Comparison of axicabtagene ciloleucel and tisagenlecleucel patient CAR-T cell products by single-cell RNA sequencing

Abstract

Background

Autologous CD19 chimeric antigen receptor (CAR) T-cell therapy leads to durable responses and improved survival in patients with relapsed or refractory large B-cell lymphoma (R/R LBCL). Among approved CAR T-cell products, axicabtagene ciloleucel (axi-cel; CD19/CD28) has greater real-world efficacy and cytokine-associated toxicity than tisagenlecleucel (tisa-cel; CD19/4-1BB), for reasons that are poorly understood.

Methods

Here we report single-cell RNA sequencing (scRNA-seq) of 57 pre-infusion CAR T-cell products from axi-cel (n=39) and tisa-cel (n=18) patients treated as standard-of-care for R/R LBCL, and their biological associations with clinical outcomes. In vitro CAR manufacturing conditions mimicking those known for axi-cel and tisa-cel were performed using CD19/CD28z or CD19/4-1BBz constructs.

Results

ScRNA-seq revealed that axi-cel and tisa-cel are markedly different products. Axi-cel is comprised of more CD4 central memory, CD8 central memory, and CD8 effectors, whereas tisa-cel is comprised of more proliferative CD4 and CD8 cells. Across multiple T-cell subsets, axi-cel had greater expression of immune response pathways and protein synthesis and trafficking pathways versus tisa-cel. On comparison of infusion product CAR transgene-positive (CAR+) cells to CAR transgene-negative (CAR−) T cells, axi-cel CAR+ cells had vastly different gene expression than axi-cel CAR− cells. Unexpectedly, tisa-cel CAR+ cells were highly similar to tisa-cel CAR− cells. Under recapitulated CAR-T manufacturing conditions known to be used for axi-cel and tisa-cel, we found that CAR+ cells differed from CAR− cells early after manufacturing yet became more similar to CAR− cells after prolonged expansion. Prolonged time in expansion culture, as used during tisa-cel manufacturing, greatly decreased naïve and central memory T-cell subsets.

Conclusions

Following manufacture, axi-cel is less differentiated and has greater immune activation compared with tisa-cel, potentially accounting for its greater efficacy and toxicity in patients. Our data support the conclusion that tisa-cel is adversely affected by its manufacturing rather than by the CAR construct.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Axicabtagene ciloleucel (axi-cel) and tisagenlecleucel (tisa-cel) are CD19-directed chimeric antigen receptor (CAR) T-cell products approved for the treatment of relapsed or refractory diffuse large B-cell lymphoma.

• In non-randomized studies including cross-trial comparisons and real-world data, axi-cel appears to have greater efficacy and toxicity compared with tisa-cel.

• CAR T-cell products with a greater proportion of naïve and central memory T cells outperform products with more effector T cells.

WHAT THIS STUDY ADDS

• Axi-cel consists of more central memory CAR T cells, whereas tisa-cel consists of more proliferative CAR T cells.

• Prolonged manufacturing, as is observed with tisa-cel, associates with a greater similarity between CAR transgene-positive and CAR transgene-negative T cells, and a decreased proportion of central memory T cells.

• Durable response to axi-cel associates with gene expression of ribosomal signatures across CAR T-cell subsets.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE, OR POLICY

• CAR T-cell manufacturing processes should aim to produce CAR T cells that have divergence between the CAR transgene-positive and CAR transgene-negative cells.

• High expression of ribosomal genes is desirable in a CAR T-cell product.

Introduction

Autologous CD19-directed chimeric antigen receptor T cell therapy (CAR-T) is now standard-of-care for relapsed or refractory large B-cell lymphoma (R/R LBCL). Three CAR-T products are US Food and Drug Administration-approved for treatment of R/R LBCL: axicabtagene ciloleucel (axi-cel), tisagenlecleucel (tisa-cel), and lisocabtagene maraleucel (liso-cel). Approximately 40–50% of patients will attain a durable remission after CAR-T, with the vast majority of relapses occurring within the first 6 months.1,4

Axi-cel is considered to be more efficacious with higher adverse event rates than tisa-cel. For example, multiple prospective national registries and large retrospective cohorts in the third or later line setting have reported better progression-free survival (PFS) for axi-cel versus tisa-cel, despite correction for baseline patient and LBCL disease characteristics.5,9 Liso-cel and axi-cel are both approved for use in the second line setting for LBCL based on randomized phase 3 clinical trials, while tisa-cel failed to improve event-free survival over existing second line standard of care.^{1 3 10} Reasons for the inferior efficacy with tisa-cel are unclear.¹¹

Axi-cel contains a CD28 costimulatory domain and its manufacture uses a gamma retroviral vector, total peripheral blood mononuclear cells as starting material, as well as anti-CD3 antibody for activation/stimulation.¹² Tisa-cel and liso-cel, by contrast, each have a 4-1BB costimulatory domain and are manufactured using a lentiviral vector, isolated CD3+ (tisa-cel) or CD4+ and CD8+ (liso-cel) T cells are starting materials, and anti-CD3/anti-CD28 coated microbeads are used for activation/stimulation.¹² In addition, the manufacturing time for tisa-cel is prolonged compared with axi-cel. The degree to which these features impact efficacy is unclear and understudied.

Autologous CAR-T cell therapy requires peripheral blood leukapheresis to obtain the T cells used for CAR-T cell manufacturing, and the resulting CAR-T cell product is unique to each patient in terms of T-cell subsets and polyfunctionality.^{13 14} Favorable outcomes with axi-cel are associated with a higher proportion of naïve-like stem central memory and central memory T cells in the infusion product, while poor outcomes are associated with a higher proportion of effector T cells.^{14 15} Other features of the manufactured axi-cel product that are associated with a favorable outcome include the capacity to excrete multiple cytokines (ie, polyfunctionality), a short in vitro doubling time, and a CD8+memory T cell phenotype.^{13 14} Further, an exhausted CD8+T cell phenotype or the emergence of CD4+T-regulatory cells (Tregs) post-infusion is associated with relapse.16,18 By contrast, features of tisa-cel CAR T-cell product associated with outcome are less clear.¹⁹

How the type of CAR constructs used and product manufacturing differences impact these consequential phenotypic findings is unclear. Furthermore, the degree to which CAR T product manufacturing may play a role in the lower efficacy of tisa-cel is unknown. Here we characterized the pre-infusion product from 57 patients with LBCL receiving axi-cel and tisa-cel using single-cell RNA sequencing (scRNA-seq). We report on product characteristics including key differences between axi-cel and tisa-cel and features associated with superior outcomes.

Methods

Patients and samples

All patients were consented to prospective sample collection and research database protocols approved by the Institutional Review Board of the University of South Florida or by Advarra (Pro00021733). All patient samples were collected at Moffitt Cancer Center. Patient CAR T cells were obtained by washes from commercial product infusion bags after patient infusion, and cryopreserved for batch analysis.

Single-cell RNA sequencing, data processing, and analysis

ScRNA-seq was performed using 10x Genomics Chromium (10x Genomics, Pleasanton, California, USA) by the Molecular Genomics Core at Moffitt Cancer Center (Tampa, Florida). Detailed methods for encapsulation, complementary DNA library preparation, scRNA-seq, data pipelining, T-cell subset annotation, and Gene Set Enrichment Analysis (GSEA) are provided in the online supplemental material.

Identification of CAR+ cells

Commercial tisa-cel products are accompanied by a manufacturing certificate that provides the % CAR+cells identified during quality control testing, which uses a flow cytometry-based assay to identify the CAR. Independently, in scRNA-seq data CAR+ versus CAR− cells were distinguished based on the presence or absence of a short sequence (CAR-CD19FvL) that is within the CAR transgene for both axi-cel and tisa-cel.

Meta-analysis of differences associated with CAR-T products and clinical outcomes

Differential expression was performed between axi-cel and tisa-cel products within each cluster. A meta p value was calculated for each gene, combining p values across clusters using Fisher’s method implemented in the sumlog function of the metap package. A meta log2(fold-change) was calculated as median fold changes across clusters. A similar meta-analysis was performed to compare durable responders versus non-durable responders in axi-cel and tisa-cel products separately. Volcano plots were used to visualize the meta-analysis results.

Additional statistical analyses

All additional statistical analyses were performed using R V.4.3.1. To compare the fraction of different T-cell subtypes between axi-cel versus tisa-cel products, between durable responders versus non-durable responders, as well as between CAR+ versus CAR− T cells, the Wilcoxon test was applied, followed by a false discovery rate(FDR) correction for multiple testing using Benjamini and Hochberg. Comparison results were shown in box plots in both main and supplementary figures with data presented as mean±SEM.

CAR T-cell manufacture in vitro

Cryopreserved peripheral blood mononuclear cells from healthy donors (LifeSouth Community Blood Centers) were thawed and either enriched for T cells using a magnetic bead T Cell Isolation Kit (STEMCELL Technologies) or left unselected. Isolated T cells were activated by resuspending 1:1 with CD3/CD28 Dynabeads (Gibco) in complete culture media (CCM) consisting of RPMI 1640 (Gibco)+10% human serum (Gemini Bio) + 1× MycoZap PR (Lonza) supplemented with 100 IU/mL rhIL-2 (R&D Systems). Unselected cells were activated by resuspending in CCM supplemented with 50 ng/mL OKT3 antibody (Invitrogen) and 300 IU/mL rhIL-2. After 24 hours of activation, cells were collected, washed and resuspended in fresh CCM as indicated above at a concentration of 1 million cells/mL. 1 mL per well of cell suspension was transferred to each well of a 6-well retronectin-coated tissue culture plate, then 1 mL of either CD19 28z or CD19 BBz viral supernatant was added to each well. Plates were centrifuged at 2000×g for 60 min at 32°C (spinoculation) and then stored in a 37°C incubator for 24 hours. After 24 hours, 800 µL was removed from each well and replaced with fresh viral supernatant as indicated, and the spinoculation step was repeated. Cells were incubated for 24 hours and then fresh CCM supplemented with rhIL-2±OKT3 antibody as indicated was added to each well. Cells were expanded with fresh CCM added every other day and collected at the described time points for antibody staining and flow cytometry analysis. Detailed methods on flow cytometry, staining, and analysis are provided in the online supplemental material.

Results

Axi-cel and Tisa-cel CAR T cell products are markedly different

To compare axi-cel and tisa-cel infusion products, scRNA-seq was performed on encapsulated residual CAR T cells from the infusion bags from 57 patients with LBCL (axi-cel, n=39; tisa-cel, n=18, figure 1A). Axi-cel treated patients had fewer prior lines of therapy, worse Eastern Cooperative Oncology Group Performance Status (ECOG PS), and were less likely to have had transformed disease (online supplemental table S1). Time from apheresis to infusion was longer for tisa-cel compared with axi-cel (median 38.5 vs 27 days, p<0.001), consistent with the longer wait for manufacturing when ordering tisa-cel. A durable remission was observed in 33% of patients overall (axi-cel 38% vs tisa-cel 22%, p=0.4; online supplemental figure S1A,B). These clinical features are consistent with the reported use and outcomes for axi-cel and tisa-cel in large registries and retrospective patient cohorts.5,9

Figure 1. Single-cell RNA sequencing of axicabtagene ciloleucel and tisagenlecleucel infusion products identifies CAR-expressing T cells. (A) Overview of study design, patient sample numbers, and manufacturing differences between axi-cel and tisa-cel. (B) UMAP demonstrating single-cell clustering based on RNA expression of 105,326 CAR-T cell infusion product cells from patients with standard care axi-cel or tisa-cel treated large B-cell lymphoma. Cluster designation is based on gene expression patterns that are known to be associated with select differentiation states (online supplemental figure S2A-D. See online supplemental methods for the full name of each cluster. (C) Total percentage of CD4+, CD8+, or myeloid cells by product type. Each dot represents a different patient/product (pink=axi-cel; olive=tisa-cel). Shown are mean+SEM. Two-sided Wilcoxon rank-sum test was used to calculate p value followed by FDR correction. (D) UMAP showing cells from axi-cel (top, pink/maroon) and tisa-cel (bottom, olive/green) infusion products that either do not express the respective CAR transgene (left, axi-cel pink, tisa-cel olive) or do express the CAR transgene (right, axi-cel maroon, tisa-cel green). (E) Correlation of tisa-cel percentage of T cells expressing CAR protein by flow cytometry (conducted by the manufacturer) and the percentage of tisa-cel product T cells expressing the CAR transgene as quantified by scRNA-seq. Each dot represents a different tisa-cel patient infusion product (n=18). (F) Stacked bar graphs demonstrating the proportion of cells within each cluster for axi-cel (left), and tisa-cel (right) products, separated by CAR+ and CAR− cells. axi-cel, axicabtagene ciloleucel; CAR, chimeric antigen receptor; DR, durable responder; FDR, false discovery rate; IACs, immune effector cell-associated neurologic toxicity-associated myeloid cells; mregDC, mature immunoregulatory dendritic cells; NDR, non-durable responder; PFS, progression-free survival; scRNA-seq, single-cell RNA sequencing; tisa-cel, tisagenlecleucel; Treg, regulatory T cells; UMAP, uniform manifold approximation and projection.

Following correction for batch effects and removal of doublets and cells with high mitochondrial content (see Methods), we analyzed a total of 105,326 cells (axi-cel, n=78,714; tisa-cel, n=26,612) that segregated into 17 clusters (figure 1B, online supplemental figure S1C,D). T-cell clusters (n=15) were consistent with known T-cell differentiation and functional states (online supplemental figure 2A–G). Two myeloid type cell clusters were also identified that classified as mature immunoregulatory dendritic cells (mregDCs) and immune effector cell-associated neurologic toxicity (ICANS)-associated myeloid cells (IACs).^{18 20}

Axi-cel and tisa-cel products were broadly different. Most tisa-cel products were composed of predominantly CD4+cells while axi-cel products had more CD8+cells and myeloid cells (figure 1C). We distinguished CAR+ from CAR− cells within each product by identifying expression of the axi-cel or tisa-cel transgene (figure 1D). Tisa-cel, but not axi-cel, products are delivered with manufacturer documentation of the CAR transduction efficiency as measured by flow cytometry. There was a strong correlation between our method of distinguishing CAR+cells within scRNA-seq versus the tisa-cel report (figure 1E), validating our identification of CAR+ versus CAR-T cells. Finally, although cells from both product types were distributed across clusters, the proportion of cells within each cluster was different for axi-cel and tisa-cel in both CAR+ and CAR− cells (figure 1F).

Axi-cel and tisa-cel CAR transgene-expressing cells have distinct gene signatures

ScRNA-seq analyses revealed that the CAR-expressing cells of the axi-cel and tisa-cel products were markedly different. For example, within CD4+T cell clusters, the axi-cel product had a significantly higher proportion of CAR+cells that were stem central memory (CD4scm), memory (CD4m), and Treg (figure 2A,B), whereas tisa-cel CAR-expressing CD4+cells were more likely to be proliferative (CD4prolif) and Th17 cells. Further, within CD8+T cell clusters, the axi-cel product had a significantly higher proportion of CAR-expressing central memory (CD8cm), cytotoxic (CD8 cytotoxic), and effector (CD8eff) T cells compared with those of the tisa-cel product (figure 2C,D), which were more likely to be effector exhausted (CD8eff-ex) and proliferative (CD8prolif) cells.

Figure 2. Axi-cel and tisa-cel CAR T cells have different phenotype and gene expression profiles. (A) Cell clusters of CAR+ CD4+ T cells between axi-cel and tisa-cel products. (B) % of CD4 CAR T cells that fall within specific clusters comparing axi-cel and tisa-cel on a patient level basis (each dot represents a different patient’s cell type as a % of total CAR+ CD4 cells). (C) Cell clusters of CAR+ CD8+ T cells between axi-cel and tisa-cel products. (D) % of CD8 CAR T cells that fall within specific clusters comparing axi-cel and tisa-cel on a patient level basis (each dot represents a different patient’s cell type as a % of total CAR+ CD8 cells). (E) Evaluation of Hallmark gene-set expression within T-cell clusters between axi-cel and tisa-cel. Red=higher in axi-cel within that cluster, blue=higher in tisa-cel within that cluster. Cells are independently clustered with Hallmark categories shown vertically on the left. Shown are normalized enrichment scores. (F) Evaluation of Gene Ontology Cellular Component pathways. Shown are gene sets with significantly different expression between axi-cel and tisa-cel within T-cell clusters. (G, H) Meta-analysis identified genes with differential expression across clusters in single-cell RNA sequencing data, between axi-cel and tisa-cel products. Genes are labeled on the volcano plot with differential expression between axi-cel and tisa-cel with a meta |log2FC|>=2, meta log(p value)>300, and expressed in >15% of CAR+ cells in axi-cel. (G) CD4+ CAR+ cells. (H) CD8+ CAR+ cells. For B and D, shown are mean+SEM; two-sided Wilcoxon rank-sum test was used to calculate p value followed by FDR correction. axi-cel, axicabtagene ciloleucel; CAR, chimeric antigen receptor; FDR, false discovery rate; tisa-cel, tisagenlecleucel.

Hallmark gene expression pathway analysis in each cluster confirmed distinct gene signatures of the axi-cel and tisa-cel products. Specifically, axi-cel CAR-T cells had broadly higher expression of immune signaling pathways across the clusters, including interferon, JAK/STAT, and allograft-rejection gene sets (figure 2E), whereas tisa-cel CAR-T cells expressed high levels of cell cycle pathways such as E2F targets and G2M signaling. Further, Gene Ontology (GO) pathways analysis revealed that axi-cel CAR-T cells expressed higher levels of genes in signaling pathways associated with cell cytotoxicity, antigen presentation and membrane-trafficking pathways, while cell cycle and mitochondrial pathways were elevated in tisa-cel CAR-T cells (figure 2E,F; online supplemental figure S3A,B).

Finally, a meta-analysis was performed to compare individual gene expression across clusters between axi-cel and tisa-cel CAR-T cells. Among genes with high expression, axi-cel expressed higher levels of lymphotoxin beta and granzyme A in both CD4 and CD8 CAR T cells (figure 2G,H). In contrast, tisa-cel had higher levels of the inducible costimulatory molecules TNFRSF4 (OX40) and TNFSRF18 (GITR), suggesting greater T-cell activation in the manufactured tisa-cel product. BATF3, a driver of CAR T-cell proliferation,²¹ was also found to be higher in tisa-cel. The differences in expression of these genes between axi-cel and tisa-cel occurred in multiple clusters (online supplemental figure S3C). Overall, these gene expression differences are consistent with the greater proportion of cytotoxic cells associated with axi-cel and the greater proportion of proliferative cells associated with tisa-cel.

CAR transgene positive and CAR transgene negative T cells are divergent for axi-cel but not tisa-cel infusion products

Manufactured CAR T-cell products consist of varying amounts of CAR+ and CAR− T cells (figure 1E,F). By scRNA-seq, axi-cel had a greater proportion of CAR+cells than tisa-cel, although there was no association between the percentage of CAR+ and clinical outcome (online supplemental figure S4A). Comparing the frequency of CAR+cells to CAR− cells in each cluster revealed that there were significantly more CAR+ than CAR− cells in axi-cel CD8eff, CD8eff_prolif, CD8 eff_ex, CD4trm, and Treg clusters, while there were fewer CAR+CD8 cytotoxic and CD4scm cells (figure 3A). In stark contrast, there were no frequency differences in the composition of tisa-cel CAR+ compared with to CAR− cell clusters.

Figure 3. ScRNA-seq reveals major differences in phenotype and gene expression profiles between CAR-expressing and non-CAR-expressing T cells within axi-cel, but not tisa-cel, products. (A) Percentage of cells that fall within specific clusters comparing axi-cel or tisa-cel product cells that do (CAR+) or do not (CAR−) express the CAR transgene by scRNA-seq. Each dot represents a different patient’s cells in a cluster within that category. Shown are mean+SEM. Two-sided Wilcoxon rank-sum test was used to calculate p value followed by FDR correction. FDR-adjusted p value: *≤0.05; **≤0.01; ***≤0.001, ****≤0.0001. (B, C) Evaluation of HALLMARK and Gene Ontology gene sets across clusters, comparing CAR+ to CAR− cells, revealed a significant difference in expression levels in many more gene sets for axi-cel products (pink) than tisa-cel products (olive). (B) Each dot represents the number of significantly different gene sets between CAR+ and CAR− cells within one cluster, for one product. (C) Each column shows a different cluster and the number of significantly different gene sets in aggregate across all patient products. (D) Previously validated gene score that quantifies tonic signaling of the CAR, comparing axi-cel (pink) to tisa-cel (olive) CAR+ cells (left), and CAR+ (red) and CAR− cells (gray) for both axi-cel (right, top) and tisa-cel (right, bottom) products. axi-cel, axicabtagene ciloleucel; CAR, chimeric antigen receptor; scRNA-seq, single-cell RNA sequencing; tsa-cel, tisagenlecleucel.

To further evaluate differences in CAR and non-CAR T cells within the infusion products, pathway analysis across clusters was performed comparing CAR+ and CAR− cells for each product. In general, the presence of the CAR transgene was associated with increased expression of gene sets across all clusters, with more gene sets increased by axi-cel than tisa-cel (online supplemental figure S4B–D). After adjusting for differences in patient and cell numbers (online supplemental methods), there were hundreds of pathways that differed significantly between the CAR+ and CAR− cells within each cluster for axi-cel, but only four pathways in one cluster that differed between CAR+ and CAR− cells for tisa-cel (figure 3B,C).

We found these results surprizing given reports that second-generation 4-1BB costimulatory CARs are more likely to exhibit tonic signaling as compared with CD28 costimulatory CARs. Assessing a gene score for tonic CAR signaling defined by comparing 4-1BB versus 28z costimulatory CAR-T cells²² revealed that, as expected, tisa-cel CAR+cells exhibited marginally higher tonic signaling. However, there were no clear differences between axi-cel and tisa-cel products, in tonic signaling differential between CAR+ and CAR− cells (figure 3D).

For axi-cel, almost every pathway across every cluster had increased gene expression in the CAR+cells compared with CAR− cells, including Myc, which is a general amplifier of gene expression (online supplemental figure 4B,D).^{23 24} Ribosomal signatures exhibited a unique pattern with increased expression among CAR+cells in some clusters but decreased expression among others, suggesting that factors other than the CAR transgene regulate ribosomal abundance differentially across T-cell types.

Manufacturing practices drive differences in commercial CAR infusion product phenotypes

While differential transcript expression between CAR+ and CAR− cells is likely associated with transgene, we felt the discrepancy in CAR+ versus CAR− differences between axi-cel and tisa-cel might also be accounted for by the manufacturing process. We previously reported that initiation of CAR-T manufacturing using frozen (tisa-cel process) or freshly collected (axi-cel process) apheresis material did not impact differentiation phenotype, cytokine release on activation, or CAR target-mediated cytotoxicity of the final CAR-T cells.²⁵ To interrogate effects of CAR-T product manufacturing conditions, and their differential impact on CAR+ and CAR− cells, we evaluated the impact of (1) starting cell material of unfractionated peripheral blood mononuclear cells (PBMC) with T-cell activation via anti-CD3+IL-2 (PBMC Ab mfg., axi-cel process) versus selected T cells with anti-CD3/CD28 bead activation (T bead mfg., tisa-cel process); and (2) shorter (early, 3–6 days, axi-cel process) versus longer (late, 12+day, tisa-cel process) ex vivo cellular proliferative expansion after viral transduction during manufacture (schema, figure 4A). Regardless of the costimulatory domain or manufacturing process, longer expansion time in culture after transduction was associated with reduced phenotypic differences between CAR+ and CAR− cells (figure 4B and online supplemental figure S9A). When evaluating multiple time points in culture, we found that much of the convergence between CAR+ and CAR− cells occurs between 6 and 10 days in culture, with differences depending on phenotypic marker (figure 4C and online supplemental figure S9B). Moreover, the expression of CCR7, a T-cell stem and central memory marker, decreased with time in culture. Thus, the limited differences between CAR+ and CAR− cells with tisa-cel are more likely due to manufacturing differences rather than direct signaling impact of the expressed CAR transgene. Further, differences in the T cell differentiation states between axi-cel and tisa-cel may also be explained by manufacturing.

Figure 4. Prolonged ex vivo CAR expansion during manufacture contributes to similarity in phenotype between CAR+ and CAR− cells. Experimentally manufactured CAR-T cell products from healthy donors were evaluated for cell surface protein expression of key differentiation and exhaustion markers by flow cytometry across time. (A) Shown is the experimental schema evaluating three different manufacturing conditions: both 4-1BBz and 28z CARs made from selected T cells stimulated with antiCD3/antiCD28 coated beads, and 28z CARs from PBMCs stimulated with antiCD3 antibody. Each of the three conditions was then expanded for a short 3–4 days duration (early) or a long 12–14 days duration (late). At each time point, the geometric mean fluorescent intensity (GMFI) for different differentiation and exhaustion markers was calculated for CD8+ cells that do (CAR+) or do not (CAR−), express the CAR protein. The delta GMFI for each manufacture condition, with both early and late expansion, was created by taking the difference between the GMFI for CD8+CAR+ and the GMFI for CD8+CAR−. (B) Shown is the delta GMFI for the indicated marker for each of the manufacture conditions with early (short 3–4 days) expansion (blue circle) and with late (long 12–14 days) expansion (red triangle). Each dot represents a different healthy donor's delta GMFI at the early (blue circle) or late (red triangle) timepoint. P values represent a paired t-test. Manufacture conditions were only significantly different for TCF7. Shown are CD8 T cells, with CD4 T cells from the same experiment displayed in online supplemental figure S9A. (C) In a separate experiment, the protein expression at four time points of ex vivo expansion was evaluated by flow cytometry. Shown is GMFI on the y axis and time in days from stimulation after viral transduction on the x axis. CAR-T cells were manufactured using a gamma-retroviral vector to express either a second generation CD19-CD28z CAR, or a CD19-41-BB, CAR construct. Cells were manufactured with PBMC as starting material stimulated with soluble anti-CD3 Ab and IL-2 (similar to axi-cel) or using a CD19-CD28z magnetically selected T cells as starting material stimulated with anti-CD3/anti-CD28 beads (similar to tisa-cel). Flow cytometry on the CAR+ (pink for CD28z construct, olive for 41-BBz construct) or CAR− (black) CD8 T cells allowed comparison of phenotype across time and by different constructs and manufacturing conditions. Shown are CD8 T cells with CD4 T cells from the same experiment shown in online supplemental figure S9B. CAR, chimeric antigen receptor; PBMC, peripheral blood mononuclear cells.

Axi-cel gene expression characteristics associated with efficacy

In our dataset, 15/39 (38%) of patients receiving axi-cel obtained a durable response, whereas only 4/18 (22%) of patients receiving tisa-cel obtained a durable response. While this is consistent with the lower reported efficacy for tisa-cel, the low number of durable responders limited our ability to analyze the association between tisa-cel and outcome (tisa-cel data presented in online supplemental figure S5A–D). We therefore focused our main analysis on axi-cel, and sought to understand the gene expression patterns of the product that are associated with efficacy.

First, we compared axi-cel CAR+cells within T-cell clusters between patients who had a durable response and those with no durable response. Although each individual patient’s infusion product showed variability, no individual cluster was significantly associated with durable response (figure 5A, and online supplemental figure S6A,B). This was surprising because some, although not all, previous scRNA-seq studies report that the memory to effector ratio and the presence of Tregs associate with clinical outcomes.16,1820 We therefore further evaluated these T-cell subsets. On combining CD4 and CD8 subsets, we found a directionally higher memory-to-effector ratio and a lower proportion of Tregs among patients with a durable response, although these differences were not significant (online supplemental figure S7A,B). Next, we evaluated GSEA from the ImmuneSigDB subset of MSigDB²⁶ for gene sets associated with durable response at the single-cell level. Among significant gene sets, we identified multiple signatures associated with Tregs (vs conventional T cells) and naïve or memory T cells (vs effector) to be significantly associated with a durable response (online supplemental figure S7C,D).

Figure 5. Different scRNA-seq gene sets and genes, independent of phenotypic clusters, associate with efficacy of axi-cel. (A) Stacked bar graphs of CAR+ cells, demonstrating the proportion of cells within each CD4 (left) and CD8 (right) cluster, separated into durable responder (DR) and non-durable responder (NDR) for axi-cel. The number of CAR+ single cells analyzed in each group is indicated in parentheses. (B) Evaluation of Hallmark gene set expression within T-cell clusters between DR and NDR patients for axi-cel. Red=higher in DR and blue=higher in NDR. Cells are independently clustered with Hallmark categories shown vertically on the left. Shown are normalized enrichment scores (NES). (C) Evaluation of Gene Ontology Cellular Component gene-set expression within T-cell clusters between DR and NDR patients for axi-cel. Red=higher in DR and blue=higher in NDR. Shown are NES. (D) Meta-analysis was performed to identify genes showing differential expression between DR and NDR CAR+ cells across clusters identified in scRNA-seq data, for axi-cel. axi-cel, axicabtagene ciloleucel; CAR, chimeric antigen receptor; scRNA-seq, single-cell RNA sequencing.

 CAR+cells were then assessed in an unbiased search for biological pathways that associated with durable response across clusters. For axi-cel, an elevated expression of Hallmark MYC target pathways was associated with a durable response, with the strongest differential expression observed in the CD4 and CD8 activated and proliferative clusters (figure 5B). In addition, the expression of cell cycle pathways “E2F targets” and “G2M checkpoint” was prominently higher in the CD4 and CD8 activated clusters in durable responders. Axi-cel treated patients who failed to obtain a durable response (non-durable responders) had elevated expression of the Hallmark tumor necrosis factor-alpha signaling pathway. Analysis of GO-annotated pathways across multiple clusters in durable responders revealed elevated expression of pathways involved in protein translation, most notably those involved in ribosomal synthesis, protein trafficking, and protein folding (figure 5C). This is in accord with elevated MYC, which is known to directly induce ribosomal synthesis.²⁷

Meta-analysis of gene expression differences between durable responders and non-durable responders was similar to these pathway findings (figure 5D). For axi-cel, high expression of mitochondrial respiratory chain genes (MT-CYB, MT-CO1, MTCO3, MT-ND3, and MT-ND4L) was associated with non-durable response, consistent with a report that the expression of glycolytic genes, as opposed to OXPHOS, is associated with superior axi-cel outcomes.²⁰ A clear outlier was elevated expression of MTRNR2L8, a nuclear-encoded mitochondrial peptide,^{28 29} that was associated with durable response for axi-cel. In addition, multiple ribosomal genes (eg, RPS18, RPS9, RPL3, RPL13) were associated with durable response.

Together, these results demonstrate that pathways associated with MYC, specifically pathways leading to capacity for protein synthesis, are associated with durable responses for axi-cel.

Finally, we evaluated canonical CAR T-cell associated toxicities including cytokine release syndrome (CRS) and ICANS. We did not find any clusters to be associated with severe CRS or ICANS for either axi-cel or tisa-cel (online supplemental figure S8).

Discussion

Here we characterized the standard-of-care commercial CD19 CAR T-cell infusion products axi-cel and tisa-cel from patients with LBCL using scRNA-seq. These analyses revealed axi-cel and tisa-cel have markedly different composition. Axi-cel has a higher proportion of CD4 and CD8 stem and central memory T cells, while tisa-cel has a greater proportion of CD8 effector T cells. Moreover, axi-cel CAR-T cells exhibit increased readiness for activation that is associated with upregulation of genes that direct protein synthesis, antigen presentation, and cellular organelle activity. In contrast, tisa-cel CAR-T cells have greater cellular division and increased capacity for cellular respiration. Intriguingly, the gene expression of manufactured axi-cel appears driven by the CAR transgene, while tisa-cel is affected to a greater extent by manufacturing, with similar expression between the CAR+ and CAR− cells.

Our studies suggest manufacturing is critical to CAR T efficacy. Specifically, the infusion product for CAR T cells consists of both CAR transgene-positive and CAR transgene-negative T cells, and scRNA-seq established that axi-cel consisted of a median 55% CAR+cells, while tisa-cel was only 20% CAR+. This is consistent with the manufacturer data provided for tisa-cel, which uses flow cytometry to distinguish CAR+ and CAR− T cells in the infusion product. For axi-cel, CAR+cells were significantly different from the CAR− cells, with broad differences in pathways and gene expression. By contrast, for tisa-cel, CAR+ and CAR− cells were much more alike. We speculate that manufacturing contributes to this difference, whereby gene expression in tisa-cel is mainly driven by manufacturing processes that affect both CAR+ and CAR− cells, whereas gene expression of axi-cel is driven by transduction of the CAR that affects CAR+cells preferentially. Interestingly, clinical data is now available for rapcabtagene autoleucel (rapca-cel), a product that uses the same CAR construct as tisa-cel but uses a rapid manufacturing platform. The manufactured rapca-cel product has a greater proportion of stem and central memory T cells and fewer effectors compared with tisa-cel, yet supported equivalent CAR T-cell expansion in patients despite rapca-cel being dosed at 1/25th of the number of cells used for tisa-cel.³⁰

While clustering of T-cell subsets identified major differences between axi-cel and tisa-cel, surprisingly, no individual cluster associated with clinical outcome. This is in accord with recent scRNA-seq atlas data for axi-cel, whereby individual clusters did not associate with outcome.²⁰ In contrast, the pivotal ZUMA-1 and phase 3 ZUMA-7 trials found that increased infusion product protein expression of the T-cell memory markers CCR7 and CD45RO, associated with higher rates of event-free survival.^{14 15} Further, others have reported the importance of effector cells with renewal capacity, as quantified by T-cell surface protein markers such as CD27 and CD28,³¹ yet flow cytometry analyses indicated that these subsets were not associated with clinical outcome on the JULIET trial of tisa-cel in LBCL.¹⁹ That said, when performing a specific analysis for these subsets we found signatures of Tregs (vs conventional T cells) and effector cells (vs naïve or memory cells) associated with poorer outcomes. It may be that in general scRNA-seq studies are underpowered to detect unbiased associations between T-cell subsets and clinical outcomes, or that they fail to detect phenotypic differences that are evident to flow cytometry. This is corroborated by our analysis of gene expression signatures across clusters demonstrating that lower Treg signatures and higher central memory signatures associated with durable response, both consistent with findings from others. That said, our cross-cluster analysis meta-analysis suggests that the strongest indicators of CAR T-cell quality are those that affect multiple T-cell subsets and clusters. Specifically, CAR T-cell expression of ribosome and protein translation pathways is associated with superior clinical outcomes for axi-cel. This is consistent with previous studies showing that polyfunctionality, or the capacity to secrete multiple protein mediators such as cytokines, associates with superior outcomes for axi-cel.¹³ Evidently, protein synthesis capacity, possibly downstream of MYC, is a critical marker of CAR T-cell health.

There are several caveats to the findings reported herein. First, while our scRNA-seq cohort consisted of a large sample of “real world” patients, with long follow-up for clinical outcomes, there was a smaller number of patients who received tisa-cel, and the outcomes of these patients were imbalanced, where most tisa-cel patients had a non-durable response. This is consistent with reports of a lower efficacy of tisa-cel than axi-cel,5,9 but fewer tisa-cel patients limit statistical power and the interpretation of tisa-cel efficacy analyses. Second, this work only assessed patient samples up to the point of CAR-T infusion, and others have demonstrated the importance of gene expression and CAR-T subsets in circulation at the time of peak expansion post-infusion.^{16 17} Third, differences in gene expression between CAR+ and CAR− cells cannot be directly extrapolated to differences in protein expression phenotype, the most studied endpoints of CAR infusion products. Despite this, we found that protein expression differences in differentiation and exhaustion markers between CAR+ and CAR− T cells decrease with increasing length of time in culture, creating a compelling argument for this driving differences in axi-cel and tisa-cel. Finally, the exact manufacturing protocols for axi-cel and tisa-cel are proprietary and thus our attempts to recapitulate these are based on existing literature from earlier development work on these products. In addition, we only used gamma-retroviral vectors in our analysis, so we cannot account for the differential impact of lentiviral vector. Therefore, conclusions about differences in manufacturing practice must be interpreted with caution.

A critical question coming from these studies is why some CAR T-cell products have higher protein synthesis capacity than others. Mechanistic studies are needed to understand this, but in the interim, the data clearly support the notion that future CAR construct design and manufacturing pathway optimization should aim at bolstering protein synthesis capacity while limiting in vitro proliferation and differentiation. Further, CAR-transgene positive cells should differ greatly from CAR-transgene negative cells, as loss of this difference may indicate an adverse manufacturing process.

Data availability statement

Data are available in a public, open access repository.