Human blood from deidentified normal donors was purchased from STEMCELL Technologies (Vancouver, Canada), which collected and distributed de-identified human blood products with consent forms, and according to the protocols, approved by the Institutional Review Board (IRB). Peripheral blood mononuclear cells were isolated by Ficoll density gradient centrifugation, and T cells were further enriched using the EasySep Human T Cell Isolation Kit (STEMCELL Technologies) as per the instructions of the manufacturer. Enriched T cells or CD4⁺CD127⁺CD25^low conventional T cells purified by fluorescence-activated cell sorting (FACS) were used for experiments. Cells were either used fresh or cryopreserved in fetal calf serum (FCS) with 10% DMSO. Frozen cells, when used, were thawed and cultured overnight in 300 IU/ml of IL-2 before editing and cell activation.

Ribonucleoprotein complexes (RNPs) were made by mixing CRISPR RNAs (crRNAs) and trans-activating crRNAs (tracrRNA, Integrated DNA Technologies, Coralville, IA) with recombinant Cas9 protein (QB3 Macrolab, UC Berkley, CA) as previously described in a study by Roth et al. (16). Guide RNA sequences used for gene editing were as follows: (1) T-cell receptor β chain constant region (TRBC): CCCACCAGCTCAGCTCCACG; (2) CD19: CGAGGAACCTCTAGTGGTGA; and (3) CD28: TTCAGGTTTACTCAAAAACG. Lyophilized RNAs were resuspended at 160 μM in 10 mM Tris-HCl with 150 mM KCl and stored in aliquots at −80°C. On the day of electroporation, crRNA and tracrRNA aliquots were thawed and mixed at a 1:1 volume and annealed for 30 min at 37°C. The resulting 80 μM guide RNA complex was mixed at 37°C with Cas9 Nucleic Localization Signal (NLS) at a 2:1 gRNA to Cas9 molar ratio for another 15 min. The resulting RNP was used for genome editing. About 1 × 10⁶ T cells were mixed with appropriate RNP and electroporated using a Lonza 4D 96-well electroporation system (pulse code EH115) to delete TCR or CD28 genes. For generating CD19⁻ variants of Raji cells, Raji cells (ATCC® CCL-86™, Manassas, VA) were electroporated (pulse code EH140) with RNP targeting CD19, and the CD19- negative fraction was purified by FACS after culturing the cells for more than 1 week.

CD4⁺ T cells were gene-edited before stimulation with anti-CD3/CD28 beads (Dynabeads Human T-Activator CD3/CD28, Thermo Fisher Scientific, Waltham, MA). Cells were cultured in RPMI supplemented with 10% FCS and 300 IU/ml of IL-2 (Prometheus laboratories, Nestle Health Science, Lausanne Switzerland) for the first 2 days, and then, the concentration of IL-2 was reduced to 30 IU/ml for CD4⁺ T cells and to 100 IU/ml for bulk T cells. Lentiviruses encoding anti-CD19-I₄-28-4-1BBζ-T2A-EGFRt and anti-CD19-I₄-28-28ζ-T2A-EGFRt were provided by Juno Therapeutics (Bristol-Myers Squibb, New York, NY). Other lentiviral constructs, present in Figure 2A, were cloned into the pCDH-EF1-FHC vector (Addgene plasmid #64874, Watertown, MA) as previously described by Hill et al. (17). Later, genes encoding CAR constructs were purchased from gBlocks™ Gene Fragments (Integrated DNA Technologies) (17, 18) and amplified by PCR and cloned into the pCDH vector using In-Fusion Cloning Tools (Takara Bio, Kusatsu, Japan). Sequences for all clones used in subsequent experiments were confirmed by sequencing. Transduction was performed on day 2 after CD4⁺ T cell activation at a multiplicity of infection of one by spinoculation (1,200 g, 30 min, 30°C) in a medium supplemented with 10% FCS and 0.1 mg/ml of protamine. For AAV production, 30 mg of helper plasmid pDGM6 (a kind gift from YY Chen, University of California, Los Angeles), 40 mg of pAAV helper, and 15 nmol PEI were utilized. AAV6 vector production was carried out by iodixanol gradient purification. After ultracentrifugation, the AAVs were extracted by puncture and further concentrated using a 50 ml Amicon column (Millipore Sigma Burlington, MA) and directly titrated on primary human T cells.

For some experiments, cells were restimulated on day 9 after primary stimulation without separating edited and transduced cells. For proliferation assays, the cell mixtures were stained with 2.5 μM carboxyfluorescein diacetate succinimidyl ester (CFDA SE, ThermoFisher, referred to as CFSE) before restimulation with anti-CD3/CD28 beads. For other experiments, cells were separated by FACS on day 9 to purify CD3⁺ and CD3⁻ T cells with or without CAR. For assessing CD25⁺CD71⁺ upregulation, purified CAR T cells were stimulated with parental CD19⁺ Raji cells or CD19^deficient Raji cells for 2 days. In some cultures, CTLA-4 Ig (provided by Dr. Vincenti, UCSF) was added at a concentration of 13.5 μg/ml. For measurements of proliferation, purified cells were stimulated with soluble anti-CD28 (clone CD28.2, 1 μg/ml, BD Pharmigen), plate-bound anti-CD28 (clone CD28.2, 10 μg/ml), or soluble anti-CD3 (clone HIT3α 2 μg/mL. BD Pharmigen). After 48 h, a portion of the supernatant was collected and analyzed for cytokine secretion using multiplexed Luminex (Eve Technologies, Calgary, Canada). The cells were then pulsed with 0.5 μCi of ³H thymidine and cultured for another 16–18 h before harvesting to determine the level of ³H thymidine incorporation using a scintillation counter.

The following antibodies were used for phenotyping and proliferation assays: anti-CD3-PE/Cy7 (clone SK7, BioLegend, San Diego, CA), anti-CD4-PerCP (clone SK3, BD Pharmigen, San Jose, CA), anti-CD4 A700 (clone RPA T4, BioLegend), anti-CD19 APC (clone HIB19, BD Pharmigen), anti-CD25 APC (clone 2A3, BD Pharmigen), anti-CD71 FITC (clone CY1G4, BioLegend), anti-Myc FITC or APC (clone 9B11, Cell Signaling, Danvers, MA), anti-FMC19 idiotype APC (Juno Therapeutics), anti-EGFRt PE (Juno Therapeutics), anti-CD28 APC (clone 28.2, Biolegend), and CD8 APC-Cy7 (clone SK1, BioLegend). DAPI (ThermoFisher, Waltham, MA) was used to stain dead cells for exclusion during analysis. Flow cytometric analyses were performed on an LSR II Flow Cytometer System (BD Biosciences). Fluorescence-activated cell sorting was performed on an FACSAria III Cell Sorter (BD Biosciences). All flow cytometry data were analyzed using the FlowJo software (Tree Star, Ashland, OR).

FACS-purified CD3⁻CAR⁺ or CD3⁻CAR⁻ CD4⁺ T cells (8 × 10⁶ each) were lysed in Pierce™ IP Lysis Buffer (ThermoFisher) supplemented with cOmplete Protease Inhibitor Cocktail (Roche, Basel, Switzerland) for 30 min using a vertical rotator. Cell lysis was completed by briefly sonicating cells using a Q500 sonicator (QSonica, Newtown, CT). Pierce^TM anti-c-Myc magnetic beads (clone 9E10, ThermoFisher) were used for immunoprecipitation of the CAR. Alternatively, rabbit anti-human CD28 (clone D2Z4E, Cell Signaling) followed by anti-rabbit IgG Pierce^TM protein A/G magnetic beads (ThermoFisher Scientific) were used for CD28 immunoprecipitation of the cell lysate according to the instructions of the manufacturer.

Equal masses of protein lysate or equal volumes of immunoprecipitation eluents were loaded into NuPAGE 4–12% Bis-Tris, 1.0 mm gels (ThermoFisher Scientific). After electrophoresis, proteins were transferred onto PVDF membranes (ThermoFisher Scientific) using an iBlot 2 Dry Blotting System. After blocking with Tris-buffered saline with 0.1% Tween-20 and 5% bovine serum albumin (TBSTB), membranes were stained with primary and secondary antibodies diluted in TBSTB. The following antibodies used were mouse anti-Myc (clone 9B11, Cell Signaling), rabbit anti-CD28 (clone D2Z4E, Cell Signaling), HRP-conjugated anti-mouse IgG (Cell Signaling), and HRP-conjugated anti-rabbit IgG (Cell Signaling).

Structural modeling of the different CARs was performed using Iterative Threading ASSembly Refinement (I-TASSER) software (19). Amino acid corresponding to the scFv was modeled on the UCHT1 scFv template (PDB ID code 1XIW) (20). The HD coordinates were recovered from the crystal structure of the pembrolizumab template (PDB ID code 5DK3) (21) and the crystal structure of human CD28 (PDB ID code 1YJD) (22) for IgG4 and CD28, respectively. Modeling of the CD8-HD was performed using the Rosetta protein modeling suite (23). Structures were assembled with PyMOL (Schrodinger, LLC). Models were further evaluated with MolProbity software (24).

To investigate the role of CAR TMD, we first generated a panel of 19-CARs differing only by their HD (CD8, CD28, or IgG₄) and their TMD (CD8 vs. CD28), all of which have been used to engineer CAR T cells for clinical applications (Figure 1A, Supplementary Figure 1). Each CAR was designed with an MYC tag on the N-terminus of the scFv and a mCherry reporter (Figure 1A). For most experiments, we selected 4-1BB as a costimulatory domain in the ICD to avoid potential interactions with the endogenous CD28. Furthermore, we disrupted the TRBC locus using CRISPR/Cas9 to prevent any potential confounding influence by the endogenous TCR (Figure 1B). The TRBC gene-disrupted human T cells retained the cell surface expression of TCR/CD3 proteins for a few days after editing and could thus be activated with anti-CD3/CD28 beads. Edited CD4⁺ T cells were transduced with various lentiviral CAR constructs by spinoculation 2 days after activation. On day 9, after stimulation, 87–98% of the cells were found to CD3-negative, demonstrating successful TCR deletion in the majority of the cells (Figure 1C). Comparable transduction efficiencies were observed across the different CAR constructs, as assessed by the mCherry expression and all CAR T cells responded to CD19 restimulation (Supplementary Figures 2A–C).

Restimulation of TCR-edited CAR-transduced T cells, containing a mixed population of CD3^+/− and CAR^+/− cells, with anti-CD3/CD28 beads on day 9, resulted in the expansion of CD3⁺ T cells that escaped TCR deletion (Figures 1D,E). However, TCR-deficient CD3⁻CAR⁺ T cells with CARs containing a CD28-TMD, but not CD8-TMD, also proliferated. Consequently, CAR⁺ T cells with a CD28-TMD, but not a CD8-TMD, were enriched at the end of the 5-day restimulation (Figure 1F). The lack of proliferation of CD8-TMD-containing CAR T cells showed that expansion was not a consequence of bystander effects, such as IL-2 production by the CD3⁺CAR⁺ T cells in the same culture. To determine if this is unique to CARs with 4-1BB-ICD, the experiment was repeated using CARs with a CD28-ICD, and a similar pattern of proliferation and enrichment of CD3⁻CAR⁺ T cells after anti-CD3/28 bead restimulation was observed (Supplementary Figures 3A–D).

Since the CD3⁻CAR⁺ T cells had no TCR expression on the cell surface, the proliferation was likely stimulated by the anti-CD28 component of the anti-CD3/28 beads. To verify the need for the endogenous CD28 receptor for proliferation in response to anti-CD3/CD28 beads, both the CD28 and TRBC genes were deleted in T cells before activation and lentiviral CAR transduction (Figure 2A). CD3⁻CAR⁺CD28⁺ T cells expressing CARs containing a CD28-TMD, but not a CD8-TMD, proliferated in response to anti-CD3/28 beads. The deletion of CD28 abrogated the ability of CD28-TMD-containing CAR T cells to proliferate in response to anti-CD3/CD28 beads, demonstrating that anti-CD28-induced activation was dependent on endogenous CD28 (Figure 2A). These results excluded the possibility that anti-CD3/CD28 beads directly bind to the CAR.

To further confirm that CD3⁻ T cells with a CD28-TMD-containing CAR could respond to anti-CD28 stimulation in the absence of other cells in the culture, we FACS-purified CD3⁻CAR⁺ cells before restimulation with plate-bound or soluble anti-CD28 antibodies (clone CD28.2). For these experiments, we excluded CD28-HD containing CARs to avoid potential interaction mediated by the CD28-HD. The results confirmed that CAR T cells engineered with a CD28-TMD, but not a CD8-TMD, proliferated in response to anti-CD28 alone (Figure 2B). The proliferative response induced by anti-CD28 alone in CD3⁻CAR⁺ T cells was similar in CAR T cells with a 4-1BB or a CD28 costimulatory domain in their ICD (Supplementary Figure 4A). Moreover, anti-CD28 induced the secretion of multiple cytokines by CD3⁻CAR⁺ T cells but not by CD3⁺CAR⁻ or CD3⁻CAR⁻ control cells (Supplementary Figure 4B). Collectively, these results show that CD28-TMD containing CARs can be activated by anti-CD28 without antigen recognition by the CAR or the TCR.

The results discussed earlier, together with recent reports of phosphorylation of endogenous CD28 upon CAR stimulation (all with a CD28-TMD domain) (25, 26), suggest interactions between CD28 and CD28-TMD-containing CARs. To directly determine if CD28-TMD-containing CAR and CD28 can physically interact, we performed co-immunoprecipitation experiments. CD28-TMD-containing, but not CD8-TMD-containing, CARs co-immunoprecipitated with endogenous CD28. Conversely, endogenous CD28 co-immunoprecipitated with CD28-TMD-containing, but not CD8-TMD-containing, CARs demonstrated that the CD28-TMD of the CAR interacted with the endogenous CD28 receptor (Figure 2C). CD8-HD/CD28-TMD CARs and CD28 co-immunoprecipitated more efficiently when compared to the IgG₄-HD-CD28-TMD construct, which is consistent with improved proliferation observed with CD8-HD/CD28-TMD CAR upon anti-CD28 stimulation (Figure 2B). Because of the difficulties in expanding CD3⁻CD28⁻ CAR T cells and the unlikelihood that anti-CD28 mAb directly binds to the CD28 TMD, we did not perform the purified CAR T-cell proliferation and immunoprecipitation studies with CD28⁻ T cells.

Next, we generated a series of CD28-TMD CAR mutants to determine the molecular basis of the CAR–CD28 interaction. We first mutated the two glycines, G160L and G161L (M1), that may function as part of a glycine-zipper motif, a process known to control TMD dimerization (27). The second mutation replaced the C165 cysteine with alanine, as cysteine can form disulfide bonds (M2). The third (M3) mutations were made on two bulky hydrophobic tryptophans at the border of the TMD (W154L and W179L), and the fourth mutation (M4) targeted four amino-acid residues (C165L, Y166L, S167L, and T171L) present at the core of the TMD, as cysteine could form a disulfide bond and others may form hydrogen bond (Figure 3A). All CARs with TMD mutants were readily expressed on the cell surface (Figure 3A). The various CD3⁻CAR⁺ cells with mutated CD28-TMD were examined for their ability to proliferate to anti-CD28 stimulation. For the analysis of these experiments, CD3⁻CAR⁺ cells were further defined as low, intermediate, or high CAR expression based on the level of the mCherry expression. CAR T cells with the wild-type CD28-TMD (CD28-TMD^WT) proliferated to anti-CD3/CD28 stimulation, regardless of the level of CAR expression (Figures 3B,C). The CD28-TMD^M4, but not the other TMD-mutants, abrogated the proliferation of CD3⁻CAR^low cells and significantly reduced the proliferation of CD3⁻CAR^int cells with either CD8-HD or IgG₄-HD (Figures 3B,C). Interestingly, CD3⁻CAR^high T-cell proliferation was only weakly affected by M4 mutations. The CD3⁻CAR^high T cells did not undergo proliferation when restimulation was not carried out, demonstrating that the activation was dependent on anti-CD28 stimulation and was not a result of autonomous CAR tonic signaling (Figure 3B).

To confirm that the CD28-TMD^M4 disrupted the interaction between CD28 and the CAR, we sorted CAR T-cells based on the Cherry expression, engineered either with CD8-HD/CD28-TMD^WT or with CD8-HD/CD28-TMD^M4, and rechallenged them with a plate-bound anti-CD28 (Figure 3D). In this assay, only CAR T cells with a CD28-TMD^WT showed significant proliferation, as measured by radiolabeled-thymidine incorporation. Importantly, co-immunoprecipitation of the endogenous CD28 and the CD28-TMD-containing CAR was abrogated by the M4-mutant, demonstrating that the four amino acids present at the core of the CD28-TMD are necessary for CAR-CD28 heterodimerization (Figure 3E).

To determine if the natural ligands of CD28, CD80, and CD86 can activate CARs by engaging CD28-CAR heterodimers, we stimulated CAR T cells with different HD and TMD with CD19-deficient Raji cells that express high levels of CD80 and CD86 (Supplementary Figure 5A). CD19^deficient Raji induced CAR T-cell activation, although at a lower intensity than that induced by the CD19⁺ Raji cells (Supplementary Figure 5B, Figure 4A). This “off-target” activation was mostly seen in T cells with a high CAR expression (Figures 4A,B). Moreover, CAR T-cell activation by CD19^deficient Raji was significantly reduced by binding CTLA-4 Ig, a high-affinity competitive inhibitor of CD28, to CD80 and CD86 (Figures 4A,B), demonstrating that the off-target activation of CAR T cells is predominantly driven by the CD28 interaction between CD80 and CD86. Importantly, CD28-TMD^M4 did not markedly change the off-target activation of either IgG₄-HD/CD28^WT or CD8-HD/CD28^WT CARs, demonstrating the inability of CD28-CAR heterodimers to respond to natural CD28 ligands.

We next examined the impact of CAR-CD28 heterodimerization on CD28 expression. Since lentiviral transduction resulted in a wide range of CAR expression levels that influenced on- and off-target T-cell activation, we expressed various CARs by knocking them into the TCR alpha constant (TRAC) gene locus using homology-directed repair that provided more homogenous expression of the CAR (Figure 5A) (28). Knock-in efficiencies ranged between 17 and 72% across the various CAR constructs, but the levels of the CAR expression were similar regardless of the differences in editing efficiency (Supplementary Figure 6A). In addition, all CAR T cells proliferated upon stimulation with CD19⁺ NALM-6 target cells (Supplementary Figure 6B), demonstrating that CARs containing M4 mutations remained functional. Six days after CAR knock-in prior to exposing the cells to target cells, we observed a 26–51% reduction in the CD28 mean fluorescence intensity on CAR⁺ T cells containing a wild-type CD28-TMD, but not an M4 CD28-TMD with either a CD8-HD or CD28-HD (Figures 5B,C). This reduction was seen in both CD4⁺ and CD8⁺ T cells with CARs containing either 28ζ or 4-1BBζ ICD. The downregulation of CD28 in CAR T cells engineered with an IgG₄-HD/CD28^WT-TMD was minimal, echoing the earlier result of inefficient CAR-CD28 heterodimerization in the context of IgG₄-HD.

Given the impact of the HD on CAR-CD28 heterodimerization, we next modeled the hinge-hinge interactions to better understand how the HD might influence the interaction of CD28 with CARs. As few TMD templates are available for modeling and their structures difficult to solve by nuclear magnetic resonance spectroscopy, we limited the modeling only to the extracellular domain of the CAR and CD28 receptors. The cysteine residue in the HD of the CD28 receptor C123 was aligned with the cysteine in the HD of CD28-HD-containing and CD8-HD-containing CARs (Figure 6). However, for IgG₄-HD-containing CARs, the cysteines in the HD could not be aligned with C123 of the CD28 (Figure 6). The modeling presented in the study demonstrated that the presence of a disulfide bound with endogenous CD28-HD is possible for the CD28-HD- and CD8-HD-containing CARs (Figure 6). However, as seen in Figures 5B,C, the lack of CD28 downregulation in CAR T cells with a CD28-HD and a M4-CD28-TMD suggests the CD28-HD alone was not sufficient to mediate the heterodimerization. Moreover, when various CD3⁻CAR⁺ T cells were stimulated with anti-CD3/CD28 beads, we did not observe enhanced CFSE dilution of T cells engineered with the CD28-HD/M4-CD28-TMD CAR construct when compared to CAR T cells with the CD8-HD/M4-CD28-TMD constructs (Supplementary Figures 7A,B). These results further support the notion that the cysteine bridge in the CD28-HD is insufficient to mediate CD28-CAR heterodimerization without interactions in the CD28-TMD. Taken together, these data suggest that cysteines and inter-molecular disulfide bonds in HDs are not the drivers of CAR-CD28 heterodimerization but can be involved in the stabilization of the CAR-CD28 heterodimers.