Understanding Mechanisms of Response to CAR T-cell Therapy through Single-Cell Sequencing: Insights and Challenges

While the categorization of CD4⁺ and CD8⁺ T cells are often considered near synonymous with “helper” and cytotoxic T-cell functionality, the genes CD4 and CD8 themselves only directly govern MHC class II and I restriction. This association relies on the typical arrangement whereby MHC II presentation occurs primarily by professional antigen-presenting cells, while antigenic MHC I complexes are found on all nucleated cells. However, it has become clear that CD4⁺ cells can express cytotoxic function toward MHC II–expressing target cells such as lymphoma and myeloma. Indeed, CD4/8 dichotomy becomes even murkier in the realm of CAR T cells, where activation is presumed to occur in a largely MHC-independent manner.

CD4⁺ cytotoxic CAR T cells have been observed across different CAR T-cell therapies in both infusion products (IP) and posttreatment patient samples, albeit at higher frequencies posttreatment, with characteristic expression of perforin (PRF1), granzymes (GZMK, GZMB), GNLY, NKG7, and other cytotoxic markers (2–4, 6). While case numbers are still too small to draw strong associations with response from individual studies, there's some indication CD4⁺ CTL polarization may be detrimental at early timepoints. Deng and colleagues clustered CD4⁺ cells in axicabtagene ciloleucel (axi-cel) infusion products, and noted a subset expressing cytotoxic markers such as GNLY mostly contributed by cells from patients with partial response/progressive disease (3). We also found a similar population, defined by the markers CD4+NKG7, to be present in both axi-cel and tisagenlecleucel (tisa-cel) IPs, significantly increasing in frequency at day 7 posttreatment and most prominent in nonresponding patients (2). Interestingly, in one patient who experienced a relapse, CAR T cells yielded a predominantly CD4⁺ CTL phenotype, yet a second infusion of the exact same product led to a durable remission and yielded CD4⁺ CAR T cells with a T helper instead of cytotoxic phenotype. Thus, it appears that either CD4⁺ CAR T cells can dynamically polarize to the CTL phenotype, or different subsets can expand, depending on yet unknown factors in the host environment. Furthering the intrigue of this population, Melenhorst and colleagues published data from two patients with chronic lymphocytic leukemia who experienced decade-long remissions, and showed that CD4⁺ cytotoxic CAR T cells were the dominant persistent cells at time points 5 to 10 years posttreatment (6). While these cells expressed strong levels of checkpoint molecules (PD-1, TIM-3, LAG-3, TIGIT), the authors performed cococulture with CD19-expressing cells and demonstrated these cells retained antitumor activity and were not exhausted in the classical sense.

The functional role of these cells is yet unclear, but phenotypically they are highly similar to CD8⁺ T cells. They tend to have granzyme profiles favoring GrK over GrB, and in some cases high expression of checkpoint molecules. Of note, more classical helper cells are still identified among CD4⁺ CAR T cells in both the IP and day 7 blood samples, and standard approaches such as flow cytometry may not capture the distinction between a “helper” CD4⁺ CAR T cell and a cytotoxic CD4⁺ CAR T cell. Prior in vivo data have supported that mixed populations of CD4⁺ and CD8⁺ CAR T cells produce superior antitumor responses than either subset individually, which motivated manufacturing cells for some clinical products at a 1:1 ratio (11). Given these deeper characterizations of CD4⁺ phenotypes now available, evaluation of the relative contributions of CD4⁺ CTL and Th subsets may yield additional insights.

The state of differentiation of CD8⁺ CAR T cells has also emerged as an important correlate of CAR T-cell response across multiple studies. T-cell differentiation trajectories are commonly put along a linear spectrum of stages with increasing cytotoxic potential and decreasing proliferative capacity. These are most commonly defined by flow cytometry, as, for example, naïve T cells (CCR7⁺CD62L⁺CD45RA⁺), central memory (CCR7⁺CD45RO⁺), effector memory (CCR7⁻CD45RO⁺), and terminal effector/Temra (CCR7⁻CD45RA⁺), though more comprehensive marker panels may yield more specific cell populations. In contrast, in scRNA-seq data, although these names are still used, a consistent standard for defining these is not applied. Because the commonly used 10X Genomics-based sequencing only capture the 3′ or 5′ end of transcripts, isoform-level expression quantification usually cannot be performed, although we have shown in the case of CD45 that the relevant exons (3 to 5) are sufficiently close to the 5′ end that these can indeed be distinguished (2). Alternatively, authors have applied complementary protein-based labeling of these isoforms to aid in characterization (1). Further complicating characterization, CD45RO and CD45RA tend to be coexpressed on cells in the infusion product, which can lead to differing conventions as to whether these cells get labeled naïve, central memory, stem central memory, or early memory. In addition, central memory/naïve markers CCR7 and CD62L are indicators of T-cell migration and homing, and thus their use as a proxy for differentiation is less clear in ex vivo T cells of infusion products that are not undergoing these processes.

Nonetheless, the overall extent of CD8⁺ cell differentiation and activation does seem to be consistently implicated in clinical response for 4-1BB–based CAR T-cell products. In this setting, 4-1BB IPs are initially dominated by CD4⁺ CAR T cells (usually ∼80%–90% of cells; refs. 1, 2, 7). However, in both studies with sequential time points, CD8⁺ CAR T cells greatly expanded posttreatment, with a significantly larger expansion occurring in responders (1, 2). CD4:CD8 ratios in the pretreatment infusion products have not been associated with response, suggesting the difference may have more to do with expansion potential rather than raw number (2, 7). Supporting this notion in TCR-seq data CD8⁺ CAR T cells have the highest diversity in infusion products, and specific clones expand posttreatment (2, 10). Meanwhile analyses of cell-cycle signatures consistently showed fewer CD8⁺ cells in the cell cycle in nonresponders (1, 2).

Signatures of activation and/or exhaustion seem to be associated with poor response. Jackson and colleagues found by flow cytometry that checkpoint molecules such CTLA4, LAG3, and TIGIT are expressed more strongly in nonresponders in both IPs and posttreatment CD8⁺ CAR T cells (1). We found through pseudobulk differential expression that IPs from nonresponders had higher expression of IFNG, LAG3, and cytotoxic genes perforin (PRF1) and granzyme H (GZMH; ref. 2). A similar IFN signature including IFNG and STAT1 was observed in data from premanufacture 4-1BB CAR T cells by Chen and colleagues (12). In contrast, analysis of bulk RNA sequencing of infusion products used to treat chronic lymphocytic leukemia (CLL) found IPs from responders enriched with memory gene signatures (13), and Haradhvala and colleagues found that responding patients had higher frequencies of cells in a cluster expressing memory genes CCR7 and LEF1 (2). Bai and colleagues used CITE-seq data to show CAR T cells from patients who experienced CD19-positive relapse had more CAR T cells with effector/effector memory and fewer with central/stem cell–like memory phenotypes (7). Dhodapkar and colleagues studied bone marrow from myeloma patients treated with BCMA CARs, also using a 4-1BB costimulatory domain, and found a similar relationship where longer progression-free survival was associated with memory rather than cytotoxic gene expression in posttreatment CD8⁺ CAR T cells (9).

CD28z products seem to follow starkly different expansion dynamics of the CD8⁺ compartment. Studies of axi-cel IPs often have majority CD8⁺ fractions, in contrast to their 4-1BB counterparts (2, 3, 5), but posttreatment frequencies of CD8⁺ CAR T cells remained fairly stable between infusion products and posttreatment samples (2), and posttreatment CD8:CD4 ratios do not appear to be associated with clinical outcomes (2, 4). Activation and/or exhaustion phenotypes have also been suggested to play a role in response: Deng and colleagues noted that cells from patients with clinical response at month 3 preferentially fall into a cluster with memory markers (CCR7, LEF1, SELL, CD27), with depletion in a cluster with highly differentiated cytotoxic and exhausted markers (GZMA, GZMA, LAG3; ref. 3). However, this result did not reach significance after collection of a larger cohort (5) and was not observed in independent datasets (2), and thus the trends for CD28z do not appear as clear as with 4-1BB products.

T regulatory (Treg) cells are an immune-inhibitory population which in normal physiology guards against autoimmunity, but in cancer are thought to play an important role in tumor immune suppression and evasion. As CAR T cells in current commercial manufacturing processes do not perform selection of specific T-cell populations, there is potential for CAR transduction in Tregs to inadvertently create a potent tumor-directed immunosuppressive population. This has indeed been observed to occur in all three axi-cel IP datasets, with this population being characterized by expression of FOXP3, IKZF2 (Helios), and CTLA4 (2, 4, 5). IL2RA (CD25) is also expressed, but is not a reliable marker in infusion products due to its presence on other activated CAR T cells.

Associations between the frequencies of these cells and response were seen by Haradhvala and colleagues in axi-cel infusion products, and by Good and colleagues in post–axi-cel treatment blood (2, 4). Curiously, the generation of these cells appeared to be product-specific, occupying approximately 5% of all CAR T cells for axi-cel, but near absent in tisa-cel. The reason for this difference is not known, although we speculate that this difference is due to the different processing of the apheresis product between the two manufacturing processes, which can include proprietary or product-specific manufacturing considerations such as cell sample handling protocols, as axi-cel is shipped as a fresh product while tisa-cel is cryopreserved before shipping (and potentially destabilizing the Treg cells preferentially), or other manufacturer-specific cell culture conditions or reagents. In vivo experiments in mice demonstrated that addition of 5% CAR Tregs of either construct was associated with increased frequency of late relapse, suggesting that the signaling domain of 4-1BB was not responsible for abrogation of Treg suppressive function. Curiously, neither study reported an association between baseline blood Treg levels and response. Thus, it remains unclear what factors, if any, predispose particular patients to CAR-Treg generation during manufacturing.

Given this data, depletion of Tregs from axi-cel products would seem a promising follow-up approach. However, in the analysis of Good and colleagues, more severe neurotoxicity was negatively correlated with CAR-Treg burden (4). Thus, it will be important to evaluate whether CAR-Tregs are also protective from severe toxicities, and ascertain whether methods to remove them during the manufacturing process may lead to worse adverse events.

Another insight into the long-term functioning of CARs has come from studies applying scRNA-seq to CAR T cells still circulating in blood years after treatment for patients remaining in durable remission. Melenhorst and colleagues performed scRNA-seq on CAR T cells from two patients with CLL treated with 4-1BB CARs spanning a decade of follow-up (6), while Anderson and colleagues studied 10 children with ALL with follow-up times ranging 3 to 5 years (8). At these late timepoints, CARs can be very infrequent (ranging ∼0.05%–1%) but still detectable. In both studies, the authors reported cases where the CAR T cells at month-to-year time scales became dominated by CD4/CD8 double-negative (DN) cells. In the study by Melenhorst and colleagues, this effect was transient, with DN cells present by 2.5 months and being replaced by CD4⁺ cytotoxic cells after a few years. In Anderson and colleagues, DN cells were not seen to disappear; however, follow-up times were significantly shorter. In Melenhorst and colleagues, scRNA-seq profiling revealed these T cells to have a γδ TCR, a subset known to be able to take on a DN phenotype and become activated in an MHC-independent manner. In Anderson and colleagues, however, DN CAR T cells were observed in patients with aβ TCR expression as well. Classification of DN cells by single-cell sequencing alone can be a challenge, particularly when detecting CD4 which is expressed nearly an order of magnitude lower than CD8A and CD8B in their corresponding subsets. However, orthogonal flow cytometry or CITE-seq further supported these observations.

The function of these cells is yet to be fully elucidated, whether they play an active role in maintaining a long-term remission, or are vestiges of a finished treatment that has already eliminated the tumor. These cells possess a cytotoxic phenotype (expressing GZMK, GZMA), showed expression of some transcription factors related to exhaustion (TOX, BATF), but at the same time retained populations of proliferating cells.

DN aβ cells are not prevalent in healthy individuals. Given the extremely low frequency of CAR T cells at late points, it is possible these persisting cells represent a population generated in normal memory responses but their enrichment is only made possible in this instance by sorting for the CAR. Another possibility is that the previously described CD4⁺ CTL subset and DN cells are one and the same, and CD4 expression becomes loosely regulated in CAR T cells due to the presence of TCR-independent activation. TCR-based tracking of CAR T-cell clones may shed light on whether these cells may shift between subtypes; however, clones identified at late and early time points are largely nonoverlapping in the existing data (8).

scRNA-seq has shown itself to be a highly useful tool to study CAR T-cell phenotypes, and these common themes have emerged across multiple studies. Nonetheless, the complexity of the data presents unique challenges that make its rigorous statistical treatment and interpretation difficult, particularly when communicating results to a broader audience. In particular:

A very common approach to testing associations between gene expression and a variable of interest, such as clinical response, is to pool together all cells from responders and nonresponders, and then test for a difference between these two groups using a t test or Wilcoxon rank sum test. However, this violates the independence assumption of these tests, and can lead to astronomically significant P values for nonexistent associations supported by few patients but many cells. Intuitively, the issue is that this tests whether the two populations of cells are different (which we trivially know they are, as they come from different patients), instead of whether two populations of patients (e.g., responders and nonresponders) are different.

Solutions to performing these analyses at the cell level are nontrivial. Mixed linear models are the standard fix for dependent observations, but generally either perform normality assumptions of the data that are generally not appropriate due the sparsity of scRNA data, or require more complicated inference techniques that have slower run times and require a high degree of familiarity to properly diagnose. One simpler approach is the use of “pseudobulk” analyses, in which the single-cell nature of the data is leveraged to sort cells into subsets but then aggregate the data into patient-level measurements, most commonly by summing the transcript counts across cells. This removes both the issues of sparsity and within-sample correlations, helping obtain P values that better portray uncertainty due to the sample size of patients and not cells (14).

Single-cell datasets often have tens of cell types, each of which can often be further subdivided into arbitrarily many subtypes. For each of these, one has many different analyses they could consider for their variable of interest (e.g., testing differences in cell type proportions, testing differences in gene expression, testing gene signatures). Given the small samples sizes of scRNA studies, running all possible tests and computing a multiple hypothesis correction would completely ablate the statistical power of the study. Thus, in practice, the decision of which hypotheses to test is inevitably swayed by exploratory analyses of which cell types have the most interesting signal related to the variable of interest. This will, however, increase the false discovery rates of the tests in a way that can be difficult to describe in a multiple hypothesis testing framework. As such, single-cell data is best used for hypothesis generation, and followed up with validation experiments to more narrowly and rigorously test the finding.

In closing, we present a summary of findings recurring from several different recent publications that have reported associations based on scRNA-seq analyses of CAR T cells in patients. These suggest actionable modifications to CAR T-cell manufacturing that may have promise for improved function including rapid manufacturing to limit ex-vivo differentiation (15), selection of less differentiated CAR T-cell subsets (11), and depletion of CAR Tregs. Further application of scRNA to more products and contexts will likely continue to lead insights; for example, sequencing of the site of the tumor as done for multiple myeloma (9), and application of spatial technologies, will likely yield additional function insights not apparent from studying circulating CAR T cells.

We believe this kind of data that emerges is complementary to traditional flow cytometry. However, although the field has extensive experience with flow cytometry protocols and analysis, the computational complexity of scRNA-seq interpretation requires particular attention to some common pitfalls in analysis and reporting, and a strong emphasis on reproducibility and validation is as important as ever.

N.J. Haradhvala reports personal fees from MorphoSys outside the submitted work. M.V. Maus is an inventor on patents related to adoptive cell therapies, held by Massachusetts General Hospital (some licensed to Promab) and University of Pennsylvania (some licensed to Novartis), has received grant/research support from Kite Pharma and Moderna, has served as a consultant for multiple companies involved in cell therapies, holds Equity in 2SeventyBio, A2Bio, Cargo, Century Therapeutics, Neximmune, Oncternal, and TCR2, is on the Board of Directors for 2Seventy Bio, and is/has been a consultant for A2Bio, Adaptimmune, Agenus/Mink Therapeutics, Allogene, Arcellx, Astellas, AstraZeneca, Atara, Bayer, BendBio, BMS, Cabaletta Bio (SAB), Cargo, Cellectis (SAB), CRISPR therapeutics, In8bio (SAB), Intellia, GSK, Kite Pharma, Neximmune, Novartis, Oncternal, Sanofi, Sobi, Synthekine, TCR2 (SAB), and Tmunity.