A membrane receptor biosensor architecture can be schematically decomposed into two elements: (1) an extracellular (EC) ligand-binding sensor and (2) an intracellular (IC) signalling responder connected by a transmembrane (TM) domain. Communication between sensor and responder (that we define as coupling below) enables signal transduction across the membrane and activation of IC functions upon ligand binding. While the underlying structural mechanisms may vary, sensing-response behaviours rely on at least the following three sequential steps: the sensor changes conformation upon ligand binding, the sensor transmits this structural change to the responder, and the responder switches to an active state conformation and triggers receptor activity (Fig. 1b).
As in any allosteric system, sensing-response properties can be achieved through diverse design scenarios that impact the receptor's basal activity, sensitivity and potency. These scenarios are outlined in Extended Data Figs. 1 and 2. Specifically, each element can independently switch between an inactive and active state and preferentially occupy one state in isolation (Extended Data Fig. 1a). When combined, different biosensor behaviours will be obtained depending on individual bias between inactive/active state and the mechanical coupling between the sensor and responder that will impact the state occupancies. For example, a programmable biosensor scaffold could involve sensor and responder elements that preferentially occupy an inactive and active state conformation, respectively, in isolation and absence of ligand (Extended Data Fig. 1b) or alternative scenarios (Extended Data Figs. 1c and 2b,c). If the sensor and responder are weakly coupled, the responder will readily access the active state and trigger high receptor basal activity (that is, without ligand). Conversely, strong coupling will maintain the responder mostly in the inactive state, turning off basal activity, while still enabling a strong ligand-induced response.
Studies of natural single-pass membrane receptors indicate that while the structural mechanisms underlying biosensing functions can be diverse, they usually involve two main structural modes of activation. In the pre-formed dimer (PFD) mode reported, for example, for the interleukin-7 receptor (IL-7R) or death receptor 5 (DR5) cytokine receptors, the receptor self-associates in the absence of ligand but mainly occupies inactive state dimer conformations. Ligand binding triggers intramolecular reorganization propagated allosterically to the cytoplasmic region through coupling, stabilizing active state conformations of the dimer structure. In the monomer-dimer equilibrium shift (MDE) mode initially described for the epidermal growth factor receptor (EGFR), the ligand-free receptor mostly occupies a monomeric inactive state (Extended Data Fig. 1). Ligand binding triggers receptor association and the formation of active state dimer conformations. However, structural insights into receptor activation are sparse and the prevalence of each mechanism remains highly debated. In fact, several lines of evidence suggest that receptor activation may involve a combination of both allosteric and binding mechanisms.
Since the structural mechanisms of activation of single-pass TM receptors remain largely elusive, we reasoned that a simplified but effective approach for designing receptor biosensors should primarily focus on building optimal active state structures. While this positive design strategy neglects alternative inactive or pre-active states, optimizing structural features for a single target state has proven effective in many protein engineering studies. To account for the diverse activation mechanisms, optimization focuses on the two main structural properties driving signal transduction: dimerization and mechanical coupling, both underlying the MDE and allosteric PFD modes in the active state. In principle, a wide range of constitutive and ligand-inducible activities can be programmed through the modulation of dimerization propensity and communication in the active state. Together, they provide a rational blueprint for engineering receptor scaffolds with diverse and precise sensing and signalling functions (Methods, Fig. 1c and Extended Data Figs. 1-3).
We developed a computational approach based on these rules to design dimeric biosensors that link the binding of a user-defined chemical input signal to modular cellular responses through genetically encoded fusions of protein domains. Overall, the approach proceeds in the following main steps (Methods, Fig. 1c and Extended Data Fig. 4): (1) selection of the structural elements defining input, signal transmission and output signals: (1.1) the sensor (EC ligand-binding domain and dimerizing domains that link to the responder) and (1.2) the responder (TM and IC signalling domains); (2) if necessary, self-association of individual domains into dimeric complexes through docking, and design of juxtamembrane linker sequence connecting the sensor and responder; (3) assembly of multi-domain dimeric scaffolds using structure prediction methods RoseTTAfold and AlphaFold2, and design protocols in Rosetta; (4) ranking of receptor scaffold structures based on their propensity for dimerization and long-range communication (that is, coupling) between ligand-binding and signalling domains that are calculated using Rosetta and Elastic Network models for investigating protein association and coupling, respectively (Methods). As mentioned above, this protocol only constructs an ensemble of conformations for the active ligand-bound state of the biosensor. It neglects the impact of the design decisions on alternative states or transitions between states that would require a detailed mechanistic understanding of the activation process. Nevertheless, this positive design selection for optimal oligomerization and coupling in active dimer structures should generate computational libraries of receptors enriched in constructs with desired sensing-response behaviours.
VEGFA and CSF1 were chosen as ligands, two soluble factors that are highly enriched in a variety of TMEs and critically involved in tumour progression. As VEGFs promote neovascularization and CSF1 supports tumour-associated macrophage development and polarization in various TMEs, these targets open broad fields of applications and offer high translational impact for our T-SenSERs. c-MPL signalling was selected as the output signal since we have previously shown that c-MPL activates beneficial co-stimulatory, cytokine and type I interferon pathways when expressed in TCR-transgenic T cells. We aimed to design robust VEGF-MPL-receptor (VMR) scaffolds that are entirely VEGF dependent, and CSF1-MPL-receptor (CMR) scaffolds that have a low but significant basal activity with full CSF1 ligand inducibility ('low constitutive-inducible'). We hypothesized that a low constitutive-inducible CMR has the capacity to counterbalance an immune-suppressive TME by enhancing T cell homeostasis and proliferative capacity in the absence of T cell stimulatory cytokines. Thus, low constitutive-inducible CMR activity is expected to sustain local CAR-T cell persistence and anti-tumour function in the TME (Fig. 1d).
We first analysed the topology and individual domain structures of the native vascular endothelial growth factor receptor 2 (VEGFR2), colony-stimulating factor 1 receptor (CSF1R) and c-MPL receptors. While the structures of the full-length receptors remain elusive, several domain structures have been characterized. Both VEGFR2 and CSF1R contain immunoglobulin-like domains, out of which domains D2 and D3 strongly bind to their cognate ligand. These two domains were selected as the input signalling region of the VMR or CMR, respectively. For the output signalling, the structure and activation mechanism of cytokine receptor homologues of c-MPL indicate that strong ligand regulation and potent JAK/STAT signalling are achieved through the intricate coupling between the cytokine TM, juxtamembrane (JM) and the cytoplasmic (CT) regions. Hence, we reasoned that an optimal biosensor scaffold should couple the native TM region of c-MPL and not that of VEGFR2 or CSF1R to the c-MPL CT domain. In the absence of structural information, we modelled the c-MPL TM domain in a dimer active signalling state from sequence using the method EFDOCK-TM and then assembled the entire signalling (TM + JM + CT) region using our assembly approach (Methods). We next curated a library of all known native VEGFR and CSF1R EC domain structures, the recombination of which could modulate coupling between the input and output signalling domains in engineered receptors. Unlike c-MPL, all seven IgG-like VEGFR2 EC domains (D1-7) have been structurally characterized and the isolated D4, D5 and D7 domains are known to homodimerize. Likewise, the D5 and D6 domains are critical for CSF1R homodimerization.
Next, we created a diverse set of chimeras to stringently test our ability to rationally design full-length receptor scaffold structures with fine-tuned signal transduction propensity. Our computational pipeline enables the arbitrary combination of VEGFR or CSF1R EC domains with pre-defined linkers to sample different densities of contacts across the receptor dimerization interface and encode different levels of mechanical coupling. Final designs are returned as a dynamic ensemble of conformations. Thus, the method offers a computationally inexpensive means of obtaining biophysically relevant states from which to assess the modulation of signal transmission triggered by VEGF or CSF1 binding (Methods and Fig. 1c).
We ultimately designed a total of 18 chimeric receptor scaffolds, with 9 constructs generated for each family of sensors (Methods, Extended Data Fig. 5a,b and Supplementary Table 1). For VMR, these included both intuitive topologies, in which the domain order was preserved as found in natural receptors, as well as non-intuitive designs, where the domain arrangement and combination differed from VEGFR (for example, VMR, VMR). For CMR, synthetic linkers were designed with properties -- such as length, structure and sequence -- that differed from those of CSF1R. We initially explored an extensive space of de novo linker structures and sequences -- examining more than 700,000 possibilities -- using advanced deep learning methods, including ProteinMPNN and S4PRED, in combination with fragment assembly approaches (Methods). From this initial in silico screening, we found that the average helicity of the linkers -- critical for dictating coupling properties -- was generally low (Methods and Extended Data Fig. 5c,d). Our nine designed CMR chimeras combined a diverse range of these ProteinMPNN de novo sequences with a more targeted, structure-informed approach that integrated fragments from native TpoR and CSF1R sequences, leading to linkers with higher helicity and the design of CMR, CMR, CMR and CMR. Except for CMR, the only chimera from ProteinMPNN to feature helicity, the calculated couplings for the ProteinMPNN linkers fell below the thresholds required to effectively programme signalling responses in the CMR constructs (Extended Data Fig. 5c,d and Supplementary Table 1).
To conduct in-depth experimental validation, we selected six constructs representative of the dataset and the range of predicted outcomes. For the VMR chimeras, we selected (1) VMR, which incorporates all native ectodomains (D1-7), and exhibited the most optimal predicted VEGF response while maintaining the lowest propensity for constitutive activity; (2) VMR, the minimal version of the chimera, where the ligand-binding domain (D1-3) is directly linked to the TM region (it was chosen as a negative control, as it is predicted to exhibit a weak signalling response to VEGF binding); and (3) VMR, a non-intuitive design where D4 is directly connected to D7 (D1-4 + D7), chosen for its intermediate properties. Our calculations predict a gradient of increasing dimerization and coupling properties from VMR to VMR to VMR, indicating a progressive enhancement in the ability of VMR chimeric scaffolds to redirect VEGF sensing into potent c-MPL signalling (Methods, Fig. 1e and Extended Data Fig. 5). Among the CMR designs, we selected (1) CMR, our top-ranked design in terms of signal transduction propensity; (2) CMR, which provided a well-balanced compromise between constitutive and ligand-induced activities, aligning with our design goals; and (3) CMR, as it exhibited one of the lowest dimerization propensities. Our calculations predicted CMR to have the weakest coupling according to our activation model (Extended Data Fig. 3), and therefore should show the highest level of basal activity (Fig. 1e). All three CMR variants had strong dimerization propensity and were predicted to provide potent signalling responses to CSF1 sensing.
To validate our design approach and the positive design hypothesis, we next characterized the impact of the ligand on the receptor structure and dynamics using molecular dynamics (MD) simulations (Fig. 2a). Since these calculations are very time-consuming, we selected the most optimal construct in terms of predicted dynamic response, VMR, and carried out a large-scale simulation both with (total 0.75 μs) and without ligand (1 μs) (Methods). While these simulations are at least one order of magnitude too short to explore the entire receptor activation process, they revealed distinct conformational properties of the ligand-free and ligand-bound forms that aligned with the expected native behaviour of c-MPL. Principal component analysis (PCA) and subsequent K-means clustering (Methods) of both the receptor and TM coordinates revealed a unique space occupied by only the ligand-bound state (Fig. 2b-d). These conformations corresponded to the receptor adopting an 'upright' position, with the representative cluster centre showing an angle of 84° between the lipid membrane and the D2's centre of mass. The remaining conformational spaces shared by both the ligand-bound and ligand-unbound tended towards lower angles of around 55° on average. Calculated mechanical coupling of these representative cluster centres correlated with these angles, with the upright position returning a much higher coupling score than the lower angle ligand-bound or unbound conformations. This finding aligns well with the consensus that inactive receptor tyrosine kinase conformations adopt a bent configuration, and indeed can form direct interactions with the membrane itself, before becoming upright when ligand bound (Fig. 2c,d). Overall, our results imply that only conformations accessible to the ligand-bound state can confer the necessary coupling required for a potent response, and that our coupling metric is sensitive enough to capture this behaviour.
Consistent with previous experimental findings on c-MPL, the ligand impacts also the conformation of the JM region that flanks the TM on the CT side of the membrane, participating in the receptor activation. For instance, a cation-π interaction between R514 and W515 known to stabilize a helical motif in the inactive state is more often observed in the ligand-free simulations (Fig. 2e, light orange square). Conversely, Q516, known to stabilize the interface of the active state, is found with higher frequencies at the interface in the ligand-bound simulation (Fig. 2e,f, dark orange rectangles). Overall, while we have not modelled the entire activation process and our simulations have unlikely reached a true inactive state, our ligand-unbound simulations revealed several known inactive-like structural features. The shifts in coupling behaviours and TM-JM motif interactions between the ligand-bound and unbound simulations therefore validate our constructed model and strongly suggest that our assembly protocol can design reasonable receptors with predictable behaviours.
To experimentally explore the computationally predicted signal transduction propensity of the designed VMR and CMR variants, we assessed baseline and VEGFA- or CSF1-dependent STAT5 phosphorylation as a surrogate for c-MPL signalling in human T cells transduced with VMR, VMR, VMR or CMR, CMR, CMR. We found that all variants were capable of transmitting signal upon VEGFA or CSF1 exposure. For VMR, VMR produced the most efficient STAT5 phosphorylation followed by VMR and VMR that were characterized by significantly lower levels of %pSTAT5 cells and lower peak mean fluorescence intensity (MFI) of pSTAT5 expression when compared with VMR or IL15 positive control (Fig. 3a,b, gating strategy in Supplementary Fig. 1). Importantly, no spontaneous pathway activation was detected in any of the three VMR variants, indicating that VMRs are fully ligand inducible. To assess whether VEGFA triggers receptor dimerization or intramolecular reorganization of pre-formed dimers, we characterized the size of VMR in transgenic T cells by western blot in the presence or absence of VEGFA. To facilitate detection of receptor oligomers, samples were treated with the crosslinking reagent BS3 (Methods). In the absence of VEGFA, we found VMR as monomers, while in the presence of VEGFA and BS3, VMR was detectable as ligand-bound oligomers (Fig. 3c, Supplementary Fig. 2 and original blot supplement file). These observations imply that VMR signalling is at least partly controlled by VEGFA-induced dimerization. For CMR, CMR had the strongest constitutive baseline activity in the absence of CSF1, while CMR and CMR had significantly lower baseline activity. In addition, all CMR variants were highly inducible in the presence of the ligand CSF1 (Fig. 3d,e, gating strategy in Supplementary Fig. 2).
Overall, the measured signal transductions are consistent with the intended designed properties. Higher coupling in VMRs locks the responder in the inactive monomeric state in the absence of ligand, hence lowering constitutive activity while promoting potent switching and activation upon ligand binding (Fig. 1f). The stronger coupling in VMR results in ligand binding driving a higher proportion of receptors into the active state than in VMR and VMR. Owing to lower communication, the c-MPL responder in CMRs often occupies the active state and triggers constitutive activity (Fig. 1g). The lower communication in CMR results in higher basal activity than in CMR and CMR while maintaining maximal activity in the presence of CSF1. In addition, within each family of sensors, we observed a linear relationship between the ligand-induced shifts in activity (that is, induced-constitutive) and the calculated coupling values (Fig. 3f,g and Supplementary Table 2). These findings indicate that our calculated coupling metrics are strong determinants of receptor signalling activity.
The first step in c-MPL signalling is activation of the JAK/STAT pathway, leading to phosphorylation of STAT5 and STAT3. In addition, c-MPL also activates the PI3K/AKT/mTOR axis as well as the MAPK/ERK1/2 pathway (Fig. 4a). To gain deeper insight into pathway activation upon VMR, CMR or c-MPL activation in transgenic human T cells by the respective recombinant ligands (VEGFA, CSF1 and thrombopoietin (TPO)), we assessed phosphorylation of STAT5, STAT3, S6 and ERK1/2. We found that the c-MPL endo-domain used in VMR and CMR reliably transmitted signals activating all evaluated components, and that the profile of ligand-dependent pathway activation was comparable between VMR, CMR and c-MPL transgenic T cells (Fig. 4b-f, gating strategy in Supplementary Fig. 3).
Both serum VEGF and CSF1 levels have been intensively studied across tumour histologies and are significantly higher in patients with cancer than in healthy individuals. Body compartment distribution of VEGFs and CSF1 is altered in cancer, indicating that T-SenSER T cells will likely encounter higher VEGF or CSF1 levels in malignant than in normal tissues, thereby mediating enhanced tumour specificity. To determine the activation threshold and the half maximal effective concentration (EC50) of both VMR and CMR in response to ligand, we quantified pSTAT5 levels in VMR and CMR T cells in response to increasing concentrations of VEGFA or CSF1 (Fig. 4g-j, gating strategy in Supplementary Fig. 4). The EC50 for VMR was 144 pg ml (Fig. 4h), while the EC50 for CMR was 315.8 pg ml (Fig. 4j).
To evaluate our strategy in different cancer models, we selected metastatic lung cancer and MM. Metastatic lung cancer is a solid tumour that is difficult to cure despite the introduction of immune checkpoint blockade therapy, and CAR-T or TCR-T cells are still in early development. MM has approved CAR-T cell therapies, but current products and indications do not confer long-lasting remissions, thus enhanced targeting moieties or combinations with novel approaches are highly warranted. To model the various VEGFA or CSF1 levels in mouse xenografts and compensate for the lack of a human TME as a source of human VEGFA and CSF1, we engineered A549.GFP-ffLuc cells with VEGFA and MM.1S.GFP-ffLuc cells with CSF1 overexpression (OE). In vitro, A549.GFP-ffLuc.VEGFA-wild-type (WT) cells produced low levels of VEGFA, contrary to A549.GFP-ffLuc.VEGFA-OE cells, which secreted VEGFA levels above the VMR EC50 threshold (Supplementary Fig. 5a). Next, A549 cells were engrafted intravenously in immunocompromised NOD-SCID-γ-chain (NSG) mice and lung tissue levels of VEGFA were determined after 3 days (Fig. 4k,l). VEGFA levels in the lung remained below the EC50 for VMR activation in A549.VEGFA-WT-engrafted mice (VEGFA model), while a median of 1.775 ng VEGFA per g total protein (range 1.725-3.061 ng g, n = 5) was reached in lung tissues of A549.VEGFA-OE-engrafted mice, significantly above the EC50 for VMR activation (VEGF model) (Fig. 4l). VEGFA tissue levels in the VEGF model were on average 126-fold lower than the levels reported in patients with lung cancer (median 224 ng VEGFA per g total protein, range 30-1,870, n = 71). The VEGF model is therefore appropriate to evaluate CAR.VMR T cell function in vivo but may underestimate VMR potency owing to lower tissue levels in the animal model compared with patient tissues. In MM, MM.1S.GFP-ffLuc.B2MKO.CSF1-WT cells did not produce detectable CSF1 in vitro. However, high levels of human CSF1 were detected from engineered MM.1S.GFP-ffLuc.B2MKO.CSF1-OE cells, above the CMR EC50 levels (Supplementary Fig. 5b). To quantify CSF1 levels in vivo, we engrafted both types of MM.1S cell intravenously in NSG mice and analysed BM lysates (Fig. 4m,n). In BM of mice engrafted with MM.1S.GFP-ffLuc.B2MKO.CSF1-OE, we detected CSF1 levels above the CMR EC50 with a median of 3.998 ng ml (range 2.360-13.878 ng ml, n = 5), while no CSF1 was detected in BM of MM.1S.GFP-ffLuc.B2MKO.CSF1-WT-engrafted mice (Fig. 4n). We termed our two models CSF1 and CSF1 models respectively and used these to characterize CMR + CAR-T cell function in vivo.
To maximize the impact of the designed T-SenSERs on CAR-T cell functions, we selected VMR, the VMR construct with the highest signalling response to VEGF, and CMR, the CMR construct combining the strong response to CSF1 with the highest basal activity to favour also constitutive homeostasis and enhanced effector function in the absence of cytokines. To evaluate T-SenSER activity in vivo in relation to the levels of ligand present in the TME, we established animal models with different levels of VEGFA or CSF1 that reflect the clinical situation of patients with lung cancer and myeloma.
VMR was efficiently co-transduced in activated human T cells with conventional second-generation (28ζ, BBζ) or non-signalling control (Δ) CARs targeting the antigen ephrin A2 (EphA2). CAR expression levels were comparable between CAR and CAR.VMR transduced cells, and VEGFA-induced STAT5 phosphorylation was comparable between VMR and CAR.VMR T cells (Fig. 5a,b, gating strategy in Supplementary Fig. 6). The 4H5 single chain variable fragment recognizes a conformational epitope of EphA2 that is exposed on a wide variety of malignant but not on normal cells, including A549 lung cancer cells (Supplementary Fig. 7). The impact of VMR activation on tumour killing and T cell expansion was assessed in sequential co-cultures where T cells were repetitively challenged with fresh tumour cells ± VEGFA (Fig. 5c,d). Tumour killing and T cell expansion were quantified after each challenge. Full T cell activation with target cell killing and sustained T cell expansion occurred in the presence of both tumour cells and VEGF (Fig. 5e-g). Cytotoxicity was sustained in vitro even in T cells transduced with CAR alone (Fig. 5e), but T cell expansion was enhanced in the presence of VEGFA and VMR signalling (Fig. 5f,g). A slight but not significant enhancement of both killing and T cell expansion with CAR.VMR T cells was observed without exogenous VEGFA addition. This is probably due to low levels of VEGFA production by A549-WT cells used in the assay (Fig. 5e-g and Supplementary Fig. 5a). VMR activity in the presence of VEGFA did not alter cytokine nor cytolytic granule production when compared with CAR alone and did not impact the T cell subset composition and differentiation status of T cells (Supplementary Figs. 8 and 9).
To assess the potential for long-term persistence of engineered T cells in the absence of EphA2 tumour, we performed a 4-week homeostatic maintenance experiment. VMR activation by VEGFA alone provided T cell homeostasis and survival. In line with our previous observations on transgenic c-MPL signalling in human T cells, overall expansion levels were significantly increased with VEGFA compared with media alone but remained below those of IL-2 control (Fig. 5h-j, gating strategy in Supplementary Fig. 10).
Lastly, we hypothesized that VMR signalling is complementary to EphA2-CAR BBζ signalling since the pathways activated by the endo-domains are largely distinct. Thus, we assessed differential gene expression analysing global immune response signatures in BBζ.VMR T cells after one in vitro tumour challenge ± VEGFA, and also compared BBζ with BBζ.VMR T cells in the presence of VEGFA. We identified several highly differentially expressed genes that are associated with enhanced T cell co-stimulation (for example, CD80, TNFRSF8 and HLA class II molecules) or enhanced effector function (for example, GNLY) in cells with VMR stimulation. VMR activation also led to a reduction in expression of genes associated with T cell exhaustion (for example, CTLA4, LAG3 and TIGIT), or factors associated with immune suppression (for example, reduced transcription of NT5E, TGFB1, increased transcription of ADA) (Fig. 5k,l).
These results suggest that, as intended, VMR activation delivered signals for CAR-T cell expansion and persistence, and involved transcriptional changes associated with co-stimulation, effector function and reduced exhaustion in combination with a BBζ CAR.
CMR was co-expressed efficiently in activated human T cells with A proliferation-inducing ligand (APRIL)-based CARs targeting two antigens expressed on MM: B cell maturation antigen (BCMA) and transmembrane activator and CAML interactor (TACI). Monomers of APRIL (m) were used as ligand-binding domains and CARs were expressed in conventional first- (mζ) or second-generation (m28ζ, mBBζ) format as previously described (Fig. 6a,b and Supplementary Fig. 11a-c). CAR cell surface expression levels were slightly lower in CAR.CMR compared with CAR-T cells, and STAT5 phosphorylation levels were slightly lower in CAR.CMR compared with CMR T cells (Supplementary Fig. 11d,e). The impact of CMR activation on tumour killing and T cell expansion was assessed in sequential co-cultures (Fig. 6c,d) with two different MM cell lines expressing different target antigen levels (NCI-H929 (BCMATACI) and MM.1S (BCMATACI)) (Supplementary Fig. 12). CMR expression provided a significant advantage for sequential killing of NCI-H929 cells only in combination with the mBBζ CAR, while killing of MM.1S cells was enhanced with all three evaluated mAPRIL-based CARs (mζ, m28ζ and mBBζ). The CMR constitutive baseline activity was sufficient to enhance killing, which was not further improved with the addition of CSF1 at the tested effector to target (E:T) ratio (Fig. 6e). CMR also boosted T cell expansion in vitro that was significantly higher in all conditions except with the m28ζ CAR targeting NCI-H929 (Fig. 6f,g). CMR activity did not alter cytokine nor granzyme production when compared with CAR alone (Supplementary Fig. 13). During sequential co-culture, the CD4/CD8 ratio changed with an enrichment in CD8 T cells (Supplementary Fig. 14). The effector/memory differentiation status was mostly dictated by the type of CAR and not significantly impacted by CMR activity (Supplementary Fig. 15). To assess the impact of CMR activity on T cell exhaustion/activation phenotype, we assessed expression of LAG3, CTLA4, TIGIT, PD1 and TIM3 at baseline and after 5 tumour challenges and compared % positive populations and marker MFI on CAR and CAR.CMR T cells. We found a trend to decreased expression of LAG3, CTLA4 and PD1 on mBBζ CAR.CMRCD4 T cells, and increased TIM3 on mBBζ CAR.CMR CD8 T cells (Supplementary Fig. 16). Lastly, in the 4-week homeostatic maintenance experiment in the absence of tumour, we found that the CMR constitutive baseline activity in T cells was sufficient to mediate significant low-level expansion and survival. Addition of CSF1 significantly enhanced expansion and survival of CMR T cells, comparable to the levels of non-transduced (NT) controls supplemented with IL2. Adding IL2 to CMR T cells significantly augmented the effects of the baseline constitutive activity, indicating that the combination of c-MPL signalling and native common γ-chain cytokine signals mediated by IL2 is complementary (Fig. 6h-j and Supplementary Fig. 17).
These results demonstrate that the constitutive baseline activity of CMR enhanced the homeostatic expansion capacity of engineered T cells and increased the in vitro sequential killing and expansion capacity of mAPRIL CAR-T cells in situations of high tumour loads.
We next evaluated the in vivo impact and VEGF dependence of VMR function in EphA2-BBζ CAR-T cells. In the systemic VEGF model, NSG mice engrafted with A549.GFP-ffLuc.VEGFA-WT cells were treated with a single limiting dose of 1 × 10 T cells. Bioluminescence imaging (BLI) revealed partial response to BBζ CAR-T cell therapy, and as expected, no impact of VMR addition was detected (Extended Data Fig. 6). In the systemic VEGF model, mice were engrafted with A549.GFP-ffLuc.VEGFA-OE cells (Fig. 7a). Mice treated with Δ or BBζ CAR-T cells had rapidly progressive disease and reached the experimental endpoint within 10 days. By contrast, mice treated with BBζ.VMR T cells mounted a potent anti-tumour response (Fig. 7b,c), associated with a significant reduction in VEGFA serum levels (Fig. 7d). Most importantly, the overall survival of mice treated with BBζ.VMR T cells was significantly enhanced compared with BBζ CAR-T cell-treated mice (Fig. 7e).
Next, we assessed whether a ligand-independent constitutively active c-MPL receptor mediated similarly potent effects in vivo as VMR in VEGFA-rich tumours. We generated a retroviral vector to express the c-MPL mutant (Extended Data Fig. 7a) previously described as a constitutively active c-MPL variant in patients with myeloproliferative neoplasms. We first confirmed that c-MPL can be expressed as transgene in T cells and mediated spontaneous STAT5 phosphorylation (Extended Data Fig. 7b, gating strategy in Supplementary Fig. 18a-c). Then, we co-expressed c-MPL with the EphA2-BBζ CAR in T cells (Extended Data Fig. 7c, gating strategy in Supplementary Fig. 18d,e) to evaluate the impact of c-MPL on in vivo anti-tumour function of T cells in the VEGFA systemic lung cancer model, using a limiting T cell dose of 1 × 10 cells per mouse (Extended Data Fig. 7d). BBζ.VMR T cells mediated a significantly stronger anti-tumour response than BBζ.c-MPL or BBζ T cells. The activity of BBζ.c-MPL T cells was comparable to T cells expressing the BBζ CAR alone, indicating that c-MPL signalling in T cells is most beneficial when activated with a ligand-dependent tumour-specific T-SenSER such as VMR (Extended Data Fig. 7e-g).
To evaluate whether VMR also enhanced BBζ CAR-T cell function in a subcutaneous solid tumour model, we engrafted A549.GFP-ffLuc.VEGFA-OE cells subcutaneously in the flanks of NSG mice. After stable tumour engraftment, mice were treated with a limiting dose of 5 × 10 T cells intravenously and anti-tumour activity was assessed by BLI. We found again a significantly better anti-tumour response in mice treated with BBζ.VMR T cells with clearance of their tumours within 2 weeks, compared with BBζ CAR-T cells or untreated controls (Fig. 7f-i), confirming the results obtained with the systemic VEGF model.
Thus, EphA2-BBζ CAR-T cells equipped with the VMR T-SenSER provided potent VEGFA-dependent in vivo anti-tumour activity in both systemic and subcutaneous lung cancer xenograft models and favoured tumour eradication.
Lastly, we assessed both the low constitutive and CSF1-dependent CMR-mediated enhancement of CAR-T cells targeting MM. In the CSF1 model, MM.1S.GFP-ffLuc.B2MKO.CSF1-WT-engrafted NSG mice were treated with a single dose of 5 × 10 T cells using the mAPRIL CAR as a model system (Extended Data Fig. 8a, gating strategy in Supplementary Fig. 19a,b). Significant but transient anti-tumour activity was observed in the mBBζ but also in the mBBζ.CMR treatment groups, with no benefit observed by the co-expression of CMR (Extended Data Fig. 8b-e). Thus, unlike the in vitro results, the constitutive baseline activity of CMR was not sufficient to enhance mAPRIL CAR-T cell potency in vivo. To test whether endogenous tissue levels of human CSF1 are sufficient to mediate enhanced anti-tumour activity to CAR.CMR T cells, we used NSG-Quad mice that express transgenic human CSF1 in tissues as recipients (CSF1 model). We systemically engrafted MM.1S.GFP-ffLuc.B2MKO.CSF1-WT cells, treated mice with fully human heavy-chain-only BCMA directed FHVH33-BBζ or FHVH33-BBζ.CMR T cells, and followed tumour growth by BLI (Fig. 8a, gating strategy in Supplementary Fig. 20). We found more potent anti-tumour responses with FHVH33-BBζ.CMR than with FHVH33-BBζ T cells at limiting dose. Tumour progression was significantly delayed, and survival of mice prolonged in the FHVH33-BBζ.CMR treatment group (Fig. 8b-e). Finally, in the CSF1 model, co-expression of CMR along with a CAR also significantly improved outcomes of mice, either when combined with the mBBζ or the FHVH33-BBζ CAR (Fig. 8f-l, gating strategy in Supplementary Fig. 21). We found significantly enhanced tumour control (Fig. 8h-j) and improved survival of mice (Fig. 8k,l) at limiting T cell doses.
Thus, our results demonstrate that the CMR T-SenSER significantly enhanced anti-tumour activity of CAR-T cells against MM in a CSF1-dependent manner. Enhanced T cell potency was achieved both in response to physiological tissue levels of human CSF1 in NSG-Quad mice and in an engineered cell line model with CSF1 overexpression by tumour cells.