CAR T-cell therapy has shown remarkable success in treating hematological cancers (Han et al., 2021). However, this success has not extended to solid cancer treatment, where efficacy is thought to be compromised by off-tumor toxicity, poor trafficking to the tumor site, limited persistence and proliferation, and functional suppression within the tumor microenvironment (Rafiq et al., 2020; Marofi et al., 2021; Sterner and Sterner, 2021). These barriers point to the need for further improvements in CAR T-cell design in order to more fully realize the clinical potential of targeted T cell therapies against more challenging tumor types.

To this end, while optimization of extracellular domains has benefitted from a diversity of high-throughput strategies used widely to affinity mature candidate receptors for antigen recognition (Rydzek et al., 2019; Bloemberg et al., 2020; Wang et al., 2021), other factors appear to limit in vivo performance. Some of these factors are beginning to be addressed by engineering expression of additional immune modulators, such as programming secretion of cytokines, and employing ON or AND switches, etc., (Rafiq et al., 2020). However, improvements in CAR activity and persistence are still likely to benefit from continued optimization of their primary sequence and structure. To efficiently explore this design space, higher throughput approaches are needed. Presently, CAR T cell design strategies generally evaluate novel CARs on a candidate-by-candidate basis, in which constructs are designed and tested one at a time based on a hypothesis about in situ performance. Candidates are then prioritized based on performance in parallel assays of varying throughput. While this rational design approach has produced the successful CAR designs used in the clinic today, more efficient and less modular discovery and optimization approaches could enable faster CAR screening, allow evaluation of greater functional diversity and T-cell fitness, and potentially accelerate CAR therapeutic design for improved clinical success rates.

As has been done for small molecule and antibody drug discovery, high-throughput screening approaches could permit deployment of directed evolution strategies. While such approaches have successfully identified novel extracellular, antigen binding domains (Han et al., 2017; Hua et al., 2017; Sharma et al., 2020; Butler et al., 2022), analogous strategies are only beginning to be developed for evaluating intracellular signaling domains in the context of a complete CAR construct, in T cells stimulated by target antigen, and which allow selection of functional phenotypes. Among CAR domains, signaling domains pose unique challenges to the binding affinity-based approaches common in high-throughput directed evolution. Unlike extracellular domains, intracellular signaling domains rely on dynamic posttranslational modifications and spatial positioning of signaling mediators unique to T-cell biology.

CARs contain the TCR-associated CD3ζ intracellular domain and are thus expected to trigger similar signaling cascades as natural TCR signaling. Indeed, both CAR and TCR activate canonical downstream pathways including Lck and Zap-70 kinases, LAT, and PLCγ1 (Salter et al., 2018). However, recent studies have revealed that CARs and TCRs exhibit distinct activation thresholds (Harris et al., 2018; Watanabe et al., 2018), downstream phosphorylation kinetics (Karlsson et al., 2015; Rohrs et al., 2018; Salter et al., 2018; Ramello et al., 2019), and recruitment of negative regulators (Hamieh et al., 2019; Sun et al., 2020). Altered signaling may be attributable to the tandem linkage of several signaling domains in CARs. This may remove normal barriers imposed by spatial separation of endogenous receptors and co-receptors but may also disrupt evolutionarily optimized spatial pairing of signaling partners. For example, fusing the CD28 co-receptor sequence to the TCR CD3ζ sequence increases the rate of CAR phosphorylation at its CD3ζ motifs (Rohrs et al., 2018). Yet, swapping domain order alters the magnitude of cytokine production and tonic signaling (Guedan et al., 2018). Additionally, domains typically expressed in other immune cells, such as the NK-cell signaling adapter DAP10, exhibit unique properties when incorporated into CAR T cell designs (Zhao et al., 2019). The cumulative result of this complexity is that the specificity and contributions of individual costimulatory domains is experimentally difficult to determine (Jutz et al., 2016).

Importantly, while modular design approaches will continue to improve our understanding and lead to more successful future CAR constructs, current strategies rely predominantly on wildtype domains, leaving the vast mutational landscape of each domain unexplored. Candidate signaling domains poised for high throughput optimization include existing T cell signaling domains used in preclinical and clinical trials, CD28, 4-1BB, OX40, and CD3ζ (Gilham et al., 2002; Carpenito et al., 2009; Savoldo et al., 2011). Among these, OX40 and 4-1BB are co-stimulatory molecules in the TNF-receptor (TNFR) family. When incorporated into CARs, OX40 promotes T cell adhesion to endothelial cells as well as proliferation and differentiation (Hombach et al., 2012), while 4-1BB prolongs survival of activated T cells and enhances IL-2 production (Long et al., 2015). Both co-stimulatory domains lack intrinsic enzymatic activity and rely on binding to TNF Receptor Associated Factor (TRAF) adapter proteins in order to transduce downstream signals (Xie, 2013). Upon receptor ligation, TRAF proteins are recruited as homo- and/or heterotrimers to the 4-1BB or OX40 intracellular domain, initiating the formation of the signalosome (Zapata et al., 2018). There are six known TRAF proteins (TRAF1-TRAF6) that are regulated by expression and subcellular localization in the cell. Thus, the composition of TRAF homo- and heterodimers is dependent on cell type and activation state. While OX40 and 4-1BB bind to multiple TRAF proteins, TRAF2 binding has been linked to canonical NF-κB signaling and cell survival (Etemadi et al., 2015). OX40 and 4-1BB interact with low affinity to TRAF2 (50 µM and 1 mM, respectively) (Ye and Wu, 2000), and while mutations that disrupt TRAF2 binding have been explored (Arch and Thompson, 1998), mutations that improve this interaction, and their effects on CAR phenotypes have yet to be described.

While affinity maturation of receptor interactions with downstream signaling partners may enhance signaling, it is also possible that affinity may be a poor proxy for activity, and that other attributes are more critical to improving CAR T cell function. Downstream of CAR signaling are the transcription factors NF-κB and NFAT, which regulate shared and distinct T cell effector functions. Moreover, NF-κB induction has been shown reflect the magnitude of T cell signaling (Berger et al., 2004; Kingeter et al., 2010; Thaker et al., 2015), while NFAT induction correlates positively with T cell stimulation (Clackson et al., 1998; Brogdon et al., 2002; Maguire et al., 2013). NF-κB and NFAT have both been used in reporter systems for the evaluation of TCRs and CARs (Jutz et al., 2016; Rosskopf et al., 2018; Rydzek et al., 2019). These studies suggest that an NF-κB and NFAT activation screen could be leveraged in a standardized directed evolution CAR discovery platform to identify novel signaling domains that induce physiologically relevant T cell effector responses. Here, we present a mammalian display approach capable of retrieving rare, functionally improved variants from among large libraries, which we anticipate will permit exploration of new terrain in the fitness landscape of CAR intracellular signaling domains.

We explored a novel platform to investigate the functional landscape of intracellular signaling domains that would allow for the discovery of mutations that improve CAR dependent transcriptional activity and function. CAR signaling domain variants have recently been explored using exciting alternative strategies. Castellanos-Rueda et al. (2022) employed CRISPr-mediated insertion of domain shuffled CAR designs in primary T cells and single cell sequencing to screen CARs to identify novel intracellular domain configurations with distinct and desirable T cell transcriptional profiles. Analogously to this work, Gordon et al. (2022) used lentiviral transduction of CARPOOL, a CAR library of shuffled intact signaling domains, to identify desirable transcriptional and functional variants.

Here, to accomplish this objective, we evaluated an in situ screening strategy based on NF-κB and NFAT activation. While a number of excellent high-throughput methods exist for the engineering of high affinity protein-protein interactions to optimize CAR recognition of antigen, we sought a means to discover CAR variants that increase function but are more agnostic to means by which those increases are achieved. For example, design criteria for CAR intracellular domains are likely to relate less to affinity for signaling partners and more to factors that are challenging to capture in standard protein engineering systems, such as spatial orientation, signaling protein abundance and stoichiometry, and post-translational modification that impact the magnitude and kinetics of signaling. Differential kinetics of activation have been shown to contribute to different signaling fates (Salter et al., 2018). Thus, while it is clear that fine-tuning of intracellular domains through a rational design approach can lead to more desirable signaling outcomes (Feucht et al., 2019), and variations in these domains have improved next-generation CAR designs, robust coverage of design space has been challenging.

While no single in vitro activity is known to be perfectly predictive of CAR T clinical efficacy, several T cell characteristics have been tied to improved clinical outcome, including adhesion, migration, polyfunctionality, persistence, and proliferation (Rafiq et al., 2020; Marofi et al., 2021). Additionally, the transcription factors NF-κB and NFAT have both been shown to be reliable surrogates of T cell persistence and survival [(Maguire et al., 2013; Thaker et al., 2015)], (Brogdon et al., 2002; Song et al., 2008; Kingeter et al., 2010; Marangoni et al., 2013). Indeed, an NF-κB and NFAT reporter Jurkat cell line has demonstrated precedent as a successful screening tool for the selection of CARs with novel scFvs and hinge construction (Rydzek et al., 2019). However, to our knowledge this approach has yet to be adapted to intracellular CAR domains and the resolution of NF-κB and NFAT activation to detect meaningful differences driven by intracellular signaling domains is unknown. Here, we determined that this screening parameter possesses the capability to discern between 4-1BB, DAP10, CD3ζ CARs, and their respective functional knockouts.

Our data demonstrated that CD3ζ first generation CARs display a relatively high level of NFAT activation, perhaps an unexpected finding considering these CAR designs were retired due to their inability to promote complete T cell activation (Brocker and Karjalainen, 1995). The platform also allowed us to make the surprising observation that insertion of a mutated DAP10 motif further enhanced NF-κB and NFAT signaling compared to the first-generation CAR or native DAP10 CAR. This supports the importance of considering spatial orientation and membrane proximity of signaling components in CAR design. We also found that the 4-1BB CAR activates NF-κB and NFAT reporter activity to different degrees—enhancing NF-κB while reducing NFAT activity in comparison to the first-generation CAR. This observation demonstrates the platform’s capacity to identify CAR variants that uncouple downstream TCR signaling pathways. The ability of a limited screen to identify sequence changes that alter function in unexpected ways demonstrates the potential value of directed evolution over rational design for the engineering of these domains. While the optimal signaling and phenotypic profile of CAR T cells has yet to be fully defined, and likely varies across targets and tumor types, an ability to drive phenotypic divergence in vitro could contribute to functional differentiation of CAR designs that could then be screened in vivo.

Having established that NF-κB and NFAT signaling can be used to discern functional differences due to signal domain modifications, we tested the ability to retrieve rare, desirable CAR constructs from a larger population of undesirable CAR constructs. We found that the higher signaling cells were consistently enriched at three different dilution ratios, providing proof-of-concept that NF-κB and NFAT activation is capable of identifying high-signaling CAR designs even in the context of significant biological variability from cell to cell.

However, numerous technical hurdles remain for the establishment of a working display system for screening new CAR signaling variants using this approach. These include design and generation of a viral library of CAR signal variants of sufficient phenotypic variation to be reliably differentiated, optimization of antigen stimulation conditions to most thoroughly and uniformly stimulate a large number of cells, and recovery of selected cells and their CAR sequences either through regrowth or immediate lysis and capture of their genotypes, which can be complicated in the context of highly homologous sequences, necessitating either single cell approaches or barcoding of the library. While a number of non-viral CAR construct delivery alternatives could be pursued, optimization of each step is likely to contribute to setting the limits on feasible library sizes. For example, reliable readout of function may be complicated by multiple sequence integrations into one cell (Rydzek et al., 2019) and will need to be assessed in future studies. While viral transduction at a low multiplicity of infection may reduce multiple CAR sequence integration, we expect that such a step would also reduce the number of designs that could be screened for a given volume and number of cultured cells, or would motivate addition of a step in which CAR-expressing cells were first enriched prior to stimulation. For this reason, a mutagenesis strategy targeting known or suspected positions or domains of importance may be more appropriate than random mutagenesis of entire CAR constructs. We have also not experimentally confirmed how finely functional variants can be resolved but expect that biological noise between cells is likely to pose a fundamental limit given that even Jurkat T cells can exhibit heterogeneous phenotypes (Snow and Judd, 1987). Given the importance of activation kinetics, it may be desirable to express fluorescent reporters that are more rapidly responsive or less persistent in order to support optimized analysis over longer or shorter timescales. Use of reporter proteins with controllable degradation tags may offer potential to this end. While the highest fidelity in situ screen would require the use of primary T cells, we expect primary cells to be more functionally heterogeneous than the J76-TPR cell line. While NF-κB and NFAT activation bears functional relevance, the genotype-phenotype coupling of this readout is rather indirect, with many possible factors capable of influencing activation phenotypes other than the CAR sequence. Collectively, these factors point to the likelihood that while in situ screens are likely to offer advantages in throughput relative to clone by clone strategies, they are unlikely to achieve the excellent performance attributes used in screens of simpler activities such as protein-protein binding affinity.

Nonetheless, our data demonstrate that NF-κB and NFAT reporter cells are able to identify high-signaling CAR candidates present at low frequencies, and highlight the high potential of in situ screening strategies for the development of novel and optimized intracellular CAR domains.