Screening for CD19-specific chimaeric antigen receptors with enhanced signalling via a barcoded library of intracellular domains

Abstract

The immunostimulatory intracellular domains (ICDs) of chimaeric antigen receptors (CARs) are essential for converting antigen recognition into antitumoural function. Although there are many possible combinations of ICDs, almost all current CARs rely on combinations of CD3ζ, CD28 and 4-1BB. Here, we show that a barcoded library of 700,000 unique CD19-specific CARs with diverse ICDs cloned into lentiviral vectors and transduced into Jurkat T cells can be screened at high throughput via cell sorting and next-generation sequencing to optimize CAR signalling for antitumoural functions. By using this screening approach, we identified CARs with new ICD combinations that, compared with clinically available CARs, endowed human primary T cells with comparable tumour control in mice and with improved proliferation, persistence, exhaustion and cytotoxicity after tumour rechallenge in vitro. The screening strategy can be adapted to other disease models, cell types and selection conditions, and could be used to improve adoptive cell therapies and to expand their utility to new disease indications.

Screening a large barcoded library of unique CD19-specific chimaeric antigen receptors with diverse intracellular domains allows for the discovery of receptors that elicit enhanced antitumoural functions.

Cancer immunotherapies that reinvigorate or reprogram antitumour T cell responses have been transformative in the treatment of a broad range of malignancies.¹ Chimeric antigen receptors (CARs) engineer T cells to leverage these mechanisms by linking intracellular immunostimulatory signalling domains to extracellular recognition domains typically derived from antibody single chain variable fragments (scFv). Upon stable introduction into T cells, the CAR redirects effector functions toward a clinically relevant target antigen. This has been most widely used in the context of B cell malignancies, where a majority of patients show complete responses within the first few months.² However, many patients later relapse, and most patients experience cytokine release syndrome or neurological impairment.² Further, CAR-T treatments have yet to meaningfully translate to solid tumours, which pose distinct immunological challenges.

Retrospective clinical studies have established which CAR-T cell phenotypes are beneficial for long-term efficacy and safety.^2,3 These phenotypes can be induced via the composition of the CAR signalling domains. While intracellular domains (ICDs) from CD3ζ in combination with CD28 and/or 4-1BB are most commonly used in current CARs,^3,4 several groups have shown that incorporating new signalling components can enhance persistence, proliferation, cytotoxicity, resistance to exhaustion, memory formation, and in vivo survival benefit.^5-7 Additionally, shortening the distance between the CD3ζ immunoreceptor tyrosine activation motifs (ITAMs) and the membrane can enhance their function, and it has been demonstrated that a CAR containing a single ITAM produced superior persistence and in vivo tumour control than the canonical sequence.⁸ The effects of distinct signalling domains can also synergize when arranged in the optimal spatial configuration.^9-11 While there is a steadily expanding compendium of CARs using new signalling domains to confer functions, these compositions undersample all possible signalling domain combinations that could be beneficial in CAR constructs, possibly impeding efficacy and translation to other diseases.

The relative scarcity of tested signalling domain combinations is in part a result of the time and effort needed to individually design and test new CARs. We hypothesized that a more systematic optimization of CAR signalling domains has the potential to produce CAR-T cell behaviours that could elicit safer, more effective therapies. To this end, we created a 700,000-member CAR library with diversified ICDs and coupled it with a selection strategy that enables enrichment of CAR-T cells exhibiting desirable antitumour functions (which, collectively, we term CARPOOL). Here, we demonstrate that CARPOOL can be used to select and identify functional ICD combinations, including combinations containing domains that have not been previously described in the context of CAR signalling. When we characterized several new CAR ICD combinations, we determined that they show distinct, and in some contexts superior, in vitro and in vivo activity relative to a conventional 19BBζ CAR (BBζ). Taken together, this evidence suggests that CARPOOL can rapidly identify CAR ICDs that more extensively explore the possible functional landscape of cellular immunotherapies, and poses a promising strategy to address current challenges faced by CAR-T therapies.²

Results

In order to design a CAR library, we identified 89 signalling domains derived from different immune cell types and functional families (Supplementary Table 1), which were then incorporated at random into each of the three intracellular positions in a 3rd generation CD19 CAR construct. Each CAR was cloned into a lentiviral transfer vector along with a randomized 18-nucleotide barcode sequence in the 3’UTR, yielding a theoretical diversity of 700,000 uniquely barcoded signalling domain combinations (Fig. 1a). We then generated lentiviruses from the CAR plasmid library, and transduced 1x10⁸ Jurkat T cells at a multiplicity of infection (MOI) of 0.5 to minimize multiple integrations per cell while maintaining reasonable transduction efficiency (Supplementary Fig. 1). This produced 3x10⁷ CAR library transduced cells, as confirmed by epitope tag staining of the CAR extracellular region, achieving approximately 40-fold coverage of our theoretical library size.

Fig. 1 ∣. Selecting for CD69^high expression enriches CARs encompassing ITAM-signalling ICDs.

a, Schematics describing the CARPOOL system. CARPOOL uses a signalling diversified library of CD19-specific chimeric antigen receptors containing 1-3 ICDs and combines cell sorting and next generation sequencing to select for function in human T cells. All CARs were bi-cistronically expressed with EGFP via an IRES sequence. b, Frequency of top 30 CAR clones throughout rounds of selection in Jurkat T cells for CD69^high and CD69^highPD-1^low expression following stimulation with 10 nM rhCD19, coloured by number of ICDs. c, Frequency of each family of ICDs throughout rounds of selection.

We next FACS sorted for cells based on EGFP expression, which was bi-cistronically expressed with each CAR construct. For the first round of selection, we stimulated the EGFP-enriched T cell pool with 10 nM of soluble recombinant human CD19 protein (a.a. 1-270) (rhCD19) and subsequently sorted the stimulated EGFP⁺ cells that expressed the highest levels of CD69, a canonical T cell activation maker with robust dynamic range in Jurkat cells (CD69^high).^12,13 We then split this selected population in two in order to conduct two parallel selections consisting of either two more rounds of selection in serial for CD69^high expression to further enrich for robustly activatable CARs, or two rounds of selection for CD69^highPD-1^low expression to potentially identify ICD combinations that render T cells less susceptible to exhaustion (Fig. 1a). The proportion of CD69⁺ cells increased substantially throughout rounds of both CD69^high and CD69^highPD-1^low selections, indicating an enrichment for functional CAR-T cells: 72% and 90% of total cell populations were EGFP⁺CD69⁺ cells after the last round of CD69^high and CD69^highPD-1^low selections, respectively (Supplementary Fig. 2).

Next, we performed next-generation sequencing (NGS) to track the prevalence of enriched barcodes in selected cells from each round of selection (Supplementary Table 2). We observed a dramatic reduction in clonal diversity following each round of both CD69^high and CD69^highPD-1^low selections, where the top 25 most enriched barcodes represented 85% and 98% of all clones from the last rounds of CD69^high and CD69^highPD-1^low selections, respectively (Supplementary Fig. 3). Given the limited read length capacity of Illumina sequencing, we used PacBio SMRT long-read sequencing of amplicons derived from CARPOOL transduced Jurkats that encompass both the CAR ICDs and barcode region in order to build a lookup table to link the enriched barcodes to the ICD combinations. Of note, only ~70% of barcodes in the top 30 most frequent CARs from all selections were identified in the PacBio data, potentially due to read depth limitations in the SMRT sequencing. Additionally, not all identified CARs encompassed 3 ICDs, likely due to infidelities in the library assembly step (Fig. 1b).

Enriched CARs encode unique combinations of activating and costimulatory ICDs.

We quantified the average frequency of each ICD at each selection step to determine which classes of ICDs were most enriched. Our result verified that ITAM-containing ICDs—which are the canonical activation motifs for antigen receptor signalling,^14,15 were more prevalent than other classes of ICDs—especially in later rounds of CD69-based selection, consistent with a requirement for ITAMs for robust CAR activation (Fig. 1c).

Analysis of our sequencing data revealed enrichment of signalling domains that have not been vetted for function in a CAR, alongside commonly used ICDs such as CD28, 4-1BB and other ICDs that have been more recently described as functional.^7,16-20 However, it is notable that we identified a BBζ CAR clone in our library that was less enriched compared to our top clones, especially in the last rounds of both CD69^high and CD69^highPD-1^low selections, indicating that these new clones possess competitive advantages over the BBζ CAR in these selection conditions (Fig. 2a). Although we did not identify a 28ζ or a 28BBζ CAR in our dataset, we found that the aggregate frequencies of all CARs containing either CD28, 4-1BB, or CD3ζ throughout rounds of selection were lower than those of the most enriched CAR clones (Fig. 2a and Supplementary Fig. 3). In order to visualize and track the extent of enrichment for each ICD, we generated heatmaps for each round of selection that depict the frequency of each ICD at each position (Fig. 2b and Extended Data Fig. 1). This analysis suggested that previously uncharacterized combinations of co-stimulatory and ITAM-containing ICDs, such as FcεR1γ, 2B4, and CD3ε ITAM or CD79β, DAP12, and CD40, were overrepresented following selection relative to more commonly used signalling domain combinations using 4-1BB, CD28, and CD3ζ (Fig. 2a). This exemplifies the utility of CARPOOL in revealing useful signalling domain combinations in a rapid, streamlined process. Notably, our enriched data identified ICDs that are contained in FDA approved 2nd generation CARs (CD28, 4-1BB, and CD3ζ), as well as domains in clinical trials or under consideration for 2nd and 3rd generation CARs (Fig. 2b and Extended Data Fig. 1).

Fig. 2 ∣. CARPOOL selections reveal new functional signalling domain configurations.

a, Clonal frequency of enriched CARs that were identified throughout rounds of CD69^high and CD69^highPD-1^low selections compared to a BBζ CAR, along with aggregate frequency of CARs containing 4-1BB, CD28, or CD3z ICDs. b, Heatmap representing log₂(fold-change) of individual ICDs found in 3rd generation CARs throughout rounds of selection, oriented by ICD position within the CAR relative to the plasma membrane. Fold-changes were normalized to frequencies in the EGFP⁺ sorted population prior to selection.

CARPOOL identifies variants with distinct functional phenotypes in human primary T cells.

In order to compare the functional characteristics of our CARs to those of a BBζ CAR, we vetted 3 highly enriched and distinct CARs—referred to as Var1 (containing CD40, CD3ε ITAM, and DAP12 ICDs), Var2 (FcεR1γ, 2B4, and CD3ε ITAM ICDs), and Var3 (FcεR1γ, OX40, and CD3ζ ITAM3 ICDs)— in Jurkat T cells. While we detected comparable surface expression and antigen sensitivity upon assessing both CAR internalization and CD69 upregulation relative to BBζ (Extended Data Fig. 2a-d), we noted that the CD69 and PD-1 expression levels of all three variants in the absence of antigen engagement were significantly lower than those of the BBζ CAR (Extended Data Fig. 2e-f). This indicates that they induce less tonic signalling in Jurkats—a feature that has been reported to lead to rapid CAR-T cell dysfunction marked by deficient IL-2 and IFN-γ production.^21,22 Upon antigen stimulation, we noted that Var1 and Var3 showed robust CD69 upregulation, while Var2 consistently produced lower CD69 upregulation (Extended Data Fig. 2d).

Thus, we chose to further characterize the two more active hits from our library selections: Var1, which was highly enriched in both CD69^high and CD69^highPD-1^low selections, and Var3, which was enriched in the CD69^high selection but not in the CD69^highPD-1^low selection (Fig. 3a). These were chosen not only for their unique signalling compositions and relative enrichment, but also to explore whether a PD-1^low selection criteria enabled identification of CARs with reduced susceptibility to T cell exhaustion.

Fig. 3 ∣. CAR variants show enhanced cytotoxicity and cytokine secretion in response to antigen stimulation.

a, Design of CD19 targeted CAR candidates. b, Cytotoxicity of human primary CD8⁺ CAR-T cells following 24-hour co-culture with FLuc⁺ CD19⁺ NALM6 cells at varying E:T ratios (n = 3 technical replicates representative of 2 biological replicates). Remaining NALM6 cells were quantified by measuring bioluminescent activity. c, Proliferation of human primary CD8⁺ CAR-T cells that were stained with CellTrace dye on day 0 prior to co-culture with CD19⁺ NALM6 cells at an E:T ratio of 1:1 (n = 3 technical replicates representative of 2 biological replicates). Extent of proliferation was assessed by degree of dye dilution and measured by FACS. d, IL-2 secretion from human primary CD8⁺ CAR-T cells following 24-hour co-culture with CD19⁺ NALM6 cells at varying E:T ratios (n = 3 technical replicates representative of 2 biological replicates). IL-2 levels were measured via ELISA. P values in panel (b) are 0.001820, 0.000140, <0.000001, and 0.001293 for Var1 vs. BBζ and 0.001860, 0.000197, 0.000007, and 0.002098 for Var3 vs. BBζ in increasing order of E:T ratio, as determined using a two-tailed unpaired student’s t-test with Benjamini, Krieger, and Yekutieli correction for multiple hypotheses (df = 4 for all). Data and error bars shown are means ± s.e.m. P values in panel (d) are <0.0001 for Var1 vs. BBζ, 0.0113 for Var3 vs. BBζ, and < 0.0001 for Var1 vs. Var3 at an E:T ratio of 1:1 (df = 4). P values are < 0.0001 for Var1 vs. BBζ and <0.0001 Var1 vs. Var3 at an E:T ratio of 1:5 (df = 4). P values are <0.0001 for Var1 vs. BBζ and 0.0004 for Var1 vs. Var3 at an E:T ratio of 1:10 (df = 4). P values are 0.0007 for Var1 vs. BBζ, 0.0024 for Var3 vs. BBζ, and 0.0003 for Var1 vs. Var3 at an E:T ratio of 1:20 (df = 4). P values for panel (d) were determined using a two-tailed unpaired student’s t-test. Data shown are individual values along with means ± s.e.m. e, Polyfunctional cytokine and chemokine secretion response following 24-hour co-culture of human primary CD8⁺ or CD4⁺ CAR-T cells (n = 3 technical replicates). Concentrations were quantified for 41 analytes. Cytokines linked to immune cell function and antitumour response that were produced at significantly higher levels relative to 19BBζ are highlighted in blue or purple, while those that were secreted at lower levels are highlighted in red.

Given that the range of functional phenotypes following CAR activation is limited in Jurkats, we characterized the function of each CAR variant in human primary T cells. Since we did not observe meaningful differences between the CAR variants and BBζ in the absence of antigen (Supplementary Fig. 4), we speculated that differences in phenotype between Var1, Var3 and BBζ may be driven by antigen-induced CAR signalling. Upon co-culturing CD8⁺ T cells expressing Var1, Var3, or BBζ with CD19-expressing NALM6 cells for 24 hours, we found significant differences in killing capacity, with Var1 and Var3 exhibiting increased cytotoxicity compared to BBζ across all E:T ratios, especially at higher E:T ratios (Fig. 3b). Additionally, Var1 uniquely showed cytotoxic function upon expression in CD4⁺ T cells (Supplementary Fig. 5a). Although we found no difference in T cell proliferation at day 7 (Fig. 3c), we did observe elevated levels of IL-2 secretion by Var1 at all E:T ratios (Fig. 3d). We also assessed IFN-γ secretion—a cytokine implicated in induction of cytokine release syndrome—and found that BBζ showed slightly higher levels of IFN-γ release at high tumour burden (Supplementary Fig. 5b).^23,24

To assess whether there were further notable differences in cytokine secretion between CARs, we conducted a 41-plex Luminex assay using supernatants collected after co-culture of either CD4⁺ or CD8⁺ CAR-T cells with NALM6 cells at an E:T ratio of 1:1. We detected increased secretion of cytokines associated with antitumour effects by Var1 relative to BBζ—such as IL-2, TNF-α, and GM-CSF—as well as Var3 relative to BBζ—such as IL-2 and TNF-α, in CD8⁺ T cells (Fig. 3e and Supplementary Fig. 6). In CD4⁺ T cells, we found elevated levels of chemokines associated with attracting immune cells from both the innate and adaptive immune system compared to BBζ, including MIP1α, MIP1β, and Flt3L in the case of Var1 and MIP1α and IP-10 in the case of Var3, indicating that modification of CAR signalling can drastically alter their cytokine secretion profiles in T cells.^25-28

CAR variants exhibit lower exhaustion following long-term tumour challenge.

In order to assess how enhanced killing and antitumour cytokine secretion would affect the function of CARPOOL-enriched CAR variants over an extended period of exposure to high tumour burden, we designed a rechallenge assay in which we repeatedly added NALM6 cells every 2-3 days at an E:T ratio of 1:10 to a 1:1 mixture of CD4⁺ and CD8⁺ T cells expressing each CAR (Fig. 4a). Under these conditions, we found that Var1-expressing CD4⁺ and CD8⁺ T cells demonstrated considerably more persistent proliferative activity at later time points compared to both Var3- and BBζ-expressing cells (Fig. 4b). This expansion pattern directly correlated with target cell killing, with Var1 showing superior tumour control at late timepoints (Fig. 4c). Var1-expressing cells also showed delayed kinetics of differentiation from T cell memory to effector phenotypes, and both Var1 and Var3 CAR-T cells showed significantly reduced exhaustion marker expression by day 22 (Fig. 4d-e and Extended Data Fig. 3). Taken together, these results imply that CARPOOL can identify CARs that are less susceptible to producing an exhausted T cell phenotype upon repeated challenge with high tumour burden relative to BBζ CAR-T cells.

Fig. 4 ∣. CAR variants exhibit distinct proliferation and exhaustion profiles upon repeated antigen exposure.

a, Schematic representation of tumour rechallenge assay, in which human primary CD8⁺ and CD4⁺ CAR expressing T cells mixed in a 1:1 ratio were stimulated every 2 days with NALM6 cells at an E:T ratio of 1:10. b, T cells and (c) NALM6 cells were quantified by FACS at each time point prior to restimulation, along with (d) T cell memory differentiation markers (n = 3 technical replicates representative of 2 biological replicates). P values in panel (b) are 0.004396, 0.001890, 0.010757, 0.000033, 0.000060, 0.000003, 0.000652, 0.003049, and 0.000080 for Var1 vs. BBζ on days 5, 7, 10, 12, 14, 17, 19, 21, and 24 of rechallenge for CD8⁺ CAR-T cells. P values are 0.000340, 0.010693, 0.000019, 0.000537, 0.000374, and 0.000409 for Var1 vs. BBζ on days 3, 7, 17, 19, 21, and 24 of rechallenge for CD4⁺ CAR-T cell expansion. P values for NALM6 cell expansion in panel (c) are 0.002063, 0.005311, 0.000043, 0.000790, and 0.000589 for Var1 vs. BBζ on days 7, 10, 12, 21, and 24 of rechallenge. P values for panel (d) are 0.011291, 0.000865, 0.000019, and 0.000014 for Var1 vs. BBζ on days 4, 6, 11, and 18 of rechallenge. All P values for panels (b—d) were determined by multiple two-sided unpaired student’s t tests with Benjamini, Krieger, and Yekutieli correction for multiple hypotheses (df = 4 for all). Data shown are means ± s.e.m. e, Exhaustion marker expression of CD4⁺ and CD8⁺ CAR-T cells on day 22 of rechallenge (n = 3 technical replicates representative of 2 biological replicates). P values are 0.0001 for CD4⁺ Var1 vs. BBζ, 0.0041 for CD4⁺ Var3 vs. BBζ, 0.0028 for CD4⁺ Var1 vs. Var3, 0.0004 for CD8⁺ Var1 vs. BBζ, 0.0080 for CD8⁺ Var3 vs. BBζ, and 0.0004 for CD8⁺ Var1 vs. Var3 for PD-1 expression. P values are 0.0019 for CD4⁺ Var1 vs. BBζ, 0.0042 for CD4⁺ Var3 vs. BBζ, 0.0017 for CD8⁺ Var1 vs. BBζ, 0.0034 for CD8⁺ Var3 vs. BBζ, and 0.0386 for CD8⁺ Var1 vs. Var3 for TIM3 expression. P values were determined using a two-tailed unpaired student’s t-test (df = 4). Data shown are individual values along with means ± s.e.m.

CAR variants induce unique transcriptional responses.

Given that the variant CARs exhibited unique functional phenotypes relative to BBζ in vitro, we asked whether the signalling perturbations introduced by the ICD combinations in Var1 and Var3 produced unique transcriptional programs in response to tumour challenge. We therefore subjected primary CD4⁺ and CD8⁺ T cells expressing each construct to three NALM6 challenges at a high-burden E:T ratio of 1:10 over a five-day period (on days 0, 3, and 5) before performing single-cell RNA-sequencing (scRNA-seq) 48 hours after the third tumour challenge. Dimensionality reduction and unsupervised clustering revealed five transcriptionally distinct cell clusters (Fig. 5a). While CD4⁺ and CD8⁺ cells cluster together to some extent (Extended Data Fig. 4), the separation between clusters was mainly defined by CAR variant, with Var1-expressing cells confined primarily to C0 and C1, and BBζ- and Var3-expressing cells localized to C2, C3, and C4 (Fig. 5a); notably, all CAR-expressing cells also clustered separately from unstimulated, untransduced controls (Supplementary Fig. 7). Chi-squared analysis confirmed significant enrichment of Var1 cells in C0 and C1 (Fig 5b), suggesting that the Var1 construct drives a unique transcriptional phenotype following repeated high-burden tumour challenge.

Fig. 5 ∣. CARs identified by CARPOOL differ in transcriptional profile and predicted efficacy correlates.

a, UMAP embeddings of merged scRNA-seq profiles coloured by cell state (left) and CAR identity (right) following rechallenge of human primary CD8⁺ and CD4⁺ CAR-T cells (mixed in a 1:1 ratio) with NALM6 cells three times at an E:T ratio of 1:10 (n = 2,082 cells). b, Chi-square enrichment values for each CAR candidate within each cluster, represented by the Pearson residuals measuring the difference between the observed and expected CAR frequencies within each cluster (df = 8). c, Heat map representing the normalized expression of the top 50 differentially expressed genes within each cluster, as determined by Wilcoxon-Rank Sum test with Bonferroni correction. d, Single-cell gene set variation analysis (scGSVA) scores measuring enrichment of a previously published BBζ-specific gene set within each T cell cluster.^33,46 Bars represent median scGSVA values. P values in panel (d) were determined using a two-sided Wilcoxon-Rank Sum test. (e) Tumour infiltrating lymphocyte (TIL) signatures correlated with tumour growth response in checkpoint blockade-treated melanoma, and CAR-T cell signatures upregulated in CAR-treated large B cell lymphoma (LBCL) patients with complete response. Bars represent median scGSVA scores. P values were determined using a two-sided Wilcoxon Rank-Sum test.

To characterize the transcriptional program associated with Var1 signalling, we performed differential expression analysis to define cluster-specific markers (Fig. 5c and Supplementary Table 3). Among the most significantly overexpressed genes in both C0 and C1 were several genes related to T cell activation (IL2RA, TNFRSF4), including a broad repertoire of MHC class II-related genes (Supplementary Fig. 8a). In addition, we observed significant upregulation of memory markers such as CCR7, as well as genes within pathways thought to promote T cell persistence and memory, including those related to non-canonical NF-kB signalling (BIRC3, TRAF1, NFKB2)^29,30 and AP-1 transcription factors (BATF3, JUNB)³¹ (Supplementary Fig. 8b-c). Although the transcriptional profiles defined by C0 and C1 were largely similar, cells within C1 uniquely overexpressed TCF7 and LEF1, which encode for transcription factors thought to be important for T cell stemness and memory (Supplementary Fig. 8c).³² Similar hits were observed when directly comparing average expression profiles of Var1 cells to BBζ (Extended Data Fig. 5 and Supplementary Table 4). Considering these differences at the gene level, we sought to characterize Var1-associated programs at the pathway level; we employed single-cell gene set variation analysis (scGSVA) to assign scores to each cell within the dataset on the basis of their relative expression of genes within canonical and curated T cell gene sets (Supplementary Table 5).³³ Interestingly, one of the most significantly upregulated pathways among Var1-expressing cells was a gene set recently reported to distinguish 19BBζ CARs from 1928ζ CARs (Fig. 5d),³⁴ suggesting that Var1 might induce similar transcriptional programs to that of BBζ but to a greater extent.

Indeed, orthogonal pathways with minimal gene overlap that have been separately demonstrated to distinguish BBζ CARs are similarly enriched in C0 and C1 (Supplementary Fig. 9a).^29,34 This response may be mediated by increased activation of transcription factor pathways known to regulate T cell differentiation and activation, such as non-canonical NF-kB,^29,30 NFAT,³⁵ and AP-1,³¹ members of which are significantly upregulated in C0 (Supplementary Fig. 9a-b). Taken together, these results suggest that Var1 triggers a coordinated transcriptional response that promotes enhanced persistence and long-lived memory formation relative to BBζ in response to antigen stimulation.

Although CAR-T cells expressing Var3 largely clustered with those expressing BBζ (Fig. 5a-b), direct comparison of pseudobulk expression profiles using scGSVA revealed subtle transcriptional shifts associated with Var3. In particular, Var3 exhibited increased expression of cytotoxicity-related gene sets³⁶ but decreased levels of T cell activation;³⁴ Var1 exhibited the opposite trend (Supplementary Fig. 10). To investigate predicted clinical benefit associated with each CAR variant, we used a similar strategy to probe our dataset with recently reported gene signatures of clinical response. Notably, Var3-expressing cells appeared to upregulate tumour-infiltrating lymphocyte (TIL) signatures correlated with melanoma regression under checkpoint blockade therapy, while downregulating correlates of melanoma progression (Fig. 5e).³⁷ In contrast, Var1-expressing cells exhibited strong upregulation of CAR-T cell signatures correlated with complete response in B cell lymphoma patients (Fig. 5e), suggesting different CAR signalling paradigms might be better suited for different tumour types.³⁸

CAR variants produce differential tumour control in liquid and solid tumour settings.

We next benchmarked the CARs discovered via CARPOOL in vivo to determine if they produced distinct outcomes. We first tested a xenograft mouse model of B cell leukaemia by injecting NOD/SCID/IL2R^null (NSG) mice with 5x10⁵ luciferase-expressing (FLuc⁺) NALM6 cells intravenously followed by treatment with a 1:1 mixture of untransduced or CAR-expressing CD4⁺ and CD8⁺ T cells 4 days later (Fig. 6a). When treated with 2x10⁵ CARs, Var1-treated mice demonstrated a subtle but not statistically significant increase in survival compared to BBζ-treated mice, with no discernible signs of toxicity (Fig. 6b and Supplementary Fig. 11). Upon collecting peripheral blood on day 27, we detected significantly elevated levels of EGFP⁺ CAR-T cells in Var3-treated mice relative to BBζ-treated mice, while the number of circulating Var1 CARs was significantly lower despite producing a similar survival benefit (Fig. 6c). All groups of mice exhibited comparable levels of circulating CD19⁺ NALM6 cells (Fig. 6d), and both Var1- and Var3-treated mice showed similar tumour progression relative to BBζ-treated mice (Fig. 6e and Supplementary Fig. 12). We additionally treated mice with a higher dose of 1x10⁶ CARs. While we observed sustained remission across all groups at this dose (Supplementary Fig. 13 and Extended Data Fig. 6), in a separate cohort of mice that received the same dose of CARs but were later rechallenged with NALM6 cells at days 44 and 58 we observed a subtle survival advantage in BBζ-treated mice relative to Var1-treated mice, possibly owing to delayed kinetics of tumour control in vivo and/or differences in CAR-T abundance after the initial tumour challenge (Supplementary Fig. 14 and Extended Data Fig. 7).

Fig. 6 ∣. CAR variants show similar tumour control to that of a BBζ CAR in vivo.

a, Experimental design: NSG mice were intravenously infused with 5x10⁵ FLuc⁺ CD19⁺ NALM6 cells, then treated with 2x10⁵ mixed CD4⁺ and CD8⁺ (1:1) CAR-T cells or untransduced control T cells (n = 8 mice for untransduced, 9 for BBζ, 8 for Var1, and 10 for Var3). b, Kaplan-Meier curve for overall survival. P values, as determined using a Log-rank test, are 0.1268 for Var1 vs. BBζ and 0.0788 for Var3 vs. BBζ (df = 1). c, Concentrations of EGFP⁺ CAR-T cells and d, CD19⁺ NALM6 cells in the peripheral blood were determined on day 27 after ACT by FACS (n = 8 mice for BBζ, 6 for Var1, and 4 for Var3). P values for panel (c) are 0.0067 for Var1 vs. BBζ (df = 12) and <0.0001 for Var3 vs. BBζ (df = 10). P values were determined using a two-tailed unpaired student’s t test. Data shown in panels (c,d) are individual values with means ± s.e.m. Tumour bioluminescence was assessed every 3-7 days by imaging for luciferase activity. e, Quantification of total photon counts are shown. Data shown are individual values. f, Experimental design: NSG mice were subcutaneously injected with 5x10⁵ CD19⁺ A375 cells, then treated with 1x10⁶ mixed CD4⁺ and CD8⁺ (1:1) CAR-T cells or untransduced control T cells (n = 8 mice for untransduced, 8 for BBζ, 7 for Var1, and 8 for Var3). g, Kaplan-Meier curve for overall survival. P values, as determined using a Log-rank test, are 0.0577 for Var1 vs. BBζ and 0.9344 for Var3 vs. BBζ (df = 1). h, Days post ACT to complete remission. The P value for panel (h) is 0.0026, as determined using a two-tailed unpaired student’s t test (n = 4 mice for BBζ, n = 5 for Var3, df = 7). Data shown are individual values with means ± s.e.m. i, Tumour area was assessed every 2-4 days by calliper measurement (arrow indicates day of ACT). P values for panel (i) are 0.0363 and 0.0307 for days 59 and 62 post tumor inoculation, as determined by a two-way ANOVA with Dunnett’s multiple comparisons correction (n = 6 mice on day 59 and 5 mice on day 62 for BBζ, n = 5 for Var3, df = 260). Data shown are individual values.

Based upon our transcriptomic analyses showing the potential for improved solid tumour control for Var3-expressing CARs, (Fig. 5e), we next assayed whether the distinct phenotypes produced by our CARs might improve their therapeutic efficacy in a solid tumour setting. To this end, we injected NSG mice subcutaneously on the right flank with 5x10⁵ A375 melanoma cells that had been engineered to produce comparable CD19 surface expression to that of NALM6 cells (Supplementary Fig. 15). We treated the animals with 1x10⁶ untransduced or CAR-expressing CD4⁺ and CD8⁺ (1:1) T cells 21 days later (Fig. 6f). In doing so, we found that Var3 produced similar survival benefit but significantly more rapid tumour regression relative to BBζ (Fig. 6g-i), with no notable toxicity (Supplementary Fig. 16). Furthermore, while a subset of both BBζ and Var3-treated mice showed complete remission by day 48, 2 out of 4 cured BBζ-treated mice later relapsed while all 5 of 5 remaining Var3-treated mice experienced sustained complete remission (Fig. 6i). Taken together, this suggests that CARPOOL can identify CAR signalling domain combinations that are capable of robust tumour control in disease indications beyond B cell malignancies.

Discussion

Despite the clinical promise of CAR-T therapies, efforts to improve CAR-T function by systematically optimizing CAR signalling remain limited. A previous effort successfully identified functionally signalling CARs from a pool of signalling domains, but was significantly smaller in scope, had a relatively high false positive rate, and lacked the sequencing analyses necessary to comprehensively examine any selected variants.³⁹ Here, we describe CARPOOL, a library-based functional screening approach for rapidly enriching and identifying new CARs with clinically useful phenotypes. As a proof of principle, we constructed a 700,000-member CD19 CAR library with diverse ICD combinations to discover new CARs based on their ability to robustly activate Jurkat T cells. We believe our observations from our selections (Fig. 1d) are representative of many library-based screening approaches: essentially every CAR included at least one ITAM domain and many contained costimulatory domains, which matches the known rules of CAR design. However, the selection, covariation, and spatial orientations of each selected construct could not have been predicted from first principles.

When we validated selected CARs, we found that Var1 CARs showed enhanced antitumour activities in vitro and comparable tumour control in vivo compared to those of the BBζ CAR currently being used in the clinic. It should be noted that the CD40 and CD3ε ICDs that comprise the two membrane-proximal ICDs of Var1 were separately reported to enhance CAR-T cell function, with CD40 augmenting CAR-T effector function and MyD88 and CD3ε promoting CAR-T cell persistence through Csk and p85 recruitment via the ITAM and basic residue rich sequence (BRS), respectively.^5,7,40 Additionally, DAP12, the membrane distal ICD of Var1, has been demonstrated as a functional ICD in both CAR-T and CAR-NK cells.^18,19 On the other hand, Var3, which harbours FcεR1γ, OX40, and CD3ζ ITAM3, shared some traits in vitro with both Var1 and BBζ, and exhibited superior long-term tumour control in an in vivo melanoma model while demonstrating faster kinetics of control. The membrane proximal ICD of Var3, FcεR1γ, was one of the earliest candidate signalling domains to be incorporated into a CAR, though its functional consequences have not been characterized extensively.^41,42 OX40 has been previously reported to enhance CAR-T cell proliferation and persistence while reducing exhaustion by activating both canonical and non-canonical NF-κB, PI3KT-AKT, and MAPK signalling,^17,43,44 while the use of CD3ζ ITAM3 in lieu of its full-length counterpart has been shown to elicit preferential differentiation of long-lived central memory subsets in a CD28-based 2^nd generation CAR, but did not improve in vivo tumour control in a B cell lymphoma model.⁸ Although each of these domains has been examined in isolation, here we have collectively characterized them. Further work may uncover the extent to which these signalling inputs synergize with each other to generate the observed T cell phenotypes, with the potential for exploring spatial configurations beyond those that were identified in the selections to offer further insight into the full spectrum of function that can be elicited by these domains.

Two key considerations for contextualizing our results are the translation between Jurkat and primary T cells, and translation to in vivo settings. While our study demonstrated that selections conducted in Jurkat cells can reliably identify new CARs, there are inherent limitations to the effector functions that can be selected for due to the physiology of the cell line. Being an immortalized cell line derived from T cell leukaemia, Jurkats continually divide, exhibit altered basal signalling and metabolism, and are not cytotoxic.^45,46 Therefore, there are desirable T cell functions such as resistance to exhaustion and efficient tumour killing that cannot be directly selected for in a Jurkat-based library. The fact that Var1 and Var3 exhibit improved phenotypes despite being identified in Jurkats may indicate that these functions correlate with features that can be selected for in Jurkat cells, or that there are a wide range of emergent properties accessible simply by altering signalling inputs.

Additionally, we consider the differences in results that were observed in vitro compared to those observed in vivo. Var1 showed superior function in many in vitro assays while appearing comparable but not significantly superior in vivo. Discrepancies in translation between the two have been well reported.^47-49 These differences are notable in that several of the differences observed in vitro involved either the production of cytokines that would not have function in a xenograft mouse model, or persistence improvements that may only be relevant upon consistent antigen exposure. It is therefore possible that the comparable results between Var1 and 19BBζ are due to commonalities in CAR function or limitations of the in vivo model itself. Furthermore, differences in both the abundance and kinetics of antigen exposure due to compartmentalization of both the CAR-T and tumour cells in an in vivo setting may add further complexity and confound translation from an in vitro setting. It is also possible that traits such as persistence and central memory formation, while important for long-term protection against recurrence, are ultimately not sufficient on their own to achieve robust tumour clearance. In the context of a solid tumour, it is notable that Var3 clears tumour significantly faster than BBζ CARs (Fig. 6h) despite only subtle differences observed between the two in vitro, while Var1 performs significantly worse than BBζ in the solid tumour paradigm. It is possible that the additional need for trafficking to and infiltration across physical barriers surrounding the tumour along with the imposition of an immunosuppressive tumour microenvironment further complicate translation between in vitro and in vivo results, amplifying some differences while dampening others.

Although the initial demonstration of CARPOOL used Jurkat cells, further extending the technique for selection in primary T cells may make a broader range of clinically relevant phenotypes available for use as selection criteria, such as persistence and maintenance of activation (4-1BB) or lack of exhaustion marker expression (PD-1, TIM3, LAG3) upon repeated long-term stimulation, formation of potent memory populations (CD62L, CD45RA), or upregulation of cytotoxicity markers (CD107a). Doing so could enable the rapid identification of CAR clones that could address challenges such as translating CAR-T cell therapies to solid tumour indications, including improving lack of persistence due to tumour-mediated immunosuppression and inefficient trafficking and tumour infiltration.^2,50-52 This may require more specialized selection strategies, such as employing T cell surface markers indicative of tumour infiltration, in addition to introducing selection conditions that mimic the microenvironment of solid tumours. As T cell-intrinsic predictors of CRS become better understood, they could also be used as selection criteria to enhance the safety profiles of these therapies. Alternatively, CARPOOL could be adapted to systemically optimize CAR ICD designs for other emerging immune cell therapy modalities such as CAR-NK cells and CAR-macrophages,^53,54 or to identify inhibitory CARs for use in logic-gated circuits or as a means of controlling cell therapies for additional safety.

We have shown that CARPOOL can identify combinations of signalling domains that engender functionally and transcriptionally distinct phenotypes to CAR-T cells that are comparable, and in some cases superior, to current standard of care. These results support the utility of an activity-based screening approach for developing the next generation of CAR-T therapies: because altering the signalling composition of CARs can have effects of comparable magnitude to changing choice of target antigen, affinity, and oligomerization state, it represents an important axis for the functional optimization of CAR activity. In summary, CARPOOL represents a versatile, streamlined method for functionally engineering synthetic receptors for use in immune cell therapies.

Methods

Cell lines.

HEK293T (CRL-3216) and Clone E6-1 Jurkat (TIB-152) lines were purchased from ATCC, while Clone G5 NALM6 (CRL-3273) cells were purchased from ATCC and transduced to stably express firefly luciferase (FLuc) along with a puromycin resistance cassette. A375 cells were purchased from ATCC (CRL-1619) and transduced to stably express the extracellular portion of CD19 along with a hygromycin resistance cassette, then sorted for similar CD19 expression as NALM6 cells (Supplementary Fig. 22). Cell lines were routinely mycoplasma-tested using the MycoAlert PLUS Mycoplasma Detection Kit (Lonza).

Plasmid construction.

The plasmid pHIV-EGFP was gifted by Bryan Welm & Zena Werb (Addgene plasmid #21373) and pMD2.G and psPAX2 were gifted by Didier Trono (Addgene plasmid #12259 and #12260). To generate 2^nd generation CD19 CAR-EGFP plasmid, a codon optimized gene encoding CD19 CAR composed of Myc-epitope tagged FMC63 scFv, IgG4 hinge, CD28 transmembrane domain, and intracellular domains derived from human 4-1BB and CD3ζ was PCR amplified from geneblocks purchased from IDT and cloned into the 3^rd generation lentiviral vector pHIV-EGFP using Gibson Assembly. In order to generate a backbone vector for CAR plasmid library, the Myc epitope tag from CD19 CAR-EGFP plasmid was changed to a Flag epitope tag, and six tyrosine residues from CD3ζ ITAM domains were mutated into phenylalanines to prevent any unmodified yet functional CARs from contaminating library selections. The signalling diversified CAR plasmid library was generated by PCR amplification of each intracellular domain (Supplementary Table 1) at each of the 3 positions, with the forward and reverse primers adding unique linkers for each position. These products were then pooled at equimolar ratios for each position and combined with a pool of randomized 18mer barcode sequences for overlap extension PCR. These were then inserted into degenerate CD19 CAR backbone vector at PacI and BamHI restriction enzyme sites to replace the tyrosine mutated BBζ intracellular signalling components via Gibson Assembly. Final products were electroporated into DH10ß electrocompetent E.coli cells (Thermo Scientific, EC0113) and purified to achieve a highly diverse plasmid library.

Lentiviral production.

Lentiviruses were generated by first transfecting 70% confluent HEKs with transfer plasmid, pMD2.g (VSVg), and psPAX2 combined at a plasmid mass ratio of 24:1:3 that was complexed with PEI at a DNA:PEI mass ratio of 1:3. For a confluent T225 flask, 60 ug of transfer plasmid was used for transfection. Media was changed 3-6 hours after transfection and lentiviral particles were harvested in the supernatant 48-120 hours after transfection. The supernatant was then filtered through a 0.45 um low protein binding filter, and centrifuged for 1.5 hours at 100,000xg. The pellet was then resuspended in serum-free OptiMEM overnight at 4°C and stored at −80°C.

Human T cell activation, transduction, and expansion.

Research using deidentified human blood was conducted as per MIT Committee on the Use of Humans as Experimental Subjects (COUHES) policies for exempt research. Peripheral blood mononuclear cells from healthy donors were purified from buffy coats purchased from Research Blood Components or leukopaks purchased from Stem Cell Technologies using Ficoll-Paque PLUS (GE Healthcare) density gradient centrifugation with SepMate tubes (Stem Cell Technologies) or EasySep Direct Human PBMC Isolation Kits (Stem Cell Technologies), respectively, as per the manufacturer’s instructions. Primary CD4⁺ or CD8⁺ T cells were isolated using EasySep Human CD4⁺ or CD8⁺ T Cell Enrichment Kits (Stem Cell Technologies) and cultured in RPMI 1640 (ATCC) supplemented with 10% fetal bovine serum, 100 U/ml penicillin-streptomycin (Corning), 100 IU/ml recombinant human IL-2 (R&D Systems), and 50 μM beta-mercaptoethanol (Fisher). Prior to transduction, T cells were activated using a 1:1 ratio of DynaBeads Human T-Activator CD3/CD28 (Thermo Fisher) for 24 hours, after which 8 μg/mL of polybrene (Santa Cruz Biotechnology) and concentrated lentivirus were added to culture at a multiplicity of infection of 10 for single lentiviral constructs and 0.5 for pooled library encoding lentivirus. After 3 days, DynaBeads and lentivirus were removed and cells were sorted for EGFP using a BD FACSAria II. Cells were rested for 4 days prior to characterization and maintained at a density of 5x10⁵-2x10⁶ cells/ml throughout.

Flow cytometry and cell sorting.

Cells were washed with 1X PBS (Sigma) supplemented with 0.5% bovine serum albumin (RPI) and 2 mM EDTA, then surface stained by incubating with antibodies for 15 minutes on ice. They were subsequently washed again prior to flow analysis on a BD Accuri C6 or Beckman Cytoflex S or cell sorting with a BD FACSAria II or Sony MA900. Anti-CD4 (clone SK3), anti-CD8 (clone SK1), anti-PD-1 (clone EH12.2H7), anti-TIM3 (clone F38-2E2), anti-LAG3 (clone 11C3C65), anti-CD3 (clone OKT3), anti-CD62L (clone DREG-56), anti-CD45RA (clone HI100), and anti-CD69 (clone FN50) antibodies were purchased from Biolegend. Anti-Myc (clone 9B11) and anti-Flag (D6W5B) antibodies were from Cell Signalling Technology.

CAR-T functional selections.

In preparation for selections, 1x10⁸ Jurkat T cells were transduced with lentivirus at an MOI of 0.5 with 8 μg/ml polybrene (Santa Cruz Biotechnology) and spinfected at 1000xg for 1.5 hours at 32°C. Virus was removed after 2 days of transduction and the cells were sorted the following day for EGFP, with 20x library coverage, which was calculated based on the theoretical maximum diversity from the previous round, being maintained throughout. For a round of selection, cells were stimulated with 1 or 10 nM rhCD19 for 4-6 hours as indicated, then stained for CD69 or CD69 and PD-1 expression. The top 5% of CD69^high or CD69^highPD-1^low expressing T cells by mean fluorescent intensity (MFI) were sorted on a BD FACS Aria II (with at least 5x10⁵ cells being collected). Cells were then rested without antigen and expanded for 7 days before subsequent rounds of selection. After each round of selection, at least 1x10⁶ cells were sampled for NGS sequencing (whereas 6x10⁷ and 2x10⁷ cells were sampled for unselected and EGFP sorted groups, respectively). NGS sequencing data was deconvoluted and analyzed using a custom package called DomainSeq, as described in the manuscript.

In vitro rechallenge assay.

CAR-transduced CD4⁺ and CD8⁺ T cells (50,000 cells each) were mixed at a 1:1 ratio, then co-cultured with target NALM6 cells at an effector to target (E:T) ratio of 1:10 in IL-2 deficient media. Every 2-3 days, approximately 10% of the culture volume was taken out for flow analysis and stained with antibodies targeting CD4, CD8, CD62L, and CD45RA. Then, 100,000 CAR-T cells were taken out from the original culture and re-plated with a fresh batch of NALM6 cells at a 1:10 E:T ratio. CAR-T cells were sampled for scRNA-seq analysis at day 7, which was 48 hours following the third NALM6 challenge. On day 22, cells were also stained for exhaustion markers (PD-1, TIM3, and LAG3).

Cytotoxicity assay.

NALM6 cells expressing firefly luciferase (FLuc) were co-cultured with CD4⁺ or CD8⁺ T cells for 24 hours in IL-2 deficient media at various E:T ratios. Cells were then harvested and washed prior to cell lysis and addition of luciferin substrate from the Bright-Glo Luciferase Assay System (Promega). The resulting luminescent signal was measured using a Tecan Infinite M200 Pro. Signals were normalized to negative controls containing only target cells.

Cytokine secretion assay.

Following stimulation of human primary CAR-T cells with NALM6 cells, the concentrations of human IL-2 and IFN-γ were measured using an IL-2 Human Uncoated ELISA Kit (Thermo Fisher) and IFN-γ Human Uncoated ELISA Kit (Invitrogen), respectively. The resulting signal was measured on a Tecan Infinite M200 Pro plate reader, and the concentrations were determined by comparison to known standards per the manufacturer’s instructions. Polyfunctional cytokine and chemokine secretion profiles in response to tumour challenge were determined using the 41-plex MILLIPLEX MAP Human Cytokine/Chemokine Magnetic Bead Panel from Miltenyi and measured on a Luminex FlexMap 3D system.

PacBio and Illumina sequencing.

Genomic DNA from selected cells was purified using the PureLink Genomic DNA Mini Kit (Thermo Fisher). For PacBio sequencing, PCR amplicons encoding CAR signalling domains and barcode regions were attached with SMRTbell adaptors using the SMRTbell Template Prep Kit 1.0 (Pacific Biosciences) and sequenced using a PacBio Sequel system. For Illumina sequencing, barcode regions were PCR amplified to conjugate P5 and P7 adaptor sequences and sequenced on an Illumina MISeq system.

Single-cell sequencing.

19BBζ, Var1, and Var3-expressing CAR-T cells were sampled from NALM6 cocultures, as well as untransduced, unstimulated T cells. Separately, samples were enriched for live cells using a Dead Cell Removal (Annexin V) Kit (Stem Cell Technologies), then labelled with unique anti-human TotalSeq-B hashing antibodies (BioLegend). Following labelling, approximately 2,500 cells from each of the four samples were pooled before encapsulation in a single channel of the Chromium Single Cell 3’ v3.1 platform (10X Genomics). Gene expression (GEX) libraries were constructed based on manufacturer’ instructions, while hashing antibody libraries were constructed as reported previously.⁵⁵ The resulting libraries were pooled at a 1:10 ratio of antibody-to-GEX before sequencing on an Illumina NextSeq500 to a depth of 60,000 reads per cell. Reads were aligned to the Genome Reference Consortium Human Build 38 (GRCh38), and a cell-gene matrix was generated using the CellRanger pipeline (10X Genomics; v4.0.0). Downstream analysis was performed using the Seurat package (v4.0.0).^55,56 In brief, cells were first assigned sample identity based on the detection of a single hashing antibody following normalization of antibody reads using the HTODemux algorithm.⁵⁵ Next, low-quality cells were filtered out on the basis of mitochondrial reads (>25%). Filtered data for each cell was normalized to total expression, and cell cycle-related genes were regressed out using the ScaleData function. To identify distinct transcriptional states, linear dimensionality reduction was performed on the scaled, normalized data, followed by shared nearest neighbours clustering on the basis of the first 40 principal components. Differentially expressed genes, both within clusters and across samples, were identified by a two-sided Wilcoxon-Rank Sum test between the populations of interest. For pathway-level analyses, individual cells were assigned scGSVA scores on the basis of their relative expression of genes within all canonical pathways, immunologic signature gene sets (Broad Institute; C2 and C7 gene sets, respectively), or T cell-specific curated pathways (Supplementary Table 4), as previously described.³³

Xenogeneic mouse models.

All animal studies were performed in accordance with guidelines approved by the MIT Division of Comparative Medicine and MIT Committee on Animal Care (Institutional Animal Care and Use Committee, protocol number 0621-032-24). Male NOD/SCID/IL2R^null (NSG) mice were purchased from Jackson Laboratory and housed in the animal facilities at MIT. For B-ALL models, mice ages 8-12 weeks old and weighing between 24 and 31 g were injected intravenously via the tail vein with 5x10⁵ FLuc⁺ NALM6 cells. CD4⁺ and CD8⁺ T cells were prepared separately as described above, then sorted for EGFP on the day of DynaBead removal; after 4 days of rest, they were then mixed at a 1:1 ratio. Mice were then treated with 2x10⁵ or 1x10⁶ CAR-T cells or untransduced control T cells intravenously via the tail vein 4 days after NALM6 injection. Tumour progression was subsequently monitored every 3-7 days using the IVIS Spectrum or IVIS Lumina S5 imaging system (PerkinElmer) to measure bioluminescent signal after intraperitoneal administration of 0.15 mg of luciferin substrate per gram of body weight (PerkinElmer 122799). Total photon counts were quantified using LivingImage software. Mice were monitored daily and euthanized upon observing signs of discomfort or morbidity, weight loss amounting to more than 20% of the original body weight, graft versus host disease, or as recommended by the veterinarian. Where indicated, peripheral blood was collected retro-orbitally to measure T cell expansion and persistence by flow cytometry. Red blood cells were lysed from the collected tissues using ACK Lysing Buffer (Thermo Fisher A1049201) and washed with 1X PBS supplemented with 0.5% BSA and 2 mM EDTA prior to antibody staining and FACS analysis.

For melanoma models, 8-week-old mice weighing between 24 and 31 g were injected subcutaneously on the right flank with 5x10⁵ CD19⁺ A375 cells. CD4⁺ and CD8⁺ T cells were prepared separately as described above, mixed at a 1:1 ratio, and then injected intravenously via the tail vein 21 days after A375 inoculation. Tumour progression was monitored every 2-3 days by calliper measurement and mice were weighed weekly. Mice were euthanized upon observing signs of discomfort or morbidity, limited mobility, or upon tumour area reaching 100 cm².

Statistical analysis.

Statistical analyses were performed using the Prism 9 (GraphPad) software, with the exception of the single-cell sequencing data which was analysed in R Studio using base packages or those described above. Sample sizes were not predetermined using statistical methods. For statistical comparisons between two groups, significance was determined using two-tailed unpaired parametric t-tests or nonparametric Wilcoxon Rank Sum tests. For in vivo experiments, differences in overall survival were analysed using a log-rank test and displayed in a Kaplan-Meier curve and differences in tumour burden were compared using two-way ANOVA. Adjusted P values < 0.05 after multiple hypothesis correction, where required, were considered statistically significant. The statistical test used for each experiment is noted in the relevant figure legend.

Extended Data

ED Fig. 1. ICD frequency post-selection for 1^st and 2^nd generation CARs.

Related to Fig. 2b. Heatmap representing log₂(fold-change) of individual ICDs, oriented by ICD position within the CAR relative to the plasma membrane. Fold-changes were normalized to frequencies in the EGFP⁺ sorted population prior to selection.

ED Fig. 2. Characterization of CAR variant expression and function in Jurkat T cells.

a, Design of CD19 targeted CAR candidates. b, CAR surface expression (n = 1), (c) internalization (n = 1), and (d) CD69 upregulation upon antigen stimulation with recombinant human CD19 (n = 6 technical replicates representative of 2 biological replicates). e, Basal CD69 (activation) and (f) PD-1 expression of unstimulated CAR-T cells (n = 7 technical replicates representative of 2 biological replicates). P values shown in panel (e) are 0.002 for Var1 vs. BBζ and 0.0149 for Var2 vs. BBζ. P values shown in panel (f) are 0.0117 for Var1 vs. BBζ, 0.0110 for Var2 vs. BBζ, and 0.0192 for Var3 vs. BBζ, as determined by two-tailed unpaired student’s t tests (df = 12). Data shown are individual values in panel (c), while panel (d) shows means ± s.e.m. Panels (e,f) show individual values along with means ± s.e.m.

ED Fig. 3. Memory and exhaustion phenotypes following tumour rechallenge.

Related to Fig. 4d,e. a, FACS plots displaying changes in CAR-T cell memory differentiation following repeated tumour challenge (days 4 and 18 shown). b, LAG-3 expression following repeated tumour challenge on day 22 (n = 3 technical replicates representative of 2 biological replicates). P values are 0.0465 for CD4⁺ Var3 vs. BBζ, and 0.0358 for CD8⁺ Var1 vs. BBζ, and 0.0112 for CD8⁺ Var3 vs. BBζ (df = 4 for all). P values in panel (b) were determined using an unpaired student’s t test. Data shown are means ± s.e.m.

ED Fig. 4. Expression of CD4 and CD8A across different phenotypic clusters.

Differential of CD4 (left) and CD8A (right) expression as determined by single cell sequencing as shown in Fig. 5. Data points shown represent individual cells.

ED Fig. 5. Differential gene expression between Var1 and 19BBζ-expressing CAR-T cells.

Volcano plot showing differentially expressed genes between Var1 (n = 863 cells) and 19BBζ-expressing (n = 637 cells) CAR-T cells 48 hours after a third NALM6 rechallenge. P values for each gene were determined using a two-sided Wilcoxon Rank-Sum test. Genes with both a Bonferroni-corrected P value less than 0.05 and an average log2FC of > 1 are highlighted.

ED Fig. 6. High dose Var1, Var3, and BBζ achieve sustained remission.

a, Experimental design. NSG mice were infused intravenously with 5x10⁵ FLuc⁺ CD19⁺ NALM6 cells, then treated i.v. with 1x10⁶ of a 1:1 mixture of human CD8⁺ and CD4⁺ CAR-T cells or untransduced control T cells (n = 5 mice for all groups). b, Kaplan-Meier curve for overall survival. Tumour burden was assessed by measuring luminescent activity. c, Quantification of total photon counts are shown. d, Weight loss following ACT was monitored routinely. Data shown in panels (c,d) are individual values.

ED Fig. 7. High dose BBζ and Var1 confer partial protection against B-ALL recurrence.

a, Experimental design. NSG mice were infused intravenously with 5x10⁵ FLuc⁺ CD19⁺ NALM6 cells, then treated i.v. with 1x10⁶ of a 1:1 mixture of human CD8⁺ and CD4⁺ CAR-T cells or untransduced control T cells (n = 5 mice for all groups). Mice were then rechallenged with 5x10⁵ FLuc⁺ CD19⁺ NALM6 cells on days 44 and 58 post ACT (indicated with arrows). b, Kaplan-Meier curve for overall survival. Tumour burden was assessed by measuring luminescent activity. c, Quantification of total photon counts is shown. d, Weight loss following ACT was monitored routinely. Data shown in panels (c,d) are individual values. Arrows in panels (b—d) indicate dates of NALM6 rechallenge. e, CD19⁺ NALM6 cells and (f) EGFP⁺ CAR-T cells were harvested from the peripheral blood (PB), spleen (SP), and bone marrow (BM) of a separate cohort of mice that received the same treatment regimen at day 14 post ACT and quantified via FACS analysis (n = 3). Data shown are individuals with means ± s.e.m.

Supplementary Material

1823414_Sup_Tab_1

Supplementary Table 1 Amino acid compositions of intracellular signalling domains incorporated into the CAR library.

1823414_Sup_Tab_2

Supplementary Table 2 Barcode frequencies and metadata from next generation sequencing of selected CAR library expressing Jurkats.

1823414_Sup_Tab_3

Supplementary Table 3 List of significantly differentially overexpressed genes in each cluster following single cell RNAseq.

1823414_Sup_Tab_4

Supplementary Table 4 Differentially expressed genes between all Var1-expressing cells and all 19BBz-expressing cells.

1823414_Sup_Tab_5

Supplementary Table 5 List of T cell-specific curated gene sets used for scGSVA analysis of transcriptional clusters.

Data availability

The NGS selection datasets have been deposited in the Sequence Read Archive and are available under the accession number PRJNA744269. The scRNA-seq data have been deposited in the Gene Expression Omnibus under accession number GSE179767. All data generated or analysed during the study (including the DomainSeq-processed CARPOOL selection data) are included in the paper or its supplementary information.