Directed Evolution of PD-L1 Targeted Affibodies by mRNA Display

Abstract

Therapeutic monoclonal antibodies directed against PD-L1 (e.g., atezolizumab) disrupt PD-L1:PD-1 signaling and re-activate exhausted cytotoxic T-cells in the tumor compartment. Although anti-PD-L1 antibodies are successful as immune checkpoint inhibitor (ICI) therapeutics, there is still a pressing need to develop high affinity, low molecular weight ligands for molecular imaging and diagnostic applications. Affibodies are small polypeptides (~60 amino acids) that provide a stable molecular scaffold from which to evolve high affinity ligands. Despite its proven utility in the development of imaging probes, this scaffold has never been optimized for use in mRNA display, a powerful in vitro selection platform incorporating high library diversity, unnatural amino acids, and chemical modification. In this manuscript, we describe the selection of a PD-L1 binding affibody by mRNA Display. Following randomization of the 13 amino acids that define the binding interface of the well-described Her2 affibody, the resulting library was selected against recombinant human PD-L1 (hPD-L1). After 4 rounds, the enriched library was split and selected against either hPD-L1 or the mouse ortholog (mPD-L1). The dual target selection resulted in the identification of a human/mouse cross-reactive PD-L1 affibody (M1) with low nanomolar affinity for both targets. The M1 affibody bound with similar affinity to mPD-L1 and hPD-L1 expressed on the cell surface and inhibited signaling through the PD-L1:PD-1 axis at low micromolar concentrations in a cell-based functional assay. In vivo optical imaging with M1-Cy5 in an immune-competent mouse model of lymphoma revealed significant tumor uptake relative to a Cy5-conjugated Her2 affibody.

Introduction

The introduction of immune checkpoint inhibitor (ICI) therapy has been highly effective for many cancers ¹. Over the past decade, ICI therapy has focused on reversing inhibition of the immune system in the tumor microenvironment by targeting cytotoxic T-cell lymphocyte-associated protein 4 (CTLA-4) and programmed cell death protein/ligand 1 (PD-1/PD-L1) ^{2, 3}. Humanized monoclonal antibodies directed against CTLA-4, PD-1 and PD-L1 (e.g., ipilimumab, nivolumab, and atezolizumab, respectively) disrupt receptor/ligand interactions and re-activate exhausted cytotoxic T-cells ⁴. While durable benefit has been demonstrated with ICI therapies, only a subset of the total patient population responds to ICI therapy.

In the clinic, immunohistochemistry (IHC) is used to determine expression of ICI proteins by biopsies in the tumor and surrounding tissues. IHC assessment of ICI target expression is not fully standardized and may not correlate with treatment efficacy ⁵. For example, clinical trials in metastatic urothelial carcinoma suggest that PD-L1 expression on infiltrating immune cells has superior predictive power for treatment efficacy relative to PD-L1 expression on the tumor ⁶. Similar data in melanoma showed PD-L1 expression in the surrounding stroma is a better predictor for PD-1 blockade efficacy than tumor PD-L1 expression ⁷. This heterogenous expression of ICI markers across the tumor, immune, and stromal microenvironments argues for a non-invasive, quantitative, image-based process to evaluate the whole-body status of ICI targets for diagnosis and treatment monitoring.

There have been numerous efforts to image systemic PD-L1 expression through positron emission tomography/computed tomography (PET/CT). While [⁸⁹Zr]-labeled antibodies have been developed as PD-L1 PET imaging agents, these tracers are suboptimal as they take multiple days to achieve sufficient systemic clearance for high-resolution imaging ⁸. In contrast, low-molecular weight, rapidly clearing tracers linked to short half-life radiotracers (e.g., ¹⁸F and ⁶⁸Ga) facilitate same-day PET imaging, providing a more attractive route for clinical translation ⁹.

PD-L1 has been targeted with multiple molecular scaffolds including Adnectins (fibronectin domain scaffolds) ¹⁰, nanobodies ¹¹, conjugated Fab fragments ¹², ankyrin repeat proteins (against PD-1) ¹³, and nonblocking single domain antibodies ¹⁴. While these mid- to high-molecular weight scaffolds often show superior affinity, their slow systemic clearance renders them susceptible to the same disadvantages as full-length antibodies in molecular imaging applications. Numerous groups have employed directed evolution of linear peptide libraries to identify PD-L1 binding peptides and peptidomimetics ^15–20. Small, macrocyclic PD-L1 binding peptides have also been developed and used in early-phase imaging clinical trials ^{21, 22}. Peptide cyclization and unnatural amino acid incorporation have been shown to increase metabolic stability ^23–26, but these compounds are often challenging and expensive to produce and may suffer from conformational heterogeneity ^{27, 28}.

Low molecular weight polypeptide scaffolds offer a compromise between small peptides and large protein scaffolds. These scaffolds can be utilized for ligand design and capture many of the affinity-based benefits of monoclonal antibodies ^{13, 29} and nonblocking single domain antibodies. The bundled triple α-helical affibody scaffold, derived from the immunoglobulin (Ig) binding Z-domain of protein A, provides a favorable combination of evolvability, size, and ease of synthesis/expression ²⁹. At 58 amino acids, affibodies are quite small at 6-7 kilodaltons (kDa), clear quickly through the kidney, and are unlikely to be non-specifically retained in tumors by enhanced permeability retention (EPR) ³⁰. Affibodies lack disulfides, can spontaneously refold, and are relatively heat stable. These properties facilitate robust E. coli production, synthetic solid phase peptide synthesis, and facile conjugation to imaging reporter groups ^{31, 32}. The three α-helices assemble into a stable three-dimensional structure with a ~1,600 Ų binding surface defined by 13 surface-exposed amino acids. These residues can be randomized to generate a library of affibodies which can be selected for binding to molecular targets ³³. The most notable and well-studied affibody was raised against Her2 (Erb-B2), an oncogenic receptor that is overexpressed in approximately 20% of breast tumors. The Her2 binding affibody was selected with phage display but underwent extensive post-selection optimization to obtain a low picomolar (pM) binding variant ^{34, 35}. A matured Her2 affibody was ultimately adapted for patient whole body PET imaging ³⁶. Similar phage selection/optimization at Merck/Affibody AB produced a PD-L1 binding affibody with nanomolar-affinity which was later affinity matured to generate a picomolar affinity ligand ^{37, 38}. The final matured affibody sequence has not been disclosed. Another affibody identified by yeast display was found to block the PD-1/PD-L1 interaction, but the final affinity was never measured ³⁹.

Directed evolution for target binding is most effective when the initial library diversity is high ⁴⁰. A common display technology, phage display, produces peptide/protein libraries with an average of 10⁹ unique sequences ⁴¹, which is orders of magnitude less than the possible combinations for the 13 randomized positions in an affibody (20¹³ ≈ 10¹⁷ possible sequences). The initial diversity of these libraries is limited, in part, by the requirement for in vivo amplification prior to the next round of selection. In contrast, mRNA display can be carried out entirely in vitro allowing for maximal diversities greater than 10¹⁵ sequences ⁴². Because it is entirely in vitro, mRNA display can also incorporate unnatural amino acids and employ pre-selections for resistance to proteolytic degradation ^{23, 43, 44}. However, the affibody scaffold has yet to be adapted for implementation in mRNA display selection experiments.

Previous work in our lab has shown it is feasible to display affibodies as affibody-mRNA fusions ^{45, 46} , but so far, no mRNA display-based selection of affibodies have been demonstrated. In this work, we describe the first affibody-based mRNA display selection against both human PD-L1 (hPD-L1) and mouse PD-L1 (mPD-L1). An affibody that binds to both hPD-L1 and mPD-L1 with low nanomolar (nM) affinity was identified in these selections. The affibody maintained high affinity for both recombinant PD-L1 and PD-L1 on the cell surface and effectively blocked PD-L1:PD-1 signaling at low micromolar concentrations. Finally, the affibody showed selective uptake in PD-L1-expressing tumors in a fully immune-competent syngeneic murine model of lymphoma. These results demonstrate the feasibility of evolving high affinity affibody-based ligands using mRNA display and identify a novel species cross-reactive ligand for PD-L1 imaging and therapy.

Materials and Methods

Materials and Reagents

All common reagents were purchased from either Millipore Sigma (Burlington, MA), Thermo Fisher Scientific (Waltham, MA) or VWR (Radnor, PA) and used as is. All DNA oligos were synthesized at Integrated DNA Technologies (Coralville, IA, USA). Constructs were synthesized at GeneART (Thermo Fisher Scientific, Waltman, MA). Additional methods can be found in the supporting information.

Target protein purification

Each target (hPD-L1 and mPD-L1) and control target (hPD-L2, Her2) was purified as previously described ^{45, 46}. Briefly, constructs (GeneART/Thermo Fisher Scientific) were designed with a C-terminal poly-histidine tag, BirA biotinylation site, and 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 ExpiFectamine^™ Transfection kit (Thermo Fisher Scientific) according to manufacturer’s directions. Conditioned media was dialyzed, and proteins purified by Ni-nitriloacetic acid (NTA) chromatography. Following buffer exchange, protein was biotinylated by recombinant BirA overnight at 4°C and exchanged into storage buffer: 50 mM HEPES-NaOH, pH 8.0, 250 mM NaCl, 5% (v/v) glycerol, and 0.01% (w/v) sodium azide.

mRNA display affibody library construction

The affibody mRNA display library was based on the H 2:2891 optimized scaffold ⁴⁷. The sequence was modified to incorporate an ApoI restriction digest site between helix 1 and helix 2 without altering codons (Supplementary Figure S1). Two oligos from the 5’ end to 4 bases after the ApoI site and two oligos from 5 bases before the ApoI site to the 3’ end were synthesized (Integrated DNA Technologies) and assembled using overlap extension polymerase chain reaction (PCR). These were then digested with ApoI and ligated with T4 DNA ligase to produce 139 pmol of affibody library, corresponding to an estimated diversity of 8.38 x 10¹³ unique molecules. Sequencing performed at MD Anderson Advanced Technology Genomics Core confirmed library assembly and randomization at the desired positions. This library was then amplified by 6 rounds of PCR to generate a working library for mRNA display selections consisting of ~40 copies of each sequence. More details on library design and assembly are provided in the supplementary methods section of the Supporting Information.

mRNA display fusion production and target selection

The mRNA-affibody fusion molecules were generated using techniques as described in previous studies ^{45, 46}. Library DNA was PCR amplified with primers designed to add a 5′ T7 polymerase sequence and a 3′ linker region. After T7 transcription and urea polyacrylamide gel electrophoresis (PAGE)-purification the DNA linker PF30P (dA₂₁-(3 x Spacer 9)-dAdCdCpuromycin) was ligated to the mRNA in the presence of a DNA splint with T4 DNA ligase and urea PAGE-purified (5% (v/v)). The ligated product was translated using a rabbit reticulocyte lysate in vitro translation kit without methionine (Thermo Fisher, AM1200). The fusion library was translated in the presence of [³⁵S]-methionine/cysteine (PerkinElmer, Neg772002MC). Following oligo deoxythymidine (dT) oligo affinity support (OAS) purification using the poly dA linker sequence, the mRNA–affibody fusions were reverse-transcribed with SuperScript^™ III Reverse Transcriptase (Thermo Fisher, 18080044) ethanol-precipitated, and resuspended in selection buffer with bovine serum albumin (BSA) (25 mM HEPES-KOH pH 7.5, 150 mM NaCl, 0.05% (v/v) Tween-20, 1 mM ethylenediaminetetraacetic acid (EDTA), 5 mM MgCl₂, 0.5 mg/mL BSA, 10 μg/mL calf liver tRNA).

For target selections, an excess of NeutrAvidin resin was pre-blocked in selection buffer with BSA for 30 min at 22 °C on an end-over-end rotator. Following washes, fusions were precleared with half of the pre-blocked resin (1 hour at 4 °C) to remove resin-binding sequences and then centrifuged through a Corning Spin-X 0.45 μm filter to recover precleared fusions in the flow-through. The other half of the resin was incubated with biotinylated hPD-L1 (5,000 picomoles in first round, down to 40 picomoles in the last round). Remaining streptavidin sites were blocked briefly with excess biotin, unbound protein and biotin were washed away, and the target-loaded resin was incubated with the pre-cleared fusions overnight at 4 °C on an end-over-end rotator for the first round. Unbound fusions were removed by centrifuging the resin and removing the supernatant. After six washes in selection buffer, the resin was incubated with thrombin (Sigma-Aldrich, SAE0006) to elute target-bound fusions at 22 °C for 1.5 hours. Flow-through, elution, and resin fractions were analyzed by PCR and visualized by ethidium-stained 4% (w/v) agarose gels. Fractions were also assessed for [³⁵S]-methionine radioactivity by liquid scintillation to provide counts per minute (CPM) values. In subsequent rounds, the selection stringency was increased by reducing target loading, increasing temperature to room temperature (~22 °C), and decreasing incubation time. After round 4, a selection against mPD-L1 was performed alongside the continued hPD-L1 for three rounds. Off-rate selections, done separately for mPD-L1 and hPD-L1, were achieved by incubating with 25x molar excess of non-biotinylated protein after the fusions bound the target loaded resin. Initial off-rate selections were sampled at 6, 12, and 18 minutes following addition of the soluble competitor. Secondary off-rate selections against both mPD-L1 and hPD-L1 were sampled at 20 and 40 minutes following incubation with the soluble competitor. Detailed selection conditions for both hPD-L1 and mPD-L1 are found in Supplemental Tables S1 and S2, respectively.

Clone sequencing and affibody clone fusion production

Final round affibody fusions from both the 20- and 40-minute final off-rate selection for both mPD-L1 and hPD-L1 targets were PCR amplified with Taq DNA polymerase (Sigma-Aldrich, D1806) according to manufacturer’s directions and cloned into a sequencing vector using the pCR^™4-TOPO^® TA kit (ThermoFisher, K457502). After DNA plasmid recovery with a mini-prep kit (ThermoFisher, K210002), plasmids were sequenced using the Sanger method. Clustal Omega sequence alignment ⁴⁸ was used to group sequences and select ten clones, five from the hPD-L1/mPD-L1 dual selection and five from the hPD-L1-only selection (Supplemental Figure S2). Selected clones were purchased from IDT as gBlock gene fragments. Following PCR amplification with primers designed to add a 5′ T7 polymerase sequence and a 3′ linker region, the purified DNA was transcribed, ligated to linker, and translated in presence of [³⁵S]-methionine/cysteine. Prior to clone screening, affibody-mRNA fusion clones were treated with 10 μg RNAse A (Millipore, 70856) for 1 hour at 22 °C to remove the RNA portion.

Affibody fusion clone screening and therapeutic antibody block

Affibody-DNA clones radiolabeled with [³⁵S]-Methionine were individually screened for binding to hPD-L1, mPD-L1, and, if not triaged, hPD-L2 and Her2. Pre-blocked NeutrAvidin resin was loaded with 40 pmol of biotinylated target protein as stated in previous sections. Affibodies were incubated at 4 °C for 1 hour in selection buffer and washed six times to remove unbound clones. Resuspended washed resin and initial flow-through were analyzed by scintillation counting to obtain a percent bound value. For the antibody blocking experiment, 20 pmol of hPD-L1 or mPD-L1 was added to each test resin as before and incubated with either mPD-L1/hPD-L1 bispecific antibody 16377 or atezolizumab (Tecentriq^®) to occupy the PD-1 binding surface. After 1 hour at 4 °C the antibody solution was removed and resin incubated with radiolabeled, RNAse A-treated affibody fusions. The percent decrease in binding resulting from antibody incubation was then calculated as compared to no-antibody controls.

Recombinant affibody production and conjugation

Production of affibody derivatives are described in detail in the Supporting Information along with purification and characterization data in Supplemental Figure S3. Briefly, pET151 bacterial expression vectors encoding the affibodies with N-terminal his tags were produced in BL21 cells. After expression and nickel-NTA purification, maleimide chemistry was used to label affibodies on the C-terminal cysteine. The N-terminal tag was removed with tobacco etch virus (TEV) protease and purified by nickel-NTA by collecting the flow-through. Following buffer exchange, and if deemed necessary by SDS-PAGE, further purification was carried out using reverse phase high performance liquid chromatography (HPLC). The identity of all constructs was confirmed by mass spectrometry.

Binding of affibodies to immobilized PD-L1

For total resin binding, streptavidin resin was blocked and loaded with 100 pmol of biotinylated protein (per replicate) in selection buffer with 0.5 mg/mL BSA. Excess streptavidin sites were blocked with D-biotin. Following resin washes, a total of 24 pmol of either sulfo-Cyanine 5 (Cy5) or AlexaFluor 488nm dye (AF488nm) conjugated affibody was incubated with the loaded resin for 1 hour at 4 °C. Supernatant was collected as flow-through. Following six washes with selection buffer, the resin was incubated with 1% (v/v) sodium dodecyl sulfate (SDS) in phosphate buffered saline (PBS; Corning, 21-040-CM) for 30 minutes to release bound affibody. Solutions were placed into opaque black 96-well plates and fluorescence intensity read with the dye’s excitation/emission using a BioTek Synergy H4 microplate reader. Solutions were normalized by volume and reported as percent bound to total.

Affibodies were tested for binding to biotinylated proteins presented on streptavidin coated wells. Biotinylated proteins were diluted to 10 μg/mL in selection buffer with 0.5 mg/mL BSA and incubated in the wells of a pre-blocked and re-hydrated streptavidin coated plate (Pierce, 15500) for 1 hour on a microtiter shaker at room temperature. Unbound protein was removed by four washes with selection buffer. Affibodies conjugated with AF488nm were serially diluted and allowed to bind the immobilized protein for 1 hour at room temperature on a microtiter shaker. Wells were washed four times with selection buffer raised to 0.1% (v/v) (vs. 0.05%) Tween-20. Conjugated affibodies were released from the surface by 1% (w/v) SDS in selection buffer for 45 min at room temperature in a microtiter shaker and volume transferred to a new black opaque plate. Wells were read at 490 nm excitation, 530 nm emission. Binding constants were obtained with GraphPad Prism 8.0 software using the one site-total binding model.

Surface plasmon resonance (SPR) analysis

Binding kinetics of M1 affibody capped with N-methyl maleimide (NMM) with PD-L1 (human and mouse) were measured by surface plasmon resonance (SPR) using a Biacore 3000 optical biosensor (GE Healthcare) at 25 °C. The PD-L1 sensor surfaces were prepared using a streptavidin sensor chip (Cytiva) by injecting biotin-tagged protein (50 nM) at a flow rate of 10 μL/min for 2 min. Approximately 700 resonance/response units (RU) of human PD-L1 and 850 RU of mouse PD-L1 were tethered to the chip. Another flow cell with only immobilized streptavidin served as a reference surface. Binding experiments were performed at a flow rate of 50 μL/min in HBSTG (20 mM HEPES-NaOH pH7.5, 150 mM NaCl, 0.02% (v/v) Tween-20, 5% (v/v) glycerol). All SPR responses were reference- and buffer-corrected.

SPR response curves were globally fit to a 1:1 Langmuir model ($\mathrm{A}+\mathrm{B}\mathrm{ }\stackrel{k_a}{\underset{k_d}{\leftrightarrow }}\mathrm{ }\mathrm{A}\mathrm{B}\mathrm{ }$) or predefined two state reaction model ($\mathrm{A}+\mathrm{B}\mathrm{ }\stackrel{k_{a1}}{\underset{k_{d1}}{\leftrightarrow }}\mathrm{ }\mathrm{A}\mathrm{B}\mathrm{ }\stackrel{k_{a2}}{\underset{k_{d2}}{\leftrightarrow }} \mathrm{A}\mathrm{B}\mathrm{x}$), where the association and dissociation rate constants (k_a1, k_d1) are for the initial binding state ($\mathrm{A}+\mathrm{B}\mathrm{ }\stackrel{k_{a1}}{\underset{k_{d1}}{\leftrightarrow }}\mathrm{ }\mathrm{A}\mathrm{B}\mathrm{ }$), and forward and backward rate constants (k_a2, k_d2) are for the conformational change state ($\mathrm{A}\mathrm{B}\mathrm{ }\stackrel{k_{a2}}{\underset{k_{d2}}{\leftrightarrow }} \mathrm{A}\mathrm{B}\mathrm{x}$). K_D values for the interactions were derived from the association and dissociation rate constants for the initial binding (K_D = k_d1/k_a1).

Flow cytometry

M1 and control Her2 affibody were tested for binding targets exogenously expressed in Chinese hamster ovary (CHO) cells. Test cells were either un-transfected (Parent), transfected with hPD-L1, hPD-L2, or mPD-L2 according to previous methods ⁴⁹. As well, Her2 high expressing SK-BR-3 cells (ATCC, HTB-30) were used as a positive control for Her2 affibody binding. All cells were cultured in DMEM (Corning, 10-013-CV) supplemented with 10% (v/v) heat inactivated fetal bovine serum (FBS, Sigma-Aldrich, F4135) and 1X penicillin-streptomycin (Sigma-Aldrich, P4333). For flow cytometry, cells were briefly washed with Mg⁺²- and Ca⁺² -free PBS, dissociated into non-adherent single cells with trypsin (0.25 (w/v) trypsin, 0.2% (w/v) -EDTA; Sigma-Aldrich, T4049), trypsin was deactivated with full media, cells centrifuged at 400 x g for 2 minutes, and pellet resuspended in serum-free (SF) media. Cells were divided into v-bottom plates (Corning, 3894) (200,000 cells/well), centrifuged, and resuspended in a dilution series of AF488nm dye-conjugated affibody diluted in cold 1:1 PBS:SF media. Cells were incubated without agitation with the affibody for 1.5 hours at 4 °C in the dark. After binding, cells were centrifuged at 400 x g and washed four times with cold PBS: SF media, resuspended and immediately acquired for live single cell fluorescent intensity in a BD LSRFortessa^™ X-20 flow cytometer (488 nm laser excitation). Gain was adjusted so that the highest observed fluorescence intensity was below maximum intensity (saturation). Parental CHO cells were used to normalize non-specific uptake/background. Dissociation constants were determined by GraphPad Prism using the one-site normalized total binding model.

PD-L1 T-cell re-activation assay

To assess the potential of PD-L1 peptides to relieve inhibition mediated by PD-1/PD-L1 interaction in T-cells, a cell-based bioluminescent assay was employed. CHO cells overexpressing an engineered surface protein that activates cognate TCRs in an antigen-independent manner (aAPC/CHO-K1) and human PD-L1 (PD-L1+ aAPC/CHO-K1) were used as artificial Antigen-Presenting Cells (APC). Jurkat T-cell line overexpressing PD-1 on the cell surface and carrying a luciferase reporter gene under the control of a NFAT-response element was used as a T-cell surrogate. CHO cells and Jurkat T-cells were cultured in Ham’s F12 Nutrient mixture and RPMI 640 medium supplemented with 10% (v/v) FBS, respectively. The cells were obtained from Promega and checked for over-expression by flow cytometry.

The aAPCs were seeded in white 96-well flat plates at the density of 40,000 cells/well in Ham’s F12 nutrient mixture 24 hours prior to co-culture with Jurkat T-cells. Next day, serial dilutions of the peptides were prepared in a final volume of 40 μL RPMI-640 medium containing 2% (v/v) FBS. Medium was removed from each well and 40 μL of the titrated peptides were added. 50,000 Jurkat T-cells in 40 μL of 2% (v/v) FBS RPMI-640 medium was added to each well and incubated at 37 °C for 6 hours. Following incubation, plates were equilibrated at room temperature for 10 minutes and 80 μL of Bio-Glo reagent (Promega, Madison, WI, catalogue no. G7940) was added to each well. After 15 minutes of incubation, bioluminescence was quantified in a Synergy HT microplate reader.

Syngeneic hPD-L1 lymphoma animal models

For EL4 lymphoma cell hPD-L1 overexpression, hPD-L1 cDNA was cloned into the pMG-Luc retroviral vector. This vector resembles pGC-IRES except the luciferase gene is used for selection. 293T cells were transfected with the retroviral vector using TransIT-293 reagent in DMEM medium containing 10% (v/v) FBS. At 24 hours post transfection, the medium was replaced with 3 mL of RPMI 1640 medium containing 10% (v/v) FBS and the incubation was continued at 37 °C and 5% (v/v) CO₂. At 72 hours post transfection supernatant containing the retrovirus was harvested for transduction. Polybrene was added to the retrovirus-containing supernatant at a final concentration of 8 μg/mL. Murine lymphoma EL4 cells were transduced with the harvested retroviral titers using the spinoculation technique. At 36 hours post-infection human PD-L1 expressing EL4 cells were sorted using flow cytometry and a stable cell line of murine lymphoma EL4 overexpressing human PD-L1 ligand (EL4-hPDL1) was established.

Male B6(Cg)-Tyrc-2J/J mice purchased from The Jackson Laboratory were housed according to the Association for Assessment and Accreditation of Laboratory Animal Care and NIH standards. In vivo experiments were conducted according to protocols approved by the University of Texas MD Anderson Cancer Center Institutional Animal Care and Use Committee. Mice were 8 weeks old at the time of implantation.

EL4-hPDL1 cells were cultured in RPMI 1640 medium supplemented with 10% (v/v) FBS. On the day of implantation adherent EL4-hPDL1 cells were trypsinized and washed twice with cold PBS. Mice were implanted with 50,000 EL4-hPDL1 cells subcutaneously on the right flank of the mouse in 100 μL of PBS containing 30% (v/v) Matrigel^® (Corning, Glendale, AZ, catalogue no. 356231).

Tumor imaging and affibody uptake

One day before imaging mice were intraperitoneally (IP) injected with either 240 μg atezolizumab or control IgG antibody (Leinco Technologies, IgG1 Clone MOPC21, LT9005). Mice were given 5 nmol of either M1-cy5 affibody or Her2-cy5 affibody (negative control) through tail vein injection. Fluorescent images were acquired 1 and 4 hours after injection with the IVIS^® Spectrum In vivo Imaging System (Perkin Elmer). Radiant efficiency of fluorescence in tumors and regions of interest (ROI) was quantified for fluorescence in Living Image^® software (Perkin-Elmer). Animals were euthanized, tumors and organs were resected and imaged ex vivo. Resected tumors were split in half, massed, either snap frozen or fixed in 10% (v/v) formalin for 2 days and switched to 70% (v/v) ethanol for paraffin embedding.

To quantify affibody Cy5 uptake, snap-frozen tumors were ground into small pieces using a stainless-steel mortar and pestle cooled by liquid nitrogen. For each tumor, the frozen pieces of tissue were homogenized in ice-cold RIPA (radioimmunoprecipitation assay) buffer (20 mM Tris-HCl pH 7.5, 150 mM NaCl, 1 mM EDTA, 1 mM EGTA, 1% (v/v) NP-40, 1% (v/v) sodium deoxycholate) with phosphatase and protease inhibitors using the Polytron PT2500E (Kinematica, Bohemia, NY). Lysing Matrix D beads (MP Biomedical, Irvine, CA) were added to the final homogenized tissue in RIPA buffer and vortexed three times along with three freeze-thaw cycles. Homogenized tissue was kept on ice throughout the experiment. To remove insoluble material, samples were centrifuged at 3,280 x g for 15 min at 4 °C. Supernatant was removed and stored at −80 °C. After another thaw cycle, solutions were centrifuged at 14,000 x g for 5 minutes, supernatants recovered, and 1:1 serial diluted in RIPA buffer to assess Cy5 content in a plate reader. A dilution standard curve of known Cy5 affibody in RIPA buffer was used to estimate the molarity of the solution. Final values were normalized to total protein content using a bicinchoninic acid (BCA) assay kit (Pierce).

Statistics

GraphPad Prism software (GraphPad, San Diego, CA) analysis was used to compare groups by one-way and two-way analysis of co-variance (ANOVA) with Tukey’s multiple comparison tests. P values less than 0.05 were considered significant.

Results and Discussion

Selection of affibodies that bind mPD-L1 and hPD-L1

An optimized affibody scaffold ⁴⁷ was adapted into an mRNA display library to produce affibody-mRNA fusion molecules connected through a puromycin linker as seen in Figure 1. The library consists of 13 randomized amino acids forming the binding surface of the affibody (magenta). The mRNA-affibody library (~8.38 x 10¹³ possible sequences) was panned over immobilized recombinant hPD-L1 ectodomain for four rounds before splitting the library to co-select for mPD-L1 binding for several rounds. This was done to identify mPD-L1/hPD-L1 cross-binding affibodies that could be validated in mouse models of ICI therapy prior to translation to humans. The bifurcated pools were subjected to off-rate selections in the presence of excess free hPD-L1 or mPD-L1. The resulting pools were sequenced and analyzed by Clustal Omega (Supplemental Figure S2). Blue (mPD-L1 dual selected) and red (hPD-L1 only selected) sequences are distinct indicating that the library had not fully converged prior to the selection split. We also note that the final diversity in both groups is still quite high as only 5 sequences were observed to occur more than once (Supplemental Figure S2, bold vertical line). These sequences comprise only 35% (19/55) of the combined selection pools with one sequence repeated seven times. The relatively high clonal diversity present in the final selection pools suggests that additional selection rounds can be carried out to further enhance affinity, stability, and selectivity.

Figure 1: Selection of PD-L1 affibodies by mRNA Display.

A) The binding surface of the affibody triple helix (green) was randomized (magenta) to create an affibody library. B) mRNA display affibody fusions were produced (>8.3E13 molecules) for selection of randomized affibodies against mouse PD-L1 (mPD-L1) and human hPD-L1 (hPD-L1). Fusions include an affibody peptide linked via a puromycin (P) poly-dA linker to its mRNA with hybridized cDNA. Four rounds of selection were performed against hPD-L1 after which, the library was split, and two parallel selections were carried out against hPD-L1 or mPD-L1. Both selections concluded with two sequential off-rate selection rounds to further enhance affinity. C) Affibody sequences described in this work were aligned. This includes the amino acid sequences of the parent Z-domain scaffold, a Her2-specific affibody ZHer2:2891, the affibody library, and the lead PD-L1 specific candidate, M1. Green indicates hydrophilic residues, blue indicates positively charged residues, red indicates negatively charged residues, and gray indicates hydrophobic residues. X indicates a randomized position in the affibody library (any of the 20 natural amino acids). V indicates a mutation from alanine to valine observed in the non-randomized region of M1.

We note that this selection employed thrombin-mediated target release to recover PD-L1-bound affibodies. This ensures that library members with affinity to the C-terminal hexahistidine tag, streptavidin, or the solid phase matrix are retained on the solid phase and do not contribute to the enriched pool. Indeed, PCR assessment of the resin bound fractions showed dramatic decreases in background resin binding after thrombin treatment (Supplemental Figure S4), suggesting that this approach enhanced the rate of enrichment of PD-L1 binding affibodies ⁴⁶.

Ten clones, five from the dual selection (M prefix) and five from the hPD-L1 only selection (H prefix) were selected for further analysis . These clones were chosen based on phylogenetic grouping and frequency with a bias towards sequences recovered in the 40-minute off rate selections (Figure 2A). The red highlighted positions correspond to the amino acids that were selected at the randomized positions. The blue colored amino acids were not randomized in the initial library but underwent mutation during the selection.

Figure 2: PD-L1 binding of selected affibody sequences.

A) Sequences of the 10 clones selected for further analysis and binding characterization. Red highlighted amino acids were randomized in the initial library. Blue bolded amino acids were spontaneous mutations arising in the non-randomized regions of the affibody scaffold. B) Sequences were expressed as mRNA-affibody fusions and panned against immobilized hPD-L1, mPD-L1, hPD-L2 or Her2. An asterisk under the affibody indicates it was not assessed for binding to hPD-L2 or Her2 due to poor hPD-L1 binding. C) The clinical PD-L1 binding antibody Tecentriq^® (atezolizumab) was used in a blocking assay to determine if selected affibody clones bound at the hPD-L1:PD-1 interface. In almost all clones, hPD-L1 binding was completely blocked relative to the no affibody control (% block binding). D) A similar experiment was carried out with a dual mouse/human PD-L1 binding antibody blocking assay to determine if mPD-L1 selected affibodies bind at the mPD-L1/PD-1 interface.

Fusions were produced in the presence of [³⁵S]-methionine and treated with RNAse A to remove the pendant mRNA. Affibodies were first tested for binding against mouse and human PD-L1 and triaged if low binding or poor translation was observed (affibodies denoted with *) (Figure 2B). The remaining affibodies were then tested against hPD-L2, resin, or Her2 to test for specificity for mPD-L1 and/or hPD-L1. None of these affibodies showed significant off-target binding, save for H32. As expected, all the dual-selected affibodies bind to both mouse and human PD-L1 with M1 and M3 showing the highest binding to each. In the case of the hPD-L1-only selection clones, only H28 and H32 translated well and showed above-background binding to hPD-L1. None of the affibodies isolated in the hPD-L1-only selection showed binding to mPD-L1 indicating that the mPD-L1 selection was necessary to identify cross-binders.

Affibodies are potentially large enough to block the PD-1:PD-L1 interaction and concomitantly abrogate signaling along the PD-L1:PD-1 axis, enhancing immunotherapy. To identify affibodies with this property, selected affibodies were tested for competition with the therapeutic antibody atezolizumab or a bispecific mPD-L1/hPD-L1 antibody. Resin-bound hPD-L1 (Figure 2C) and mPD-L1 (Figure 2D) were pre-treated with atezolizumab or the bispecific mPD-L1/hPD-L1 antibody, respectively, at two concentrations, briefly washed, and incubated with affibody fusions. Binding of each affibody to hPD-L1, with the exception of M7, was nearly completely blocked by atezolizumab. Binding of affibodies with mPD-L1 affinity was likewise inhibited by the bispecific antibodies, except for M7. These data suggest that affibodies selected for PD-L1 binding are likely to interact with the PD-1 binding surface and that clone M7 may have an orthogonal binding site.

To evaluate the functional consequences of interaction at the PD-L1:PD-1 interface, we employed a CHO/T-cell (Jurkat) re-activation assay (Supplemental Figure S5) ⁵⁰. Jurkat cells with an NFAT driven bioluminescence reporter were incubated with either parental CHO cells expressing nothing (K1), hPD-L1, hPD-L2, or dual hPD-L1/2 in the presence of recombinant affibodies (H28, H32, M1, M3) at varying concentrations. This assay requires that an affibody physically disrupts the engagement of hPD-L1 (on CHO cells) and PD-1 (on Jurkat cells) to activate the bioluminescence reporter. Only two affibodies, M1 and M3, were able to abrogate binding of PD-1 expressing Jurkat cells to hPD-L1 expressing CHO cells as seen by a recovery in bioluminescence signal. However, only M1 was able to block the hPD-L1/PD-1 interaction with statistical significance. The other affibodies in the screening panel failed to produce any appreciable effect despite ostensibly binding to the PD-1 interacting surface. Indeed, H28 failed to have a functional effect despite high apparent PD-L1 affinity. Based on its affinity, cross-reactivity, and effect on PD-L1 signaling, M1 was chosen for further characterization.

To establish an approximate EC₅₀ for the disruption of PD-L1:PD-1 signaling, the T-cell reactivation assay was employed again, but with a larger range of M1 concentrations (Figure 3). The Her2 affibody was used as a negative control and to normalize the resulting data. The M1 affibody was found to disrupt signaling through the PD-L1:PD-1 axis with an EC₅₀ of 2 μM, which is roughly in-line with inhibitory potencies of most PD-L1 binding peptides in the literature ¹⁹. These data also suggest a sub-micromolar dissociation constant (K_D) as the EC₅₀ values obtained in this functional assay are often multiple logs higher than their dissociation constants. For example, atezolizumab shows an EC₅₀ in the mid-low nM range in this bioluminescence assay despite is low picomolar dissociation constant ⁵¹. Based on these results, we predicted a low nanomolar dissociation constant for M1 which was borne out in the experiments described below.

Figure 3: M1 affibody blocks PD-1:PD-L1 signaling in cell culture.

M1 affibody and control Her2 affibody were incubated with Jurkat T-cells expressing PD-1and CHO cells expressing hPD-L1. Blocking the PD-1:PD-L1 signaling axis in T-cells increases bioluminescence. The M1 affibody blocks this interaction with an IC₅₀ of 2 μM (0.7 to 6 μM for 95% confidence interval). The Her2 affibody was used for background normalization.

Dual specific M1 affibody binds PD-L1 in many contexts

To assess the potential of M1 as an imaging agent, the affibody was labeled on a C-terminal cysteine with one of two fluorescent dyes: AF488nm and sulfo-Cy5. The Her2 affibody was similarly labeled in order to function as a negative control in these experiments. Based on our preliminary screening data, we predicted that the PD-L1 selection resulted in full phenotypic conversion from Her2 binding to PD-L1 binding. As shown in Figure 4A and B, the M1 and Her2 affibodies were incubated with either empty resin or resin loaded with excess mPD-L1, hPD-L1, or Her2. As expected, the Her2 affibody showed no detectable binding to PD-L1 nor to resin, yet retains high affinity for immobilized Her2. In contrast, The M1 affibody shows no binding to Her2 and strong binding to both mouse and human PD-L1. In these experiments, the effect of the dye label (AF488nm or Cy5) on binding appears to be quite modest. Titration of affibody-dye conjugates against PD-L1 immobilized on streptavidin plates was used to obtain an estimate of the dissociation constants of both M1 (K_D = 89 nM for hPD-L1 and K_D = ~47 nM for mPD-L1) and the Her2 affibody (K_D = 11 nM for Her2) (Figure 4C-E).

Figure 4: Binding of the M1 Affibody to immobilized PD-L1.

A) Binding of AlexaFlour488nm and B) SulfoCy5 dye-labeled affibodies (M1 and Her2 affibody control) to streptavidin immobilized hPD-L1, mPD-L1, or Her2. The binding is measured as the % of resin-associated fluorescence relative to the total input fluorescence. These experiments demonstrate that the dye-labeled affibodies bind selectively to their immobilized targets. The M1 affibody bound to both C) mPD-L1 and D) hPD-L1 in a concentration-dependent fashion (K_D= 47 (38 to 58) nM and 89 (68 to 120) nM, respectively) but not to HER2 or PD-L2. E) As expected, the control HER2 affibody bound only to immobilized Her2 (K_D= 11 (7.3 to 16) nM). K_D values are provided along with the 95% confidence interval in parentheses (GraphPad Prism).

We next sought to verify the affinity of M1 using SPR. In these experiments (Figure 5) we evaluated NMM-capped M1 affibody (to block the free C-terminal cysteine) as well as its AF488nm and Cy5 conjugates. Biotinylated target protein was captured on the sensor via streptavidin and exposed to varying concentrations of affibody. Neither the M1 affibody, nor any of its fluorescent derivatives showed detectable binding to hPD-L2 or Her2 which is consistent with our previous experiments (Supplemental Figure S6A, B). None of the Her2 affibody conjugates bound to mPD-L1 or hPD-L1 (Supplemental Figure S6C, D). As expected, the Her2 affibody bound strongly to Her2 although its slow off-rate prevented a precise K_D calculation (Supplemental Figure S6E). The M1 affibody was found to bind human (Figure 5A) and mouse PD-L1 (Figure 5B) with kinetic-derived dissociation constants of 68 nM and 9.5 nM, respectively, suggesting a higher affinity for the mouse isoform. Steady state measurement gave similar affinity values (Figure 5C). In Figure 5B, a two-state model of binding produced the best fit of mPD-L1 binding to M1 affibody, suggesting a receptor conformational change and/or dimerization event following M1 binding. Further experimental studies are required to support or reject this hypothesis.

Figure 5: Surface plasmon resonance (SPR) analysis of binding kinetics of M1 Affibody.

Two-fold serial dilutions of M1 affibody were flowed over a biotin-tethered PD-L1 sensor surface on a streptavidin chip. A) SPR sensorgrams for the binding of N-methylmaleimide (NMM)-capped affibody to human PD-L1 surface (~ 700 RU) are shown in black and data were fit to a 1:1 Langmuir binding model (red overlay line). The derived dissociation constant K_D was calculated from the rate constants (K_{D =} k_d/k_a). B) M1 Affibody and mouse PD-L1 (~850 RU) binding was fitted to a two-state model (blue overlay line) in order to describe the complex interaction observed. The apparent dissociation constant K_D value was estimated using the binding rate constants (K_D = k_d1/ k_a1), while the conformational change rates (k_d2 = 6.94 × 10⁻³ 1/s and k_a2 = 1.91 × 10⁻² 1/s) were omitted (see materials and methods). C) The average binding response (RU) at/close to steady state (110-113 s) as a function of affibody concentration is shown.

While the M1 affibody sequence was recovered after 20 minutes under off-rate selection conditions, the SPR data show background response signal within several minutes in the off-rate phase. How then were high-affinity sequences identified at the conclusion of the off-rate selections? This incongruity can be explained by extraordinary sensitivity of PCR and its utility in recovering extremely low quantities of genetic information at the conclusion of a high stringency selection steps. Indeed, re-amplification of the off-rate selected fusions required 18 cycles of PCR, resulting in a 2¹⁸ (>200,000-fold) increase in DNA concentration relative to the original fusion material remaining on the resin after 20 or 40 minutes. Furthermore, SPR measurements were performed under a laminar flow and a high flow rate was used to prevent mass transfer limitations. Constant shear generated on the PD-L1 protein sensor surface could contribute to the fast unbinding of M1 affibody during the dissociation phase where off-rates were measured. In the resin-based selection experiments, the binding and washing steps are closer to static conditions. These results indicate that the affinity maturation step worked as expected and, given the substantial sequence diversity in the final pool, suggest that further improvements in affinity may be obtainable. Finally, we also observed that the Cy5 and AF488nm versions of M1 affibody did not dramatically affect their kinetic-derived or steady state final K_D measurements (Supplemental Figure S7).

M1 Affibody binds to PD-L1 on the surface of living cells.

With promising in vitro results, the M1 affibody was tested for its ability to bind PD-L1 on living cells. Fluorescently labeled M1 affibody or Her2 affibody was incubated with CHO cells exogenously expressing nothing (parent), hPD-L1, mPD-L1, or hPD-L2, or to SKBr3 cells that overexpress Her2. The cells were then washed and analyzed by flow cytometry. The CHO parent (black) was used to subtract background (non-specific uptake) of the M1-AF488nm affibody. As seen in Figure 6, the dose-dependent binding of M1 affibody to hPD-L1 (Figure 6A) and mPD-L1 (Figure 6B) expressing cells confirms a complete phenotypic switch from Her2 to PD-L1 target specificity. No binding of M1 to PD-L2 expressing cells was observed (Figure 6C). These experiments reveal an apparent K_D of 95 nM for hPD-L1, which is consistent with SPR experiments and inhibition of PD-L1 signaling in the T-cell re-activation assay from Figure 3. However, M1 shows an order of magnitude decrease in affinity for mPD-L1-expressing cells relative to hPD-L1-expressing cells. This may be attributable to alternate conformations of cell surface mPD-L1 or improper presentation of the binding site relative to the human isoform. We note that mPD-L1 was produced in HEK293F cells while flow cytometry was performed on CHO cells expressing mPD-L1. Previous studies suggest extensive differences between CHO and HEK293 post-translational glycosylation ⁵². It is possible that altered glycosylation of mPD-L1 on the surface of CHO cells resulted in an suboptimal binding surface, leading to the observed decrease in affibody binding. Further experimental work is needed to confirm this. As expected, the Her2 affibody strongly binds to SKBr3 cells (high Her2), but not to hPD-L1 expressing CHO cells (red in Figure 6D). These flow cytometry experiments demonstrate robust and selective binding of M1 to PD-L1 in its biologically relevant state and in the presence of multiple non-target proteins. Table 1 summarizes the affinity measurements for all conjugates of the M1 and Her2 affibodies in each of the binding experiments described above.

Figure 6: Affibody binding to CHO cells by flow cytometry.

A) M1-AF488nm affibody binding to CHO cells expressing hPD-L1 and normalized to CHO parental cells (K_D = 95.5 (81.7 to 111) nM) . B) M1-AF488nm affibody binding to CHO cells expressing mPD-L1 and normalized to CHO parent cells (K_D=587 (458 to 751) nM) . C) M1-AF488nm affibody binding to hPD-L2 and normalized to CHO parent cells; no binding was observed. D) HER2-AF488nm affibody binding to SKBr3 (Her2 overexpressing breast cancer cell line) and hPD-L1 expressing CHO cells lines. HER2 affibody bound strongly to SKBr3 cells (K_D = 7.12 (6.11 to 8.30) nM) but not to hPD-L1 expressing CHO cells. K_D values are provided along with the 95% confidence interval in parentheses (GraphPad Prism)

Table 1:

Summary of M1 and Her2 affibody binding data with different conjugates.

Affibody	Conjugate	Assay	Target	KD (nM)
M1	NMM	SPR	hPD-L1	68 (± 9.9)
M1	NMM	SPR	mPD-L1	9.5 (± 5.8)
M1	AF488nm	SPR	hPD-L1	66.8 (± 11.5)
M1	AF488nm	SPR	mPD-L1	16.7 (± 1.2)
M1	Cy5	SPR	hPD-L1	47.1 (± 7.9)
M1	Cy5	SPR	mPD-L1	33.0 (± 2.9)
M1	AF488nm	Plate	hPD-L1	89 (68-120)
M1	AF488nm	Plate	mPD-L1	47 (38-58)
M1	AF488nm	FC	hPD-L1	95.5 (81.7-111)
M1	AF488nm	FC	mPD-L1	587 (458-751)
M1	NMM,Cy5,AF488nm	SPR,FC, Resin, Plate	hPD-L2, Her2	None detected
Her2	NMM	SPR	Her2	0.013*
Her2	AF488nm	SPR	Her2	0.115*
Her2	Cy5	SPR	Her2	0.046*
Her2	AF488nm	Plate	Her2	11 (7.3-16)
Her2	AF488nm	FC	Her2	7.12 (6.11 to 8.30)
Her2	NMM,Cy5,AF488nm	SPR,FC, Resin, Plate	hPD-L1, hPD-L2, mPD-L1	None detected

For SPR, shown are averages with standard deviation (n=3). Ranges were based on 95% confidence intervals (CI) calculated in GraphPad Prism.

SPR=surface plasmon resonance; FC=flow cytometry; NMM=N-methylmaleimide; AF488nm=AlexaFluor 488nm dye; Cy5= sulfo-cyanine-5;

^*could not regenerate sensor due to low off-rates affecting final measurements.

The M1 Affibody shows selective uptake in PD-L1 expressing tumors

Affibody molecules as imaging agents have shown robust uptake and specificity in mouse tumor models and in human patients ^{32, 36}. We sought to determine if the M1 affibody could accumulate in tumors expressing hPD-L1 in the presence of endogenous mPD-L1 expression in normal tissue. Immune-competent mice were implanted with EL4 lymphoma cells expressing hPD-L1 to generate tumors with both mPD-L1 and hPD-L1 expression, which was confirmed by flow cytometry pre-implantation (Supplemental Figure S8). Before IV injection of M1-Cy5 affibody, mice were IP injected with either control IgG1 or atezolizumab at 10 mg/kg, conditions similar to effective therapeutic treatment in mice ⁵³. Her2-Cy5 affibody with IgG1 pre-injection was used as a non-specific uptake control. As seen in Figure 7A, tumors exposed to M1 show higher radiant efficiency (excitation normalized fluorescent signal) compared to tumors exposed to the Her2 affibody control at 1 hour and 4 hours. Quantification of tumor regions of interest (ROIs) (Figure 7B) show statistically significant differences between control Her2 affibody and the M1 affibody (both with and without atezolizumab pre-injection) at both time points. The average tumor signal for the M1 affibody groups drops from 1 to 4 hours but remains well above pre-scan levels. The Her2 affibody begins to approach pre-scan levels in the tumor at 4 hours, consistent with complete probe washout in the absence of target binding.

Figure 7: M1 affibody uptake of hPD-L1 expressing tumors.

A) Following IP pre-injection with either atezolizumab (ATZ) or control IgG1 antibody, mice with EL4 lymphoma subcutaneous flank tumors expressing hPD-L1 were imaged on the IVIS^® Spectrum In vivo Imaging System to establish background signal arising from autofluorescence. Mice were then injected with either 5 nmol M1-Cy5 affibody or HER2-Cy5 affibody (negative control) and imaged 1 and 4 hours later. Mice were sacrificed to expose tumor and organ uptake (T is tumor, H is heart, L is lungs). Affibody uptake in side-facing tumors B) and ex vivo resected tumors C) was quantified by placing regions of interest (ROI) over the tumor and measuring radiant efficiency. Two-way ANOVA with multiple comparisons show M1 affibody has statistically higher uptake than control HER2 in tumors over all time points and states. D) Lysed and processed tumor tissue evaluated for Cy5 presence show increased amounts of fluorescence (pM Cy5) with M1 affibody in both conditions versus the Her2 affibody. P value *<0.05, **<0.01,***<0.001, ****<0.0001.

After four hours the animals were sacrificed with tumors and organs resected for imaging and quantification (Figure 7A, ex vivo; quantified in Figure 7C). As expected, EL4 tumors treated with M1-Cy5 retained Cy5 signal at levels significantly higher than EL4 tumors exposed to the Her2 affibody-Cy5 conjugate. Following ex vivo analysis, tumors were halved, frozen, lysed in RIPA buffer, and analyzed for Cy5 content. Values were normalized to protein content and used to calculate the concentration of Cy5 in the tumor (Figure 7D). The values ratiometrically mirror the live and ex vivo optical imaging quantification. Cy5 concentrations are significantly higher in both M1 affibody groups relative to those in the Her2 affibody group. Notably, the Her2 affibody is not completely blank indicating a low level of non-specific retention. The affibody is well below the purported 45 kDa size for the EPR effect found in tumors with hypervascularization and malformed vessels ³⁰, so the dye itself could be causing non-specific uptake. Additionally, tumors were quite large (ranging from ~2 to 4 grams) with large necrotic cores indicative of blunted or dead-end vasculature. These physiological features may have contributed to non-specific uptake and/or retention. Even with these caveats, M1 affibody was specifically concentrated in the PD-L1 expressing tumors. In fact, accounting for tumor volume and uptake, M1 concentration in the tumor was approximately 25 nM, a value close to the K_D values measured above. This suggests that higher doses (readily achievable) could induce affibody loading sufficient for PD-L1 receptor saturation.

Contrary to expected results, pre-injection with atezolizumab at therapeutic concentrations resulted in a slightly higher accumulation of M1 affibody than control pre-injection. Indeed, live imaging shows a modest increase in affibody retention in tumors treated with atezolizumab relative to IgG controls (* p value <0.05). Based on the in vitro blocking data, we expected atezolizumab to completely block M1 binding to the PD-1 binding site on PD-L1 (Figure 2C). Given the low pM affinity of atezolizumab, we assumed a ratio of 1.6 nmol atezolizumab to 5 nmol M1 affibody would be sufficient to achieve partial affibody blocking within the tumor. Although this was not observed, there are several possible explanations. Atezolizumab binds to both human and mouse PD-L1 with high (pM) affinity ⁵¹, and PD-L1 expression is not limited to the tumor. Other organs like the liver, spleen, and thymus express mPD-L1 ⁵⁴. This endogenous pool of mPD-L1 may deplete the antibody from circulation and prevent significant accumulation and blocking in the tumor compartment. Indeed, we observe significant antibody-mediated blocking of M1 accumulation in the spleen, heart, and lungs. None of the differences in organs were significant between atezolizumab and IgG1 pre-injections for M1 affibody, except for the heart where radiant efficiency values were suppressed to near background levels. Spleen uptake with M1 seemed to be partially blocked with atezolizumab, with a P value at the boundary of statistical significance (0.07). This indicates atezolizumab may be blocking M1:mPD-L1 binding in visceral organs, thus allowing for increased affibody loading in the tumor. This phenomenon has also been seen in studies where blocking endogenous protein targets increases imaging probe accumulation in the tumor ⁵⁵. Indeed, when used as a PET imaging agent, the accumulation of Her2 affibody in Her2-expressing tumors was enhanced when low specific-activity injections were performed ⁵⁶. This suggests that saturating endogenous target protein (off-target or undesired on-target) with unlabeled affibody can facilitate increased on-target uptake of labeled affibody in the tumor. Finally, the presence of circulating soluble PD-L1 (sPD-L1) may also provide an additional sink for the therapeutic antibody and prevent on-target blocking in the tumor compartment ⁵⁷. This may explain the absence of antibody-mediated blocking in the tumor and the indications of antibody-mediated blocking in the normal tissue.

Differences in tumor penetration may also account for the anomalous blocking results. The antibody is significantly larger than the affibody (150 kDa versus 7 kDa) and is predicted to show reduced tumor penetration ⁵⁸ and incomplete saturation of PD-L1 distant from the vasculature. However, these sites may be accessible by the affibody which is predicted to diffuse further into the tumor bed. This could result in equal or increased observed tumor uptake by the affibody in the blocking experiment. This effect could be further exacerbated by the higher affinity and lower off-rate of atezolizumab relative to the affibody which may lead to peri-vascular localization rather than deep tumor penetration. This effect has been observed in previous studies where reagents with higher off rates and lower affinities are better able to penetrate the tumor compared to high affinity antibodies which stay near the blood vessels ⁵⁹. Finally, in a recent PET-based imaging study using radiolabeled PD-L1 binding peptide WL-12, it was found that maximal PD-L1 receptor occupancy by atezolizumab resulted in only a 2-fold reduction in the %ID/cc of [⁶⁴Cu]-WL12 even at doses as high as 24 mg/kg ⁶⁰. This is a relatively modest blocking effect which may be challenging to resolve in a fully immune competent mouse model (the previous study was carried out in NOD SCID γ mice). These effects may account for the absence of significant antibody-mediated blocking of affibody uptake in the tumor although additional imaging studies are required to confirm this hypothesis.

Conclusion

The primary goals of this work were to validate the affibody scaffold in the context of mRNA display and to develop a high affinity, species cross-reactive ligand for PD-L1. Other folded polypeptide scaffolds have been used in mRNA display, notably Adnectins (fibronectin-based) and single chain variable fragments (scFv) with variable success ⁴². Fibronectin-based scaffolds are only ~10 kDa but require extensive manipulation to utilize them in mRNA display selections. In addition, many of the selected fibronectin ligands are unstable or cannot be robustly expressed in E. coli. The many advantages of affibodies mentioned in the introduction are also coupled with the ease of library production and design. Affibody libraries for selection were produced easily with standard techniques and selected affibodies were readily obtained in high yield in standard E. coli expression systems. This technology should be immediately deployable for any other ICI targets, known and unknown, which is further facilitated by our ectodomain expression/selection system.

Affibody development with mRNA display provides the requisite diversity required to identify low nM target binding ligands. Despite multiple rounds of selection against mouse and human PD-L1, the end diversity of both libraries was still quite high, and the most promising candidate was found to exhibit single-digit nM binding. This suggests that affibodies with even higher PD-L1 affinities are present in the final pools and that this, and other properties (e.g. serum stability, clearance rate) can be further optimized by additional rounds of directed evolution.

The mRNA display selection resulted in complete phenotypic switch from a Her2-specific affibody to a PD-L1-specific affibody. The Her2 affibody scaffold was optimized for a variety of characteristics including E. coli production and solid phase synthesis while maintaining Her2 specificity ⁴⁷. It remained an outstanding question as to whether a library based on this highly optimized sequence would yield a full phenotypic conversion for PD-L1 binding following selection. Fortunately, this fear was unfounded, as all selected clones completely lost detectable affinity for Her2. This selection also produced a dual-specific binding affibody, which was achieved by selection against human and mouse PD-L1. A dual binding reagent can facilitate pre-clinical imaging and therapeutic trials in mouse models prior to moving into primate models and human trials. The dual isoform binding M1 affibody not only has high affinity for mouse and human PD-L1 but seems to bind at the interface of PD-1 and PD-L1. Therefore, M1 affibody could act be used as a non-antibody therapeutic to disrupt PD-L1:PD-1 signaling. If, as we suspect, the affibody will show higher tumor penetration than antibody-based therapeutics, this would be a significant advantage, particularly in highly desmoplastic tumors (e.g. pancreatic tumors).