T-cell Receptors for Clinical Therapy: In Vitro Assessment of Toxicity Risk

Therapeutic use of antitumor T cells has proven feasible in a multitude of trials over the past decade. Alongside the demonstration of clinical benefit and enthusiasm about the therapeutic efficacy, the occurrence of toxicities has stimulated awareness of safety pitfalls. Although initial studies demonstrated recognition, and sometimes destruction of healthy tissues (1–3), later studies demonstrated lethal adverse effects in individual patients (4, 5). These studies highlight the two main challenges facing the safety of T-cell receptor (TCR) gene therapy: on- and off-target toxicities. With a quickly expanding panel of TCRs that have generally been selected for their ability to provide T cells with high avidities toward tumor antigens, there is an urgent need for streamlining the safety assessment of these TCRs prior to clinical usage.

An ideal target antigen for adoptive T-cell therapy (AT) displays two important features: It is immunogenic, and it shows selective and homogenous expression in tumor tissue. Immunogenicity is best explained as an antigen's ability to be recognized by and sufficiently activate T cells, a feature that is generally well addressed when selecting a target antigen and its corresponding TCR (reviewed in ref. 6). For example, characteristics related to the immunogenicity of two MAGE-C2 (MC2) antigen epitopes are summarized in Table 1. Selective expression in tumor tissue, and hence its absence in healthy tissues, would reduce the risk for on-target toxicities. Differentiation, overexpressed, or oncofetal antigens are not absent from healthy tissues and the targeting of these antigens, for example, MART-1 or carcinoembryonic antigen, by TCR-transduced T cells has resulted in severe destruction of melanocytes in the eyes, ears, and skin (1) or inflammation of the colon (2, 3). Candidate antigens that are exclusively expressed by tumor tissues include neoantigens and oncoviral antigens. Neoantigens are derived from somatic DNA alterations, and their identification requires analyses of mutations, gene expression, and algorithms that predict antigen processing and presentation by MHC (7–9). AT studies with tumor-infiltrating lymphocytes (TIL) in patients with melanoma and cholangiocarcinoma showed that clinical benefit was associated with T-cell responses against neoantigens (9–11). Current exploitation of neoantigens in AT, however, is challenged by the uncertainty of current algorithms accurately predicting immunogenicity and the fact that neoantigens are usually specific per patient (12–14). Viral antigens are present in more than 10% of human cancers and are often the result of viral insertion into the genome and subsequent reactivation in tumors. AT studies using TILs reactive against either human papillomavirus (HPV) or Epstein–Barr virus (EBV) have shown clinical successes in patients with cervical cancer or nasopharyngeal carcinoma, respectively (15, 16). Similar to neoantigens and oncoviral antigens, certain cancer germline antigens (CGA) also demonstrate tumor-selective expression. In fact, CGAs are expressed in gonadal tissues and some in thymus (17), and certain CGAs are considered to be selectively derepressed in tumor tissuess (for detailed reviews, see refs. 6, 18, 19). MAGE-A3 and NY-ESO1 are examples of CGAs that have already been targeted by TCR-engineered T cells in patients with metastatic melanoma, metastatic synovial sarcoma, or multiple myeloma (4, 20). Although off-target toxicities were observed with the targeting of the former antigen (most likely an issue of the TCRs; see next section), the safe use of selected CGAs was suggested by the targeting of NY-ESO1 demonstrating clinical benefit without toxicities (20, 21).

We recommend for any antigen, with the exception of neoantigens, to test the antigen's absence from a large panel of healthy organs. Online databases, such as the protein atlas (http://www.proteinatlas.org/) or the CGA database (www.cta.lncc.br), combine extensive data from transcriptomic analyses and antibody stainings from numerous normal, noncancerous cell lines and tissues. When applying these tools to assess the expression of MC2, we observed that mRNA expression is restricted to cells from cancers and testes—the latter considered to possess an immune-privileged status (no MHC expression; thus no detection by T cells). The use of commercially available cDNA libraries of a large series of healthy tissues enables researchers to extend online analyses and quantify antigen expression with a laboratory assay. When performing qPCR using such a cDNA library, we demonstrated absence of MC2 mRNA in healthy tissues, as illustrated in Fig. 1A. In case specific antibodies are available, we would recommend to follow up qPCR with IHC. Using an MC2-specific antibody, we confirmed the presence of MC2 protein in testis and melanoma as well as its absence in multiple healthy tissues, such as brain, heart, intestine, and lung (see Fig. 1B). However, both qPCR and IHC cannot formally exclude the presence of rare antigen-positive cells within a tissue, for example, stem cells. For example, mRNA of certain CGAs has been detected only in medullary thymic epithelial cells, but not total thymus (17). As an additional means to exclude target antigen expression, sophisticated in vitro cell cultures have been developed to closely mimic complete tissues or organs (commented on in the next section; refs. 5, 22).

Once the safety of the target antigen has been assessed, one can start selecting TCRs. Procedures to obtain tumor-reactive T cells and hence TCRs can generally be divided into those that rely on tolerant and those that rely on nontolerant repertoires of T cells. Tolerant repertoires, where deletion of T cells with an avidity outside the thymic selection window has occurred, have been used to obtain T-cell clones from patients following successful TIL therapy, following peptide vaccination, or using in vitro pulsing of autologous dendritic cells (23, 24). Notably, the thymic selection of T cells is most likely a trade-off between producing a self-tolerant yet sufficiently diverse and responsive TCR repertoire, and escape of self-reactive TCRs cannot be negated. Indeed, TCRs specific for overexpressed antigens and obtained from the native repertoire have been shown to initiate autoimmune side effects (1). The use of nontolerant repertoires, with the rationale of allowing the generation of high-avidity T cells, has been applied in allogeneic in vitro as well as in vivo systems. For example, HLA-mismatched antigen-presenting cells (25) or artificial antigen-presenting cells pulsed with peptides of interest facilitate the in vitro generation of tumor-reactive T cells (26). In vivo, mice transgenic for human HLA as well as mice transgenic for human TCR and HLA-A2 genes (27) have been immunized and used as a source of TCRs. After having obtained antigen-reactive T cells, sequences of the TCRα and β chains can be determined by molecular techniques such as 5′ RACE (23), enhanced PCR methods, capturing and indexing of genomic DNA-encoding TCR chains (28), or sequencing and pairing of TCR chains based on combinatorial algorithms (29).

It is important to note that TCRs derived from the aforementioned repertoires, in particular the nontolerant repertoir, have not been selected in the presence of all patient MHC alleles and in the case of mice neither against human peptides present in the thymus, and may show allo- and noncognate reactivity. Moreover, TCRs, even though obtained from highly tumor-reactive T cells, have an inherent degeneracy for peptide recognition and are able to recognize more than a single peptide. This dynamic flexibility in antigen recognition is in part credited not only to the bending ability of the TCR–CDR domains (30, 31) but also to the dominant interaction of the TCR with a restricted number and order of amino acids present in the MHC-presented peptide. To minimize the risk of selecting a TCR that recognizes noncognate self-peptides, we recommend a series of assays that assess the risk of this so-called off-target toxicity. These assays are illustrated with two patient-derived, MC2-specific TCRs, 6 and 16, with details and evaluation of antitumor T-cell responses mediated by TCR6 and TCR16 summarized in Table 1.

Initial assessment of a TCR's self-reactivity can be done by testing the responsiveness of TCR-transduced T cells toward random peptides known to be presented by the respective HLA restriction allele. Mathematical projections indicate that among a pool of approximately 10¹² peptides, a single TCR may react with >10⁶ peptides, supporting the notion of TCR degeneracy for peptide recognition, which potentially contributes to a more diverse TCR repertoire (32). We cocultured TCR6 and TCR16 T cells with antigen-presenting cells loaded with saturating concentrations of >100 common, HLA-A2–eluted self-peptides (4). As depicted in Fig. 2, both T-cell populations mediated a T-cell response to their respective cognate peptides but to none of the other peptides. Importantly, these data hint at a lack of cross-reactivity of these two TCRs, but we cannot fully exclude recognition to random peptides. Another assessment of self-reactivity can be conducted by testing TCR-transduced T cells toward allogenic HLA molecules. To exclude activation of T cells upon recognition of foreign HLA molecules, panels of lymphoblastoid B cell lines with various HLA allotypes have proven valuable (33, 34).

Further assessment of a TCR's self-reactivity, and a key assay in this communication, is the testing of a TCR's intrinsic capacity to recognize peptides highly homologous to its cognate epitope (35, 36). To this end, one can determine the recognition motif, that is, the position and sequence of amino acids within the cognate epitopes that are crucial for binding to the TCR. This motif is unique per TCR and can be considered a surrogate measure for the extent of cross-reactivity of TCRs (22, 33). The importance to assess such motifs became apparent from two recent clinical trials using AT with TCR-engineered T cells. The first trial targeting MAGE-A3 and A9 (MA3/9) in the context of HLA-A2 utilized the TCR 9W11 and reported neurologic toxicities in two patients with metastatic melanoma (4). The second trial targeting MA3 in the context of HLA-A1 utilized the TCR a3a and reported cardiac toxicities in one patient with metastatic melanoma and one patient with multiple myeloma (5). Both TCRs were affinity enhanced in vitro and mediated toxicity by recognizing peptides highly similar to the cognate peptide, namely peptides derived from MAGE-A12 and Titin present in brain and heart tissue, respectively. These studies clearly underline the need to assess recognition motifs and search for T-cell reactivities against homologous self-peptides prior to clinical application. The importance of recognition motifs is timely, and its assessment has only occurred for a limited number of TCRs to date, which is summarized in Table 2.

Using a set of altered peptide ligands (APL), peptides containing individual alanine replacements at every single position in the cognate peptide (in case of an endogenous alanine→glycine), we conducted stimulation assays with TCR6 and 16T cells. Critical amino acids are defined as those that, when testing the respective APLs, result in a drop of the T-cell response (generally using IFNγ production as a readout) of >50% when compared with the response toward the cognate peptide. Following the determination of these motifs, as exemplified in Fig. 3A, we used ScanProSite (http://prosite.expasy.org/scanprosite/; ref. 37) and identified target antigens, listed in Fig. 3B, that harbor the recognition motif and represent potential cross-reactive targets for TCRs 6 and 16. Subsequently, these self-peptides were tested for their ability to induce T-cell responses toward HLA-A2–postive antigen-presenting cells loaded with saturating concentrations (1 μmol/L) of peptide. These experiments yielded a short list of self-peptides that are actually recognized by the TCRs under study, defined as those that resulted in at least a T-cell IFNγ response >2.5% of the response to cognate peptide (Fig. 3B: underlined antigens).

Once self-peptides that can be recognized by TCR T cells have been identified, we recommend executing two additional tests to more stringently assess the risk for self-reactivity. These tests aim to provide measures for T-cell avidity as well as efficiency of cellular processing and presentation. Toward the first test, one can titrate amounts of noncognate peptides and determine the concentration of these peptides that elicits 50% of the maximal T-cell response (EC₅₀). For both TCR6 and 16 T cells, we found that only a single self-peptide, namely a peptide from the antigen MAGE-B4 (MB4) or MAGE-C1 (MC1; see Fig. 3B for homology and predicted peptide–HLA binding), respectively, revealed detectable EC₅₀ values that were only 2- to 5-fold lower than those of the cognate peptides (Fig. 4). Extent of homology and predicted peptide–MHC binding of peptides that induce T-cell IFNγ are listed in Fig. 3B, which shows that loss of homology and peptide–MHC affinity was least affected for MB4 and MC1 peptides. All other self-peptides revealed no T-cell reactivities at titrated doses, or at the very best 5 log scales lower compared with cognate peptide. These were considered not recognizable by the TCRs under study and excluded from further assays. Toward the second test, one can predict whether self-peptides are the result of antigen processing and presentation to enable T-cell recognition in vitro. NetCTLpan (http://www.cbs.dtu.dk/services/NetCTLpan/; ref. 38) takes proteasomal C terminal cleavage, TAP transport efficiency, and peptide MHC class I binding of peptides into account and can be employed to obtain an initial score for antigen processing and presentation. Analysis of the MB4 peptide, but not MC1 peptide, yielded a high score according to this Web-based tool. Such predictions may not be fully accurate and should be verified using cells known to express the antigen or antigen-presenting cells (e.g., dentritic cells) transfected with antigen-encoded RNA followed by cocultivation with TCR-transduced T cells. To this end, we stimulated TCR T cells with the esophageal cancer cells line OEC-19, which natively expresses the MB4 and MC1 antigens but is devoid of the MC2 antigen, and observed that both TCRs failed to initiate T-cell activation against either MB4 or MC1 (see Fig. 4B). When using cell lines that natively express the MC2 antigen, as a control for the processing and presentation of MAGE antigens, we observed that both TCRs did initiate T-cell activation. It is noteworthy that standard tissue culture systems may not always accurately reflect antigen processing and presentation. This was evidenced by the recognition of Titin by the TCR a3a that could only be observed in more elaborate tissue culture systems, such as 3D cultures of beating cardiomyocytes derived from induced pluripotent stem cells (22). In case the aforementioned two assays do not exclude self-reactivity of TCR T cells, one could pursue assessment of the tumor-selective expression of such new antigens. In case the expression of new antigens is not selective for tumors, the corresponding TCR should be excluded. Using online tools (39) as well as qPCR (Fig. 1A), MB4 showed expression within epididymis and vagina, whereas MC1 showed no expression in any of the healthy non-gonadal tissues. These data highlight the stringent safety profile of TCR 16, the TCR selected for a clinical trial to treat melanoma and head and neck carcinoma, currently prepared at Erasmus MC (Rotterdam, the Netherlands).

The proposed collection and sequence of in vitro assays to assess risks for toxicities are presented in Fig. 5. We advocate this testing for TCRs with clinical intent, in particular those TCRs reactive against a self-peptide and derived from a nontolerant repertoire and/or following gene enhancement. In extension to gene enhancement, introduction of TCR–CDR mutations has been a commonly used tool to generate high-affinity TCRs (2, 4, 5, 20). Although such gene-enhanced TCRs recognize target peptides at increased affinities when compared with the corresponding wild-type TCRs, consequently, such TCRs are also at risk to recognize noncognate peptides. To test whether affinity enhancement led to an increase in degeneracy for peptide recognition, we made use of a panel of eight TCRs specific for the same cognate peptide gp100_181-188:HLA-A2) but harboring two to three mutated amino acids in either their CDR2β, CDR3α, or CDR3β domains (C. Govers; submitted for publication). Upon assessment of the recognition motifs and search for motif-harboring self-peptides, it became apparent that enhanced affinity was accompanied by a drastic increase in the TCR's ability to recognize self-peptides (Table 3). These data extend earlier findings regarding a correlation between affinity enhancement and loss of TCR specificity (40) and warrant caution when trying to change the TCR–CDR structure, as it compromises the stringet recognition of cognate peptide (31, 41).

Taken together, here, we propose a platform of in vitro assays that in combination with available online databases and tools allows for optimal toxicity risk assessment for target antigens and TCRs currently under consideration for clinical trials.

No potential conflicts of interest were disclosed.

The authors would like to thank Mandy van Brakel and Mieke Timmermans for technical support toward Table 1 and Fig. 1B, respectively.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.