CRISPR Screening of CAR T Cells and Cancer Stem Cells Reveals Critical Dependencies for Cell-Based Therapies

Glioblastoma (GBM) ranks as one of the most lethal of human cancers, with current therapy offering only palliation. Standard-of-care therapy consisting of maximal surgical resection followed by combined radiation and chemotherapy extends median survival by less than three months. The activation of antitumor immune responses may provide new opportunities to augment tumor control. As such, immunotherapies have been extensively investigated with positive results in preclinical studies, yet broad antitumor efficacy has not occurred in patients (1). The adoptive transfer of chimeric antigen receptor (CAR) engineered T cells has shown promising clinical activity in a subset of cancers, particularly B-cell malignancies (2, 3). To target GBM, CAR T cells have been engineered to recognize selected tumor antigens and have demonstrated cytolytic activity against GBM cells, including GBM stem cells (GSC; refs. 4–6). In patients with GBM, CAR T-cell therapies have shown early evidence of activity, clinical feasibility, and safety (7–10). However, the overall outcomes of CAR T-cell treatment remain unsatisfactory, prompting efforts to enhance the antitumor potency of GBM-targeting CAR T cells (11, 12). The functional potentiation of CAR T cells, though attractive due to the modifiable nature of these cells, requires a comprehensive understanding of the molecular events regulating CAR T-cell activation, exhaustion, and tumor-induced immune suppression (11, 13). Aside from CAR recognition of tumor antigens, the complicated and dynamic interaction between CAR T cells and their target tumor cells remains poorly characterized. Thus, there is a need for new strategies to further enhance CAR T-cell potency.

Gene editing using CRISPR/Cas9 is a promising approach to enhance cancer immunotherapy (14). Directed CRISPR/Cas9 gene knockout (KO) of checkpoint and other immune-regulatory receptors has shown utility for adoptive T-cell therapy (15, 16); however, this approach has focused on a limited set of known pathways. By contrast, large CRISPR-KO screens are an effective platform for unbiased target discovery and have been successfully used to identify genes in tumor cells which when deleted synergize with various types of immunotherapeutics (17–19). CRISPR screens in T cells identified modulators of T-cell receptor (TCR) activation in response to stimulation with CD3/CD28 agonistic beads, viruses, or tumor cells (20–22). Although CAR constructs are synthetic TCR-like receptors incorporating CD3ζ and costimulatory domains, the molecular events are not identical between TCR and CAR T-cell activation signaling pathways (23). Thus, unbiased CRISPR screen presents an attractive strategy for discovering key regulators of CAR–tumor interaction.

GSCs represent a potentially important cellular target in GBM, as they have been linked to therapeutic resistance, invasion into normal brain, promotion of angiogenesis, and immune modulation (24, 25). We hypothesized that systematic interrogation of molecular regulation of CAR T-cell efficacy against GBM could be optimized by screening both CAR T cells and GBM cells, thereby informing the interplay between a cell-based therapy and its target population. Here, we developed a robust method for performing whole-genome CRISPR-KO screens in both GBM cells and human CAR T cells. Using our well-established CAR T-cell platform targeting the tumor-associated surface marker IL13 receptor α2 (IL13Rα2; refs. 7, 8, 26), we identified novel CAR T cell– and tumor-intrinsic targets that substantially improved CAR T-cell cytotoxicity against GSCs both in vitro and in vivo. Targeted genetic modification of identified hits in CAR T cells potentiated their long-term activation, cytolytic activity, and in vivo antitumor function against GSCs, demonstrating that CRISPR screen on CAR T cells leads to the discovery of key targets for augmenting CAR T-cell therapeutic potency. In parallel, KO of identified targets in GSCs sensitized them to CAR-mediated killing both in vitro and in vivo, revealing potential avenues for combinatorial inhibitor treatment to augment CAR T-cell efficacy. Our findings represent a feasible and highly effective approach to discovering key targets that mediate effective tumor eradication using CAR T cells.

The fitness of CAR T-cell products correlates with clinical responses (27, 28), indicating that key regulators of CAR T-cell function can be targeted to potentiate therapeutic efficacy. T-cell exhaustion resulting from chronic tumor exposure limits CAR T-cell antitumor responses (29). To identify the essential regulators of T-cell functional activity in an unbiased manner, we performed genome-wide CRISPR screen adapting our previously developed in vitro tumor rechallenge assay, which differentiates CAR T-cell potency in the setting of high tumor burden and reflects in vivo antitumor activity (30, 31). IL13Rα2-targeted CAR T cells from two healthy human donors were lentivirally transduced to express the Brunello short-guide RNA (sgRNA) library (32) and the CAR construct and then electroporated with Cas9 protein. CAR T cells harboring CRISPR-mediated knockouts were recursively exposed to an excess amount of PBT030-2 GSCs (Fig. 1A), an IDH1 wild-type patient-derived GSC line that highly expresses IL13Rα2 (33). After tumor stimulation, CAR T cells were sorted from coculture and classified into subsets based on expression of the inhibitory receptor PD-1, which is associated with T-cell exhaustion (Fig. 1A). To identify gene knockouts that augment CAR T antitumor activity, we identified sgRNAs enriched in the less-exhausted PD-1–negative versus PD-1–positive CAR T-cell compartments (Fig. 1B; Supplementary Fig. S1A; Supplementary Table S1). To eliminate targets that nonspecifically impaired CAR T-cell proliferation or viability, we excluded sgRNAs depleted (β < −1) in CAR T-cells after 72-hour coculture with GSCs or 72-hour monoculture (Fig. 1C). Two hundred twenty genes were common hits in both T-cell donors (Fig. 1D). Many of these 220 genes are induced by the IL4 receptor (IL4R), which suppresses T-cell activity (34), as well as type I IFN, NFAT, TCF4, and JAK1/2, which all play complex roles in T-cell activation and mediate T-cell exhaustion and inhibition under some circumstances (refs. 35–38; Fig. 1E). In addition, these genes were enriched for pathways that contribute to T-cell exhaustion, including nuclear receptor transcription and cholesterol responses (refs. 39, 40; Supplementary Fig. S1B). In contrast, genes preferentially depleted in PD-1–positive cells included pathways associated with T-cell activation, including amide metabolism and NFκB signaling (41, 42), as well as negative regulation of oxidative stress–induced cell death (Fig. 1E). Together, these data verify that our screen identified genes involved in T-cell effector activity, providing candidate genes that can be modulated to prevent exhaustion and enhance effector function of CAR T cells.

We interrogated the 220 targets enriched in PD-1–negative cells common between two T-cell donors (Supplementary Table S2), focusing on four representative genes identified in the top third of hits, which have not been previously explored for their role in enhancing CAR T-cell function. These included the high-ranking hits: eukaryotic translation initiation factor 5A-1 (EIF5A), transcription factor transducin like enhancer of split 4 (TLE4), Ikaros family zinc finger protein 2 (IKZF2), and transmembrane protein 184B (TMEM184B; Fig. 1F). Most sgRNAs targeting these genes (≥2 out of 4 in both replicates) were enriched in PD-1–negative CAR T cells (Supplementary Fig. S1C). To verify the function of these targets by CRISPR-mediated KO on CAR T cells, we leveraged the challenging in vitro killing assay (CAR:tumor = 1:40), confirming that targeting TLE4, IKZF2, TMEM184B, or EIF5A improved in vitro killing potency of CAR T cells against GSCs, as well as their expansion potential, although sgEIF5A-3 effects were more modest (Fig. 2A and B). Mechanistically, KO of these genes reduced PD-1 expression on CAR T cells following tumor stimulation (Supplementary Fig. S2A). We and others have shown that CAR T-cell exhaustion is associated with coexpression of PD-1, LAG3, and TIM3 (43, 44). All four KOs reduced CAR T-cell exhaustion, TLE4-KO and IKZF2-KO most effectively (Fig. 2C). KO of these genes minimally affected initial CAR T-cell activation upon tumor cell recognition (Supplementary Fig. S2B), suggesting that these KOs improved T-cell fitness and long-term function instead of initial activation. Targeted KOs did not affect the expression and stability of the CAR in T cells (Supplementary Fig. S2C and S2D). As validation, we performed independent studies with a HER2-targeted CAR model that also demonstrated improvements in CAR killing and expansion, suggesting that genetic screens of CAR T cells may yield broadly effective molecular strategies (Supplementary Fig. S2E and S2F).

TLE4 is a transcriptional corepressor of multiple genes encoding inflammatory cytokines (45) and IKZF2 is upregulated in exhausted T cells (37, 46, 47), supporting potential roles in inhibiting CAR T-cell function. To elucidate molecular mechanisms underlying the regulation of CAR T-cell activity, we compared the transcriptomes of CAR T cells with individual knockouts against cells transduced with nontargeted sgRNA (sgCONT). TLE4-KO in CAR T cells upregulated critical regulators of T-cell activation, including the transcription factor EGR1, which promotes Th1 cell differentiation (48), and the metabolic regulator BCAT, which mediates metabolic fitness in activated T cells (ref. 49; Supplementary Fig. S2G). IKZF2-KO in CAR T cells upregulated proinflammatory cytokines and pathways, including CXCL8, CCL3, and CCL4 (50–52), as well as EGR1, similar to TLE4-KO (Supplementary Fig. S2H). We next compared transcriptional profiles of TLE4-KO or IKZF2-KO CAR T cells to the signatures of known T-cell subsets and pathways (35, 53, 54). TLE4-KO or IKZF2-KO induced molecular signatures representing activation over memory T cells, together with key T-cell activation signaling pathways (TCR signaling, T-cell activation, AP-1, and ZAP; Fig. 2D and E; Supplementary Fig. S2I and S2J). T-cell activation characteristics in TLE4-KO or IKZF2-KO cells were uncoupled from exhaustion (Fig. 2D and E), suggesting retention of CAR T-cell function. TLE4-KO cells downregulated an apoptosis signature (Fig. 2D and F) and upregulated AP-1 signaling, which maintains CAR T-cell function (ref. 55; Fig. 2G). In particular, the AP-1 family transcription factor FOS was enriched after TLE4-KO, together with many of its downstream targets (Fig. 2G). As overexpression of the AP-1 family member JUN prevents CAR T-cell exhaustion (55), we investigated whether TLE4-KO phenocopied transcriptional changes of JUN overexpression, revealing that genes upregulated with TLE4-KO overlapped with genes with upregulated following JUN overexpression (Fig. 2H). IKZF2-KO upregulated pathways involving interactions between cytokines and their receptors, as well as NFAT signaling, which regulates key molecular signals following T-cell activation (ref. 56; Fig. 2I and J). As EGR1 was upregulated after IKZF2-KO, many genes in these pathways were likely downstream targets (Fig. 2I and J).

Whole-transcriptome analyses following TMEM184B or EIF5A KO revealed convergence of altered pathways, similar to those induced by TLE4-KO or IKZF2-KO, including the upregulation of BCAT1, EGR1, and IL17RB (Supplementary Fig. S3A and S3B) and the acquisition of memory or effector over naïve T-cell signatures (Supplementary Fig. S3C and S3D). However, targeting TMEM184B or EIF5A did not enrich for cytokine secretion and response pathways in CAR T cells (Supplementary Fig. S3E and S3F), which were found in TLE4-KO or IKZF2-KO CAR T cells. As these cytokines (CCL3 and CCL4) maintain T-cell function during chronic viral infection and in the tumor microenvironment (57, 58), our results indicate that TMEM184B-KO and EIF5A-KO CAR T cells might be prone to terminal effector differentiation and subsequent exhaustion, thereby compromising their overall functional capability despite their potent in vitro cytotoxicity. Overall, KO of these genes in CAR T cells also maintained transcriptional profiles of T-cell activation, which are associated with effector potency.

To determine the impact of TLE4-KO or IKZF2-KO on specific subpopulations of CAR T cells, we performed comparative single-cell RNA sequencing (scRNA-seq) on KO and control CAR T cells with or without stimulation by tumor cells. Comparing TLE4-KO cells with control CAR T cells by unbiased clustering of pooled data identified 10 different clusters, the distribution of which was greatly influenced by stimulation (Fig. 3A–C; Supplementary Fig. S4A). CD4⁺ and CD8⁺ CAR T cells were well delineated (Supplementary Fig. S4B). Stimulated cells downregulated naïve/memory-related markers (e.g., IL7R and CCR7) and upregulated activation-related markers (e.g., MKI67 and GZMB; Supplementary Fig. S4C). Stimulation enriched clusters 0, 1, and 10 (showing high expression of activation or exhaustion markers) and depleted clusters 3, 5, 7, and 9 (expressing naïve/memory markers; Fig. 3D). The impact of TLE4-KO on the overall distribution CAR T cells was more significant after stimulation; notably, cluster 8 was depleted after stimulation only in control, but not in TLE4-KO cells (Fig. 3C and D). This cluster represented a subset of CD4⁺ T cells expressing multiple costimulatory molecules, including CD28, ICOS, CD86, and TNFRSF4 (OX40), as well as the cytokine IL2 (Fig. 3D; Supplementary Fig. S4D). Although no proliferative activity was detected in this cluster (indicated by low Ki-67), preservation of this cluster in TLE4-KO cells was maintained post-stimulation (Fig. 3D). In TLE4-KO cells, cluster 8 also showed expression of the immune stimulatory cytokine CCL3 (Fig. 3E), costimulatory molecule TNFRSF4 (Fig. 3F), and AP-1 transcription factors FOS and JUN (Supplementary Fig. S4E and S4F), which were minimally expressed in control cells. Cluster 10 was an activated CD4⁺ subset expressing multiple cytokines, including IL2 and TNF, and this cluster displayed greater post-stimulation expansion in TLE4-KO cells (Fig. 3D). In the clusters with activation signatures (0, 1, and 10), TLE4-KO upregulated IFNG, BCAT, GZMB, CCL3, and CCL4 (Fig. 3G and H; Supplementary Fig. S4G–S4I). Combining the transcriptome readouts from all single cells revealed that tumor stimulation in TLE4-KO cells induced T-cell stimulatory and cytotoxic factors (e.g., GZMB, CCL3, CCL4, and IFNG) to a greater degree than control CAR T cells (Supplementary Fig. S4G–S4J). Taken together, the enhanced cytotoxicity of TLE-KO CAR T cells could result from the preservation of specific T-cell subsets after tumor stimulation.

Comparison between IKZF2-KO cells and control CAR T cells identified 10 clusters using unbiased clustering of pooled data (Fig. 4A). In parallel with the comparisons between TLE4-KO versus control cells, we observed a dramatic change in cluster distribution and gene expression after stimulation, with moderate changes from IKZF2-KO (Fig. 4B and C; Supplementary Fig. S5A–S5D). Given a role for IKZF2 in regulatory T cells (Treg; ref. 59), cluster 0, characterized by Treg signatures (e.g., CLTA4, FOXP3, and IL2RA), was reduced in IKZF2-KO cells (Fig. 4B–D). Cluster 10 was induced after stimulation, enriched in IKZF2-KO cells, and expressed elevated levels of the AP-1 signaling molecules FOS and JUN (Fig. 4B–D). These cells expressed a limited repertoire of cytokines beyond TNF, but had medium-to-high levels of Ki-67, high expression of EGR1 and IL8, and exclusively expressed CXCL10 and CCND1 (Fig. 4D–F; Supplementary Fig. S5D). Upregulated genes in cluster 10 were enriched for transcriptional regulation by ATF3 and JUN (Supplementary Fig. S5E and S5F). This subset contained a very limited number of cells and was present only upon stimulation, potentially explaining the lack of differential expression of FOS and JUN in bulk RNA-seq analysis in IKZF2-KO cells. Cluster 2 was also expanded after stimulation in both IKZF2-KO and control cells (Fig. 4D; Supplementary Fig. S5A). However, induction of activation-associated genes in this cluster, including IFNG, CCL3, and CCL4, was more robust in IKZF2-KO versus control cells upon tumor stimulation (Fig. 4G and H; Supplementary Fig. S5G). In IKZF2-KO cells, CCL3 was expressed at higher levels in clusters 0, 1, and 9 (Fig. 4G). As a result, IKZF2-KO cells exhibited an augmented responsiveness to tumor stimulation, illustrated by the upregulation of activation-associated cytokines (Supplementary Fig. S5H). Overall, scRNA-seq analysis revealed that TLE4-KO or IKZF2-KO resulted in the preservation or expansion of certain CAR T-cell subsets after tumor stimulation. These cellular subsets displayed transcriptional signatures of T-cell cytotoxicity and/or immune stimulation, providing some underlying mechanisms of their superior effector function against tumor cells.

Augmenting the efficacy of CAR T cells against GBM can be approached by studying T cells themselves, as above, which may inform targeted KOs in addition to CAR engineering for enhancing CAR activity. Reciprocal screening of GBM cells, especially GSCs, potentially informs interactions with CAR T cells to predict clinical responsiveness to CAR T-cell therapy. To identify potential genes in GSCs that promote resistance to CAR-mediated cytotoxicity, we performed genome-wide CRISPR screens on two independent patient-derived GSC lines (PBT030-2 and PBT036), both derived from primary GBM tumors with high expression of IL13Rα2 (33). To identify tumor cell targets that rendered GBM cells more susceptible to T-cell immunotherapy, we subjected GSCs to two rounds of coculture with IL13Rα2-targeted CAR T cells (Fig. 5A; Supplementary Fig. S6A). We identified sgRNAs that were enriched (β > 1) or depleted (β < −1) in the surviving GSCs compared with GSCs in monoculture for the same amount of time (Fig. 5B). The genes with sgRNAs depleted in coculture (β < −1) represented targets that promoted CAR killing upon KO (Fig. 5C). To exclude sgRNAs that nonspecifically targeted essential genes for GSC survival, we removed gene hits that were depleted in GSCs after 48-hour culture without CAR T cells (Fig. 5C). A total of 228 CAR-modulating genes were identified as hits in either GSC line, with only four overlapping targets common to both lines (Fig. 5D). Enriched pathways included tumor immune modulation, such as MHC I antigen presentation, IL1 signaling, and NFκB activation (Fig. 5E), indicating that sgRNAs depleted in surviving GSCs targeted genes responsible for resistance to T-cell killing.

Next, we sought to confirm and further characterize the function of common top hits whose deletion promoted CAR killing (Fig. 5D). V-Rel reticuloendotheliosis viral oncogene homolog A (RELA) and nuclear protein localization protein 4 homolog (NPLOC4) were selected for further validation as all sgRNAs targeting these two genes in the screen showed depletion in GSCs cocultured with CAR (Fig. 5F). As expected from our selection process, CRISPR-mediated KO of either RELA or NPLOC4 caused limited reduction in the growth of GSCs in vitro compared with GSCs transduced with control nontargeted sgRNAs (sgCONT; Supplementary Fig. S6B). When cocultured with CAR T cells in a challenging in vitro model at low T-cell ratios [effector:target (E:T) = 1:40; 48 hours], RELA or NPLOC4 KO in GSCs increased susceptibility to CAR T cell–mediated killing (Fig. 6A), which was also associated with increased expansion of CAR T cells (Fig. 6B). Thus, KO of either RELA or NPLOC4 in GSCs enhanced the cytotoxic and proliferative potency of CAR T cells.

RELA (also known as p65) is an NFκB subunit that regulates critical downstream effectors of immunosuppressive pathways in tumors (60, 61). NPLOC4 mediates nuclear pore transport of proteins, but its role in cancer or immune modulation remains unclear. To elucidate the mechanism by which these genes mediate GSC sensitivity to CAR T-cell killing, we performed in-depth characterization of GSCs harboring KO of each gene. The increased sensitivity was not a result of alterations in target antigen expression on GSCs (Supplementary Fig. S6C). CAR T cells induced PD-L1 in GSCs, which was not altered by depletion of either RELA or NPLOC4 (Supplementary Fig. S6D). Likewise, CAR T cells cocultured with GSCs transduced with sgCONT, sgRELA, or sgNPLOC4 did not show differences in activation after stimulation as indicated by markers CD69 and CD137, or exhaustion measured by levels of exhaustion markers, including PD-1, LAG3, and TIM3 (ref. 30; Supplementary Fig. S6E). Whole-transcriptome analysis of GSCs after RELA-KO showed downregulation of immunosuppressive cytokines, including CXCL3, CCL20, and IL32 (Fig. 6C), all of which suppress antitumor immune responses (62, 63). Downregulated genes were highly enriched for known direct transcriptional targets of RELA, and RELA-KO reduced NFκB signaling, as well as the immunosuppressive effectors of TNF responsiveness and IL10 signaling (Fig. 6D and E; Supplementary Fig. S6F). Targeting NPLOC4 in GSCs downregulated genes mediating rearrangement of extracellular matrix (ECM), including proteoglycans, integrins, and collagens (Fig. 6F–H). Reactome analysis revealed the involvement of specific tumorigenic factors, such as EGFR and PDGFA (Fig. 6G). Pathways downregulated after NPLOC4 depletion were highly enriched for ECM remodeling and cell adhesion (Fig. 6H; Supplementary Fig. S6G). Although tumor ECM remodeling has been reported to suppress antitumor immune responses by preventing T-cell trafficking into the tumors, ECM-associated factors may directly repress T-cell activity (64, 65). To interrogate NPLOC4 interactions, we performed immune precipitation followed by mass spectrometry (IP/MS), revealing that NPLOC4 bound multiple targets in immune-related pathways (IL1, Fc receptor, antigen presentation), WNT signaling, and protein synthesis/degradation pathways (Supplementary Fig. S7A). These mechanisms may regulate the immune-related profiles of GSCs, where NPLOC4-KO led to the upregulation of immune stimulatory cytokines (Supplementary Fig. S7B and S7C). Together, we found that the tumor-intrinsic regulators RELA and NPLOC4 mediate GBM resistance to CAR T-cell cytotoxicity via mechanisms distinct from induction of CAR T-cell exhaustion.

Next, we used an orthotopic intracranial patient-derived xenograft (PDX) model to evaluate whether modulating the identified targets on GSCs enhanced the antitumor function of CAR T cells in a preclinical setting. Established GBM PDXs were treated with CAR T cells delivered intracranially into the tumors, mimicking our clinical trial design of CAR T-cell administration to patients with GBMs (7, 8). First, we used CAR T cells without CRISPR KO to treat control, RELA-KO, or NPLOC4-KO tumors. A limited number of CAR T cells (50,000/mouse) completely eradicated xenografts derived from RELA-KO or NPLOC4-KO GSCs, whereas the same CAR T cells were only partially effective against tumors established with sgCONT GSCs (Fig. 7A and B; Supplementary Fig. S8). These results suggest that tumors with low expression of RELA and/or NPLOC4 are more sensitive to CAR T therapy.

To further dissect the roles of RELA and NPLOC4 in immune modulation in GBM, we analyzed 41 GSC samples and found that high RELA- or NPLOC4-expressing GSCs showed enrichment in immune-suppression signatures (Supplementary Fig. S9A and S9B). Interrogating The Cancer Genome Atlas (TCGA) GBM data set revealed that RELA and NPLOC4 both positively correlated with TGFβ signaling, a key pathway mediating immune suppression in GBMs and many other types of tumors (66). RELA was also positively correlated with immunosuppressive regulatory T-cell signatures and negatively correlated with the signatures of antitumor T-cell responses (lymphocyte infiltration, TCR richness, Th1 and CD8 T cells; Fig. 7C). NPLOC4 was negatively correlated with the immune stimulatory IFNγ responses (Fig. 7D). The infiltration signature of CD4⁺ and CD8⁺ T cells in GBM inversely correlated with RELA or NPLOC4 expression (Supplementary Fig. S9C and S9D). These results suggest that high expression of RELA and NPLOC4 in GBM is indicative of a more suppressive tumor immune microenvironment and repressed antitumor T-cell responses.

We next evaluated the molecular targets identified in our CAR T-cell screen in vivo, with the goal of establishing clinically translatable strategies to improve CAR T-cell function. The antitumor function of different CAR T cells was tested against tumors without CRISPR knockouts, with a further limited CAR T-cell dose (20,000/mouse) showing enhanced survival benefit as compared with the control CAR T cells that failed to achieve long-term tumor eradication (Fig. 7E and F). Consistent with improved maintenance of T-cell effector activity and decreased exhaustion, targeting either TLE4 or IKZF2 augmented in vivo antitumor activity of CAR T cells against PDXs, as measured by extension of survival in tumor-bearing mice (Fig. 7E and F; Supplementary Fig. S10). Depletion of TMEM184B or EIF5A in CAR T cells showed a trend toward improved efficacy in increasing the survival of tumor-bearing mice (Supplementary Fig. S11A and S11B). Therefore, these targets on GSCs and CAR T cells can be exploited to advance the efficacy of CAR therapy against established GBM tumors.

We then investigated whether the CAR T-cell targets indicate the potency of clinical therapeutic products. We mapped upregulated genes in IKZF2-KO CAR T cells compared with control CAR T cells after tumor stimulation, with the transcriptomes of CAR T-cell products from patients with chronic lymphocytic leukemia (CLL) achieving complete responses (CR) or no responses (NR; ref. 27). Supporting our results, these genes were induced to a greater degree after CAR stimulation in the products from patients achieving CR (Fig. 7G). Similarly, genes enriched in cluster 10, whose expansion was induced by tumor stimulation and further augmented with TLE4-KO, were also highly expressed in the products from patients with CR (Fig. 7H). Further, both TLE4-KO and IKZF2-KO led to gene upregulation similar to comparisons of products from patients with CR and NR (Fig. 7I and J).

To further understand how TLE4 and IKZF2 contribute to the function of clinical CAR T-cell products, we analyzed scRNA-seq from 24 patient-derived CD19-CAR T-cell products (67). An unbiased clustering of the scRNA-seq data revealed that IKZF2 expression was highly enriched in cluster 7 (Supplementary Fig. S12A and S12B), overlapping with key markers of immune-suppressive Tregs (CTLA4, FOXP3, IL2RA; Supplementary Fig. S12C and S12D). Cluster 7 was more frequently detected in patients with progressive disease (PD) than those with CR (Supplementary Fig. S12E). In these same cells, cluster 11 represented exhausted T cells, as indicated by the markers TOX and TOX2 (Supplementary Fig. S13A–S13C; ref. 37). This cluster showed low expression of TLE4-repressed genes, indicating high TLE4 activity (Supplementary Fig. S13D and S13E). Further, TLE4 was upregulated in CAR T cells undergoing extended ex vivo culture (Supplementary Fig. S13F), a process associated with impaired effector function (68). Together, these observations establish that the targets identified from CRISPR screening have clinical implications for both tumor immunoreactivity and CAR T-cell functional potency (Supplementary Fig. S14).

T cell–based therapies may offer several advantages in GBM therapy. T cell–based therapies, especially when delivered into the cerebrospinal fluid (CSF), traffic to multifocal tumor populations within the central nervous system (CNS; refs. 8, 69–71), thus overcoming challenges associated with the blood–brain barrier that limit the CNS penetration of most pharmacologic agents. T-cell therapies compensate for cellular plasticity within brain tumors more effectively than traditional pharmacologic agents. GBMs display striking intratumoral heterogeneity, and tumor cells readily compensate for targeted agents against specific molecular targets. With T-cell therapy targeting different antigens, personalized treatments based on the antigen expression profile of individual tumors may be designed. T cell–based therapies induce secondary responses that augment endogenous antitumor responses. Adoptive cell transfer, especially CAR T therapies, has been investigated in clinical trials for patients with GBM, but efficacy has been restricted to limited cases (11). Our focus on CAR T cells was prompted not only by the potential value for clinical translation, but also because our findings inform a broader understanding of T-cell function in brain tumor biology.

Previous genetic screens used to identify interactions between immune cells and tumor cells have largely focused on the tumor cells (18, 19, 29), as these cells are easier to manipulate genetically. Screens on tumor-reactive mouse T cells have also been reported (20, 72, 73) given the establishment of Cas9-knock-in mouse strain (74), as well as the convenience to acquire large numbers of these cells. Here, we interrogated both the human CAR T-cell and tumor cell compartments. The screening strategy on CAR T cells was greatly facilitated by the development of the nonviral Cas9 expression system in primary human T cells (21). Here, the screening on tumor cells was performed on two independent GSCs, displaying a relatively narrow range of shared molecular targets involved in mediating responses to CAR T cells in our studies, which might be a consequence of subtype difference between these GSC lines (33). The screening identified both rational targets (RELA/p65) and novel targets (NPLOC4) in immune regulation, which were not restricted to a specific GBM molecular subclass. NPLOC4 displayed unexpected associations with GBM-targeting immune cell activity, as NPLOC4-KO in GSCs led to enhanced potency of CAR T cells and increased cytokine production in GSCs, although the detailed mechanism awaits further investigation. In the analyses of GSC models and the TCGA database, high RELA and NPLOC4 expression was associated with immuno-suppressive signatures. More specifically, higher expression of RELA and NPLOC4 in GBMs correlated with low infiltration of both CD4⁺ and CD8⁺ T cells, indicating that targeting these genes may confer immune modulatory effect and enhance antitumor T-cell responses in GBMs.

The assay used for CRISPR screening in T cells is crucial for reliable readouts and is required for its sensitivity to differentiate effective versus noneffective therapies. Although the in vivo antitumor efficacy in mouse models has been the standard to evaluate the functional quality of T cells in adoptive transfer, the utilization of this system in screening has been controversial. Tumor-infiltrating T cells harvested after the injection of therapeutic cells display signatures of tumor reactivity (72) or, conversely, T-cell exhaustion (40). The differential results appear model-dependent, leading to mixed interpretation of the results. The coculture assays that we used in this study identified key regulators by creating challenging screening environments. For the screening on GSCs, two rounds of short-term (24 h) killing with a relatively large number of T cells (total E:T = 1:1) were performed and GSCs were harvested immediately after the second round of killing, minimizing the effect of knocking out genes essential for the GSC growth. For the screening on CAR T cells, a repetitive challenge assay was used with an excessive number of GSCs (total E:T = 1:12), which we have shown to induce CAR T-cell exhaustion (30). The screen was performed by comparing a less exhausted (PD-1 negative) with a more exhausted (PD-1 positive) subset, informing prioritization for maintenance of recursive killing function, while reducing the noise from tumor cell or T-cell growth. The screening was performed with two independent CAR T-cell donors, and the relatively small proportion of overlapping hits between the two donors was expected and consistent with previous studies (21, 75), due to the variation in T-cell populations between individuals. The target validation was done with different T-cell donors and CAR platforms; therefore, the discovered immunotherapy targets may be generalizable to multiple CAR designs. Although we validated four representative genes, the screening on CAR T cells resulted in more than 200 potential targets involved in critical pathways of T-cell biology and activation, offering additional targets for future investigation of CAR refinement. One limitation of our approach, however, is the exclusion of apoptosis pathways in tumor cells due to their critical role in tumor cell growth, which have been demonstrated as important regulators of CAR T-cell–mediated tumor killing as well as tumor-induced CAR T-cell exhaustion (29).

T-cell exhaustion has been considered one of the major hurdles for reducing CAR T-cell potency (76–78). Blocking/KO of inhibitory receptors is being rigorously investigated to augment CAR activity or other tumor-targeting T cells (29, 79, 80). T-cell exhaustion is a feedback mechanism after activation, occurring upon recursive exposure to antigens in the contexts of chronic infection or the tumor microenvironment (77, 81), compromising their antitumor potency (78). Here, we observed that TLE4-KO or IKZF2-KO resulted in unstimulated CAR T cells to express transcriptional profiles of activation, while prohibiting exhaustion. The AP-1 family transcription factors FOS and JUN, which were induced after both TLE4-KO and IKZF2-KO, provide a possible mechanism by which CAR T-cell fitness was protected. The protein c-JUN forms homodimers or c-FOS/c-JUN heterodimers to initiate transcription of proinflammatory cytokines, and heterodimers with other cofactors (including BATF, IRF4, JUNB, and JUND) induce inhibitory receptors or suppress transcriptional activity of c-JUN (82–85). FOS was more upregulated than suppressive cofactors after TLE4-KO, therefore driving T-cell activation together with a protection from exhaustion, which was reminiscent of the effect after expressing c-JUN in CAR T cells with tonic signaling (55). In IKZF2-KO cells, however, the uncoupling of activation from exhaustion signatures was likely influenced by the upregulation of cytokines CCL3 and CCL4, which inversely correlated with PD-1 expression during T-cell exhaustion (86). Both TLE4-KO and IKZF2-KO in CAR T cells upregulated essential regulators for Th1 cell differentiation (BCAT and EGR1, respectively), consistent with a previously identified role of this T-cell population in mediating antitumor immunity (87, 88). Consequently, targeted KOs in CAR T cells enhanced not only killing, but also expansion potential, which is correlated with clinical responses (89). Although it remains unresolved if these KOs potentiate CAR activity in immune-competent settings, our results have revealed the feasibility that CAR T cells can be modified for their activation/exhaustion signals to achieve functional improvement in clinically relevant models. Consistent with these findings, we explored public databases of scRNA-seq on patient-derived CAR T-cell products and discovered that high IKZF2 expression and TLE4 activity were associated with other suppressive/exhaustion signatures of CAR T cells as well as poor clinical responses.

Single-cell analyses reveal subset composition within a mixed cell sample, such as CAR T cells, in which minority populations serve critical roles. scRNA-seq revealed that CAR activation, rather than genetic modification of CAR T cells (TLE4-KO or IKZF2-KO), resulted in a major cluster switch, which is consistent with the observation that TLE-KO or IKZF2-KO in monoculture CAR T cells did not dramatically alter transcriptional profiles, as suggested by bulk RNA-seq. Following tumor challenge, KO of targeted genes upregulated T-cell activation markers and proinflammatory cytokines across different clusters, especially IFNG and CCL3, which showed similar induction by both TLE-KO and IKZF2-KO. Further, after CAR activation, TLE4-KO maintained a specific cluster, which existed preactivation, and IKZF2-KO led to the emergence of a new cluster. The transcriptional signature of these clusters (expression of several costimulation molecules and cytokines) indicated their critical role in mediating effector function of CAR T cells. Therefore, the superior functions of TLE4-KO or IKZF2-KO CAR T cells were likely the result of a generally elevated activation state, as well as the stimulatory effect from critical subsets. Our scRNA-seq results also suggested the existence of Treg-like populations, the expansion of which was seen after CAR activation and can be reduced by IKZF2-KO. The suppressive function of these cells still requires further investigation, but these results indicate the potential of enhancing CAR function through inhibiting differentiation toward Treg-like cells. Both TLE4-KO and IKZF2-KO CAR T cells appear to modify specific CD4⁺ T-cell subsets, which supports our previous observation that CD4⁺ CAR T cells play a critical role in mediating potent effector function (30).

In summary, whole-genome CRISPR KO screens identified a limited set of key targets on GSCs that promote resistance to CAR killing. Modifying these genes before CAR T-cell infusions, with genetic or chemical approaches, might be a potential strategy to enhance CAR T-cell antitumor function and also guide patient selection. The spectrum of identified CAR T targets in our screen was broad, but manipulating expression of several of these targets appeared to regulate shared transcriptional programs, including cytokines and AP-1 regulation. Given the exciting advances in CAR T engineering with logic gates, it should be possible to interrogate multiple molecular targets within CAR T cells to determine potential synergy or additional benefit. Although our screens and evaluations were mainly focused on GSCs, the results are also expected to be consistent non-GSC populations, as CAR-mediated cytotoxicity has been demonstrated not to bias between antigen-positive GSCs and more differentiated tumor cells (4–6). Although further investigation is needed to demonstrate the function of these targets in T cells harboring other CAR constructs, our results found that most of the targets selected for deeper studies regulate core pathways in immune cell function, suggesting that these targets likely have broad regulatory function in immune cell antitumor function in various contexts. Future studies will also permit interrogation of critical targets in an immune-intact environment, but our current focus was to interrogate molecular regulators directly involved in CAR T killing of the therapeutically resistant human GSC population. Overall, the bidirectional screening provides a platform to robustly identify targets with the potential toward clinical benefit in CAR T and other adoptive cell transfer therapies.

GSCs were acquired from patient specimens at City of Hope under protocols approved by the Institutional Review Board (IRB), and maintained as tumorspheres in GSC media as previously described (4, 90). GSC lines used in this study to test CAR T-cell function are IDH1/2 wild-type. The sgRNA library and single-targeted sgRNA lentiviral plasmids (containing a puromycin-resistant gene) for GSC transduction were purchased from Addgene (#73179 and #52961, respectively). Lentiviral particles were generated as previously described (91). For lentiviral transduction, GSC tumorspheres were dissociated into single cells using Accutase (Innovative Cell Technologies), resuspended in GSC media, and lentivirus was added at a 1:50 v/v ratio. GSCs were then washed once after 12 hours, resuspended in fresh GSC media, and cultured for three days. To ensure that only transduced cells were expanded for further assays, GSCs were selected by puromycin (Thermo Fisher Scientific) for seven continuous days, with a 1:10,000 v/v ratio into GSC media.

Naïve and memory T cells were isolated from healthy donors at City of Hope under protocols approved by the IRB (26, 30). The constructs of IL13Rα2-targeted and HER2-targeted CARs were described in previous studies (8, 26, 92). Procedures of CAR-only transduction on primary human T cells were previously described (44). The sgRNA library and single-targeted sgRNA lentiviral plasmids for T-cell transduction were purchased from Addgene (#73179 and #52961, respectively). All sgRNA plasmids contain a puromycin-resistant gene. Dual transduction of CAR and sgRNA was performed using modification of previously reported procedures (21). In brief, primary T cells were stimulated with Dynabeads Human T expander CD3/CD28 (Invitrogen; T cells:beads = 1:2) for 24 hours and transduced with sgRNA lentivirus (1:250 v/v ratio). Cells were washed after 6 hours and then transduced with CAR lentivirus [multiplicity of infection (MOI) = 0.5]. Four days after CAR transduction, CD3/CD28 beads were removed and cells were resuspended in Lonza electroporation buffer P3 (Lonza, #V4XP-3032; 2 × 10⁸ cells/mL). Cas9 protein (MacroLab; 40 μmol/L stock) was then added to the cell suspension (1:10 v/v ratio), and electroporation was performed using a 4D-Nucleofactor Core Unit (Lonza, #AAF-1002B). Cells were recovered in prewarmed X-VIVO 15 media (Lonza) for 30 minutes before proceeding to ex vivo expansion. All T-cell transduction and ex vivo expansion experiments were performed in X-VIVO 15 containing 10% FBS, 50 U/mL recombinant human IL2 (rhIL2), and 0.5 ng/mL rhIL15, at 6 × 10⁵ cells/mL. To ensure that only sgRNA-transduced cells were expanded, puromycin (1:10,000 v/v ratio) was added to the media three days after electroporation, and puromycin selection was performed for six continuous days before CAR T cells were used for further assays. CRISPR screening was performed on two independent donors, and two other donors were used to generate IL13Rα2-targeted and HER2-targeted CARs, respectively.

GSCs transduced with the CRISPR KO library were dissociated into single cells, and cocultured with CAR T cells at an E:T ratio of 1:2 in culture plates precoated with Matrigel. After 24 hours, the media containing CAR T cells and tumor debris were removed, and the same number of CAR T cells were added in fresh media. Twenty-four hours after the second CAR T-cell addition, the media were removed and the remaining GSCs were washed with PBS and harvested. Genomic DNA was isolated from the remaining GSCs after coculture with CAR T cells, as well as GSCs harvested before coculture and GSCs after monoculture for 48 hours.

T cells transduced with CAR and the CRISPR KO library were cocultured with GSCs at an E:T ratio of 1:4 in culture plates precoated with Matrigel. After 48 hours, CAR T cells were rechallenged by GSCs doubling the number of the initial coculture. Twenty-four hours after the rechallenge, the coculture was harvested and stained with fluorescence-conjugated antibodies against human CD45 (BD Biosciences; catalog no. 340665, RRID:AB_400075), PD-1 (BioLegend; catalog no. 329922, RRID:AB_10933429), and IL13 (BioLegend; catalog no. 501914, RRID:AB_2616746). Different subsets were sorted using an Aria SORP (BD Biosciences): total CAR T cells (CD45⁺, IL13⁺), PD-1⁺ CAR T cells (CD45⁺, IL13⁺, PD-1⁺), and PD-1⁻ CAR T cells (CD45⁺, IL13⁺, PD-1⁻). Genomic DNA was isolated from the sorted subsets of cells, as well as CAR T cells harvested before coculture and CAR T cells after monoculture for 72 hours.

FASTQ files were trimmed to 20-bp CRISPR guide sequences using BBDuk from the BBMap (https://jgi.doe.gov/data-and-tools/bbtools; RRID:SCR_016965) toolkit, and quality control was performed using FastQC (RRID:SCR_014583, https://www.bioinformatics.babraham.ac.uk/projects/fastqc/). FASTQs were aligned to the library and processed into counts using the MAGECK-VISPR “count” function (https://bitbucket.org/liulab/mageck-vispr/src/master/). β values were calculated using a Maximum Likelihood Estimation model generated independently for each comparison. Nontargeting sgRNAs were used to derive a null distribution to determine P values.

For in vitro cytotoxicity test, CAR T cells were cocultured with GSCs at an E:T ratio of 1:40. After 48 hours of coculture, the numbers of CAR T cells and GSCs were evaluated by flow cytometry. Flow cytometry assays were performed on GSCs, CAR T cells from monoculture or coculture with procedures described previously (30). For coculture, anti-CD45 (BD Biosciences; catalog no. 340665, RRID:AB_400075) staining was used to distinguish GSCs with T cells, and CAR T cells were identified by anti-IL13 (BioLegend; catalog no. 501914, RRID:AB_2616746) staining. Other antibodies used for flow cytometry are listed as follows: PD-L1 (Thermo Fisher Scientific, catalog no. 17-5983-42, RRID:AB_10597586), TIM3 (Thermo Fisher Scientific; catalog no. 17-3109-42, RRID:AB_1963622), LAG3 (Thermo Fisher Scientific; catalog no. 12-2239-41, RRID:AB_2572596), PD-1 (BioLegend; catalog no. 329922, RRID:AB_10933429), CD69 (BD Biosciences; catalog no. 340560, RRID:AB_400523), CD137 (BD Biosciences; catalog no. 555956, RRID:AB_396252), and IL13Rα2 (BioLegend; catalog no. 354404, RRID:AB_11218789). All samples were analyzed via a Macsquant Analyzer (Miltenyi Biotec) and processed via FlowJo v10 (RRID:SCR_008520).

Total mRNA from GSCs or CAR T cells was isolated and purified by the RNeasy Mini Kit (Qiagen Inc.) and sequenced with Illumina protocols on a HiSeq 2500 to generate 50-bp reads. Trim Galore (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/; RRID:SCR_011847) was used to trim adaptors and remove low-quality reads. Reads were quantified against Gencode v29 using Salmon (RRID:SCR_017036, https://combine-lab.github.io/salmon/) with correction for fragment-level GC bias, positional bias, and sequence-specific bias. Transcripts were summarized to gene level and processed to transcripts per million (TPM) using the R/Bioconductor (https://www.bioconductor.org/) package DESeq2 (RRID:SCR_000154, https://bioconductor.org/packages/release/bio-c/html/DESeq2.html). Comparisons were performed using contrasts in DESeq2 followed by Benjamini–Hochberg adjustment to correct for false discovery rate (FDR).

ClueGO gene-set enrichment plots were generated using the ClueGO plugin (http://apps.cytoscape.org/apps/cluego, RRID:SCR_005748) for GO BP, KEGG, or Reactome gene sets and visualized in Cytoscape v3.7.2 (https://cytoscape.org/). Gene set enrichment analysis (GSEA; RRID:SCR_003199) plots were generated from preranked lists using the mean β value as the ranking metric. Reactome networks were created using the Reactome FI plugin (https://reactome.org/tools/reactome-fiviz) with network version 2018 and visualized in Cytoscape. Networks were clustered using a built-in network clustering algorithm, which utilizes spectral partition-based network clustering, and node layout and color were determined by module assignment. GSEA plots from RNA-seq data were generated from preranked lists. Weighting metrics for preranked lists were generated using the DESeq2 results from the gene knockdown versus nontargeting control and applying the formula: −log₁₀(FDR) × log₂(fold change). ssGSEA scores for specific immune or functional pathways were generated using the ssGSEA function from the R/Bioconductor package GSVA (https://bioconductor.org/packages/release/bioc/html/GSVA.html; ref. 93) and plotted using pheatmap (https://cran.r-project.org/web/packages/pheatmap/). ChEA enrichments were performed using Enrichr (https://amp.pharm.mssm.edu/Enrichr/). Bar plots for positive or negative gene set enrichments were performed using Metascape (https://metascape.org/gp/index.html) for significantly upregulated or downregulated genes (FDR <0.05 and log₂ fold change >1 or < −1).

Reactome networks were derived from RNA-seq data using the Cytoscape Reactome FI plugin (RRID:SCR_003032). A gene list of upregulated (FDR <0.05 and log₂ fold change >1) or downregulated (FDR <0.05 and log₂ fold change < −1) genes plus the target gene (as KO by CRISPR/Cas9 would not be detected by RNA-seq) was input into Reactome FI, and all genes with at least one edge were included in the network plot. Node color (light to dark) and size (small to large) are proportional to node degree. Pathway enrichment was performed on this network of genes using the Reactome FI enrichment option. Box plots for genes from selected pathways were generated using RNA-seq TPM data. KEGG pathway visualizations were generated using the R/Bioconductor package pathview (https://www.bioconductor.org/packages/release/bioc/html/pathview.html) for selected pathways, and genes were colored on the basis of the log₂ fold change KO versus control.

scRNA-seq files were processed using the Cell Ranger workflow (https://support.10xgenomics.com/single-cell-gene-expression/software/overview/welcome). FASTQ files were generated using the Cell Ranger “mkfastq” command with default parameters. FASTQs were aligned to the hg19 genome build using the “count” function and aggregated using the default Cell Ranger “aggr” parameters with normalization performed by subsampling wells to equalize read depth across cells. Downstream analyses were performed using the R/Bioconductor package Seurat (https://satijalab.org/seurat/; ref. 94). Specifically, data sets of stimulated and unstimulated cells in KO or control populations were merged using the “FindIntegrationAnchors” Seurat function. Clustering was performed using UMAP using PCA for dimensional reduction and a resolution of 0.6 from 1 to 20 dimensions. Dead cell clusters were determined by high expression of mitochondrial genes and removed. Samples were then reclustered. Clusters with similar CD4 or CD8, Ki-67, and marker expression, determined using the “FindAllMarkers” function that were proximal on the UMAP projection, were merged. All plots for gene expression were generated using normalized data from the default parameters of the “NormalizeData” function. Gene expression was visualized on the UMAP projection using the “FeaturePlot” function with a maximum cutoff or gene expression determined on a gene-by-gene basis.

All mouse experiments were performed using protocols approved by the City of Hope Institutional Animal Care and Use Committee. Orthotopic GBM models were generated using 6- to 8-week-old NOD/SCID/IL2R^−/− (NSG) mice (IMSR; catalog no. JAX:005557, RRID:IMSR_JAX:005557), as previously described (95). Briefly, ffLuc-transduced GSCs (1 × 10⁵/mouse) were stereotactically implanted (intracranially) into the right forebrain of NSG mice. Randomization was performed after 8 days of tumor injection based on bioluminescence signal, and mice were then treated intracranially with CAR T cells (2 × 10⁴ or 5 × 10⁴/mouse as indicated for each experiment). To ensure statistical power, all treatment groups included ≥6 animals. Mice were monitored by the Department of Comparative Medicine at City of Hope for survival and any symptoms related to tumor progression, with euthanasia applied according to the American Veterinary Medical Association Guidelines. Studies were done in both male and female animals. Investigators were not blinded for randomization and treatment.

Analysis of genes in the TCGA data set was performed using RNA-seq TCGA GBM data. Immune infiltration signatures were previously reported (96). GSEA plots for each gene in the context of TCGA GBM data were generated using the normalized gene expression as a continuous phenotype.

Gene sets derived from TLE4-KO or IKZF2-KO were analyzed in the context of CAR T-cell nonresponders versus responders from a previous report on patients with CLL (27). Genes upregulated in bulk RNA-seq of CAR T cells following KO of TLE4 or IKZF2 (FDR <0.05 and log₂ fold change > 1) were plotted by their fold change expression in stimulated versus unstimulated CAR T cells for responders or nonresponders. Fold change was calculated using DESeq2 for stimulated versus unstimulated cells independently for each group (nonresponder or complete responder). Cluster 10–enriched genes in the TLE4-KO and control sc-seq data, identified by the “FindAllMarkers” function in Seurat subsetted for overexpressed genes, were plotted similarly. Genes upregulated (>0.4 log₂ fold change of normalized counts) in sc-seq for IKZF2-KO versus control in stimulated CAR T cells were plotted similarly.

CAR T-cell functional data (tumor killing, expansion, survival of tumor-bearing mice) were analyzed via GraphPad Prism. Group means ± SEM were plotted. Methods of P value calculations are indicated in figure legends.

Both bulk RNA-seq and scRNA-seq data sets have been deposited to the Gene Expression Omnibus (GSE163409) and will be accessible to the public at the time of publication.

D. Wang reports grants from NCI during the conduct of the study; in addition, D. Wang has a patent for 63/117439 pending. R.C. Gimple reports grants from NIH during the conduct of the study. M.H. Lorenzini reports personal fees from Kite, a Gilead company, outside the submitted work. B. Badie reports grants from NIH and grants and personal fees from Mustang Therapeutics during the conduct of the study; in addition, B. Badie has a patent for CAR T-cell delivery pending. C.E. Brown reports grants from Ivy Foundation, CIRM, and grants from NIH during the conduct of the study; personal fees from Mustang Bio outside the submitted work; in addition, C.E. Brown has a patent for 63/117439 pending. J.N. Rich reports grants and personal fees from Synchronicity and other support from Function Oncology outside the submitted work. No disclosures were reported by the other authors.

D. Wang: Conceptualization, data curation, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. Q. Wu: Validation and methodology. L.J.Y. Kim: Software, formal analysis, and visualization. Z. Qiu: Software, formal analysis, and visualization. P. Lin: Methodology. M.H. Lorenzini: Writing–original draft. B. Badie: Resources. S.J. Forman: Resources and supervision. Q. Xie: Conceptualization, data curation, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. C.E. Brown: Conceptualization, resources, data curation, supervision, funding acquisition, validation, investigation, visualization, writing–original draft, project administration, writing–review and editing. J.N. Rich: Conceptualization, resources, data curation, supervision, funding acquisition, investigation, visualization, methodology, project administration, writing–review and editing. B.C. Prager: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. R.C. Gimple: Data curation, validation, and methodology. B. Aguilar: Data curation, validation, investigation, and visualization. D. Alizadeh: Validation, investigation, and visualization. H. Tang: Data curation, software, formal analysis, and methodology. D. Lv: Data curation, investigation. R. Starr: Data curation, validation, and methodology. A. Brito: Data curation and methodology.

This study was supported by the following funding sources: NIH: CA234923 (to D. Wang), CA217066 (to B.C. Prager), CA217065 (to R.C. Gimple), CA203101 (to L.J.Y. Kim), CA236500 (to C.E. Brown), CA238662, CA197718, NS203434 (to J.N. Rich); Ivy Foundation (to C.E. Brown); and Westlake Education Foundation (to Q. Xie). C.E. Brown receives royalty payments from Mustang Bio. Patent associated with this study has been submitted (application serial number: 63/117439).

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