Preclinical development of gene-engineered CAR- and TCR-modified T cells is truly a “first in class” living medicine. Traditional evaluations applied to pharmacologic small molecule inhibitors or even biologics including vaccines and mAb therapies do not apply. Neither do usual evaluations of cellular therapies such as blood transfusions or bone marrow transplants. A third class of medicines that have fairly recently made an impact include gene therapies, generally used to treat single-gene defects in patients with rare diseases. Immunotherapy with CAR and TCR T cells involves aspects of all of these, brought together for the first time. Biologic pharmaceuticals provide the mAb binding region, while gene engineering borrows the use of transforming vectors such as gamma-retro or lentivirus to combine targeting receptor with functional genes. All of these come together to engineer patient T lymphocytes ex vivo to generate a new “living drug” that relies upon millennia of built-in evolution to combat disease inside the patient, in this case, cancer.

There are several steps in common to developing any preclinical CAR or TCR therapy. First consideration is the target. Engineered T cell targets should be expressed on a reasonable proportion of tumors and should not be expressed on critical normal tissues. This became frankly apparent in the first-in-human treatment of a 39-year-old woman with metastatic melanoma who received a high number (10¹⁰) of autologous HER2 CAR T cells [36]. Although to date, over 400,000 women with breast cancer have received trastuzumab anti-HER2 mAb in the adjuvant setting, with minimal side effects, the first patient to receive the same HER2 mAb-redirected CAR T cells succumbed to severe respiratory distress, followed by multiple organ failure starting 15 minutes after infusion [36]. This is a clear demonstration of the difference in impact of a naked antibody compared with an antibody tied to the signaling region of a T cell. While there may be mitigating circumstances, whereby utilizing a limited route of administration [49], lowering the affinity of a CAR receptor may allow for selective destruction of high-antigen expressing tumors while sparing low-antigen bearing normal tissues [50, 51], the safest approach is to target truly tumor-specific antigens.

Several methods have been employed to identify preclinical targets on tumors. Initially, many institutions had to rely on their own available tumor banks for antigen detection by immunohistochemical (IHC) or RT-PCR for gene expression. More recently, however, numerous in silico bioinformatic resources have become widely accessible, in particular evaluation through publicly available Human Protein Atlas (proteinatlas.​org) or via The Cancer Genome Atlas (TCGA) (cancergenome.​nih.​gov).

Upon selecting a target, a binder is required, usually in the form of the scFv of a mAb. Once selected, the binder/scFv is engineered into a DNA construct including a linker, hinge/transmembrane domain (typically derived from CD4, CD8, or an IgG), and followed by a co-stimulatory molecule and CD3ζ at minimum, though new additions are continually being produced. The difference between a CAR and a TCR is that a TCR requires expression of a matching full-length CD3 α and β chain comprising the constant and variable regions of each. Only upon appropriate α/β pairing is antigen recognition conferred, implementing the induction of the full CD3 signaling mechanism.

In vitro, recombinant TCR or CAR are expressed in donor T cells, and upon confirming surface expression, these cells undergo specificity recognition and functional testing. At minimum, known antigen-positive and antigen-negative target cells are needed, though in practice, often several panels of different tumor types are assessed for targeting. While there are many variations of experimental methods to determine function, they fall in one of two forms: (1) effector T cell stimulation or (2) target cell destruction. Once they have passed this initial specificity test, they must be evaluated for safety. Two types of safety predominate, specifically (1) on-target, off-tumor, whereby the antigen being targeted is also expressed on normal tissues, or (2) off-target toxicity, in this case the receptor recognizes something other than the specific target. Each type of safety testing poses its own set of challenges, including the need to profile each cell/tissue type with regard to gene expression. In the case of “on-target, off-tumor,” essentially a screening panel of live normal human cell types and tissues are needed for use as experimental targets for the CAR in question. Some groups have utilized human-derived primary cell lines for this [52]; while others have obtained direct primary cell and tissue grafts from donors for this purpose [20, 53].

Off-target toxicity is more difficult to predict and has been observed more for TCRs than for CARs. Part of the reason for this may be (1) the potential for more than one gene/protein to share sequences that can be presented in the context of MHC in a similar conformation to tumor-derived epitopes and (2) the act of “affinity tuning” TCR toward higher affinity for a specific peptide-MHC complex which may introduce unforeseen cross-reactive binding to peptide-MHC complexes from other proteins expressed in normal tissues that would otherwise have been screened out biologically during thymic selection. One such example of this was observed in the affinity maturation of a TCR to a MAGE-A3 peptide-HLA-A*01 complex [41]. Subsequent analyses suggested that the fatalities were likely resulted from recognition of an off-target peptide from the muscle protein titin, highly expressed in cardiac muscle [34]. New methods to predict cross-reactivity profiles for new TCRs such as the regular use of 3D culture models for toxicity-relevant tissues and systematic scans of permutated peptides and use of in silico predictions to predict cross-reactivity profiles for novel TCRs have now become a standard requirement of preclinical development programs. While normal cell/tissue toxicity detection methods still continue to evolve, the progress in developing new models and safety technologies for this type of predictive evaluation will prevent unforeseen adverse events in patients in the future.

In vivo models have some, albeit limited, benefit for translation of CAR and TCR therapies, in that immune-compromised inbred mouse models such as the NOD-SCID-common gamma chain k/o (NSG) or NOG mouse can engraft both human tumors and human TCR or CAR T cells. These animals act as tiny incubators complete with functioning cardio-vasculature and an extensive network of blood vessels to transport the T cells to the tumors. Generally, however, these models are limited to showing antitumor efficacy and homing to tumor sites, and most recently have shown some promise in modeling T cell exhaustion and checkpoint blockade at the tumor site. Unfortunately, few tumor target epitopes are shared between mouse and human, making it difficult to determine information on safety or potential normal tissue toxicity [54]. Syngeneic mouse models with congenic tumors may provide a unique ability to evaluate both scientific questions about the efficacy of CAR therapy, and the specific mechanism of action of downstream cancer “cures,” in the context of an animal model with an intact immune system and physiologic expression of tumor antigen in both tumors and normal tissues, if present [55–58]. However, these models are far from perfect: (1) they need separate binders that likely have different affinity for target than human counterparts; (2) they require completely different vector constructs encoding murine genes for CD3ζ, 4-1BB, and CD28 in place of their human orthologs; and last but not least, (3) mice are not people. There has been much divergent evolution in the development of their immunity and physiology, so results in a mouse model are by no means a guarantee of what will work for patients. To date the use of nonhuman primates (NHP) has been of minimal/negligible benefit in determining safety or toxicity to this field, primarily due to the lack of similarity of action and reagent comparability between human and NHP and also due to the lack of any NHP tumor models. Intriguingly, according to the US Food and Drug Administration (oncology), they currently do not require any animal modeling to support CAR or TCR first time in human phase I clinical trials.