing the lowest affinity (2.5 nM).

FIGURE 6.

Durvalumab inhibition with SPAM2, SPAM3, and SPAM4. Fluorescein‐labeled durvalumab was inhibited with increasing concentrations of SPAM2, SPAM3, and SPAM4 or control‐matched % (v/v) DMSO on (a) CHO hPD‐L1 cells, (b) D423 cells, or (c) MDA‐MB‐231 cells. (d) Competition experiments IC₅₀ values with absolute‐value hill slopes (|h|) ([Inhibitor] vs. Response–Variable slope 4 parameters, GraphPad). CHO, Chinese hamster ovary; hPD‐L1, human programmed death‐ligand 1.

To further explore this interaction, we also performed the competition assay in reverse with SPAM2‐fluorescein binding in competition with unlabeled durvalumab on CHO hPD‐L1 cells (Figure S16A). SPAM2‐fluorescein was incubated with CHO hPD‐L1 cells in the presence of increasing concentrations of durvalumab or trastuzumab (Herceptin, anti‐Her2) as a control. As expected, trastuzumab shows no effect on the binding of SPAM2‐fluorescein, while durvalumab is a potent competitor (K _i  = 0.4 nM, 95% CI −0.12 to 1 nM). This provides additional evidence that SPAM2 binds at the PD‐1/hPD‐L1 interface and therefore potentially disrupts PD‐1 engagement.

If SPAM peptides block the PD‐1 interface, they should be able to reactivate T‐cells, which we tested using a cell‐based bioluminescent T‐cell reactivation assay. In this assay, compounds that block the interaction between hPD‐L1 on the antigen‐presenting cell and PD‐1 on the T cell (Jurkat cell) result in re‐activation of the NFAT reporter and induction of luciferase activity. Increasing amounts of SPAM2 and SPAM3, along with vehicle controls, were added, and the bioluminescence at each concentration was measured and normalized as relative luminescence units (RLUs). As seen in Figure S16B, SPAM2 and SPAM3 reactivated T cells in a concentration‐dependent manner, restoring more than 50% of signal at 5 μM, further confirming that SPAM2 and SPAM3 block PD‐1 binding. The positive control atezolizumab (Figure S16C) demonstrates the assay works as intended. Atezolizumab has an IC₅₀ of 0.76 nM (0.61–1.03 nM, 95% CI). Vehicle CHO‐K1 indicates the maximal RLU values where PD‐L1 is absent and unable to interfere with T‐cell activity. Taken together, these results suggest that later generation SPAM peptides could function as standalone PD‐1/PD‐L1 checkpoint inhibitors.

4. CONCLUSION

Many peptide development approaches rely on nonproteinogenic peptides to impart conformational constraint and enhanced metabolic stability. While these strategies have proven successful (Fiacco et al., 2016; Fujino et al., ²⁰¹⁶; Goto et al., ²⁰⁰⁸; Gray et al., ²⁰²¹; Howell et al., ²⁰¹⁴; Kawakami et al., ²⁰⁰⁸; Maini et al., ²⁰¹⁹; Passioura & Suga, ²⁰¹⁷; Subtelny et al., ²⁰⁰⁸), nonproteinogenic functionalities and architectures increase the cost and complexity of synthesis and preclude genetic encoding of the resulting sequence into higher‐order biologics. The primary goal of this work was to design high‐affinity, proteinogenic peptides using an extension mRNA display selection based on a previously identified lead sequence (SPAM). This approach yielded SPAM2 and SPAM3 as fully proteinogenic peptides that bind recombinant hPD‐L1 in vitro and hPD‐L1 on the cell surface with high affinity. Indeed, the affinities of SPAM2, SPAM3, and SPAM4 are 100‐fold higher than the parental SPAM peptide and are comparable with the affinities of clinical anti‐hPD‐L1 antibodies. This improvement in affinity was achieved with a modest increase in molecular weight, demonstrating the molecular efficiency of the extension selection approach. SPAM2 and SPAM3 also showed relatively high stability to serum‐mediated hydrolysis in vitro suggesting, in addition to our previous work (Kamalinia et al., 2020) and other works, that proteinogenic, linear peptides may be sufficiently stable for in vivo applications.

Our preliminary spectroscopic and computational analyses suggest that the enhanced PD‐L1 affinity of SPAM2 and SPAM3 may arise from increased structural pre‐organization. If so, the N‐ and C‐terminal extensions may act to stabilize a β‐hairpin conformation rather than introducing additional contacts with the hPD‐L1 target. Indeed, antiparallel β‐sheets are a common structural motif in endogenous and designed ligands that recognize the PD‐1 binding site on hPD‐L1 (Figure S11). Although high‐resolution structural studies are required to confirm this binding model, it does suggest additional design strategies to further enhance affinity and stability.

While the SPAM3 sequence could theoretically have been present in the extension selection library, it was not found in the final pool despite having a higher affinity for hPD‐L1 than SPAM2. We calculate that the probability of SPAM3 being present in the original synthetic library was ~5 × 10⁻¹⁶, and with an mRNA display library size of 10¹², roughly a 1 in 2000 chance of being found by our mRNA display selection. SPAM3 shows that even if selections are performed using the largest experimentally accessible libraries (10¹² to 10¹⁴ sequences), these selections only search a very small fraction of all the theoretical sequences possible (20²¹ to 10²⁷ possible sequences). Based on the findings of this work, we propose that higher affinity binders could be identified by applying aggregate analyses to high‐throughput sequencing data. Indeed, we were able to design a sequence with modestly higher affinity by simply using the most frequent amino acid at each position in the final selection pool. More powerful analyses could be implemented by applying machine learning models to next‐generation sequencing data, an approach which has been explored in other works (Liu et al., 2020; Makowski et al., ²⁰²²; Parkinson et al., ²⁰²³; Saka et al., ²⁰²¹; Schmidt & Hildebrandt, ²⁰²¹; Vinogradov et al., ²⁰²²).

We note that all three types of hPD‐L1 ligands (antibodies, SPAM peptides, and WL12) behave differently in different cell lines. Although we would predict that B _Max (the maximal signal) would vary across cell lines due to differential hPD‐L1 expression levels, we also observed large changes in K _D and K _i for all three ligands. This argues that the structure and behavior of hPD‐L1 itself may depend on the cellular context. Indeed, previous experiments have shown that glycosylation of hPD‐L1 has a significant effect on the affinity of clinical hPD‐L1 antibodies and that dimerization may also play a critical role in governing PD‐L1 behavior (Benicky et al., 2021; Gao et al., ²⁰²⁰; Zhou, Chai, et al., ²⁰²²). Future work will focus on addressing the effects of these properties on ligand affinity and selectivity and the effects of ligand binding on hPD‐L1 function.

AUTHOR CONTRIBUTIONS

Justin N. Ong: Formal analysis; validation; methodology; conceptualization; investigation; writing – original draft; writing – review and editing. Brian J. Grindel: Conceptualization; investigation; validation; writing – review and editing; writing – original draft; formal analysis; methodology. Scott A. Rankin: Validation; methodology. Sarah H. Naylon: Methodology; writing – review and editing. Anupallavi Srinivasamani: Methodology; validation. Guillaume J. Trusz: Methodology; writing – review and editing. Xiaowen Liang: Methodology; validation; writing – review and editing. Md. Nasir Uddin: Writing – review and editing; methodology; validation. Lauren Fuller: Methodology; validation; writing – review and editing. Michael Curran: Investigation; funding acquisition; writing – review and editing; supervision. Stephane P. Roche: Investigation; methodology; writing – review and editing. Terry T. Takahashi: Conceptualization; investigation; funding acquisition; writing – review and editing; validation; supervision. Richard W. Roberts: Conceptualization; funding acquisition; investigation; writing – review and editing; project administration; supervision. Steven W. Millward: Conceptualization; investigation; funding acquisition; writing – original draft; writing – review and editing; formal analysis; project administration; supervision.

CONFLICT OF INTEREST STATEMENT

JNO, BJG, TTT, SWM, and RWR are co‐inventors on US Patent Application 17/997,507, which incorporates some of the work described in this manuscript.

Supporting information

Data S1. Experimental details.

Ong JN, Grindel BJ, Rankin SA, Naylon SH, Srinivasamani A, Trusz GJ, et al. Using extension‐based mRNA display to design antibody‐like proteinogenic peptides for human PD‐L1 . Protein Science. 2025;34(9):e70268. 10.1002/pro.70268

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.