Programming CAR T Cell Tumor Recognition: Tuned Antigen Sensing and Logic Gating

The advent of T cell engineering and CD19 chimeric antigen receptor (CAR) therapy has opened a new field of cancer immunotherapy (1, 2). Premised on genetic programming and synthetic receptors for antigen, T cells can be targeted to antigens other than HLA–peptide complexes and repurposed at will for a multitude of tasks. CARs typically engage cell-surface molecules through immunoglobulin-derived polypeptides, such as a single-chain variable fragment (scFv) and single-domain heavy chains (VHH), or other ligand–receptor interactions. Upon target engagement, CAR signaling not only activates but augments T cell functions through composite signaling modules (3, 4). CAR T cells targeting CD19, a cell-surface antigen found in most lymphomas, leukemias, and also normal B cells, represent the paradigm for this cellular therapy (1, 5). A series of remarkable clinical results obtained a decade ago in patients with refractory B-cell malignancies, including non-Hodgkin lymphoma (NHL), chronic lymphocytic leukemia (CLL), and acute lymphocytic leukemia (ALL), fostered worldwide interest in CAR T cell engineering and resulted in the regulatory approval of CD19 CAR therapies within a few years (6–12). Rapid progress has since been made in developing similarly conceived CAR T cells for the treatment of refractory multiple myeloma (MM) by targeting B-cell maturation antigen (BCMA; refs. 13–16). There are, as of this writing, six FDA-approved CAR products, four of which target CD19 and two BCMA (Table 1), and over 1,300 clinical trials utilizing a CAR molecule listed on the ClinicalTrials.gov website.

The prevailing CAR structures in use comprise an activation domain, typically the cytoplasmic domain of the CD3ζ chain, fused to the cytoplasmic domain of either CD28 (17) or 4-1BB (18), two costimulatory receptors that affect activation strength, effector and proliferative capabilities, apoptosis, and metabolism in the engineered T cells (3, 19). We hereafter refer to these two canonical structures as 28ζ and BBζ. In the context of CD19 CAR therapy, they provide overall comparable clinical outcomes (20). Most CAR T cells are generated by transducing the CAR cDNA into autologous T cells using γ-retroviral or lentiviral vectors (21).

Although this CAR design and manufacturing approach has yielded extraordinary results in some refractory hematologic malignancies, the direct application of the same CAR T cell recipe to solid tumors has not been as compelling (19, 22). Satisfyingly, however, a number of limitations to the effectiveness or applicability of CAR T cells have been identified, including antigen escape, which is to be expected in the face of tumor heterogeneity, insufficient T cell persistence, T cell functional decline, on-target/off-tumor toxicities, and the elevated cost of manufacturing autologous CAR T cells (23–26). These limitations provide a road map for research efforts aiming to improve future CAR therapies for both hematologic and solid cancers.

We focus here on how to enhance tumor recognition by CAR T cells and produce such cells. The targeting of a single antigen is fraught with limitations and pitfalls, exposing to primary resistance or relapse of tumor cells that either fully lack or express low levels of the target antigen. Strategies to enhance recognition of targets of low abundance and to enable T cells to safely and effectively engage more than one antigen are direly needed.

Considerable knowledge has accumulated over the past years on the 28ζ and BBζ foundational CAR designs, both of which achieve impressive activity when targeting CD19 (1, 20). These two canonical structures nonetheless differ in how they support T cell function, including the induction and maintenance of effector functions as well as T cell persistence and metabolism, resulting in differences in CAR T cell kinetics and toxicity profiles. These features have been amply analyzed in multiple reviews (3, 19).

More recently, it has been recognized that 28ζ and BBζ CARs differ in their antigen sensitivity (Fig. 1A; refs. 27–30). In experimental models of ALL, Hamieh and colleagues observed leukemia relapses exhibiting reduced CD19 levels, which were more effectively treated with CD19.28ζ (19.28ζ) than CD19.BBζ (19.BBζ) CARs (27). In this model, the decreased CD19 levels measured at the time of relapse (<5,000 molecules per cell) were not due to the selection of CD19-null variants (31, 32) but were reversible and the consequence of CAR-mediated trogocytosis, a process whereby the target antigen is actively transferred from the tumor cell to the CAR T cell (27, 33, 34). When facing leukemia cells expressing 1,500 CD19 molecules per cell, 28ζ CARs showed superior in vivo leukemia control relative to BBζ CARs (27). When comparing the antigen sensitivity of 28ζ and BBζ CARs targeting either CD19 or GPC2, Majzner and colleagues and Heitzeneder and colleagues found that the 28ζ CAR better controlled in vivo tumor cells that expressed 2,000 CD19 or 6,000 GPC2 molecules per cell, respectively (28, 29). These findings are consistent with faster and larger changes in protein phosphorylation following antigen binding mediated by the 28ζ endodomain compared with BBζ (35). Although there may be exceptions depending on the epitope or the binder's affinity, these studies collectively establish that 28ζ CARs are better suited to target tumors with antigen levels <6,000 molecules per cell, whereas BBζ CARs are better suited for targeting tumor cells expressing high antigen levels and averting reactivity to normal cells only expressing few thousand target molecules per cell. Although 28ζ CARs have a lower target antigen density requirement, they too eventually struggle in vivo when target antigen density drops below 1,000 molecules per cell (36).

The sensitivity limits of canonical CARs raise the specter of tumor antigen escape and call for strategies to adapt CAR T cell design to specified ranges of target antigen density. Hereafter we discuss two promising remedies, one consisting in regulating the threshold antigen density for eliciting effective T cell activation and the other in generating T cells capable of productively engaging two or more target antigens.

CARs that fail to effectively control antigen-low tumors can be rendered more sensitive by modifying the CAR structural design or by parallel engineering to amplify their signaling. Thus, BBζ CAR sensitivity can be augmented by duplicating the CD3ζ chain segment to double the number of immunoreceptor tyrosine-based activation motifs (ITAM; ref. 28). Deleting ITAMs in BBζ and 28ζ CAR T cells decreases in vitro cytolysis of antigen-low tumors (28). Interestingly, however, point mutations in distal ITAMs in 28ζ do not diminish T cell function but to the contrary extend T cell persistence and delay terminal T cell differentiation in leukemia, pancreatic cancer, ovarian cancer, and melanoma models (37–39). The antigen threshold for eliciting cytolysis by BBζ CARs can be lowered to 5,000 molecules per cell by incorporating CD3ε (BB-εPRS-ζ) or the LCK binding site GRB2 (BBζ-GRB2; ref. 40). CAR components other than the signaling domain may also regulate the threshold of T cell activation. Thus, substitution of the CD8α hinge-transmembrane (H/TM) region of a BBζ CAR with the CD28-H/TM segment (17) further lowers the threshold for CAR reactivity, owing, at least in part, to the establishment of a more effective immunologic synapse (28, 29). Nonetheless, 28ζ CARs comprising CD28-H/TM still show superior control of antigen-low tumors relative to BBζ CARs reinforced by the CD28-H/TM domain (29).

CAR antigen sensitivity may also be modulated through the CAR's binding affinity (41). Cryogenic electron microscopy structural studies performed on the FMC63 and SJ25C1 CD19-specific scFvs have pinpointed contact residues with CD19 and guided the rational design of 28ζ CARs with increased or decreased sensitivity to CD19 (42).

CAR antigen sensitivity may also be increased without a structural modification by augmenting downstream activation signaling. Genome-wide CRISPR knockout identified RAS GTPase–activating protein (RASA2) as a checkpoint in T cells. Ablation of RASA2 expression in CAR T cells enhanced MAPK signaling and T cell cytolytic activity against a range of CD19-low tumors (43).

Much like the physiologic T cell receptor (TCR), CARs can display enhanced sensitivity by increasing their functional avidity for their target cell. This may be achieved through antigen-specific or antigen-agnostic pathways. The coexpression of a companion scFv in a CAR T cell can lower the antigen threshold for tumor lysis imparted by the CAR. Thus, expressing a cell-surface binder for CD38, with or without a signaling domain, lowers the threshold for cytolysis of BCMA or CD19 CARs in the face of low-abundance target antigen (44). Expression of ICAM-1 in T cells and their target has also been found to stabilize CAR synapse formation, much like for the TCR (45), enabling better lysis of some solid tumors (46).

Although tumor lysis in vivo is affected by additional tumor-intrinsic and microenvironmental factors, calibrated in vitro cytoxicity assays provide a useful metric for characterizing CAR T cell-intrinsic tumor recognition properties. Using tumor cells with graded CD19 levels, we further found that CAR T cells at an effector:target (E:T) ratio of 2 were more likely to lyse tumor cells with low target abundance than at a ratio of 1, suggesting the potential for cooperative killing when higher intratumoral E:T ratios are achieved (27).

Virtually all CARs in present clinical development operate independently of the multimeric CD3 complex, which provides the signaling machinery to support TCR-mediated antigen recognition and signaling (47). In αβ-T cells, the antigen-binding αβ heterodimer associates with the signaling CD3 subunits δε, γε, and ζζ, which in aggregate provide 10 ITAMs per αβ heterodimer. In vitro T cell activation occurs upon the engagement of only a few cognate MHC–peptide complexes, underscoring the formidable signal transduction afforded by the TCR–CD3 complex (48). We investigated whether incorporating the V_H and V_L regions used in a sensitive 19.28ζ (27) into a TCR–CD3 complex increases antigen sensitivity relative to the V_H–V_L matched CAR. We generated an HLA-independent TCR termed HIT receptor (Fig. 1A and B) by substituting chimeric V_L–Cα and V_H–Cβ chains for the endogenous TCR (36). Peripheral blood T cells expressing the HIT receptor displayed ∼10-fold greater antigen sensitivity than matched 28ζ CAR T cells and achieved control of leukemia expressing ∼200 CD19 molecules per cell in a mouse xenograft model. Using a comparable HLA-independent TCR comprising V_H–Cα and V_L–Cβ subunits called synthetic T cell receptor and antigen receptor (STAR; Fig. 1B), Liu and colleagues showed greater antigen sensitivity compared with 28ζ and BBζ T cells containing the same V_L and V_H elements (49), although CAR target densities were not quantified. Birtel and colleagues recently reported a TCR-like receptor containing V_H–V_H–Cα and V_L–V_L–Cβ subunits called TCAR (TCR-like CAR; Fig. 1B; ref. 50), which suggested increased sensitivity compared with a 28ζ and BBζ in terms of IFNγ secretion.

Even though chimeric receptors based on the TCR were proposed over 30 years ago (51), interest in CD3-based CARs is recent. In addition to the HIT and STAR receptors, TCAR, antibody-TCR (AbTCR; Fig. 1B), TCR fusion constructs (TRuC; Fig. 1B), and T cell antigen coupler (TAC; Fig. 1B)receptors also co-opt the CD3 complex in different configurations. In the AbTCR, V_L is fused to the TCRγ constant region and V_H to that of TCRδ (52). The TAC receptor fuses two scFvs with the CD4 hinge, transmembrane, and intracellular domains (53), interacting with a target cell antigen through one scFv and the ε subunit of the endogenous TCR–CD3 complex via the second scFv. The ε-TRuC receptor is formed via the incorporation of an overexpressed scFv–CD3ε fusion into the endogenous TCR–CD3 complex (54). The ε-TRuC receptor is the only of these to have been investigated for its antigen sensitivity, which showed sensitivity to be superior to a 28ζ CAR but inferior to the STAR receptor, all bearing the same antigen specificity (55). Altogether, the above studies suggest that some CD3-based receptors have superior sensitivity to canonical CARs, which increases their ability to lyse targets with low antigen densities, on the order of a few hundred in the case of HIT receptors. Importantly, the HIT receptor is the only one to replace the endogenous TCR, whereas all others generate dual-specific T cells as they retain their TCR.

Simultaneously targeting more than one antigen is a potential remedy to prevent antigen escape, but how to do so effectively remains to be determined. One approach is to infuse multiple T cell products, each one targeting a different antigen; another is to engineer multispecific T cells.

The administration of multiple validated CAR T cells either simultaneously or sequentially is attractive for mixing and matching different cell products. Several clinical trials, mostly in B-cell malignancies, have shown the feasibility of these approaches with an acceptable safety profile (56). However, long-term responses have not improved substantially relative to single-cell product infusion, corroborating preclinical findings that predicted this outcome, especially when targeting antigens of low abundance or using less sensitive CAR designs (27, 57–59). The coinfusion of combined CAR T cell products will require further investigation of infusion dose, order, and timing and careful evaluation of synergy or dominance. These challenges support the rationale for engineering multispecific CAR T cells, especially when individual targets are of low abundance.

The first embodiment of concomitant antigen targeting, in which either one in a pair of antigens can serve as target, is defined as an OR-gate in Boolean terminology. This goal may be achieved by constitutively coexpressing two independent CARs (hereafter referred to as Dual-CAR; Fig. 2A) or by expressing a single bivalent CAR comprising tandem binding domains (hereafter referred to as Tan-CAR; Fig. 2B). How do these approaches compare, and are there emerging rules for effective multitargeting?

This approach is attractive owing to the facility of coexpressing two independent, functionally validated CARs in T cells. Using a bicistronic vector encoding CD19 and CD123 BBζ CARs, Ruella and colleagues demonstrated the superiority of the Dual-CAR over pooled CD19 and CD123 CAR T cells (57). Another advantageous feature of Dual-CARs is to mix and match costimulatory domains. In leukemia models of antigen escape, dual targeting with two CARs respectively targeting CD19 and CD22 not only outperformed the sequential infusion of CD19 or CD22 monospecific CAR T cells but was superior when combining a 28ζ and a BBζ CAR as compared with coexpressing two 28ζ CARs and, even more so, two BBζ CARs. Moreover, assigning the 28ζ CAR to CD19 and the BBζ CAR to CD22 resulted in the most effective treatment (27). Shalabi and colleagues corroborated the superiority of combining 19.28ζ and 22.BBζ (60).

The Tan-CAR design requires careful examination of scFv order, hinge domains, and linker length to preserve efficient recognition of both antigens. Thus, cotargeting CD19 and CD22 with an optimal loop structure 19V_L–22V_H–22V_L–19V_H to design Tan-19.22.BBζ allowed better T cell/tumor cell conjugate formation and provided higher tumor control than other variable chain orientations (58). Cotargeting CD19 with a short extracellular domain (EC) and CD20 with an extended EC (Tan-19.20 CAR) was more efficient at controlling heterogeneous tumors in vitro and in vivo, further underscoring the need to adapt Tan-CAR design to antigens topology (61). In an MM model, De Larrea and colleagues combined BCMA and GPRC5D CARs to target a heterogeneous tumor mix including a BCMA^−/− tumor fraction and found that pooled BBζ and Dual-BBζ CAR T cells outperformed their Tan-BCMA.GPRC5D.BBζ CAR (62). In preclinical models of glioblastoma, HER2.28ζ and IL13Rα2.28ζ Dual-CAR T cells outperformed pooled monospecific CAR T cells (63) but were later superseded by an IL13-mutein/HER2.scFv Tan-CAR that allowed for more efficient synapse formation and superior in vivo tumor control (64).

In a recent study, Leung and colleagues found that CD19 and CD79a cotargeting with Tan- or Dual-BBζ CARs superseded pooled monospecific CARs in preventing antigen escape (65). Despite improved tumor control in vivo, the Tan-CAR configuration compromised antigen binding compared with monospecific CARs, consistent with previous observations (61, 66), whereas Dual-CAR exhibited diminished signaling, possibly owing to competition for downstream signaling molecules (65). Dual-CARs that employ a common hinge/transmembrane component may heterodimerize and function as parallel CARs (see below).

Altogether, the Dual-CAR and Tan-CAR studies support the benefit of OR-gate combinatorial antigen targeting to offset antigen escape, but no single approach has proven to be uniformly superior. Antigen topology and density on tumor cells, together with CAR structural design, including binding domain affinity, hinge length, and signaling domain selection, are all parameters that require careful optimization.

Several Dual-CAR and Tan-CAR T cells are already in the clinic in the setting of B-cell malignancies and MM (Table 2), with more to begin clinical investigation in solid tumors (56). Although published clinical trials are still few and lack single-target comparisons, the feasibility and safety profile of dual-targeting are encouraging, even though early results have not proven so far to be superior to historical controls obtained with single-antigen targeting (Tables 1 and 2; refs. 56, 67, 68). Despite its superiority to pooled and Dual-CARs in preclinical models, Tan-19.22.BBζ CAR treatment has still resulted in disease relapses, which have been pinned on suboptimal CD22 responses (58, 60, 66). The single configuration of BBζ may not be best suited to target tumors with low abundant antigen densities, particularly in patients who manifested a low level of CD19 after previous axicabtagene ciloleucel treatment. Shalabi and colleagues, thus, proposed to adopt dual 19.28ζ and 22.BBζ for future clinical trials (27, 60). Tumor relapses have also occurred following Tan-19.20 CAR treatment in B-cell lymphoma, although owing to poor in vivo T cell expansion rather than antigen downregulation (69).

In ALL patients, Cordoba and colleagues evaluated Dual-CAR T cells expressing a 19.OX40ζ CAR and a novel 22.BBζ CAR using a pentameric coiled-coil hinge domain to increase sensitivity to CD22 (70). Despite encouraging initial remissions, 70% of the patients relapsed within a year of the treatment, attributed mainly to lack of T cell persistence but also CD19 loss.

Interestingly, the biepitope targeting of BCMA via two tandem VHH elements (Fig. 2B) and Tan.BCMA.CD38 CAR has shown high complete response rates (Table 2; refs. 15, 71), leading to the former's rapid FDA approval for adults with relapsed or refractory MM (Tables 1 and 2). Relapses occurring in both clinical trials have been attributed to insufficient T cell persistence (15, 71).

Altogether, these early trials on dual-targeted T cells in hematologic malignancies are encouraging but highlight the need to stringently optimize antigen binding and CAR signaling to achieve the required antigen sensitivity while providing a sustained tumor response.

Engineering trivalent CAR T cells has been achieved using tricistronic vectors to target HER2.IL13Ra2.Epha2 or CD19.CD22.CD20 (72–74). From a technical standpoint, large vector inserts encumber efficient gene delivery, whereas cotransduction of single CAR vectors generates T cell mixtures with variegated specificity and an increased risk of insertional mutagenesis. Polypeptides that bind antigens with high affinity, known as designed ankyrin repeat proteins (DARPin), have been successfully employed instead of scFvs to generate less bulky trispecific CARs targeting EGFR, EpCAM, and HER2 (Fig. 2B; ref. 75). In another strategy, ligands that bind to several receptors on tumor cells can be incorporated into CARs, as recently shown using a BAFF–ligand CAR, which engages BAFFR, TACI, and BCMA, although the efficacy of this design compared with conventional scFv-based CARs is still unknown (76).

Another OR-gate design is based on the heterodimerization of two chains, wherein a CAR chain specific for one antigen dimerizes with a chimeric costimulatory receptor (CCR) chain specific for a second antigen facilitated by the intentional choice of the same H/TM domains for each receptor (Fig. 2C). Thus, MUC-1.28ζ+ERbb2.28, GD2.28ζ+B7H3.BB, or MSLN.28ζ+CSPG4.BB provided optimal tumor-dependent costimulation and enhanced tumor control in several models of solid tumors, further supporting the benefit of providing both CD28 and 4-1BB support to engineered T cells (77, 78).

Modular Zip-CAR systems that utilize a leucine zipper adapter expressed on the T cells to bind to Zip-Fv (Zip-scFv) offer a versatile OR-gate tool that can be adapted to several antigens, whereas the dose and binding strength of Zip-Fv can be tuned to regulate T cell activity or can be terminated by a competitive nonantigen-binding Zip-Fv (Fig. 2D; refs. 79, 80). Multiantigen targeting can also be achieved by engineering CAR T cells to secrete a dual-engager protein [e.g., EGFRvIII-CAR and EGFR-bispecific T cell engager (BiTE); Fig. 2E; ref. 81].

Altogether, a rapidly expanding number of studies demonstrate the potential of OR-gates to broaden antigen coverage and mitigate antigen escape. However, how to effectively achieve multiantigen recognition and ensure functional CAR T cell persistence has not yet been established. Moreover, the repurposing of an intact TCR–CD3 complex to increase antigen sensitivity and the targeting of multiple antigens via an OR-gate may increase the risk of on-target/off-tumor toxicity, thus requiring careful antigen selection, especially in solid tumors in which target antigens are often shared with normal tissues, and suitable preclinical modeling.

CD19 was selected as a CAR target in part for its limited expression in normal cells (5), which is largely confined to B cells and thus exposes CD19 CAR T cell recipients to B-cell aplasia, a clinically manageable on-target/off-tumor toxicity. Targeting CD20, CD22, and BCMA likewise exposes recipients to restricted on-target tissue damage. However, many potential CAR targets, especially in solid tumors, are expressed in cell types that cannot sustain an immune assault, calling for the need to detarget those cell types. Short of identifying tumor-specific epitopes, one needs to design T cells that preferentially engage tumor cells and spare normal cells. Differential levels of expression alone provide, in some instances, a therapeutic window that can be exploited to spare low-antigen-expressing normal cells based on CAR antigen sensitivity alone. T cell engineering, however, opens many more opportunities for discerning malignant from normal cells. An early embodiment of this concept was to design two antigen-reactive receptors, respectively, specific for antigens A and B that alone did not lyse normal cells expressing either A or B but in concert killed tumor cells that expressed both A and B (82).

Such combinatorial strategies to design discriminatory, tumor-selective T cells are now flourishing and can be classified using Boolean terminology (Fig. 3). We focus here on strategies that require combinatorial antigen input to direct T cell activity. The OR-gate strategies (Fig. 3A) reviewed above are only one of several ways to gate tumor recognition. Other approaches to manage T cell specificity and limit on-target toxicities such as remote controls, suicide genes, or binding domain affinity tuning are reviewed elsewhere (83–85).

One tumor-targeting concept that aims to create tumor specificity from a set of nonspecific targets is based on the Boolean AND-gate (Fig. 3B). In this instance, productive T cell activation depends on combined inputs emanating from two antigens coexpressed in the tumor but not in normal cells (the latter may express either one alone). Each separate input must therefore not suffice to trigger target cell lysis but exceed an activation threshold upon their simultaneous engagement. One early embodiment was exemplified by Kloss and colleagues, who combined a defective ζ chain–based CAR specific for prostate-specific membrane antigen (PSMA) with a CCR with CD28 and 4-1BB endodomains to rescue the poor activation signal provided by the CAR (82). A related design combining a first-generation CAR lacking a costimulatory endodomain with a CCR providing costimulation has been touted as an AND-gate (86–88) but rather functions as an IF-BETTER-gate (see below). Under certain conditions, IF-BETTER- and OR-gates may operate as an AND-gate, but an authentic AND-gate requires minimal or absent T cell activation in the presence of either target alone.

Tousley and colleagues recently described the Logic-gated Intracellular NetworK CAR (LINK-CAR), wherein the split CD3ζ and costimulatory domains are replaced with LAT and SLP-76 molecules. To optimize LINK-CAR specificity, point mutations were introduced into LAT (del171–233) and SLP76 (del224–244) to minimize the engagement of the adapter protein GADS essential for LAT and SLP76 interaction. In a proof-of-concept model targeting CD19 and ROR1 in NALM6 cells, only the optimized LINK-CAR achieved tumor control in the absence of ROR1 on-target/off-tumor toxicities in a mouse model. In this study, the LINK-CAR was less toxic than a CD19-synthetic Notch receptor (SynNotch)→ ROR1-CAR design (see IF-THEN-gate, below), whereas the split CAR (ROR-1.ζ and CD19.28) was ineffective (89).

AND-gates might be achieved using modular Zip-CAR strategies if CD3ζ and CCR modules can be suitably affinity-tuned (80), but their in vivo safety and efficacy remain to be determined. Stringent AND-gates are challenging to achieve and have not yet been tested clinically, but their potential to unleash T cell potency limited to tumor cells is attractive.

This form of gating depends on an inhibitory CAR (iCAR) to turn off CAR T cell activity upon encountering unintended target cells. In this instance, antigen A on tumor cells and normal cells is targeted by an activating CAR but is impeded upon engaging antigen B that is present on the normal cells only (Fig. 3C). As a proof of concept, Fedorov and colleagues engineered a PSMA.iCAR incorporating the endodomain of the inhibitory molecule PD-1 to reversibly restrict CD19 CAR T cell activity against CD19⁺ PSMA⁺ cells without interfering with antitumor CD19 CAR activity against CD19⁺PSMA⁻ cells (90). PD-1–based iCARs also abated TCR-mediated allogeneic response with an iCAR directed to HLA molecules (90). In another example, Richards and colleagues combined CD93 targeting with a CD19.iCAR based on PD-1 or TIGIT inhibitory domains to spare CD93⁺CD19⁺ cells (91). Exploiting the frequent loss of HLA in some cancer cells, NOT-gates utilizing a PD-1–based iCAR against HLA-A*02:01 or A*03:01 have been paired with a 28ζ CAR targeting a different HLA allele retained by tumors (92). A related approach has been devised to selectively target CEA⁺/HLA⁻ tumor cells and spare HLA⁺ cells using the inhibitory leukocyte Ig-like receptor 1 (LIR1) that binds to the HLA-A*02 molecule (93).

Much like AND-gate, the level of expression of the CARs and the antigen pairing are critical for optimal NOT-gate function, which depends on a balance between activation and inhibition strength of signaling (90). Future studies using tuned receptor affinity, balanced CAR and iCAR signaling, and optimal target selection may allow for the translation of NOT-gates to the clinical setting.

Spatiotemporal regulation of CAR expression is an attractive concept to restrict a conditionally expressed CAR that poses safety concerns. In this instance, engagement of antigen A by a CAR or another sensor triggers transient expression of a CAR specific for antigen B, thus restricting B CAR expression to the tumor (space) but not beyond when (time) the T cell reaches a normal tissue (Fig. 3D). The Lim lab combined the use of a cleavable Notch receptor with the induction of CAR expression under the control of a synthetic transcription factor (TF) released upon cleavage of SynNotch (94, 95). The released TF then binds to a responsive promoter located upstream of a specific transgene such as CARs, thus making this system an IF-THEN-gate. Using such an on-off switch circuit, the authors showed that CAR expression is only induced upon activation of the SynNotch receptor leading to the eradication of tumors coexpressing the SynNotch and CAR ligands (96, 97). This approach has since been widely applied to a range of preclinical mouse models to treat, for example, mesothelioma coexpressing ALPPL2 and MCAM antigens (ALPPL2-SynNotch and MCAM-CAR) or glioblastoma expressing either EphA2 or IL13Ra2 present in the vicinity of EGFRvIII⁺ tumor cells (SynNotch-EGFRvIII and EphA2.IL13Ra2 Tan-CAR; refs. 98, 99). In these models, SynNotch-regulated CAR expression outperformed single CAR strategies by limiting tonic signaling leading to less exhausted and long-lived memory T cells (98, 99). Hernandez-Lopez and colleagues further harnessed the SynNotch system to increase specificity for HER2⁺ tumors devising a low-affinity HER2-SynNotch that gated expression of a high-affinity HER2 CAR, showing effective in vivo induction by tumor cells expressing HER2 in the range of 10⁷ molecules per cell but not 60,000 molecules per cell (100). The effectiveness of this sensing system at low antigen densities remains to be evaluated.

As mouse SynNotch and synthetic TFs may be immunogenic, Zhu and colleagues described humanized SynNotch-like receptors named SyNthetic Intramembrane Proteolysis Receptors (SNIPR) by assembling a library of human EC, transmembrane (TMD), and juxtamembrane domains (JMD) that allow for ligand-dependent release of orthogonal synthetic TF to deliver therapeutic payloads such as IL2 or CARs. Combination of human CD8α-EC, Notch-TMD/JMD, and HNF1A (DNA-binding domains) fused to the transactivation domain of human NF-κB p65 induced CAR expression similarly to conventional SynNotch, successfully eradicating tumors expressing antigens in preclinical animal models (101). The SynNotch system can also be adapted to function as a NOT-gate (OFF-Notch) by inducing proapoptotic factor tBID (truncated BH3-interacting domain death agonist) upon off-tumor recognition, leading to rapid T cell death. Although this system curbs toxicity, it functions similarly to suicide genes or antibody-mediated depletion leading to T cell loss, which also curtails therapeutic efficacy (102).

Antigen selection for IF-THEN gating requires careful pairing, as the kinetics of CAR induction and decay play a critical role in avoiding on-target/off-tumor toxicities. Srivastava and colleagues found that EPCAM-SynNotch–regulated ROR1-CAR T cells failed to protect ROR1⁺ normal stroma cells when tumors were in close proximity or disseminated (103). In a different antigen pairing, controlling ROR1-CAR with SynNotch B7H3 (B7H3 is absent on normal ROR1⁺ stroma cells) was efficient at mitigating on-target/off-tumor toxicities (103). No SynNotch system is yet in the clinic, but the recent development of humanized systems and further kinetic optimizations are poised to accelerate the clinical application of this technology.

An IF-BETTER-gate is one in which a CAR-engaging antigen A performs better in the presence of antigen B. B is not obligate for CAR T cell function (for example, if A is abundant) but the presence of B helps activation following recognition of A when the latter's abundance is limiting (Fig. 3E). IF-BETTER thus differs from the obligate dual-requirement for A and B in an AND-gate and from the temporally regulated IF-THEN-gate. Recognition of B is not mediated by an activating receptor and may consist in a CCR or simply a cell-surface-anchored scFv. It, therefore, does not incur the potential on-target/off-tumor toxicity imparted by combining a second CAR, as is done in an OR-gate. Wilkie and colleagues coexpressed a ζ chain–based CAR (HER2.ζ) and a MUC1.28 CCR to target HER2-high but not HER2-low cell lines. Complementary signaling from CAR and CCR upon dual-antigen recognition showed enhanced in vitro cytokine secretion and T cell proliferation (88). A similar concept was applied by combining a mesothelin CAR and a Folate receptor CCR to target ovarian cancer or combining a CD13 CAR and a TIM3 CCR to target acute myeloid leukemia (AML). With the CAR alone, tumors were transiently controlled in vivo, in contrast to the conditions where CAR and CCR antigens are present (83, 84). Katsarou and colleagues combined a fully functional 28ζ CAR targeting either BCMA or CD19 with a CCR binding to CD38, an antigen highly expressed in B-cell malignancies but also present in normal immune cells including T cells (44). In order to not direct cytolysis to cells expressing antigen B, the CAR and CCR used to create an IF-BETTER-gate must differ structurally so as to not allow heterodimerization as in the parallel CAR design (77, 78). Use of a CCR rather than a cell-bound scFv supports CAR signaling and T cell persistence through its costimulatory function, which should be selected to complement the paired CAR (44). Thus, combining a CD19 or BCMA.28ζ CAR with a high-affinity CD38 CCR increased cytokine production, T cell persistence, and in vivo control of tumors, including CD19⁺ leukemia with <1,000 molecules per cell (44).

IF-BETTER-gates yield a CAR T cell that can modulate antigen sensitivity to A in the presence of B and thus imparts T cell “preference” for cells expressing B, but with less stringency than an AND-gate and allowing broader target selection for B than an OR-gate. IF-BETTER-gates have only been reported for preclinical B-cell malignancies and myeloma models but may be especially useful for AML and solid tumors, for which highly expressed antigens with restricted systemic expression are scarce.

Gating strategies that aim to augment CAR T cell tumor selectivity are a promising approach to mitigate unwanted on-target/off-tumor toxicities. These strategies are often initially tested in proof-of-principle models wherein tumors express abundant antigen levels and animals lack faithful tumor microenvironments. Defining clinically relevant antigen pairs and realistic parameters for their targeting (e.g., level of antigen expression on tumors vs. normal tissues, scFv affinity, tumor accessibility and tumor environment), particularly in solid tumors, are important follow-up studies to guide successful clinical development.

The successful introduction of more complex T cell engineering strategies in the clinic will also depend on advances in T cell manufacturing. At present, CAR T cells are most often manufactured using viral vectors (21). After chromosomal integration, CAR expression is driven by the 5′ long terminal repeat in γ-retroviral vectors or by an exogenous promoter, usually long EF1α, in lentiviral vectors. Although CAR expression is variegated in T cells due to the semirandom integration pattern (104), CAR T cells have shown remarkable therapeutic activity against hematologic malignancies using either γ-retroviral or lentiviral vectors (Table 1). Nonviral approaches have also been used to generate clinical-grade CAR T cells, primarily utilizing the Sleeping Beauty or piggyBac transposons (105, 106). Here, T cells are electroporated with two plasmids, one encoding a transposase and the other the transposon that encodes the CAR, also resulting in variegated CAR expression owing to the semirandom integration of transposon DNA (107).

The benefit of tightly controlling CAR expression has been demonstrated by targeting the CAR cDNA to the TCR alpha constant (TRAC) locus using sequence-specific chimeric nucleases (108). Integration of the 19.28ζ cDNA in exon 1 of the TRAC locus improved antitumor T cell activity relative to retroviral-encoded CAR, owing to reduced tonic signaling and delayed T cell differentiation upon repeated antigen stimulation (Fig. 4A; ref. 108). TRAC-CAR T cells that control CAR expression from the endogenous TCRα promoter show homogeneous and consistent CAR expression levels across multiple T cell donors, which contrasts with virally modified CAR T cells (108). Moreover, due to high TCR knockout (KO) efficiency, TRAC-CAR T cells could be used in both autologous and allogeneic settings. The superior antitumor activity of TRAC-CAR versus retroviral CAR T cells has also been shown in a syngeneic mouse model (109). Alternatively, CAR cDNAs have also been integrated at other loci under control of either endogenous or exogenous promoters. Zhang and colleagues further found that CAR T cells bearing an EF1α-controlled 19.BBζ CAR transcription unit integrated into the PDCD1 exon 1 show improved antitumor activity compared with lentivirally modified CAR T cells (Fig. 4B; ref. 110). Another key feature of this design is the reduction of PD-1 expression, which may decrease CAR T cell exhaustion and enhance functional T cell persistence. However, not all integrated CAR transcription units lead to optimal antitumor activity (108, 111, 112), requiring careful evaluation of promoter selection at any given chromosomal site.

T cells expressing CD3 complex–based receptors have been engineered using site-specific integration (HIT receptor) or lentiviral/γ-retroviral vectors (STAR, TCAR, AbTCR, ε-TRuC, and TAC receptors). To engineer HIT T cells, we relied on our TRAC-CAR strategy (108) to insert the V_H–Cβ–P2A–V_L–TRAC exon1 donor sequence to express the chimeric V_H–Cβ and V_L–Cα chains under the control of the endogenous TCRα promoter (36). Importantly, this strategy also results in the elimination of the endogenous αβ TCR, thus abolishing the potential alloreactivity of HIT T cells and the competition between chimeric and endogenous TCRs. As for STAR, TCAR, and AbTCR T cells, the chimeric receptors are expressed by using lentiviral/γ-retroviral vectors and thus show variegated expression (49, 50, 52). To circumvent genetic disruption of the endogenous αβ TCR, these strategies made use of different C domains to minimize TCR chain mispairing. The STAR receptor contains mutated mouse Cα and Cβ regions (49), the AbTCR receptor contains human Cγ and Cδ domains (52), and the TCAR receptor contains human Cα and Cβ regions fused to V domains in tandem (V_H–V_H–Cα and V_L–V_L–Cβ; Fig. 1B; ref. 50). These modifications do not eliminate the interaction of the endogenous αβ TCR and CD3 complex, which is needed to assemble STAR, TCAR, and AbTCR receptors at the surface; this also leads to the retention of their potential alloreactivity. In ε-TRuC and TAC T cells, the scFv–CD3ε and scFv1–scFv2–CD4 fusions, respectively, are overexpressed using lentiviral vectors (53, 54). These fusions do not contain TCR elements and therefore depend on expression of the endogenous αβ TCR to properly assemble at the surface. This results in ε-TRuC and TAC T cells possessing dual specificity and potential alloreactivity as well.

Many of the above studies underscore the broad potential of site-specific genome engineering to develop improved CAR T cells for a variety of purposes (113, 114). For example, TRAC or TRBC KO eliminates alloreactivity of CAR T cells (115); PDCD1 KO may reduce T cell exhaustion (116); CD52 KO eases the use of TRAC KO-CAR T cells for allogeneic applications in combination with anti-CD52 antibodies to deplete host T cells (117–119). However, gene editing approaches can lead to genetic abnormalities in T cells. Triple CRISPR/Cas9-mediated KO ablating TRAC, TRBC, and PDCD1 can induce frequent chromosomal translocations (120). Translocations have also been detected when targeting TRAC and CD52 loci with either TALEN or CRISPR/Cas9 (119, 121), and aneuploidy when editing three loci with CRISPR/Cas9 RNPs (122). These observations put a note of caution when multiplexing double-strand breaks and increase interest in alternative DNA editing methods based on CRISPR/Cas9 nickases, such as base editing and prime editing (123).

Development of CAR T cells with additional functionalities will require more complex genetic modifications. Though γ-retroviral and lentiviral are sufficient to deliver AND, OR, IF-THEN, and IF-BETTER gate constructs, and other simple genetic constructs, their limited cargo capacity will constrain delivery of more sophisticated circuits. Transposon-mediated delivery could potentially address this limitation. However, approaches requiring homogeneous gene expression may not be advanced by this method. PASTE technology enables integration of large cargos (up to ∼36 kb) into specific loci in human cells, including primary T cells. The platform uses a CRISPR/Cas9 nickase fused to both a reverse transcriptase and a serine integrase, a pegRNA, and a minicircle cargo plasmid. The pegRNA contains the serine-integrase site, which is inserted into the target DNA via prime editing; then the serine integrase introduces the cargo plasmid at the integrase site (124).

Another important variable for effective T cell engineering is the selection of loci or extragenic chromosomal regions where not only CAR genes but more complex genetic circuits could be integrated and reliably expressed (125). We recently identified an extragenic region on human chromosome 7, termed GSH6, that supports CAR expression as effectively as the TRAC locus (126). With these novel gene editing/targeting tools to target gene loci or extragenic genomic safe harbors, one can expect rapid progress in engineering T cells endowed with precisely calibrated functions to achieve greater T cell specificity and potency.

Integrating next-generation CAR T cell designs targeting multiple antigens together with the progress in understanding the basis for T cell differentiation states at the transcriptional and epigenetic levels should inform improved methods to generate better autologous and allogeneic CAR T cells. In concert with fine-tuning combinatorial receptor signaling, novel manufacturing processes can be adopted to modulate transcriptional, epigenetic, and metabolic pathways.

From a production standpoint, Dual- and Tan-CARs expressed from a single vector maintain the cost-effectiveness and relative ease of manufacturing of single-CAR T cell products, averting the need to produce two cell products or two separate vectors and ensuing product heterogeneity. Cell dosing in Tan-CAR T trials tends to mimic those dispensed in single-CAR T trials (15, 60, 66, 69, 70, 127). New CAR designs (37, 108) and rapid manufacturing platforms (128), however, offer the prospect of lowering effective cell doses. For example, therapeutic doses as low as 25 × 10⁶ autologous CAR T cells were recently reported by Park and colleagues in adult patients with relapsed/refractory (R/R) diffuse large B-cell lymphoma (NCT04464200), using a human CD19–targeted calibrated 19.28ζ-1xx (37). The dose-expansion phase of this trial is now proceeding with only 25 × 10⁶ CAR T cells per infusion (129). Svoboda and colleagues reported therapeutic doses in the range of 3 × 10⁶ to 30 × 10⁶ CAR T cells in patients with R/R NHL (NCT04684563) using IL18 secreting autologous 19.BBζ CAR T cells (huCART19-IL18; ref. 130). The FasT CAR-T (F-CAR-T) manufacturing platform in which T cell culture time is less than 24 hours was evaluated in two phase I clinical studies in patients with B-ALL using either CD19 or 19.22 Dual-CAR T cells (NCT03825718 and NCT04129099). F-CAR T cells were successfully manufactured, and cell doses as low as 0.3 × 10⁵ to 1.5 × 10⁵ CAR⁺ T cells/kg enriched in stem cell memory T (T_SCM) and central memory T (T_CM) cells as shown in preclinical studies were infused in 25 patients (NCT03825718). The safety of this approach, its neurotoxicity in particular, as well as its efficacy, will need to be further evaluated, as most responders to CD19 F-CAR T therapy subsequently received allogeneic hematopoietic stem cell transplantation therapy (131, 132). The ability to rely on low cell infusion doses raises the prospect that blood draws rather than leukapheresis products may be sufficient to manufacture autologous T cell products. Increased T cell potency and shorter culture times that require reduced amounts of consumables, reagents, and manpower bestow logistics and financial advantages that could broaden patient access, provided that release testing with adequate sensitivity can be adapted to processes necessitating a low number of cells.

CAR T cells derived from naive T cells as well as CAR T cells endowed with T_CM and T_SCM phenotypes may further increase antitumor efficacy by providing greater functional persistence. Biasco and colleagues analyzed CAR T cell phenotypes in pre- and postinfusion samples and established the critical role of T_SCM in mediating early antileukemic responses and long-term persistence of CAR T cells (133). Enriching for T_CM cells and T_SCM cells prior to ex vivo expansion or adoptive transfer can also improve high frequencies of persistent cells with stem cell–like characteristics. Arcangeli and colleagues have shown that CAR T cell manufacturing from naive/stem memory T lymphocytes enhances in vivo antitumor responses in a leukemia-bearing humanized mouse model while curtailing cytokine release syndrome (134). Cytokines such as IL7, IL15, and IL21, small-molecule treatments, and antioxidants such as N-acetylcysteine (NAC) have the potential to enhance the ex vivo maintenance of T_CM and T_SCM cell subsets and to enable the expansion of more potent antitumor T_SCM cells (NCT04464200; refs. 129, 135). Further studies will be required to evaluate the relevance and functionality of T_SCM cells in patients with cancer as well as the optimal conditions for their manipulation.

Building on the approval by the FDA of six CAR T cell therapies since 2017, manufacturing platforms are evolving toward closed and automated systems (recently reviewed in ref. 127) in order to bolster reproducibility and patient access. Control of T cell activation is key. Interestingly, Shalabi and colleagues (60) and Spiegel and colleagues (66) reported that earlier removal of Transact beads during ex vivo expansion shortens the time to reach the cell dose, suggesting that prolonged CD3/CD28 activation is detrimental. In addition, Ghassemi and colleagues have recently shown that this step is dispensable in a short manufacturing protocol (128).

Cryopreservation enables the storage and distribution of drug products. The efficacy of cryopreserved CAR T cells was demonstrated to be comparable to fresh CAR T cells upon measuring in vivo expansion, persistence, incidence of toxicities, and disease response (136). Based on this study and data from two clinical trials (NCT02315612 and NCT03448393), CAR T cells can be cryopreserved without altering their functionality, providing greater flexibility for scheduling infusions and delivery to CAR T cell administration sites.

Cargo delivery through viral vectors can be complemented or replaced by gene editing platforms. In a clinical trial for R/R B-cell NHL, Zhang and colleagues used a homology-directed repair template in the form of a linear dsDNA containing the CD19 CAR to target the CAR cDNA to the PDCD1 locus (Fig. 4), achieving a high rate of complete remission (87.5%) and durable responses without serious adverse events (ref. 110; NCT04213469). Gene editing is attractive in the context of multiplexed genetic modifications involving KO and knockin. Clonal screening following induced pluripotency reprogramming and multiplexed editing allows for selecting safe T cell reservoirs without translocations, aneuploidy, or mutations in addition to enabling mass production of CAR T cells (137). Finally, the emergence of in vivo T cell engineering opens a new set of possibilities and challenges. Polymer or lipid nanoparticles, retroviral particles, and redirected viral vectors derived from HIV-1 are being tested for the in vivo generation of CAR T cells. Their clinical application will require optimized delivery and close monitoring of off-target effects (138–140).

The CD19 CAR therapy paradigm has spawned a torrent of CAR T cell innovation with potential applications in virtually any cancer. These endeavors will need to overcome the common challenge of antigen expression heterogeneity. Furthermore, many attractive tumor targets may be found in indispensable, normal cell types, calling for the need to maximize tumor specificity to minimize collateral on-target toxicities. New chimeric receptor designs are poised to improve CAR T cell efficacy against tumor cells expressing <1,000 target molecules per cell. Several strategies are emerging to increase tumor specificity and safety based on dual-antigen targeting and logic gating (OR, AND, NOT, IF-THEN, and IF-BETTER gates), which may be further combined (e.g., OR-NOT-gate). The targeting of more than one antigen may be achieved through multispecific CARs, coexpressed CARs, or reversibly targetable CARs. The first Tan- and Dual-CAR clinical trials in B-cell malignancies, however, illustrate the complexity of multitargeting. In solid tumors, multitargeting is likely to be critical as well and will require the development of more probing preclinical models. Although genome editing is useful to expand the realm of T cell engineering, the risks of genomic abnormalities following induced double-strand breaks should not be underestimated. Finally, although the horizons for T cell engineering will further expand with the emergence of allogeneic approaches, induced pluripotent stem cells, and in situ engineering, autologous manufacturing remains the cornerstone of current clinical exploration.

M. Hamieh reports patent 11377637 issued and licensed to Takeda Pharmaceuticals, patent 11377637 issued and licensed to Atara Biotherapeutics, patent 11377637 issued and licensed to Mnemo Therapeutics, and patent 11377637 issued and licensed to Minerva Biotechnologies. J. Mansilla-Soto reports patent 11377637 issued and licensed to Takeda Pharmaceuticals, patent 11377637 issued and licensed to Mnemo Therapeutics, patent 11377637 issued and licensed to Minerva Biotechnologies, and patent 11229669 issued and licensed to Mnemo Therapeutics. I. Rivière reports grants from Takeda Pharmaceuticals and Atara, personal fees from Mnemo Therapeutics, Akron, the Centre for Commercialization of Cancer, and Oribiotech, and other support from Bristol Myers Squibb outside the submitted work. M. Sadelain reports grants from Atara Biotherapeutics outside the submitted work, as well as patent 8389282 issued and licensed to Juno Therapeutics, patent 11242375 issued and licensed to Atara Biotherapeutics, patent 10370452 issued, licensed, and with royalties paid from Fate Therapeutics, patent 11377637 issued and licensed to Takeda Pharmaceuticals, patent 11377637 issued and licensed to Mnemo Therapeutics, and patent 11377637 issued and licensed to Minerva Biotechnologies.

The authors thank Sascha Haubner for his comments and insightful discussion. The authors acknowledge the Memorial Sloan Kettering Cancer Center Core Grant P30 CA008748.