Pooled screening for CAR function identifies novel IL-13Rα2-targeted CARs for treatment of glioblastoma

Abstract

Background

Chimeric antigen receptor (CAR) therapies have demonstrated potent efficacy in treating B-cell malignancies, but have yet to meaningfully translate to solid tumors. Nonetheless, they are of particular interest for the treatment of glioblastoma, which is an aggressive form of brain cancer with few effective therapeutic options, due to their ability to cross the highly selective blood-brain barrier.

Methods

Here, we use our pooled screening platform, CARPOOL, to expedite the discovery of CARs with antitumor functions necessary for solid tumor efficacy. We performed selections in primary human T cells expressing a library of 1.3×10⁶ third generation CARs targeting IL-13Rα2, a cancer testis antigen commonly expressed in glioblastoma. Selections were performed for cytotoxicity, proliferation, memory formation, and persistence on repeated antigen challenge.

Results

Each enriched CAR robustly produced the phenotype for which it was selected, and one enriched CAR triggered potent cytotoxicity and long-term proliferation on in vitro tumor rechallenge. It also showed significantly improved persistence and comparable tumor control in a microphysiological human in vitro model and a xenograft model of human glioblastoma, but also demonstrated increased off-target recognition of IL-13Rα1.

Conclusion

Taken together, this work demonstrates the utility of extending CARPOOL to diseases beyond hematological malignancies and represents the largest exploration of signaling combinations in human primary cells to date.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Chimeric antigen receptor (CAR)-T cell therapies, while transformative for the treatment of B-cell malignancies, remain ineffective in treating solid tumor diseases; all solid-tumor targeting CAR-T cell therapies used in the clinic thus far have relied on the same signaling machinery to elicit antitumor functions as those used to treat blood cancers, despite their vastly different disease presentation.

WHAT THIS STUDY ADDS

• This study explores CAR signaling as it relates to cytotoxicity, proliferation, and memory formation in human primary T cells in the context of targeting a commonly expressed solid tumor antigen, IL-13Rα2. In doing so, we discovered CAR signaling that produces enhanced cytotoxicity in vitro and substantially improved persistence both in vitro and in vivo against human glioblastoma, a feature that is likely necessary to combat antigen-positive disease recurrence. While its off-target activity precludes it from clinical consideration without further optimization, we demonstrate the utility of CARPOOL in discovering CARs that elicit improved function.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• We posit that translating the success seen using CAR-T cell therapies in treating B-cell malignancies to solid tumors and beyond will require precise tuning of both the signaling architecture and structural elements of the CAR in physiologically relevant conditions; our CARPOOL platform is ideally suited for such a task and this study demonstrates the largest exploration of signaling function to date in human primary T cells.

Background

Glioblastoma (GBM) is the most common form of primary malignant brain tumor and has a very poor prognosis, with the current standard of care offering only marginal survival benefit.¹ There are very few therapeutic options for this disease, owing in part to the tightly regulated blood-brain barrier (BBB) impeding drug trafficking.2,4 T cells are one of the few cell types that are capable of trafficking across this barrier in order to carry out immune surveillance.^{5 6} This, along with T cells’ ability to establish persistent memory populations and to trigger epitope spreading against multiple antigens with heterogenous expression in coordination with the rest of the endogenous immune system,7,16 makes chimeric antigen receptor (CAR) T-cell therapies a uniquely promising treatment strategy for GBM. However, while CARs have been transformative in the treatment of relapsed and refractory B-cell acute lymphoblastic leukemia (B-ALL), large B-cell lymphoma, and multiple myeloma (MM), as evidenced by the Food and Drug Administration (FDA) approval of 6 different CAR-T cell products targeting CD19 or B cell maturation antigen (BCMA) for these indications, they have not yet shown evidence of long-term durable responses in the context of solid tumors.17,21 This is largely due to the unique challenges posed by solid tumors, including tumor antigen heterogeneity, barriers to physical trafficking and infiltration into the tumor, and immunosuppression within the tumor microenvironment that impedes T-cell persistence and antitumor function. Accounting for these factors in CAR design and discovery may inform better design principles and aid in the translation of solid tumor-targeted CARs.

CARs redirect native T-cell function towards cells expressing a target antigen of interest by linking extracellular recognition domains to intracellular immunostimulatory signaling domains, most commonly derived from 4-1BB and/or CD28 and CD3ζ, as is the case for CD19-targeted and BCMA-targeted therapies. For GBM-targeted CARs, target antigens that have been pursued in the clinic include EGFRvIII, HER2, IL-13Rα2, B7-H3, CD147, GD2, and chlorotoxin, while others such as CSPG4 and CAIX have been reported in the literature.22,30 Additional targeting approaches have been shown to enhance function and overcome antigen heterogeneity, including implementing IF/OR gated logic and secreting bispecific antibodies.^{15 31} Strikingly, ~95% of CARs under consideration in the clinic, including those that target solid tumors, share the same signaling domains as those employed against hematological malignancies.³²

While effective for treating blood cancers, we posit that these signaling domains may induce suboptimal responses in the context of solid tumors, which may require different effector programs from those employed against B-cell malignancies given their distinct immunological challenges. We propose that a more extensive exploration of signaling programs may unlock useful T-cell functions and expedite the translation of CAR-T cell therapies to solid tumors. Until recently, large scale signaling explorations employing selections for function in mammalian cells have been difficult. Within the last few years, several groups, including our own,³³ have pioneered pooled screening strategies in which immune cells are engineered to express a CAR and tagged with an untranslated DNA barcode, subjected to antigen exposure, and screened for a phenotype of interest, after which the barcode frequencies can be readout via next generation sequencing (NGS).34,38 Collectively, we refer to this approach as CARPOOL. While other groups have employed this technique on the scale of hundreds to thousands of library members,34,38 these libraries are largely composed of a limited set of co-stimulatory domains in combination with CD3ζ, only representing a fraction of the comprehensive diversity of signaling inputs and combinations that could be useful in eliciting antitumor function.

Furthermore, while our initial demonstration of CARPOOL used Jurkat cells due to the ease with which they can be engineered and expanded, primary T cells offer a broader and more nuanced range of clinically relevant phenotypes available for use as selection criteria. Thus, we speculated that designing our screens to select for antitumor function in primary T cells may enable the discovery of effective solid-tumor targeted CARs. To test this theory, we employed a 1.3×10⁶-member library of third generation CARs targeting IL-13Rα2 and expressed in human primary T cells, from which we identified a novel CAR that elicits potent antitumor cytotoxicity and long-term persistence following selections for cytotoxic degranulation, persistence, and memory formation—all of which are expected to be critical features in targeting solid tumors. To our knowledge, this work represents the largest exploration of CAR signaling in human primary T cells described to date.

Methods

Cell lines

HEK293T (CRL-3216) and Clone E6-1 Jurkat (TIB-152) lines were purchased from American Type Culture Collection (ATCC), while the U87 MG (HTB-14) line was generously gifted by Professor Forest White’s laboratory and transduced to stably express human IL-13Rα2 with an N-terminal Flag tag along with firefly luciferase (FLuc) via an internal ribosome entry site (IRES) sequence, then sorted for Flag and IL-13Rα2 expression to produce the IL-13Rα2⁺ U87 cell line (online supplemental figure 4F). U87 MG IL-13Rα2⁺ RFP⁺ cells were generated by stable expression of human IL-13Rα2 and RFP (Addgene plasmid #26001) and sorting for RFP and IL-13Rα2 expression. Target Jurkat T-cell lines were similarly produced by transduction to stably express IL-13Rα1 or IL-13Rα2 and sorting for antigen expression. Similarly, mCherry was introduced into these lines through transduction and sorting for fluorescence. HEKs and U87-derived cell lines were maintained in Dulbecco's Modified Eagle Medium (DMEM) (ATCC) supplemented with 10% fetal bovine serum and 100 U/mL penicillin-streptomycin (Corning), while Jurkats were maintained in Roswell Park Memorial Institute (RPMI) 1640 medium (ATCC) supplemented with 10% fetal bovine serum and 100 U/mL penicillin-streptomycin (Corning). Human brain microvascular endothelial cells were obtained from Angio-Proteomie (cAP-0002), maintained in EGM-2 MV BulletKit, and used in passage 6 (Lonza, CC-3156 and CC-4147). Primary human brain pericytes were obtained from Cell Systems (ACBRI 498), maintained in The System (Cell Systems, CSS-A101), and used in passage 6. Human astrocytes were obtained from ScienCell (1800), maintained in Astrocyte Medium from ScienCell (1801), and used in passage 4. Cell attachment factor (Cell Systems) was used to coat flasks for culturing of brain microvascular endothelial cells, pericytes, and astrocytes. Cell lines were routinely Mycoplasma-tested using the MycoAlert PLUS Mycoplasma Detection Kit (Lonza).

Plasmid construction

The plasmid pHIV-EGFP was gifted by Bryan Welm and Zena Werb (Addgene plasmid #21373), and pMD2.G and psPAX2 were gifted by Didier Trono (Addgene plasmid #12259 and #12260). To generate second-generation IL-13Rα2 CAR-EGFP plasmid, a codon-optimized gene encoding IL-13Rα2 CAR composed of E13Y mutated IL-13 cytokine, L235E and N297Q-mutated IgG4 hinge, CD4 transmembrane domain, and intracellular domains (ICDs) derived from human 4-1BB and CD3ζ was PCR amplified from geneblocks purchased from IDT and cloned into the third-generation lentiviral vector pHIV-EGFP using Gibson Assembly. To generate a backbone vector for the CAR plasmid library, the intracellular signaling domains of the IL-13Rα2 CAR-EGFP plasmid were replaced with the LacZ gene flanked by BsmBI restriction sites. The signaling-diversified CAR plasmid library was generated by PCR amplification of each ICD (online supplemental table 1) at each of the three positions,³³ with the forward and reverse primers adding unique linkers for each position. These products were then pooled at equimolar ratios for each position and combined with a pool of randomized 18-mer barcode sequences. These were then inserted into a pGGA vector using BsaI restriction sites, with the insert also incorporating flanking BsmBI restriction sites. The final lentivirus plasmid library was generated by ligating the pGGA ICD library into the IL-13Rα2 CAR backbone vector at the BsmBI restriction enzyme sites to replace the LacZ gene via Golden Gate Assembly, which has lower proclivity to recombination due to its ability to ligate 4 bp segments of homology and lack of reliance on exonuclease activity and long regions of homology, as is the case with conventional methods such as Gibson Assembly. Final products were electroporated into recombinase (recA1) deficient DH10ß electrocompetent Escherichia coli cells (Thermo Scientific, EC0113), cultured overnight at 32°C to reduce recombination, and purified to achieve a highly diverse plasmid library at~11.8×coverage of the theoretical diversity.

Lentiviral production

Lentiviruses were generated by first transfecting 70% confluent HEKs with transfer plasmid, pMD2.g (VSVg), and psPAX2 combined at a plasmid mass ratio of 5.6:1:3 that was complexed with polyetherimide (PEI) at a DNA:PEI mass ratio of 1:3. For a confluent T225 flask, 42 µg of transfer plasmid was used for transfection. The medium was changed 3–6 hours after transfection, and lentiviral particles were collected in the supernatant 48–96 hours after transfection. The supernatant was then filtered through a 0.45 µm low-protein-binding filter, and centrifuged for 1.5 hours at 100,000 g. The pellet was then resuspended in serum-free OptiMEM overnight at 4°C and stored at −80°C.

Human T-cell activation, transduction, and expansion

Research using de-identified human blood was conducted as per Massachusetts Institute of Technology (MIT) Committee on the Use of Humans as Experimental Subjects policies for exempt research. Peripheral blood mononuclear cells from healthy donors were purified from leukopaks purchased from STEMCELL Technologies using EasySep Direct Human PBMC Isolation Kits (STEMCELL Technologies) as per the manufacturer’s instructions. Primary CD4⁺, CD8⁺, or CD3⁺ T cells were isolated using EasySep Human CD4⁺, Human CD8⁺, or Human T Cell Enrichment Kits (STEMCELL Technologies) and cultured in RPMI 1640 (ATCC) supplemented with 10% fetal bovine serum, 100 U/mL penicillin-streptomycin (Corning), 30 IU/mL recombinant human IL-2 (R&D Systems), and 50 µM β-mercaptoethanol (Fisher). Before transduction, T cells were activated using a 1:1 ratio of DynaBeads Human T-Activator CD3/CD28 (Thermo Fisher) for 24 hours, after which 8 µg/mL of dextran (Sigma) and concentrated lentivirus were added to the culture at a multiplicity of infection of five for single lentiviral constructs and two for pooled library encoding lentivirus, where human primary CD4⁺ T cells were transduced at an efficiency of~10% to limit multiple integration events. Transduced cells were then sorted for Enhanced Green Fluorescent Protein (EGFP) expression. In the case of CAR library production, this was performed for two biological replicates using different donor Peripheral Blood Mononuclear Cells (PBMCs) to source CD4⁺ T cells. After 3 days, DynaBeads and lentivirus were removed and cells were sorted for EGFP using a BD FACSAria II. Cells were rested for 4 days before characterization and maintained at a density of 5×10⁵ to 2×10⁶ cells/mL throughout.

Flow cytometry and cell sorting

Cells were washed with 1×phosphate-buffered saline (PBS) (Sigma) supplemented with 0.5% bovine serum albumin (Research Products International) and 2 mM EDTA, then surface stained by incubating with antibodies for 15 min on ice. They were subsequently washed again before flow analysis on a BD Accuri C6 or Beckman CytoFLEX S or cell sorting with a BD FACSAria II or Sony MA900. Anti-IL-13 (clone JES11-5A2), anti-CD213a2 (clone SHM38), anti-Flag (clone L5), anti-CD4 (clone SK3), anti-CD8 (clone SK1), anti-CD3 (clone OKT3), anti-CD107a (clone H4A3), anti-PD-1 (clone EH12.2H7), anti-TIM3 (clone F38- 2E2), anti-LAG-3 (clone 11C3C65), anti-CD62L (clone DREG-56), anti-CD45RA (clone HI100), and anti-CD69 (clone FN50) antibodies were purchased from BioLegend. For flow cytometry analysis of CAR-T cells isolated from microtumors, CAR-T cells in the side channels were flushed out of the devices and collected before microtumors were removed from microfluidic devices and incubated in a digestion mix consisting of 0.25% Trypsin, Nattokinase, collagenase IV, and DNAse for 30 min at 37°C, pipetting every 15 min. Liberated cells were then spun down and resuspended in 1×PBS (Sigma) supplemented with 0.5% bovine serum albumin (RPI) and 2 mM EDTA, and stained accordingly. A BD FACSymphony A5 Cell Analyzer was used for acquisition. Anti-4-1BB (clone 4B4-1, BioLegend), anti-CD107a (clone H4A3, BioLegend), anti-IFN gamma (clone 4S.B3, eBioscience), anti-TNF-alpha (clone Mab11, BD), anti-CD62L (clone SK11, BD), anti-TIM-3 (clone 7D3, BD), anti-LAG-3 (polyclonal goat IgG, R&D Systems), anti-PD-1 (clone EH12.1, BD), and fixable viability dye eFluor 455UV (eBioscience) antibodies and dyes were used for staining. Data were exported and analyzed with FlowJo V.10.10.0.

CAR-T functional selections

In preparation for selections, 1.4×10⁸ human primary CD4⁺ T cells were transduced with lentivirus at a multiplicity of infection of two with 8 µg/mL dextran (Sigma). Virus and DynaBeads were removed after 3 days of transduction, and the cells were sorted for EGFP, with~10×library coverage, which was calculated based on the theoretical maximum diversity from the previous round, being maintained throughout. For a round of selection, cells were stimulated with IL-13Rα2⁺ Jurkat cells which were previously incubated for 1 hour in 50 µg/mL mitomycin at 37°C, then stained for 4-1BB after 24 hours or CD107a expression after 6 hours. For CD107a staining, cells were stimulated in the presence of Brefeldin (BD 555029), monensin (BD 554724), and anti-CD107a antibody (clone H4A3) per the manufacturer’s instructions. The 4-1BB⁺ cells, gated relative to the unstimulated library as a negative control, were collected after the first round of stimulation. After the second round, the top 5% of CD107a expressing T cells by mean fluorescent intensity were sorted on a BD FACS Aria II. Cells were then rested without antigen and expanded for 4 days before subsequent rounds of selection. After the final round of stimulation, 30–50% of cells were stained for CD62L and CD45RA and sorted for memory populations while the remaining unsorted cells were frozen for NGS. NGS sequencing data was deconvoluted and analyzed using a custom package called DomainSeq, as described in the manuscript.

In vitro rechallenge assay

1×10⁶ CAR-transduced CD4⁺ T cells were co-cultured with mitomycin-treated target U87 cells, prepared by incubation in 50 µg/mL mitomycin for 1 hour at 37°C, at an effector to target (E:T) ratio of 1:1 in interleukin (IL)-2 deficient medium in triplicate. Every 2–3 days, approximately 5% of the culture volume was taken out for flow analysis for EGFP following the addition of CountBright Plus Absolute Counting Beads (Thermo Fisher C36995). Then, 1×10⁶ CAR-T cells were taken out from the original culture and re-plated with a fresh batch of mitomycin-treated U87 cells at a 1:1 E:T ratio. CAR-T cells were sampled for single-cell RNA sequencing (scRNA-seq) analysis on day 14, which was 48 hours following the fourth U87 challenge. On day 13, cells were also stained for memory markers (CD62L and CD45RA) and exhaustion markers (Programmed cell death protein 1 (PD-1), T-cell immunoglobulin and mucin domain 3 (TIM3), and Lymphocyte activation gene 3 (LAG-3)).

Cytotoxicity assays

U87 cells expressing FLuc were co-cultured with CD4⁺ or CD8⁺ T cells for 24 hours in an IL-2-deficient medium at various E:T ratios. Cells were then collected and washed before cell lysis and the addition of luciferin substrate from the Bright-Glo Luciferase Assay System (Promega). The resulting luminescent signal was measured using a Tecan Infinite M200 Pro. Signals were normalized to negative controls containing only target cells. For IncuCyte experiments, human primary CD3⁺ CAR or mock T cells were co-cultured with Jurkats expressing mCherry and either IL-13Rα1 or IL-13Rα2 at varying E:T ratios in IL-2-deficient medium. The red object area was subsequently monitored over time to assess target cell growth.

Cytokine secretion assay

Following 24 hours of stimulation of human primary CAR-T cells with U87 cells, polyfunctional cytokine and chemokine secretion profiles in response to tumor challenge were determined using the 41-plex MILLIPLEX MAP Human Cytokine/Chemokine Magnetic Bead Panel from Miltenyi and measured on a Luminex FLEXMAP 3D system.

PacBio and Illumina sequencing

Genomic DNA from selected cells was purified using the PureLink Genomic DNA Mini Kit (Thermo Fisher). For PacBio sequencing, PCR amplicons encoding CAR signaling domains and barcode regions were attached with SMRTbell adaptors using the SMRTbell Template Prep Kit V.1.0 (Pacific Biosciences) and sequenced using a PacBio Sequel system. For short read (Illumina) sequencing, barcode regions were PCR amplified to conjugate P5 and P7 adaptor sequences and sequenced on an Illumina NextSeq 500 or Element AVITI system.

Single-cell sequencing

Primary human T cells expressing CARs 3, 4, and 13BBz were sampled from U87 co-cultures following four tumor re-challenges, as described above. Samples were separately stained with a panel of CITE-seq antibodies (BioLegend) specific for a panel of T-cell phenotypic markers, described previously,³⁹ along with three unique hashing antibodies to label each CAR sample (BioLegend). Following staining and washing according to the manufacturer’s protocols, live CAR⁺ cells were enriched from each sample on the basis of viability dye staining and Green Fluorescent Protein (GFP) expression. Sorted samples were then pooled at approximately equal proportions, then encapsulated in two channels via Chromium Next GEM Single Cell 5’ Kit V.2. Gene expression (GEX) and feature barcoding (FB; for CITE-seq antibodies) libraries were constructed based on the manufacturer’s instructions, then pooled to achieve approximately 12.5% FB and 87.5% GEX, before sequencing on an Illumina NextSeq500 to a depth of 34,380 reads per cell. Reads were aligned to the Genome Reference Consortium Human Build 38 (GRCh38), and a cell-gene matrix was generated from the aligned reads using the CellRanger software (10x Genomics; V.7.0.1). Analysis of this matrix was performed with the Seurat package (V.4.3.0.1; in R V.4.2.2).^{40 41} To filter out low-quality or dying cells, cells with >5% of reads relating to mitochondrial genes were removed, as well as any cells with less than 1,000 unique genes detected. Filtered cells were then assigned sample identity using hashtag antibody reads via the HTODemux algorithm⁴¹; only cells confidently assigned to a single sample were used in downstream analysis. Although we aimed to encapsulate approximately equal numbers of cells from each sample, the data showed unequal cell calls; to address this, we downsampled cells to gain approximately equal cell depth per CAR, yielding 1471, 1471, and 1194 cells expressing CAR 3, 4, and 13BBz, respectively. This downsampled data set was normalized and scaled, using the SCTransform function in Seurat. Linear dimensionality reduction was then performed on both GEX and FB data, followed by weighted nearest neighbors clustering on the basis of the first 35 Principal components (PCs) for GEX and the first 10 PCs for FB using a resolution of 0.5, allowing identification of distinct cell states while incorporating both transcriptional and protein-level data.⁴² Cell cycle scores were calculated using the CellCycleScoring function in Seurat. Differentially expressed genes were calculated by the Wilcoxon rank-sum test. Χ² analysis to determine sample enrichment within each cluster was performed via the chisq.test R function.

In vitro model generation and CAR-T assay

Microtumors were generated according to Lam et al.⁴³ Briefly, 1.5×10⁴ U87 MG IL-13Rα2⁺ RFP⁺ cells were seeded in a hanging drop and allowed to form dense spheroids over 4 days before resuspending in a fibrin gel with a mix of human primary brain endothelial cells, pericytes and astrocytes, and injected into the central well of the OrganiX microfluidic device (AIM Biotech). Over 7 days, a perfusable vasculature formed around the tumor spheroid, and the tumor cells invade into the perivascular space. 1×10⁴ transduced CAR 4 or 13BB𝜁 were introduced into one side channel of the device, and microtumors with no CAR-T were used as control. 5 IU/mL IL-2 was supplemented in the media, and 50 µL of media was topped up every 3 days. Tumor spheroids and CAR-T infiltration were imaged on a Zeiss LSM880 inverted microscope with an environmental chamber and a 10×lens just before CAR-T addition (day 0) and on days 3, 6 and 9. Images were imported into FIJI and radial signal intensities were measured using the ImageJ built-in Radial Plot plugin.

Xenogeneic mouse models

All animal studies were performed in accordance with guidelines approved by the MIT Division of Comparative Medicine and MIT Committee on Animal Care (Institutional Animal Care and Use Committee, protocol number 0621-032-24). Male NOD/SCID/IL-2Rnull (NSG) mice were purchased from Jackson Laboratory and housed in the animal facilities at MIT. For in vivo studies, 6–8-week-old mice weighing between 24 and 31 g were injected subcutaneously on the right flank with 2×10⁶ IL-13Rα2⁺ U87 cells, followed 21 days later by intravenous administration of 1×10⁶ CD3⁺ CAR-T cells or untransduced T cells, which were prepared as described above, via the tail vein. Tumor progression was subsequently monitored every 2–3 days using caliper measurement and the IVIS Spectrum imaging system (PerkinElmer) to measure bioluminescent signal after intraperitoneal administration of 0.15 mg of luciferin substrate per gram of body weight (PerkinElmer 122799). Total photon counts were quantified using Living Image software. Mice were weighed every 2–3 days for signs of toxicity. Mice were euthanized on observing signs of discomfort, morbidity, or limited mobility, or on tumor area reaching 100 cm² for untreated mice and 300 cm² for CAR-treated mice.

Statistical analysis

Statistical analyses were performed using the Prism (V.10) software, with the exception of the single-cell sequencing data, which were analyzed in R Studio using base packages or those described above. For microphysiological in vitro models, signal intensities were exported to Prism (V.10.1.1) and the area under the curves were calculated. Sample sizes were not predetermined using statistical methods. For statistical comparisons between multiple groups vs control, significance was determined using a two-way analysis of variance. Adjusted p values<0.05 after multiple hypothesis correction with Dunnett’s multiple comparisons correction were considered statistically significant. The statistical test used for each experiment is noted in the relevant figure legend.

Results

Library construction

A second generation 4-1BB-based zetakine CAR targeting IL-13Rα2 (which will be referred to as 13BB𝜁) has been reported to produce complete remission in patients with recurrent multifocal lesions that were heterogenous in antigen expression, with three experiencing transient complete remission.^{11 44} The target antigen IL-13Rα2 is a high-affinity IL-13 decoy receptor and cancer testis antigen that drives tumor invasiveness and is commonly overexpressed in GBM in addition to a variety of other cancer types, including pancreatic cancer, breast cancer, and melanoma, while its endogenous expression is largely restricted to the testis.45,50 Given its demonstrated clinical efficacy, IL-13Rα2 was selected as the target antigen against which we chose to employ our CARPOOL selection strategy.

To target IL-13Rα2, we constructed a third generation CAR library that uses a tethered IL-13 cytokine with an E13Y mutation (to abrogate binding to IL-4Rα, the co-receptor for the more commonly expressed IL-13 receptor, IL-13Rα1) as the extracellular targeting moiety, as described previously.^{11 51 52} The CAR also included an IgG4 hinge with mutations in the Fc binding region (L235E, N297Q) and a CD4 transmembrane domain,^{11 53} after which the intracellular signaling domains were inserted. This was performed using Golden Gate assembly to shuffle 110 different signaling domains into three intracellular positions within the CAR at random, yielding a theoretical library size of 1.3×10⁶ unique CARs—each of which included an 18-nucleotide barcode in the 3’ untranslated region followed by an IRES and EGFP sequence. To unbiasedly explore a large space of potential signaling inputs, the receptor signaling domains were derived from the same pool that was described previously,³³ with the addition of viral immunoreceptor tyrosine-based activating motifs (ITAMs) and cytokine receptor signaling domains (online supplemental table 1). Viral ITAMs have evolved to enhance the propagation of infected host cells and are capable of initiating activation, proliferation, and differentiation, all of which are known requirements for CAR signaling.⁵⁴ Additionally, cytokine receptor signaling is one of three signals required for T-cell activation and plays an important role in directing T-cell differentiation, expansion, and persistence⁵⁵; indeed, one study showed that incorporating truncated IL-2Rβ and STAT3 signaling could enhance proliferation, polyfunctionality, and cytotoxicity, and reduce terminal differentiation.⁵⁶ Figure 1A shows the architecture and composition of this library.

Figure 1. CARPOOL screening strategy. (A) Construct design for 1.3×10⁶-member library of IL-13Rα2 targeted third generation CARs. (B) Rechallenge and sorting timeline for CAR enrichment. CARs were stimulated at each time point with growth-arrested IL-13Rα2⁺ Jurkat T cells at a 1:1 effector to target ratio in 30 IU/mL IL-2 with or without 5 ng/mL TGF-β supplementation. Active CARs were sorted by collecting EGFP⁺ 4-1BB⁺ CAR-T cells, gated relative to unstimulated controls. Cytotoxic CARs were collected by sorting the top 5% of CD107a-expressing cells by MFI 6 hours after stimulation on day 4. Memory populations were sorted from the stimulated CAR library on day 13 via CD62L and CD45RA expression. FACS plots are shown for sorted CD4⁺ library populations, with and without antigen stimulation. Data shown for donor 1. CAR, chimeric antigen receptor; EGFP, enhanced green fluorescent protein; FACS, Fluorescence-activated cell sorting; ICD, intracellular domain; IL, interleukin; ITAMs, immunoreceptor tyrosine-based activating motifs; MFI, mean fluorescence intensity; TGF-β, transforming growth factor beta.

While CD8⁺ T cells are known for their cytolytic functions, CD4⁺ T cells have been found to be more effective in treating GBM in vivo and clinical evidence suggests that long-lived CD4⁺ memory T cells with sustained cytotoxic function have driven long-term remissions in patients treated with CD19-targeted CAR-T cell therapies.^{10 57} CD4⁺ T cells may also traffic across the BBB more efficiently due to the lack of class I major histocompatibility complex-mediated retention.⁵⁸ Thus, we decided to pursue selections in CD4⁺ human primary T cells. The resulting CAR-T cell libraries consisted of 1.0×10⁷ and 1.3×10⁷ CD4⁺ EGFP⁺ CAR-T cells, representing 7.7-fold and 10-fold library coverage, respectively. Notably, ~26% of EGFP⁺ cells expressed CAR at the surface, as assessed by staining for the targeting ligand IL-13, likely due to many ICD combinations that do not properly fold and express (online supplemental figure 1A).

Library screening

To enrich persistent and proliferative CARs, 3.1×10⁷ CARPOOL transduced CD4⁺ T cells were stimulated with growth-arrested IL-13Rα2⁺ Jurkats every 4 days. 24 hours after the initial antigen exposure, CAR-T cells exhibiting active responses were identified by the upregulation of the co-stimulatory molecule 4-1BB, which constituted 3.16% of the EGFP⁺ population for a total of 2.21×10⁶ CAR-T cells, as demonstrated in figure 1B. These were expanded in vitro, after which cells displaying cytotoxic potential were enriched by sorting for the T-cell degranulation marker CD107a following the second stimulation cycle of 3×10⁶ CARs (refer to figure 1B). The remaining cells were rechallenged with or without transforming growth factor β (TGF-β) supplementation—which is commonly overexpressed in GBM—to select for survival under immunosuppressive conditions (3×10⁶ CARs each).^{59 60} At the third round of rechallenge, 4.68×10⁶ CARs were stimulated with and without TGF-β, while 2.94×10⁶ and 6.18×10⁶ CARs, respectively, were available for the fourth round of rechallenge. Finally, T cells that sustained proliferative and persistent responses were collected. A subset of these cells (n=4.08×10⁶ rechallenged and 2.68×10⁶ TGF-β rechallenged cells) was further sorted based on the expression of memory-associated markers CD62L and CD45RA, suggesting varying enrichment of memory T-cell subsets (depicted in figure 1B). This selection scheme—with the exception of CD107a sorting—was repeated with CD4⁺ T cells sourced from a second healthy donor (online supplemental figure 2). All total CAR-T cell numbers harvested for genomic DNA extraction prior to barcode amplification can be found in online supplemental table 2.

We used NGS to determine the frequency and identity of our selected CARs using previously described methods,³³ where long-read PacBio sequencing was performed on day 13 rechallenged populations to maximize our ability to map ICD composition to barcodes in enriched pools.⁶¹ More than 50% of identified barcodes mapped to ICDs, with the majority containing three ICDs (online supplemental figure 1B). Furthermore, ~95% of barcodes in the unselected population fell between a frequency of 10^–6 and 10^–4 (online supplemental figure 1C). Encouragingly, sequencing 20% of the pre-stimulation population yielded 205,027 unique barcodes or 16% of the theoretical library diversity. Selected populations showed substantial enrichment in the most abundant barcodes relative to the unselected population (figure 2A), with the majority of enriched CARs containing ITAMs (figure 2B). Co-stimulatory domains were also prevalent, while cytokine and inhibitory domains were relatively less enriched (figure 2B and online supplemental figure 2C). While 4-1BB and CD3𝜁 were among the top 30 ICDs for all selected populations, several others, including CD40, FCεRIγ, and DAP12, consistently enriched to a greater extent across biological replicates (figure 2C and online supplemental figure 2D). Notably, CD28 was not among the most frequent ICDs, which may be due to its proclivity to produce short-lived effector T-cell responses.^{52 62} TGF-𝛽 treated stem cell memory and effector populations—which are grouped by CD45RA⁺ expression—shared many patterns of enrichment on assessing the frequency of the top 30 ICDs, irrespective of position, in the selected populations (figure 2C), but TGF-𝛽 supplementation had little effect on the relative abundance of particular signaling domains despite suppressing control CAR-T cell proliferation under selection conditions (figure 2C,D and online supplemental figure 1D).

Figure 2. Sequencing analysis of enriched CD4⁺ library barcodes and signaling domains. (A) Barcode enrichment for unselected and selected populations, with barcode frequency on the y-axis and barcode rank (by frequency) on the x-axis. (B) Frequency of each family of signaling domain throughout different selected populations. (C) Heat maps showing bulk log₁₀ frequencies of ICDs (irrespective of position) for CD107a sorted CARs and rechallenged CARs, with or without TGF-β. (D) Log₂ frequencies of ICDs at each intracellular position relative to the transmembrane domain for selected populations. Data shown for donor 1. Barcode and ICD enrichment data can be found in online supplemental files 1 and 2. CAR, chimeric antigen receptor; CM, central memory; EFF, effector; EM, effector memory; ICD, intracellular domain; ITAM, immunoreceptor tyrosine-based activating motif; SCM, stem cell like memory; TGF-β, transforming growth factor beta.

During our examination of the spatial orientation of enriched ICDs, we discovered that most of them strongly prefer either one or two positions, with the exception of CD40. Additionally, we observed that 4-1BB is enriched in the membrane-proximal position, which aligns with the current design of CAR constructs that include 4-1BB. Other ICDs, some of which have not previously been tested in a CAR, consistently enriched in the same positions across CD4⁺ replicates, including CD79α, PILRβ, CD8α, and BLV gp30 (figure 2D and online supplemental figure 2F). While the top enriched ICD combinations differed between biological replicates, the patterns of ICD enrichment were largely reproducible and generally well correlated between donor replicates of CAR library selections (online supplemental figure 3 and table 3). While 100×library coverage is a more typical scale at which to conduct library selections, such a scale would present technical limitations at the given library size and low multiplicity of Infection (MOI). Thus, barcode undersampling and stochastic dropouts during the manufacturing process likely explain the variance observed between individual barcode frequencies between donor replicates.

CAR 4 exhibits potent antitumor function in vitro

Next, we selected six CARs that enriched in different selection criteria and harbored diverse ICD compositions (figure 3A) to test for expression in Jurkat T cells. Five of the six—excluding a CAR with a combination of GITR, CD3ε ITAM, and CD28 signaling domains—showed surface expression (online supplemental figure 4A). We then tested the five selected CARs—which we named CARs 1, 2, 3, 4, and 5—along with a 13BB𝜁 control for basal CAR expression, memory, and activation phenotype. We first characterized these constructs in human primary CD4⁺ T cells (online supplemental figure 4B). We found that all CARs produced higher tonic basal activity than 13BB𝜁, with CARs 2, 4, and 5 showing the highest activity; CARs 2 and 5 also produced a lower proportion of naïve cells, possibly due to tonic signaling (online supplemental figure 4C–E). Given that pan CD3⁺ T cells are used in the clinic, we also hypothesized that these signaling perturbations may produce useful but distinct effector functions in CD8⁺ T cells, which are generally short-lived but known to mediate target cell lysis and cytokine secretion. On expression in human primary CD4⁺ T cells and antigen challenge with U87 cells transduced to exogenously overexpress Flag-tagged IL-13Rα2 at varying E:T ratios (online supplemental figure 4F), CAR 4—which was the 8th most enriched CAR in 4-1BB sorted cells, 1st in CD107a sorted cells, 8th in effector memory sorted cells, and 13th in rechallenged cells—consistently showed robust activation via CD69 and 4-1BB upregulation as well as potent tumor cell killing in both CD4⁺ and CD8⁺ primary T cells (figure 3B,C and online supplemental figure 5A,B). In fact, CAR 4, which was selected for validation due to its enrichment in CD107a selected cells, showed the highest level of cytotoxicity of all novel CARs; notably, this was also accompanied by higher off-target cell killing against IL-13Rα1⁺ target cells, possibly owing to a lower activation threshold which has been shown to affect the target affinity needed to elicit a effector functions (online supplemental figure 5C).⁶³ CARs 1, 2, and 5 showed lower activity, with CARs 2 and 5 being the only candidates to harbor immunoreceptor tyrosine-based inhibitory motif (ITIM)-containing domains from the inhibitory proteins KIR3DL1 and TIM-3. While CAR 4 produced similar patterns of proinflammatory cytokine secretion to that of 13BB𝜁, including interferon (IFN)-γ and IL-12p70 in both CD4⁺ and CD8⁺ T cells and Flt-3L and IL-12-p40 in CD8⁺ T cells, CAR 3 was generally much less polyfunctional at an E:T ratio of 1:1 (online supplemental figure 6A). Interestingly, both CAR 3 and CAR 4 showed consistently lower levels of Granulocyte-Macrophage Colony-Stimulating Factor (GM-CSF) secretion which is associated with CRS,^{64 65} in both CD4⁺ and CD8⁺ T cells and IL-1β in CD4⁺ T cells compared with 13BB𝜁 (online supplemental figure 6). At a higher E:T ratio of 1:10, CAR 4 produced lower levels of GM-CSF in both CD4⁺ and CD8⁺ T cells, along with similar levels of IFN-γ, IL-17, and FLT3L in CD4⁺ T cells and similar levels of IFN-γ in CD8⁺ T cells (online supplemental figure 6).

Figure 3. Enriched CARs show antitumor functions in response to glioblastoma in vitro. (A) Composition of enriched CARs selected for characterization, along with the populations in which they were enriched. (B–C) Dose response curves following 24-hour co-culture of human primary CD4⁺ T cells with IL-13Rα2⁺ U87 cells at varying E:T ratios (n=3 technical replicates). Resulting (B) CD69 and 4-1BB upregulation in CD4⁺ T cells and (C) tumor cell killing in CD4⁺ and CD8⁺ T cells were measured after 24 hours. Data shown in (B,C) depicts means±SEM (n=3 technical replicates). P values were determined using two-way analysis of variance with Dunnett’s multiple comparisons test. For CD69 expression, p values were 0.0384 and <0.0001 for CAR 4 versus 13BB𝜁 at E:T ratios of 1:5 and 1:10 (n=3, df=60). For 4-1BB expression, both p values were <0.0001 for CAR 4 versus 13BB𝜁 at E:T ratios 1:5 and 1:10, respectively (n=3 technical replicates, df=59). Data in (B) is representative of two biological replicates, while data in (C) is representative of three biological replicates. CAR, chimeric antigen receptor; E:T, effector to target; ICD, intracellular domain; IL, interleukin; MFI, mean fluorescence intensity; TGF-β, transforming growth factor beta.

Enriched CARs show enhanced proliferation relative to 13BBζ on rechallenge

In order to assess the long-term behavior of our selected CARs, we subjected CD4⁺ CAR T cells to four serial antigen challenges with IL-13Rα2⁺ U87 cells using the experimental design and schedule described in figure 1B with TGF-𝛽 supplementation and tracked CAR-T cell number at each time point, assessing memory and exhaustion phenotype on day 14. CAR 4 showed enhanced proliferation relative to 13BB𝜁 by day 14 in two biological replicates, with no notable differences in memory phenotype (figure 4A,B and online supplemental figure 7A); however, CAR 4 showed higher levels of exhaustion marker expression (figure 4C and online supplemental figure 7B). On rechallenge with wild-type U87 cells expressing endogenous levels of IL-13Rα2, we were intrigued to find that each CAR was relatively enriched in the memory phenotype from which it was originally selected, with the exception of short-lived effectors (online supplemental figure 8B).

Figure 4. Selected CARs persist and proliferate in response to tumor rechallenge. (A) CAR fold-change following rechallenge with IL-13Rα2⁺ U87 cells at a 1:1 effector to target ratio on days 0, 4, 8, and 12. (B) Memory and (C) exhaustion phenotypes on day 14 of rechallenge. Data for (A–C) shows means±SEM (n=3 technical replicates). The p values in (A) were determined using two-way ANOVA with Dunnett’s multiple comparisons test and are 0.0129, 0.0153, and 0.0141 for CAR 4 versus 13BBζ on days 8, 12, and 14 (n=3, df=2). P values in (C) were determined using one-way ANOVA with Dunnett’s multiple comparisons test. In (C), the p values for PD-1 expression are 0.0024 for CAR 1 versus 13BBζ, 0.0015 for CAR 2 versus 13BBζ, 0.0004 for CAR 4 versus 13BBζ, and 0.0038 for CAR 5 versus 13BBζ. The p values for LAG-3 expression are 0.0466 for CAR 2 versus 13BBζ and 0.0002 for CAR 4 versus 13BBζ (n=3 technical replicates, df=12). Data is representative of two biological replicates. ANOVA, analysis of variance; CAR, chimeric antigen receptor; CM, central memory; EFF, effector; EM, effector memory; LAG-3, lymphocyte activation gene 3; MFI, mean fluorescence intensity; PD-1, programmed cell death protein 1; SCM, stem cell like memory; TIM3, T-cell immunoglobulin and mucin domain 3.

CAR 4 induces proliferative and cytotoxic transcriptional programs

Given that CAR 3 and CAR 4 showed the highest levels of activation and cytotoxicity, we next sought to determine whether their signaling inputs produced distinct transcriptional states in a clinically relevant context, where synergistic effects between CD4⁺ and CD8⁺ T cells might compound over time. Thus, we subjected human primary donor CD3⁺ CAR-T cells to tumor rechallenge with IL-13Rα2⁺ U87 cells, according to the timeline shown in figure 1B. Then, we sorted the rechallenged cells for EGFP 48 hours after the final stimulation and performed single-cell RNA sequencing.⁶⁶ Dimensionality reduction and unsupervised clustering of the resulting transcriptional data yielded seven distinct cell clusters, with CAR 3 most enriched in clusters 1 and 7, CAR 4 enriched in clusters 2, 3, and 4, and 13BB𝜁 enriched in clusters 4, 5, and 6—as confirmed by χ² analysis (figure 5A, B and D). Notably, the CAR 4-enriched and 13BB𝜁-enriched clusters 3 and 4 were enriched in dividing cells, which were largely CD8⁺ (figure 5C and online supplemental figure 9). Comparing average expression profiles for the top 10 genes driving each cluster, shown in figure 5E, revealed that cluster 2, which was heavily enriched for CAR 4, contained genes related to cytotoxicity such as PRF1, GZMK, GZMA, GNLY, and NKG7 along with CCL5, which distinguishes glioma infiltrating lymphocyte expansion within patient tumor samples when present in combination with a cytotoxic gene signature.^{67 68}

Figure 5. CAR 4 produces a cytotoxic gene signature that is associated with expansion within patient with glioma samples. (A) UMAP embeddings of merged scRNA-seq profiles colored by cell state, (B) CAR identity, and (C) proliferative state following rechallenge of human primary CD3⁺ CAR-T cells with IL-13Rα2⁺ U87 cells four times at an effector to target ratio of 1:1 (n=1,471, 1,471, and 1,194 cells for CARs 3, 4, and 13BBζ, respectively). (D) Χ² enrichment values for each CAR candidate within each cluster, represented by the Pearson residuals measuring the difference between the observed and expected CAR frequencies within each cluster. (E) Average expression of top 10 differentially expressed genes by cluster. CAR, chimeric antigen receptor; scRNA-seq, single-cell RNA sequencing; UMAP, uniform manifold approximation and projection .

CAR 4 elicits improved persistence and comparable tumor control in vitro

Given that CAR 4 exhibited both persistence and cytotoxicity, it became our lead candidate for subsequent characterization. We next sought to more accurately validate its antitumor activity compared with 13BB𝜁 in a micro-physiological human in vitro model for GBM with a representative tumor microenvironment.⁴³ Briefly, tumor spheroids were generated with IL-13Rα2⁺ RFP⁺ U87 cells at~500 $\mathrm{\mu }$m in diameter and seeded into a commercially available multi-well insert microfluidic device⁶⁹ in a fibrin gel with brain endothelial cells, pericytes, and astrocytes, and allowed to develop over 7 days as detailed in Lam et al.⁴³ The microtumor model consists of a tumor spheroid that is co-cultured with a perfusable vasculature whose permeability has been validated to be in the same physiological range as the BBB in vivo. The model is able to recapitulate gross clinical features of GBM tumors and tumor spatial heterogeneity, as the microtumor consists of a dense tumor core surrounded by a diffused invasive front. This allows for testing of CAR-T cell targeting of different subpopulations of tumor cells, similar to in vivo. The perfusable vasculature in the microtumor model allows for the testing of CAR-T cell targeting and extravasation, while the presence of astrocytes and pericytes in the model allows for the observation of off-target cytotoxicities, providing a more accurate assay of CAR 4 function in vitro. Furthermore, it is able to capture CAR-T and tumor cell spatial localization over time—a feature that is absent in most in vivo workflows.

After the formation of the GBM-BBB model, which we termed microtumors, 1×10⁴ CD3⁺ CAR 4 or 13BB𝜁 CAR-T cells were perfused into the microfluidic devices. Tumor growth, CAR-T position, and proliferation were subsequently monitored over 9 days with confocal microscopy (figure 6A–D and online supplemental figure 10). Both CAR 4 and 13BB𝜁 controlled tumor growth over 9 days. While 13BB𝜁 exhibited better tumor control at day 6, this regressed by day 9, at which point both CAR 4 and 13BB𝜁 performed similarly in controlling tumor growth (figure 6C,D and online supplemental figure 10C). Both CAR 4 and 13BB𝜁 expanded in situ, as observed from an increase in signal over time, with CAR 4 expanding more by day 9 (figure 6D and online supplemental figure 10B,D). CAR 4 appeared to have better local tumor control at the invasive front, with a corresponding increase in CAR-T signal, suggesting local CAR-T function in controlling tumor growth. In contrast, 13BB𝜁 tumor control did not correlate with locations of high CAR-T signal, suggesting a more systemic mechanism of tumor control (figure 6C,D and online supplemental figure 10C,D). On day 9, T cells were retrieved from inside or outside the microtumors (figure 6E) and stained for T-cell activation, exhaustion, and memory markers.

Figure 6. CAR 4 controls IL-13Rα2⁺ U87 cells and exhibits increased persistence in vitro. (A) Representative image of IL-13Rα2⁺ U87-RFP tumor spheroids (magenta) in microdevices at day 3 and day 9 after CAR-T (cyan) addition. (B) Illustration of radial plot analysis carried out in (C) and (D). (C) Increase in U87 cell signal intensity at day 9 relative to day 0. (D) CAR-T signal at day 9. Data shown in (C,D) are means±SEM. (E) Illustration of regions inside and outside the microtumors. (F) CAR-T cell counts inside and outside the microtumors. (G) Activation, (H) exhaustion and (I) memory phenotypes of CAR-T cells retrieved from inside and outside the microtumors. P values in (C) were determined by comparing the area under the curve with one-way analysis of variance with Tukey’s multiple comparisons test and are <0.0001 (n=5 for untreated, 13BB𝜁 and CAR 4, df=12). P value in (D) were determined by comparing the area under the curve with two-way unpaired t-test and are <0.0001 (n=5 for 13BB𝜁 and CAR 4, df=4). Data in (F–I) are pooled from three technical replicates. CAR, chimeric antigen receptor; CM, central memory; EFF, effector; EM, effector memory; IFN, interferon; LAG-3, lymphocyte activation gene 3; MFI, mean fluorescence intensity; PD-1, programmed cell death protein 1; SCM, stem cell like memory; TIM3, T-cell immunoglobulin and mucin domain 3.

Consistent with what was observed with live confocal microscopy, cell counts for CAR 4 were higher than for 13BB𝜁, especially for T cells retrieved from inside the microtumors (figure 6F). IFN$\mathrm{\gamma }$ expression was low outside the microtumors for both CAR 4 and 13BB𝜁, but increased after entering the microtumors, as would be expected on encountering antigen, with 13BB𝜁 showing a larger increase in expression (figure 6G). CD107a expression was high for both CAR 4 and 13BB𝜁, with 13BB𝜁 increasing and CAR 4 expression decreasing slightly on entering the microtumors (figure 6G); the reduced CD107a expression by CAR 4 could indicate a higher susceptibility to immune suppression present within the tumor microenvironment. In contrast, exhaustion markers PD-1 and TIM3 were highly expressed in both CAR 4 and 13BB𝜁 outside of the microtumors, but had lower expressions inside the microtumors, possibly as a result of immunosuppression limiting upstream activation. This decrease was dramatic for CAR 4 but slight for 13BB𝜁 (figure 6H). On the other hand, LAG-3 expression was low outside of microtumors, and increased significantly for 13BB𝜁, but was almost absent in CAR 4 inside microtumors (figure 6H). On assessing memory phenotype, 13BB𝜁 retained less differentiated stem cell and central memory subtypes compared with CAR 4, while CAR 4 heavily skewed towards effector memory populations—a population in which it was enriched on selection—both inside and outside of the microtumor (figure 6I). Taken together, this suggests that CAR 4 may be capable of persisting within a tumor microenvironment, although at a slightly reduced functionality relative to 13BB𝜁 once subjected to immune suppression.

CAR 4 elicits improved persistence and comparable tumor control in vivo

Finally, we sought to assess the in vivo persistence and tumor control of our lead CAR candidate, CAR 4, relative to 13BB𝜁 in a subcutaneous xenograft model of human GBM. Mice were injected with 2×10⁶ Flag⁺ IL-13Rα2⁺ FLuc⁺ U87 cells subcutaneously, followed by intravenous administration of 1×10⁶ CD3⁺ CAR or mock transduced T cells 21 days later (figure 7A). Tumor area and flux along with weight change relative to pretreatment as a proxy for toxicity were assessed for 20 days following adoptive cell therapy (ACT) (figure 7B, online supplemental figures 11A,B 12, 13A–C, 14). On day 20 post ACT, CARs from tumors and spleens were harvested and stained for T-cell memory and exhaustion phenotypes, along with antigen for tumor samples. This was performed using two different sources of donor T cells, designated as donors 3 and 4.

Figure 7. CAR 4 controls IL-13Rα2⁺ U87 cells in vivo. (A) Experimental timeline. (B) Tumor flux of Flag⁺ IL-13Rα2⁺ FLuc⁺ U87 cells implanted subcutaneously in NSG mice. Data shown are for individual mice (n=6 for mock, CAR 4, and 13BB𝜁). (C) CD4⁺ and CD8⁺ CAR abundance in the spleen and (D) tumor. (E) Antigen-positive tumor cell abundances. Data shown in (C–E) are means±SEM The p values in (B) were determined using two-way ANOVA with Dunnett’s multiple comparisons test and are 0.0272, 0.0453, and <0.0001 for CAR 4 versus 13BB𝜁 on days 15, 17, and 20, respectively (n=6 mice each, df=180). P values in (C) and (D) were also determined by two-way ANOVA and are <0.0001 and 0.0003 for CD4⁺ CAR 4 versus 13BB𝜁 in the spleen and tumor, respectively (n=6 each, df=36). The p value in (E) is 0.0039 for CAR 4 versus 13BB𝜁 (n=6 each, df=5) as determined by an unpaired t-test. Data is representative of three biological replicates from donor 3. ACT, Adoptive cell therapy; ANOVA, analysis of variance; CAR, chimeric antigen receptor; FLuc, firefly luciferase; NSG, NOD/SCID/IL-­ 2Rnull; s.c., subcutaneous.

Despite showing similar tumor control to 13BB𝜁—possibly owing to the counteracting persistence, intravasation, and resistance to immune suppression, as was observed in the microtumor model—CAR 4 showed greater abundance in both the tumor and spleen for both donors (figure 7C,D and online supplemental figure 13A,D,E), alongside lower IL-13Rα2⁺ tumor cell abundance in donor 3-treated mice (figure 7E); in donor 4-treated mice, U87 cells following CAR treatment showed reduced antigen expression following CAR treatment relative to surviving mock-treated mice, suggesting potential antigen downregulation (online supplemental figure 13F,G). Notably, no CAR-treated mice experienced toxic side effects due to systemic CAR-induced inflammation, as assessed by weight loss (online supplemental figures 11B and 13C). In donor 3-treated mice, all CD8⁺ tumor-infiltrating CARs largely skewed towards effector phenotypes, while—similarly to what was observed in the selection data and microtumor model—CD4⁺ CAR 4 produced a higher proportion of effector memory T cells relative to 13BBζ (online supplemental figure 11C,E). In the spleen, CAR 4 showed greater persistence relative to 13BBζ for both donors, with a majority being CD4⁺ T cells that were overall less terminally differentiated than those found in the tumor (figure 7C and online supplemental figures 11D,F and 13D,H,I). In donor 4-treated mice, CAR 4 and 13BBζ produced similar proportions of memory populations in both niches (online supplemental figure 13H,I). All tumor-infiltrating CARs showed similar levels of exhaustion marker co-expression, though CAR 4 showed significantly higher levels PD-1 in CD4⁺ and CD8⁺ T cells—which is thought to be an activation marker in the absence of LAG-3 and TIM-3—while 13BB𝜁 showed higher expression of TIM-3 in CD8⁺ T cells in donor 3 (online supplemental figure 11G,H). However, in donor 4-treated mice, CAR 4 showed comparable TIM3 and LAG-3 expression (online supplemental figure 13J,K).

Discussion

CARs have shown incredible efficacy in treating B-ALL, diffuse large B cell lymphoma (DLBCL), and MM, but have yet to translate to disease indications beyond hematological malignancies successfully. In this work, we used high throughput screening methods to select CD4⁺ CARs that persist, proliferate, differentiate, and kill tumor cells from a pool of signaling diversified CARs targeting the cancer testis antigen IL-13Rα2. Lead hits were chosen for validation from each selected population while attempting to avoid highly similar signaling combinations, of which 8/12 were successfully cloned and 6/8 produced high titer virus. Of these viruses, five produced CAR surface expression. CAR 4, consisting of PILRβ, DAP12 ITAM, and CD3𝜁 ITAM 2, was identified following selection for the cytotoxicity marker CD107a as well as persistence on rechallenge and effector memory formation. It was found to produce significantly enhanced long-term proliferation and persistence on tumor rechallenge compared with 13BB𝜁 accompanied by elevated activation and cytotoxicity in vitro at the tradeoff of higher off-target activity against IL-13Rα1. Given the high rate of recurrence in patients with GBM, long-term persistence could prove to be pivotal in GBM-targeted CAR-T cell therapies. It showed a similar depletion of antigen-positive tumor cells in an IL-13Rα2⁺ xenograft model of GBM with markedly improved persistence in both the spleen and tumor—a feature that was also observed in a microphysiological human in vitro model of GBM with a tumor microenvironment. We also found that it robustly recapitulated effector memory formation throughout validation studies. This work demonstrates that our strategy is capable of identifying functional CARs from a large library that faithfully recapitulate the qualities for which they were selected.

Notably, the reduction in antigen load in vivo was somewhat subtle and did not correlate with a reduction in tumor size, possibly due to the outgrowth of antigen-negative tumor cells; this phenomenon of downregulation is largely due to inherent variation in antigen expression within the tumor population where low antigen expression—possibly through differences in epigenetic state—can be selected for on CAR-mediated tumor cell killing. In addition to the FDA-approved treatments targeting CD19 facing similar challenges, the original case study using 13BB𝜁 also showed recurrence of antigen-negative tumor lesions, and several groups have employed dual targeting of both IL-13Rα2 and EGFRvIII to overcome antigen heterogeneity and improve overall outcomes in treating GBM^{16 31 70}; recent application of this dual targeting approach in phase I clinic trials has led to reduction in tumor burden.^{71 72} Our in vitro rechallenge studies employed mitomycin pretreatment of the U87 cells in order to use them solely as a source of antigen and would not have captured this phenomenon, but future selections employing low antigen density target cells could increase CAR sensitivity in the context of antigen downregulation. The observed difference in translation between an in vitro and in vivo setting may be partly due to limitations in tumor infiltration as well as differential sensitivity to tumor immunosuppression, as seemed to be the case for CAR 4 on assessment in the microtumor model. Additionally, our immunocompromised xenograft model does not capture the interplay between the CAR-T cells and host immune system, a feature that is known to support CAR-T cell function as well as promote epitope spreading; the latter may be critical for targeting IL-13Rα2 and treatment of GBM in general and has been observed using a murine IL-13Rα2-targeted zetakine, where IFN-γ was found to be critical.⁷³ In both the microtumor and in vivo models, CAR 4 consistently demonstrated enhanced persistence and a skew toward an effector memory phenotype. However, while the microtumor model provided a controlled environment to quickly assess CAR-T efficacy in a three-dimensional structure mimicking solid tumor architecture with high spatial resolution, the in vivo model introduced additional complexities such as compartmentalization within other lymphoid tissues and stromal interactions. These differences highlight the robustness of CAR 4 in overcoming physical barriers similar to a BBB in the microtumor, but also underscore the unique challenges of overcoming immunosuppressive factors and heterogeneity in a more complex, in vivo setting. Notably, survival tumor studies in an orthotopic glioma model, which were not conducted in this study, would more strengthen our understanding of CAR 4’s antitumor potential.

Additionally, the signaling inputs of CAR 4, which were selected for cytotoxic function, also increase the cytotoxicity against IL-13Rα1, which has also been observed in comparing CD28 versus 4-1BB-based second generation IL-13Rα2-targeted zetakine CARs.^{51 74} Given that signaling perturbations can affect affinity requirements necessary to elicit T-cell activation,⁶³ future selections against target antigens with known homologs could be vetted for off-target toxicity through negative selection. While its specificity may preclude this design from clinical consideration, its affinity for IL-13Rα1 could be further reduced by introducing additional mutations known to selectively hinder IL-13Rα1 versus IL-13Rα2 binding in the IL-13 binding domain, which have been previously characterized.⁷⁵

CD3𝜁 ITAM 2 is commonly included in CARs via the full-length CD3𝜁. While it has been examined in isolation,75,83 it has not been tested for function in a CAR at the membrane distal position of a third generation CAR. DAP12 is less commonly incorporated in the context of T cell-based therapies and its signaling is typically associated with natural killer (NK) and myeloid cells, where it plays an activating role in the context of high avidity ligands by phosphorylating LAT and NTAL and activating PI3K, PLCγ, VAV, ERK, and CBL signaling; it also initiates cytoskeletal reorganization and may enhance CD8⁺ T-cell cytotoxicity.76,84 PILRβ has not—to our knowledge—been reported as functional in the context of a CAR, but its putative function is to counteract inhibitory signals from its ITIM-bearing isoform PILR⍺ in order to activate myeloid and NK cells via association through positively charged residues near the transmembrane domain with DAP12 in response to CD99 ligation^{85 86}; it is also thought to regulate T-cell activation, but its exact role remains unclear.⁸⁷ Further CAR domain deletion and protein–protein interaction studies may elucidate the exact mechanism of action and identify any synergistic function of this combination of signaling domains. Additionally, in-depth transcriptomic and proteomic profiling could reveal signaling mechanisms employed by CAR 4 that may be of clinical interest. While each of these domains would be plausibly useful in a CAR-T cell therapy on examination in isolation, this exact combination and spatial configuration would be unlikely to emerge to the forefront through conventional design approaches.

While several groups have recently reported employing similar strategies for library-based screening of CARs,34,36 our study represents the most extensive exploration of signaling diversity in human primary T cells described to date in terms of signaling domain identity, functional family, immune cell source, and spatial arrangement; though screens of this scale in primary cells present unique technical challenges, their representative physiology enables the use of a broader and more complex variety of screening criteria. While this large library size complicates comprehensive sampling of the represented molecular diversity, which would benefit the development of predictive machine learning algorithms and the establishment of structure-function relationships for CAR design, it enables a more extensive search of the theoretical signaling landscape with fewer biases in inferring design principles and should still reveal instances of strong motif or barcode enrichment despite sampling limitations. Increased sequencing coverage of the signaling domain diversity in the unselected population using higher read depth long-read sequencing techniques or defined ICD barcodes would enable deeper statistical insight into domain enrichment, but was not possible in this study given the read-depth limitations of PacBio sequencing. Furthermore, while only a fraction of the library variants was able to express at the cell surface, which reduces the overall coverage of the explored sequence space, it represents which CARs might be functional proteins. We propose that this large library format may be a better tool for novel CAR discovery that can be broadly applied to many different disease targets and selection criteria. Altogether, the modular nature of high throughput CAR screening makes it uniquely poised to reinvigorate CAR discovery and expedite clinical translation in a streamlined process and its principles have broad implications for therapeutic modalities beyond CARs, making it a powerful tool for unlocking next-generation cell therapies.

Data availability

Data and materials availability

The NGS selection data sets have been deposited in the Sequence Read Archive and are available under the accession number PRJNA957857.⁶¹ The scRNA-seq data have been deposited in the Gene Expression Omnibus under accession number GSE259352.⁶⁶ All data generated or analyzed during the study (including the DomainSeq-processed CARPOOL selection data) are included in the paper or its online supplemental information.

Code availability

The code used to analyze the domain composition of selected CARs can be accessed in the DomainSeq repository at https://github.com/birnbaumlab/ Gordon et al 2022.

Data availability statement

Data are available in a public, open access repository. Data are available upon reasonable request.