Engineering a natural ligand-based CAR: directed evolution of the stress-receptor NKp30

Abstract

B7H6, a stress-induced ligand which binds to the NK cell receptor NKp30, has recently emerged as a promising candidate for immunotherapy due to its tumor-specific expression on a broad array of human tumors. NKp30 can function as a chimeric antigen receptor (CAR) extracellular domain but exhibits weak binding with a fast on and off rate to B7H6 compared to the TZ47 anti-B7H6 single-chain variable fragment (scFv). Here, directed evolution using yeast display was employed to isolate novel NKp30 variants that bind to B7H6 with higher affinity compared to the native receptor but retain its fast association and dissociation profile. Two variants, CC3 and CC5, were selected for further characterization and were expressed as soluble Fc-fusion proteins and CARs containing CD28 and CD3ς intracellular domains. We observed that Fc-fusion protein forms of NKp30 and its variants were better able to bind tumor cells expressing low levels of B7H6 than TZ47, and that the novel variants generally exhibited improved in vitro tumor cell killing relative to NKp30. Interestingly, CAR T cells expressing the engineered variants produced unique cytokine signatures in response to multiple tumor types expressing B7H6 compared to both NKp30 and TZ47. These findings suggest that natural CAR receptors can be fine-tuned to produce more desirable signaling outputs while maintaining evolutionary advantages in ligand recognition relative to scFvs.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00262-021-02971-y.

Introduction

Adoptive cell transfer is a rapidly expanding field of immunotherapy showing remarkable success in treating certain cancers. Chimeric antigen receptor (CAR) T cell therapy is a form of adoptive transfer that increases the frequency of tumor-reactive T cells by genetically modifying the patient’s T cells to express engineered CARs that recognize a tumor antigen. Designing safe and effective CARs by achieving the proper balance of promoting tumor elimination while avoiding severe host pathology is an area of intense investigation. Traditional CARs contain an extracellular antigen-recognition domain, hinge domain, transmembrane domain, and intracellular signaling domain(s). The most commonly used antigen-recognition domain is a single-chain variable fragment (scFv) that is constructed from the variable chains of the antigen-binding fragment of an antibody [1]. Natural receptors that recognize stress ligands, such as those expressed by natural killer (NK) cells, are less frequently employed as CARs [2–8]. Thus, it remains unclear to what extent scFv-based and natural receptor-based CAR recognition of tumor antigens differ in their abilities to mount effective tumor immunity. Further, while scFvs are frequently affinity matured and engineered in vitro, there is limited precedence to conducting the same type of directed evolution of a natural receptor for CAR T cell engineering.

There are unique challenges to repurposing tumor-specific scFvs as a CAR-recognition domain. The elimination of stabilizing interactions between CH1 and the light chain, and the requirement for fusion of a non-native linker between the VH and VL domains often results in decreased thermal stability, which is associated with suboptimal scFv folding, expression, and solubility [9]. In addition, some studies suggest that scFv-based CARs are more prone to receptor aggregation that results in tonic signaling and poor CAR T cell persistence in vivo [10–13]. Indeed, recent work has aimed to ameliorate this phenotype by modifying the scFv framework sequence [14] and separating the VH and VL domains into separate constructs [15]. In contrast, native receptors, whose sequences have been honed over evolutionary timescales, appear to be less likely to exhibit these undesirable properties. Thus, engineering limited modifications to natural receptors may allow improved CAR activity without the increased risk of tonic signaling, instability, and potential immunogenicity of scFv-based CARs.

To begin to explore this possibility, we considered a ligand of B7H6, a stress-induced ligand that is expressed in multiple cancer cell types [16, 17] and upregulated in other tissues in rare cases of autoimmunity [18, 19]. B7H6 recognition by the NK cell surface receptor NKp30 initiates NK cell-mediated cytotoxicity [20]. We have previously co-opted NKp30 as a CAR extracellular-recognition domain and demonstrated its ability to induce IFN-γ production in NKp30-CAR T cells when incubated with B7H6-expressing cells and to promote tumor lysis both in vivo and in vitro [4]. In addition to NKp30 CARs, we generated multiple scFv-based CARs that specifically target B7H6 and mediate anti-tumor activity [21, 22]. Interestingly, NKp30 CARs performed similarly to these scFvs in vitro, despite NKp30 having a reported affinity considerably weaker than B7H6-specific scFvs PB6 and TZ47, suggesting that natural receptor-ligand pairs may have been evolutionarily optimized for sensitive ligand recognition [21]. The extent to which natural receptors can be further improved for overall activity or selective activation of a subset of effector functions through directed evolution remains to be determined.

Given the positive association between scFv affinity to its target ligand and improved CAR function [23, 24], we sought to determine whether increasing the natural affinity of NKp30 to B7H6 through directed evolution could improve NKp30-CAR functionality. We report the generation of a new panel of NKp30 variants that bind to B7H6 with increased affinity compared to native NKp30. These variants showed distinct cytokine expression profiles against tumors expressing varying levels of B7H6. Despite more similar killing ability among NKp30-, NKp30 variant-, and TZ47 scFv-based CARs, each CAR elicited distinct cytokine profiles. Most notably, the newly engineered higher affinity NKp30 variants, in some instances, exhibited a decrease in IL-6 expression, which has been associated with cytokine release syndrome (CRS), while exhibiting retained or elevated expression of desirable cytokines. Our findings provide proof-of-concept for fine-tuning the affinity of existing natural receptors to function more optimally in a CAR T cell setting.

Materials and methods

Constructs, protein production, and purification

Soluble monomeric B7H6-His constructs were generated by cloning the extracellular B7H6 sequence into the pCMV expression plasmid, containing both a C-terminal Avi-tag and 6xHis-tag. B7H6-Fc, TZ47-Fc, NKp30-Fc, CC3-Fc, and CC5-Fc were generated by fusing the extracellular domains to a mouse IgG2a Fc domain in pCMV expression vectors. Constructs were verified by Sanger sequencing. Fc-fusion proteins were expressed in expi293 HEK cells (Thermofisher) following the manufacturer’s protocol. Soluble B7H6 with a 6xHis-tag was purified via nickel-charged immobilized metal affinity chromatography (Thermofisher) according to the manufacturer’s protocol; Fc-fusion recombinant proteins were purified from cell supernatants via Protein A resin-based affinity chromatography as previously described [25]. Eluates were then buffer exchanged into PBS using Amicon-30KD ultracentrifugation filters (Millipore).

Directed evolution of NKp30 by yeast surface display

Generation of NKp30 library and re-diversification

The extracellular domain of NKp30 was cloned as a C-terminal Aga2 fusion protein into the yeast expression vector pCHA [26]. The NKp30 extracellular domains in the pCHA vectors were mutagenized by salt-based error-prone PCR [27]. Error-prone PCR was carried out with Taq polymerase (NEB) using the manufacturer’s protocol. The generated PCR products were set up for a reamplification process via PCR to generate enough DNA for the yeast transformation process, again using Taq polymerase, with the same cycling conditions as for the initial error-prone PCR. Yeast libraries were created by electroporation of the reamplified error-prone PCR products and digested pCHA vector. The error-prone PCR products were designed in a way such that the 5′ and 3′ ends of the PCR products shared homology with the digested vector. The digested vector and PCR products were transformed into EBY100 yeast following the electroporation transformation protocol, and cultured and induced as previously described [28]. Library diversity of the resulting g1.0 population was estimated by counting colonies from plated dilutions of the transformed yeast. Plasmids recovered from the g1.4 population using the ZymoPrep Yeast Plasmid Miniprep II Kit (Zymo Research) served as the template for another round of error-prone PCR as described above to introduce two to four random mutations for generation of population g2.0, which was transformed and enumerated as above.

Magnetic-activated cell sorting (MACS) of yeast libraries

Magnetic bead-based sorting of yeast libraries was conducted as described previously [29]. Briefly, B7H6-His was site-specifically biotinylated on the N-terminal avi-tag using the BirA biotin ligase and biotinylation kit, following instructions from the manufacturer’s protocol (Avidity) for capture on Streptavidin beads. Streptavidin-coated magnetic beads (DynaBeads M-270) were washed five times with 0.1% PBSB (PBS + 0.1% BSA) prior to incubating overnight with saturating amounts of biotinylated B7H6-His as determined through titration experiments. After saturation with biotinylated B7H6-His, the beads were washed 5 × with PBSB to remove excess soluble protein prior to co-incubation with yeast libraries for 1.5 h at 4 °C. Each MACS round included one negative selection using bare streptavidin magnetic beads and one positive selection using biotinylated B7H6-His-coated beads. To increase the likelihood of sampling the complete diversity available, a minimum of 10 × the theoretical diversity of each population was employed in all culturing, induction, and selection steps. After each round of selection, beads were washed 1 × with 1 mL PBSB for 30 min at 4 °C to separate bead-bound yeast from those that were trapped between beads (wash fraction). For g1.1 and following, aliquots of bead-bound and washed yeast were counted separately, but populations were regrown and combined in subsequent rounds of selection.

Fluorescence-activated cell sorting (FACS) of yeast libraries

For each round of FACS, at least 10 × the estimated population diversity, or a minimum of 1 million, induced yeast was handled. Cells were spun down at 3000 xg for 3 min and washed three times with 0.1% PBSB to remove cellular debris. Primary incubations comprised of an incubation with chicken anti-cMyc (Gallus) or mouse anti-HA (Invitrogen) to monitor full-length protein expression. Cells were also incubated with biotinylated B7H6-His at concentrations ranging from 1 μM to 1 nM depending upon the stringency and stage of selection, in a total volume of 100 μL. After primary incubation, cells were washed 3 × with 200 μL 0.1% PBSB. Fluorescent staining with secondary reagents included a 30-min incubation with goat anti-chicken conjugated to AF488, Streptavidin-PE, or goat anti-mouse conjugated to AF647 (Life Technologies). After 30 min, the yeast were washed three times with 200 μL 0.1% PBSB before being resuspended in a final volume of 200 μL 0.1% PBSB. The labeled cells were either analyzed on the Miltenyi MACSQuant analyzer or sorted on a Sony iCyt Sorter. The top 0.5–1% of cells falling into a diagonal gate with enhanced B7H6 binding relative to their level of expression were selected using FlowJo.

Structure visualization and manipulation

Receptor alignments were generated using Geneious. The sequence for human NKp30 was retrieved from the Protein Data Bank entry 3NOI. For NKp30 and B7H6 interaction, the protein data bank entry PDB:3PV6 was used. Structure visualizations were performed with UCSF Chimera, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH P41-GM103311[30].

Affinity and avidity assays

BioLayer interferometry for affinity K_D measurements

BioLayer Interferometry (BLI) measurements of affinity K_D were obtained using the ForteBio Octet RED96 system as previously described [25]. Briefly, Protein A-coated biosensor tips were “activated” by incubation with 41 nM TZ47-Fc, NKp30-Fc, or CC3-Fc for 10 min to load tips. Binding to soluble B7H6-His were assessed by allowing TZ47-Fc, NKp30-Fc, or CC3-Fc coated tips to equilibrate in PBS + 0.1% Tween-20 (PBST) for 5 min, followed by association steps in 200 µL of each antigen at concentrations ranging from 0 to 3000 nM for 10 min, and dissociation steps in 200 µL of PBST for 5 min. Data were analyzed using Octet® System Data Analysis software.

Fluorescent staining of B7H6-expressing cells

NKp30-Fc, CC3-Fc, CC5-Fc, or TZ47-Fc binding to B7H6-expressing cell lines was analyzed using a Miltenyi Biotech MACSQuant Flow Cytometer. Three different tumor types were stained with Ig-fusion proteins to determine avidity on varying levels of B7H6 expression. 96-well plates containing 2.5 × 10⁵ cells/well of K562 (ATCC® CCL-243™), A375 (ATCC® CRL-1619™), or Panc1 (ATCC® CRL-1469™) were washed twice with FBS Stain Buffer (BD #554,656) before a primary incubation for 1 h with varying concentrations of Fc-fusion proteins at concentrations ranging from 0 to 500 nM. Cells were then washed 3 times with Stain Buffer (PBS + 2% FBS) before incubation with anti-Mouse Ig antibody conjugated to AF647 for 30 min at room temperature followed by a final wash. Live cell gates were drawn based on FSC vs SSC profiles and were used to calculate mean fluorescence intensity (MFI) values using FlowJo software. For staining of B7H6 expression, K563, A375, and Panc1 were stained with phycoerythrin-conjugated anti-human B7H6 (R&D #FAB7144P). PC3 (ATCC® CCL-1435™) and A549 (ATCC® CCL-185™) cell lines do not express B7H6 and were stained with anti-B7H6 as negative controls. 96-well plates containing 2.5 × 10⁵ cells/well of each tumor cell line were washed twice with FBS Stain Buffer before incubating with anti-B7H6 for 30 min. Cells were then washed three times with Stain Buffer before flow cytometry analysis. Staining was overlayed with unstained samples of each cell type in FlowJo software.

Construction and transduction of CAR T cells

The NKp30 CARs were constructed by PCR amplification of the natural receptor human NKp30, then subcloned using restriction cloning into pFB-Neo retroviral vectors containing the sequence for human CD28 cytoplasmic domain, human CD3ζ cytoplasmic domain, a T2A sequence, and a truncated form of mouse CD19. Vector control constructs were constructed similarly, with truncated mouse CD19 only in the pFB-Neo retroviral backbone. Constructs were verified by Sanger sequencing. NKp30 variants were generated as described above and subcloned into pFB-Neo retroviral vectors using restriction cloning.

Retrovirus production

To package retrovirus, 2.5 × 10⁶ HEK-293 T cells were plated on a 10-cm plate in Dulbecco’s modified Eagle’s media (DMEM) 18 h before transfection. To produce ecotropic virus, cells were transfected using the calcium phosphate method with 10 μg of a B7H6-specific CAR plasmid and 10 μg of ψpcl-eco packaging plasmid. Media was replaced with fresh media ~ 8-h post-calcium phosphate treatment. Viral supernatants were harvested 48 h post-transfection and filtered through a 0.45-μM filter before use for mouse CAR T cells or flash freezing and storing at -80 °C. To make stable cell lines which produce amphotropic virus used for T cell transduction, ecotropic virus made from Hek-293 T cells was used to infect PG13 packaging cells, followed by G418 selection (1 mg/mL) for 5 days. Amphotropic virus was collected, filtered, and used to generate human CAR T cells or stored as described above. All cell lines were cultured under 5% CO₂ at 37 °C.

Human CAR T cells

Human PBMC from a deidentified and coded healthy donor cohort approved by the Dartmouth College Institutional Review Board were activated using anti-CD3 (40 ng/ml; OKT3 Biolegend) and cultured in complete RPMI media plus 100 U/ml of recombinant human IL-2 for 48 h. Complete RPMI media is Hyclone RPMI-1640 media supplemented with 10% heat-inactivated FBS (Hyclone), 10 mM HEPES (Gibco), 0.1 mM non-essential amino acids (Gibco), 1 mM sodium pyruvate (Gibco), 100 U/mL penicillin (Hyclone), 100 μg/mL streptomycin, and 50 μM of 2-mercaptoethanol (Gibco). To prepare for retroviral transduction, non-tissue culture-treated 24 well plates (Greiner Bio-One) were coated with 10 μg/mL retronectin (Takara) for 24 h at 4 °C and washed with PBS. To transduce activated T cells, 900 μL of viral supernatant containing 100 U/mL of IL-2 and 40 ng/ml OKT3 antibody were aliquoted into rectronectin-coated wells prior to addition of 1 × 10⁶ T cells per well. Plates were spinoculated by centrifugation at 1500 xg for 60 min at 30 °C. Plates were then incubated at 37 °C for 24 h. The following day, 500 μL of media was removed from each well and replaced with new viral supernatant containing 100 U/mL of IL-2. Cells were then spinoculated again as above. After 24 h, T cells were resuspended in complete RPMI media containing 100 U/mL and cultured at 0.5–1 × 10⁶ cells/mL for 24 h. Two days post-initial transduction, cells were analyzed by flow cytometry to detect CAR expression before experimental assays were performed.

Mouse CAR T cells

Mouse T cells from the spleens of C57Bl/6 mice were transduced 18–24 h after Concanavalin A (1 μg/ml; Sigma) stimulation and cultured in Complete RPMI media containing 25 U/ml of IL-2. Mouse T cells were transduced with the cytokine expressing backbone by resuspending activated T cells in ecotropic virus and spinning in 24 well plates with 8 million cells per well. Cells were spun at 1500 g for 90 min at 37 °C. Three days post-activation, cells were analyzed by flow cytometry to detect CAR expression before experimental setup.

CAR T cell functional assays

Luciferase-based cytotoxicity

Luciferase-expressing human tumor cell lines A375, Panc1, and K562, and mouse cell lines B16-F10 and RMA with and without B7H6 expression were plated at 5 × 10³ cells per well in a black, tissue culture-treated, and 96 well flat-bottom plate. CAR T cells were added at various T cell effector to target ratios (E:T) of 1:1 and 0.5:1. Cells were co-cultured at 37 °C for 24 h, followed by addition of 50 μL of luciferin (200 μg/mL) (Goldbio), and incubated at 37 °C for 30 min before analyzing luminescence via a Centro LB960 Berthold Technologies luminometer. Supernatant was collected for multiplex cytokine analysis.

Multiplex cytokine array

Levels of Interleukin (IL)-2, IL-4, IL-5, IL-6, IL-10, IL-13, IFN-γ, Granzyme B, Macrophage Inflammatory Proteins (MIP)-1a, and tumor necrosis factor (TNF)-α were determined using a MILLIPLEX MAP Human CD8 + T Cell Magnetic Bead Panel (Millipore) and measured using a Flexmap Luminex instrument and xponent software (version 4.2, Luminex) according to manufacturer’s instructions. Based on previous optimization experiments, supernatant from cytotoxicity assays was analyzed at a 1:4 dilution.

Data analysis

A paired one-way ANOVA using Dunnett’s multiple test correction was performed to analyze differences in target cell killing between NKp30 and other CARs. For analysis of cytokine expression, a paired one-way ANOVA with false discovery controlled using the Benjamini, Krieger, and Yekutieli two state step up method was performed. Statistical analysis was conducted in GraphPad Prism.

Results

Isolation and characterization of yeast-displayed NKp30 variants

Directed evolution of NKp30 variants with enhanced binding affinity to B7H6 was conducted via yeast surface display, employing iterative diversification by error-prone PCR and magnetic and fluorescent cell sorting (Fig. 1a). Neither wildtype NKp30 (Fig 1b) nor the initial library (not shown) showed detectible binding to B7H6-Fc by flow cytometry. Consequently, a bead-based magnetic-assisted cell sorting (MACS) approach that leverages high avidity was used for initial selection, followed by fluorescence-activated cell sorting (FACS) with decreasing concentrations of either bivalent B7H6-Fc or monovalent B7H6-His to gradually increase selection stringency until clones with binding profiles similar to the TZ47 scFv were observed (Fig. 1b-c). Sequencing of selected clones identified a consensus mutation (S82P) at a position defined as a contact residue [31] that was conserved among all sequenced clones (Fig. 2a). Other mutations were observed both within and outside of the B7H6 binding interface (Fig. 2b), and one (T70A) eliminated an N-linked glycosylation motif.

Fig. 1.

Directed evolution of NKp30. a Mutation and selection strategy for engineering NKp30 variants with enhanced binding to B7H6. Error-prone PCR (ePCR), magnetic-activated cell sorting (MACS) for target ligand (B7H6) or lack of binding to bare streptavidin (SA) beads, and fluorescence-activated cell sorting (FACS) against B7H6 were applied with increasing stringency across library generations (g). Population sizes following ePCR diversification are indicated in inset. b. Flow cytometry biplots showing clonal construct or population expression levels and binding of bivalent B7H6-Fc when yeast were stained at the concentration indicated for NKp30, generation 2.0, clone CC3, and TZ47, a mouse-derived antibody fragment. c. Binding of B7H6 relative to expression across a titration of B7H6 for native NKp30, TZ47, selected populations, and clone CC3 when displayed on the yeast cell surface. Error bars represent standard deviation across experimental triplicates from one of three repeated experiments

Fig. 2.

Positions and identities of mutations in engineered NKp30. A. Amino acid sequence alignment of selected clones from generation 2.3 with reported contact residues highlighted in dark orange and N-linked glycosylation sites highlighted in blue. B. Model of the NKp30 (orange): B7H6 (gray) cocrystal structure (PDB:3PV6). The side chains of amino acids that varied among the select sequenced variants are shown (green, stick figure), and the contact interface residues (dark orange), illustrated on the NKp30 ribbon backbone in close up (left) and to view the complete cocrystal (right)

Affinity, avidity, and specificity of soluble NKp30-Fc variants

Clone CC5, which contains only the consensus S82P mutation, along with CC3, which had two additional amino acid substitutions, were chosen for further characterization. To define binding affinities and kinetics with improved resolution, CC3, CC5, NKp30, and TZ47 sequences were expressed recombinantly as Fc-fusion proteins. Affinity to monovalent B7H6 was determined by Biolayer Interferometry (BLI). Both CC3 and CC5 bound to B7H6 with a steady-state affinity (K_D) of 680 nM and 580 nM, respectively, representing a twofold improvement compared to NKp30 (1.3 µM) (Fig. 3a). Interestingly, TZ47 bound B7H6 with a similar affinity (580 nM). However, this scFv exhibited drastically different B7H6 binding kinetics (k_d and k_a) compared to NKp30 and variants, with considerably slower on- and off-rates (Fig. 3b). The fast on and off rate profile of NKp30 was maintained in the variants despite the increase in overall affinity.

Fig. 3.

Binding kinetics and affinities. a. Representative BLI sensorgrams (color) and curve fits (black) of binding of each indicated variant to monovalent B7H6 in solution. Affinities (K_D) and fit qualities (R²) are indicated in inset. b. Kinetic association (k_a) and dissociation (k_d) rate constant biplot

Next, we sought to measure the avidity and specificity of the newly generated NKp30 variants by testing their ability to bind to tumor cell lines expressing varying levels of B7H6. To characterize binding to native membrane-bound B7H6 antigen, three tumor cell lines which naturally express B7H6 at varying levels were stained with soluble NKp30-Fc, CC3-Fc, CC5-Fc, or TZ47-Fc. K562, a human leukemic cell line, expresses a high amount of B7H6 (B7H6^high), while A375, a human melanoma cell line, and Panc1, a human epithelial carcinoma cell line, express a lower amount of B7H6 (B7H6^low) [4] (Supplemental Fig. 1). Despite showing a difference in monovalent affinity when measured with BLI, NKp30-Fc bound to K562 tumor cells with a similar avidity as the CC3-, CC5-, and TZ47-Fc-fusion proteins (Fig. 4a). This result suggests that in the context of a bivalent Fc fusion, the avidity from high antigen expression compensates for low affinity interactions. While it is unclear whether this phenomenon was related to differences in epitope [21], binding kinetics, or other factors, TZ47-Fc was unable to bind to A375 and Panc1 tumor cells, which express low levels of B7H6, whereas NKp30-, CC3-, and CC5-Fc were able to bind to both low B7H6 expressing tumor cell lines (Fig. 4a, b).

Fig. 4.

Tumor recognition and killing. A. Staining of K562, A375, and Panc1 cells with native NKp30, engineered variants, and TZ47, each expressed as a bivalent Fc-fusion protein. B. Representative histograms of staining each target cell line at a 500 nM concentration of fusion protein. C. Killing of each target cell line when co-incubated with primary human T cells transduced to express native NKp30, engineered variants, or TZ47 CD28 + CD3z CARs or a vector control. Reduction in viability (luminescence) of target tumor cells when incubated with varying effector (T cells):target (tumor) ratios is reported relative to that observed when target cells were cultured alone. Data are representative of three independent replicates from different human donors. D. Statistical significance of differences between NKp30 and its variants or TZ47 in cytotoxicity toward each cell line as determined using a one-way ANOVA adjusted for multiple comparisons across all three donors (p < 0.05*, p < 0.005**). Error bars indicate mean and standard deviation of triplicate samples

NKp30 variant CAR T cell cytotoxicity

We next hypothesized that the difference in binding seen in the B7H6^low cell lines may confer a difference in CAR T cell anti-tumor function. CAR constructs composed of NKp30, CC3, CC5, or TZ47-based extracellular domains and the intracellular domains of CD28 and CD3ζ were generated and expressed in primary human T cells from three donors (Supplemental Fig. 2). These human CAR T cells were co-cultured with K562, A375, and Panc1 tumor cell lines expressing luciferase. CAR cytotoxicity was defined by the reduction of luminescence associated with target cell death after 24 h of co-culture (Fig. 4c). While CC3 and CC5 CAR T cells demonstrated similar killing potency as NKp30 against B7H6^high target cells, they demonstrated a modest increase in killing of B7H6^low tumor cells (Fig. 4d). This suggests that engineered natural receptors that bind with higher affinity to their tumor ligands can exhibit improvement in cytolytic activity against targets expressing low tumor antigen. Despite its lack of cell surface staining of low B7H6 cell lines as an scFv-Fc, TZ47 CAR T cells showed similar cytotoxicity as the engineered NKp30 variants, again suggesting the important role of avidity in CAR T cell activity. The killing activity of each CAR construct was confirmed to be antigen-dependent based on additional killing assays conducted using primary mouse T cells and targeting two mouse cell lines (RMA and B16) with and without engineering to express B7H6 (Supplemental Fig. 3).

NKp30 variants drive distinct cytokine profiles

In addition to direct cytolysis, CAR T cells produce an array of soluble effectors including cytokines and chemokines that are important for CAR T cell efficacy. To determine the effect of increased affinity on relative expression of soluble cytokines, we performed a multiplex cytokine assay on the supernatant collected from the killing experiments for each CAR construct and each of the B7H6-expressing tumor cell lines (Fig. 5). Cytokines were grouped into functional, stimulatory, regulatory, and inflammatory categories, essentially as defined by Xue et al.[32]. Among functional cytokines (Fig. 5a), although Granzyme B production was constitutively high in CAR T cells alone, IFNγ, TNFα, and to a lesser extent, MIP-1α, were all induced by co-culture with tumor cells. NKp30 variants generally exhibited higher IFNγ production than either NKp30 or TZ47. For B7H6^low Panc1 and A375 target cells, the differences between NKp30 and its variants were more apparent, with higher MIP-1α, TNFα and IFNγ secreted by the NKp30 variants compared to wildtype. However, for B7H6^high K562 target cells, the variants showed similar or somewhat decreased production of functional cytokines compared to wildtype. This trend was also observed with the stimulatory cytokines, GM-CSF and IL-2 (Fig. 5b), and to a lesser degree with the regulatory cytokines IL-5 and IL-13 (Fig. 5c). In contrast, IL-10 and IL-4 remained largely unchanged with or without tumor cell co-culture. This pattern was not observed with the single inflammatory cytokine tested, IL-6 (Fig. 5d). IL-6 was constitutively expressed by K562 and A375 target cells and varied considerably across different target cell co-cultures, with CAR T cells cultured with Panc1 cells showing limited to no expression, A375 cells inducing high amounts of IL-6 across all CAR constructs, and K562 showing an intermediate phenotype and greater expression induced by NKp30 than either of its engineered variants or TZ47. Collectively, these cytokine expression profiles highlight the role that target cell diversity can play in cytokine expression beyond the impact of CAR-dependent factors. To this end, target cells expressing low levels of B7H6 drove increased expression of cytokines considered to be desirable from T cells expressing the engineered variants relative to NKp30 (Fig. 6). With the exception of the Panc1 cells that had low levels of basal IL-6 expression, this elevation was not observed for the cytokine release syndrome-associated IL-6, suggesting that desirable and undesirable T cell profiles may be separable and that even subtle changes in antigen recognition have the potential to influence soluble mediators of CAR T cell activity.

Fig. 5.

Cytokine secretion profiles of CAR T cells expressing engineered NKp30 variants. a–d. Functional (a), stimulatory (b), regulatory (c), and inflammatory (d) cytokines secreted by primary human T cells transduced to express engineered variants, native NKp30, the TZ47 antibody fragment as a positive control were co-cultured with K562, Panc1, and A375 cells, which express varying levels of B7H6. Cytokine expression of T cells alone and tumor cells alone are provided as controls. Co-culture results presented are geometric mean values of cytokine expression from T cells from three independent donors are presented

Fig. 6.

Statistical significance of differences in cytokine profiles. Adjusted p values resulting from comparisons between each variant and NKp30 by one-way ANOVA across all three donors (p < 0.05*, p < 0.005**, p < 0.0005***)

Discussion

We sought to determine the effect of directed evolution of a natural receptor for improved antigen recognition on in vitro CAR T cell function. A set of single- and double-point mutations to the stress-receptor NKp30 that permitted a high level of antigen binding in the context of yeast display resulted in a twofold increase in binding affinity to recombinant tumor-associated antigen B7H6. This affinity increase was associated with little difference in the ability to bind B7H6 expressed on tumor cells when tested in the context of an Fc fusion. In contrast, some improvement in killing of tumor cells with low but not high antigen expression was observed when tested in the context of a CAR T cell. These observations suggest that high avidity can substitute for high affinity in the context of high antigen expression, but that affinity may play a role in the efficiency of killing target cells with low levels of antigen expression. To this end, it has been shown previously that low affinity interactions can allow for discrimination between tumor cells with high antigen expression and normal tissues with lower expression, thus decreasing on-target, off-tumor toxicity [33]. In instances of high CAR or high tumor antigen expression, avid interactions can overcome low affinity, effectively neutralizing small receptor:ligand affinity differences [34]. Indeed, it has long been theorized that in the case of TCR:MHC-peptide binding, there is an optimal threshold of affinity, and exceeding this threshold does not benefit T cell activation or signaling [35]. However, it has been shown that in some instances CARs with µM affinity can be superior to those with nM affinity [36, 37]. Our data suggest that the native NKp30 affinity to B7H6 is sufficient for cytotoxic functionality in T cells, but can still be modestly improved by modest increases in affinity.

Beyond killing, CAR signal strength has been shown to be a key determinant of T cell fate [38]. To this end, we measured a panel of cytokines in order to determine if tumor epitope and affinity differences could affect functional outputs in T cells. We found notable differences in cytokine output between NKp30 and its variants. While no single cytokine signature is yet predictive of in vivo outcomes, evidence for some relationships has accrued. For example, IL-2 is well characterized as a T cell growth factor and CARs expressing higher levels of IL-2 are likely to survive in the tumor microenvironment [39, 40]. We found higher affinity of NKp30 variants to B7H6 led to higher production of IL-2. We also observed increases in induction of TNFα and IFNγ, both of which have been shown to correlate with delayed tumor relapse in vivo [41], in CC3 and CC5 compared to wildtype in B7H6^low tumor cell lines. In contrast, expression of IL-6, a cytokine found to be predictive of severe cytokine release syndrome (CRS) [42, 43], was not robustly increased for engineered variants as compared to NKp30. However, IL-6 has also been found to be produced by myeloid cells in CRS; in vivo experiments in models capable of observing this toxicity [44] are needed to investigate whether these constructs impact this aspect of IL-6-mediated CRS. These results suggest that while a modest increase in affinity had a small effect on binding and killing functionality, it resulted in more marked differences in cytokine output. Most notably, these differences in cytokine output were most distinct with the B7H6^low cell lines, underscoring that increased affinity of NKp30 may be more advantageous against tumors with low B7H6 expression.

Notable distinctions between NKp30-based recognition of B7H6 and that of the TZ47 scFv were observed. Though this antibody fragment has a similar equilibrium binding affinity as the engineered NKp30 variants, it recognizes a different epitope. Whereas NKp30 binds to a distal site on B7H6, TZ47 recognizes a membrane proximal site [21], which has been reported to enhance the anti-tumor activity of CARs [45]. An Fc fusion of TZ47 failed to stain two tumor cell lines with low B7H6 expression; NKp30-based Fc fusions reliably stained these targets. This difference raises the possibility that the NKp30 epitope on B7H6 may be more favorable or accessible to binding in some contexts, and that bivalent binding may be achieved by recognition of some but not other epitopes of B7H6 when expressed at low levels. For a number of tumor-associated antigens, recognition of cells with low expression can lead to on-target, off-tumor toxicity. While it is possible that effective stimulation of CAR T cells by low levels of B7H6 could increase the risk of toxicity, this concern is tempered by its biological role as a stress marker and the absence of evidence of expression on healthy cells. Despite the lack of binding of bivalent TZ47-Fc, a TZ47-based CARs showed excellent tumor killing, in some cases exhibiting the greatest cytotoxicity. This observation, too, points to the complex interplay of affinity, avidity, and epitope specificity in tumor recognition and clearance,

As compared to scFvs, which can be isolated in principle against any tumor antigen epitope, and are often affinity matured in vitro [46], natural receptors bind to established epitopes which have been evolutionarily selected for functional ligation, but are typically treated as modular CAR extracellular domain units and not modified in sequence. Here, by evolving a natural receptor for use in CAR platforms, some limitations, such as the instability and tonic signaling that can be associated with scFvs may be avoided, while advantages of natural receptor recognition could be maintained. One of these advantages may be the fast on- and off-rates we observed for NKp30 and its variants, which are in stark contrast to the approximately 100-fold slower on and off-rates observed for TZ47, a characteristic most CAR scFvs exhibit [47]. Despite similar equilibrium affinity and somewhat poorer in vitro cytotoxicity, engineered NKp30 variants showed elevated expression of diverse functional and stimulatory cytokines as compared to TZ47, especially when B7H6 expression on target cells was low. These results suggest that natural receptor- and scFv-based CARs that bind the same tumor antigen with similar affinity can produce different signaling outputs, allowing for the potential to fine-tune this parameter during CAR domain development. Whether an optimal activity has been reached, or whether continued increases in affinity would result in further improvement in CAR T cell function is not known, but is of interest given the complex relationships known to exist between affinity and activity. Additionally, further studies are necessary to investigate the relevance of these phenotypes to tumor clearance in vivo.

Like other natural receptors, NKp30 has been reported to interact with multiple ligands. Beyond B7H6, NKp30 is known to interact with several other stress-induced ligands. While we cannot exclude that these ligands play a role in the phenotypes that we have observed, at least one, BAT3 (BAG6) is reported to interact with a different site of NKp30, and is therefore not expected to show the same enhancement of binding. Unlike B7H6, BAT3 is not a well characterized tumor antigen, and it’s ligation has been reported to be alternatively immune stimulating or immune inhibitory [48, 49]. Whether or not the potentially promiscuous binding of B7H6 to other NK cell ligands is beneficial or detrimental is unknown. Future investigation into this area could include mutational studies to eliminate interactions with receptors to which binding is undesirable.

In sum, we observed the impacts of modifying the affinity of tumor recognition can extend beyond killing to drive differential cytokine expression profiles suggesting that further engineering can be attempted to improve CAR therapy to more broadly influence cytokine, chemokine, or regulatory interactions. These observations suggest the value of thorough exploration of cytokine profiles in other contexts. Reducing or enhancing interactions to fine-tune both recognition of target cells and manipulate signaling has the potential to significantly affect the outcome of cell-based therapies. As molecular engineering strategies are deployed to improve upon evolved interactions, we anticipate that additive advantages can be combined to propel future therapeutics forward.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

B7H6

B7 homolog 6

BLI

Biolayer interferometry

CAR

Chimeric antigen receptor

CRS

Cytokine release syndrome

FACS

Fluorescence-activated cell sorting

Fc

Fragment crystallizable

IL

Interleukin

MACS

Magnetic-activated cell sorting

MFI

Median fluorescence intensity

MHC

Major histocompatibility complex

MIP

Macrophage inflammatory protein

NK

Natural killer

scFv

Single-chain fragment variable

TNF

Tumor necrosis factor

Authors' contributions

SEB, RAB, and CHC performed experiments. MEA, YHH, and CLS supervised experiments. SEB and RAB drafted and all authors reviewed and revised the manuscript.

Funding

This work was supported in part by the National Institutes of Health T32-AI007363 to SEB and RAB, the Munck-Pfefferkorn Novel and Interactive Grant and Norris Cotton Cancer Center Prouty Developmental Grant to MEA and YHH, Tom and Susan Stepp to YHH, and P30-CA023108, which supports the Norris Cotton Cancer Center’s flow cytometry cores.

Availability of data and materials

Source data in numerical form is available upon request.

Code availability

Not applicable.

Declarations

Conflicts of interest

Authors are co-inventors on a patent application on the affinity maturated NKp30 mutants. CLS and MEA are inventor on other patents and patent applications related to targeting NKp30 and/or B7H6. This work is managed in compliance with the policies of Dartmouth College.

Ethics approval

These studies were reviewed and performed with the approval of the Dartmouth College Institutional Review Board (IRB).

Consent to participate

No identifiable subject information has been published that warrants individual consent, which was waived as per the regulations covered under IRB.

Consent for publication

Authors consent to publication of this original manuscript.