Abstract Advancements in chimeric antigen receptor (CAR) T-cell therapy for treating diffuse large B-cell lymphoma (DLBCL) have been limited by an incomplete understanding of CAR T-cell differentiation in patients. Here, we show via single-cell, multi-modal, and longitudinal analyses, that CD8^+ CAR T cells from DLBCL patients successfully treated with axicabtagene ciloleucel undergo two distinct waves of clonal expansion in vivo. The first wave is dominated by an exhausted-like effector memory phenotype during peak expansion (day 8–14). The second wave is dominated by a terminal effector phenotype during the post-peak persistence period (day 21–28). Importantly, the two waves have distinct ontogeny from the infusion product and are biologically uncoupled. Precursors of the first wave exhibit more effector-like signatures, whereas precursors of the second wave exhibit more stem-like signatures. We demonstrate that CAR T-cell expansion and persistence are mediated by clonally, phenotypically, and ontogenically distinct CAR T-cell populations that serve complementary clinical purposes. Subject terms: Immunotherapy, B-cell lymphoma, CD8-positive T cells, RNA sequencing __________________________________________________________________ Diffuse large B-cell lymphoma (DLBCL) is a heterogeneous disease group with CAR T cells offering therapeutic success in otherwise hard-to-treat cases. Here, authors study the in vivo expansion and persistence of CAR T cells in the peripheral blood of successfully treated DLBCL patients, demonstrating that two different CD8+ precursor phenotypes in the initial cell product give rise to two independent waves of clonally expanded CAR T cells with distinct phenotypes in peripheral blood. Introduction Diffuse large B-cell lymphoma (DLBCL), the most common non-Hodgkin’s lymphoma in the United States, is characterized by diffusely proliferating and malignant B cells at nodal or extranodal sites^[45]1. Although up-front chemoimmunotherapy is often curative, patients with DLBCL that is refractory to up-front treatment or relapse following remission (r/r DLBCL) have limited treatment options and poor outcomes^[46]2,[47]3. Effective treatment options were limited until the United States Food and Drug Administration approved autologous CD19-directed chimeric antigen receptor (CAR) T-cell therapies for r/r DLBCL in 2017. Autologous CD19-directed CAR T-cell therapy involves virally transducing a patient’s T cells ex vivo with a CD19-directed CAR—an engineered receptor consisting of an extracellular anti-CD19 single-chain variable fragment, a hinge/transmembrane region, and intracellular costimulatory (CD28 or 4-1BB) and activation (CD3ζ) domains. CAR-transduced T cells (i.e., CAR T cells) are cultured and infused into the patient, where they lyse CD19^+ lymphoma cells^[48]4. CD19-directed CAR T-cell therapy has achieved complete response rates of 40–54%^[49]5–[50]8 for r/r DLBCL, but non-response rates^[51]6,[52]9 and treatment toxicities^[53]10 remain as challenges^[54]11. Development of CAR T-cell formulations with higher response rates and fewer toxicities requires a thorough understanding of how CAR T cells differentiate in patients with r/r DLBCL. The pioneering ZUMA-1 trial ([55]NCT02348216) for axicabtagene ciloleucel (autologous CD28-costimulated CAR T cells) demonstrated that peripheral blood CAR T cells expand, contract, and sometimes persist.^[56]5 Greater expansion predicted higher response rates, but also higher likelihood of developing treatment toxicities^[57]5,[58]12. Longer persistence predicted durable remission and long-term immunosurveillance with leukemias, but its relevance for preventing DLBCL relapse remains obscure^[59]9,[60]13. Although many factors are associated with expansion and persistence (including CAR design^[61]14,[62]15 and CAR T-cell phenotypes^[63]16–[64]20, among others^[65]21–[66]23), how and why CAR T cells differentiate into expansive or persistent phenotypes in vivo is still an open question. Addressing this question requires a single-cell approach that integrates CAR T-cell phenotypes and clonal kinetics over longitudinal timepoints. However, existing studies have either focused on CAR T-cell phenotypes^[67]19,[68]24,[69]25 or clonal kinetics^[70]26 without integrating both data modalities, or lacked the temporal resolution required to decipher cell fates over longer periods^[71]12,[72]27. Consequently, a complete and longitudinal understanding of CAR T-cell differentiation in vivo has remained elusive. To comprehensively elucidate CAR T-cell differentiation in vivo, we perform single-cell, multi-modal (paired RNA-seq/CITE-seq/TCR-seq), and longitudinal analyses of CD28-costimulated CAR T cells from infusion product and peripheral blood of seven patients with r/r DLBCL who were complete responders under treatment with axicabtagene ciloleucel. Peripheral blood CAR T cells are sorted using CD19 antigen-tetramers^[73]28. Importantly, we report that the CD8^+ CAR T cells observed during the peak expansion period (day 8–14) have distinct clonotypic repertoires compared to the later CD8^+ CAR T cells observed during the post-peak persistence period (day 21–28). We further analyze how these two CAR T-cell populations differ with respect to phenotypes, transcriptional profiles, regulatory networks, and infusion product precursors. Our findings not only offer a finer understanding of CAR T-cell biology in vivo, but also inform efforts to develop CAR T cells with improved expansion and persistence. Results Study design and clinical findings To interrogate CD28-costimulated CAR T-cell differentiation in vivo, we longitudinally interrogated the phenotypes and clonal dynamics of CAR T cells from seven patients (P1-7) who achieved complete responses under treatment with axicabtagene ciloleucel (Fig. [74]1a). Patients were diagnosed with r/r DLBCL and treated at the University of Chicago Medicine between 2019 and 2021 (Table [75]1). Clinical response was determined by positron emission tomography/computed tomography imaging 30 days after infusion product administration. Complete response was defined as no detectable lymphoma (Fig. [76]S1). To capture phenotypic heterogeneity and clonal dynamics longitudinally, we performed single-cell multi-modal analyses (paired RNA-seq/CITE-seq/TCR-seq via the 10x Genomics platform) on infusion product and peripheral blood biospecimens at three timepoints: peak expansion (T[exp], day 8–14), early post-peak persistence (T[per1], day 21), and late post-peak persistence (T[per2], day 28) (Table [77]2). The three peripheral blood timepoints were available for all patients, except P3 (only T[exp] and T[per1]). CAR T cells were sorted from peripheral blood using CD19 antigen-tetramers (representative staining in Fig. [78]1b, Fig. [79]S2)^[80]28. Compared to CD19 antigen-tetramer-negative T cells, CD19 antigen-tetramer-positive T cells specifically expressed the CAR transgene, which validates our sorting strategy (Fig. [81]1c). Fig. 1. CD8^+ CAR T cells from complete responders undergo a clonotypic shift in vivo. [82]Fig. 1 [83]Open in a new tab a Schematic depicting sorting strategy, data generation, and single-cell multi-modal analysis of CAR T cells in peripheral blood from seven CAR T-cell therapy patients (P1-7) who exhibited complete responses with axicabtagene ciloleucel. Created in BioRender. Hu, Y. (2025) [84]https://BioRender.com/hqo12oz. b Representative flow plots depicting anti-CD3 and CD19 antigen-tetramer staining of peripheral blood mononuclear cells from P3 versus a healthy donor. CD3^+Tet^+ (“CAR”) and CD3^+Tet^− (“ENDO”) patient cells were sorted from the indicated gates. c Violin plots depicting normalized CAR transgene mRNA expression of sorted CAR and ENDO T cells, split by CD4^+ (left) and CD8^+ (right) subsets. Expression levels were compared by two-sided Wilcoxon Rank-Sum test with Bonferroni correction, whereby **** indicates p < 0.0001. d Line plots depicting expansion and contraction of peripheral blood CAR abundance over the course of therapy. e, f Bar graphs and heatmaps depicting overlap coefficients for TCR clonotypes comparing CAR (left) and ENDO (right) repertoires between T[exp], T[per1], and T[per2]. Overlap coefficients were compared by two-sided t test, whereby * indicates p < 0.05 and ns indicates not significant. Each dot represents a measurement from a single patient (n = 7). Data are presented as mean values ± SEM. Exp expansion stage, Per persistence stage. Source data are provided as a Source Data file. Table 1. Patient characteristics and relevant clinical data Characteristics P1 P2 P3 P4 P5 P6 P7 Age Decade 70s 60s 50s 50s 40s 60s 60s Sex Male Male Female Female Male Male Female Disease stage II IV II I IV IV IV No. of prior therapies 2 5 2 4 5 3 3 Prior lines of therapy R-CHOP R-CHOP DA-EPOCH w/ IT MTX Radiation R-CHOP R-CHOP R-CHOP R-ICE Benda-obinutuzumab ICE R-CHOP R-ICE R-ICE R-ICE R-ICE R-ICE BEAM + ASCT BEAM + ASCT BEAM + ASCT Hu5F9-G4 + rituxumab BEAM + ASCT FCR + AlloSCT R-GEMOX R-GEMOX Disease status Relapsed Relapsed, then refractory Primary refractory Relapsed Relapsed Relapsed Relapsed Baseline (Ref. range) LDH, U/L (116-245 U/L) 165 311 312 186 203 1276 297 CRP, mg/dL ( < 0.5 mg/dL) 0 ND 3.5 0.6 0.3 4.1 1.4 Ferritin, ng/mL (20-300 ng/mL) 110 ND 576 215 544 1228 229 ECOG PS 1 1 2 0 1 1 0 Bridging therapy No No Yes No Yes Yes No Maximum Grade CRS 2 1 2 1 1 1 1 Maximum Grade ICANS 2 3 3 4 1 1 0 Tocilizumab or steroids Both Both Both Steroids Tocilizumab Tocilizumab Tocilizumab Survival status Deceased Alive Deceased Alive Alive Alive Alive Cause of Death Second Malignancy --- Progressive lymphoma --- --- --- --- [85]Open in a new tab ASCT autologous stem cell transplant, AlloSCT allogeneic stem cell transplant, BEAM BCNU (carmustine), etoposide, cytarabine, melphalan, CRP C-reactive protein, CRS cytokine release syndrome, DA-EPOCH dose-adjusted etoposide, prednisone, oncovin, cyclophosphamide, hydroxydaunorubicin, ECOG PS Eastern Cooperative Oncology Group Performance Score, FCR fludarabine, cyclophosphamide, rituximab, Hu5F9-G4 magrolimab, ICANS immune effector cell-associated neurotoxicity syndrome, ICE ifosfamide, carboplatin, etoposide, IT MTX intrathecal methotrexate, LDH lactate dehydrogenase, ND not determined, R-CHOP rituximab, cyclophosphamide, hydroxydaunorubicin, oncovin, prednisone, R-GEMOX rituximab, gemcitabine, oxaliplatin, R-ICE rituximab, ifosfamide, carboplatin, etoposide. Table 2. Number of sequenced CD4^+ and CD8^+ CAR T cells for each sample Patient Timepoint # of CD4^+ CAR T cells # of CD8^+ CAR T cells P1 IP 1021 2375 P1 T[exp] 13 7 P1 T[per1] 431 3347 P1 T[per2] 166 1009 P2 IP 1734 1716 P2 T[exp] 789 3383 P2 T[per1] 542 2227 P2 T[per2] 567 5370 P3 IP 754 3468 P3 T[exp] 51 172 P3 T[per1] 331 1597 P4 IP 1214 1081 P4 T[exp] 1299 977 P4 T[per1] 695 3633 P4 T[per2] 224 1144 P5 IP 284 3656 P5 T[exp] 205 1030 P5 T[per1] 878 958 P5 T[per2] 395 431 P6 T[exp] 17 1041 P6 T[per1] 126 4602 P6 T[per2] 144 3805 P7 IP 1810 1360 P7 T[exp] 1402 3043 P7 T[per1] 53 80 P7 T[per2] 100 169 [86]Open in a new tab CD8^+ CAR T cells undergo a clonotypic shift between T[exp] and T[per] To analyze CAR T-cell population dynamics, we tracked CAR abundance (% CAR^+ of CD3^+ T cells) in peripheral blood throughout the course of therapy. CAR abundance at peak expansion ranged from 11% to 73% (Fig. [87]1d). Peak expansion occurred at day 8–14, which is consistent with prior clinical findings^[88]5. For each timepoint, we quantified proportions of CD8^+ and CD4^+ T cells within the total CAR T-cell population by single-cell RNA-seq and CITE-seq. CAR T cells were predominantly CD8^+ across most patients and timepoints (Fig. [89]S3). We next analyzed the dynamics of the CAR T-cell clonal repertoire across longitudinal timepoints (T[exp] at day 8–14, T[per1] at day 21, T[per2] at day 28) using single-cell TCR-seq. Clone sizes were similar across timepoints (Fig. [90]S4a). Clonotypes did not overlap between patients (Fig. [91]S4b). Repertoire overlap analysis indicated that T[exp] clonotypes were significantly distinct from T[per1] or T[per2] clonotypes (Fig. [92]1e, f, left). In sharp contrast, T[per1] and T[per2] clonotypes overlapped substantially more. As a control, we also analyzed endogenous (non-CAR) T cells from matched timepoints. Unlike with CAR T cells, endogenous T cells did not show distinctive clonotypic patterns (Fig. [93]1e, f, right), indicating that the shift in clonotypes between T[exp] and T[per] is CAR-specific. This CAR-specific clonotypic shift was consistent across patients (Fig. [94]S4b). Moreover, the distinction between T[exp] and T[per] clonotypes is driven by CD8^+ T cells, and not by CD4^+ T cells (Fig. [95]S4c). Collectively, these findings indicate that CD8^+ CAR T cells undergo a clonotypic shift between T[exp] and T[per]. CD8^+ CAR T cells undergo a phenotypic shift from exhausted-like effector memory to terminal effector Having shown a shift in CD8^+ CAR T-cell clonotypes between T[exp] and T[per], we hypothesized that unique CD8^+ CAR T-cell phenotypes dominate T[exp] and T[per]. To test this hypothesis, we filtered CD8^+ CAR T cells for Uniform Manifold Approximation and Projection (UMAP) and identified six T-cell clusters (Fig. [96]2a) based on gene and protein markers (Fig. [97]2b, expanded marker set in Fig. [98]S5a). No cluster was patient-specific (Fig. [99]S5b). All clusters expressed CAR transgene and CD8α, validating our sorting and filtering processes, respectively. All clusters expressed CXCR3, a chemokine receptor that demarks activated T cells. Fig. 2. Phenotypic heterogeneity of peripheral blood CD8^+ and CD4^+ CAR T cells. [100]Fig. 2 [101]Open in a new tab UMAPs depicting single-cell transcriptomes of CD8^+ (a) and CD4^+ (e) CAR T cells colored by cell cluster. Inset depicts distribution of transcriptomes across timepoints. Violin plots depicting normalized expression levels of key genes and proteins for annotating and phenotyping CD8^+ (b) and CD4^+ (f) CAR T cells. For extended versions, see Fig. [102]S5a and [103]S6a. Stacked bar graphs depicting proportions of each CD8^+ (c) and CD4^+ (g) CAR T-cell phenotype at different timepoints. Boxplots depicting proportion of CD8^+ (d) and CD4^+ (h) CAR T cells of a given phenotype at different timepoints. Each dot represents a measurement from a single patient (n = 7). The central line indicates the median. The bounds of the box indicate the 25th–75th percentile. The length of the whiskers indicates 1.5 times the inter-quartile range from the first and third quartiles. Proportions are compared between timepoints by two-sided t test, whereby *** indicates p < 0.001, ** indicates p < 0.01, * indicates p < 0.05, and ns indicates not significant. CM central memory, EM effector memory, TE terminal effector, Mem memory, ISG interferon stimulated genes, Th1 type 1 helper-like. Source data are provided as a Source Data file. Examination of gene and protein markers revealed that one of the six clusters represented proliferating T cells (MKI67^+TOP2A^+). The remaining five non-proliferating clusters were annotated as central memory (CM, TCF7^+TBX21^–), effector memory (EM, TCF7^+TBX21^+), or terminal effector (TE, TCF7^-TBX21^+) T cells. The TCF7^+TBX21^– CM cluster exhibited markers of stemness (IL7R, high CD127) and minimal markers of effectorness (GZMB, CX3CR1). The two TCF7^+TBX21^+ EM clusters occupied the lower half of the UMAP and uniquely expressed CXCR6, a chemokine receptor that facilitates trafficking into solid tumors^[104]29. One of these EM clusters upregulated markers consistent with early exhaustion (NR4A2, TOX, GZMK, low TIM-3), hence it was designated “exhausted-like EM”. Lastly, the two TCF7^-TBX21^+ TE clusters occupied the upper half of the UMAP and uniquely downregulated GZMK. One of these TE clusters upregulated markers consistent with late exhaustion (TOX, PDCD1, high TIM-3), hence it was designated “exhausted-like TE”. The other TE cluster was highly clonal (some clone sizes >100), suggesting expansion through proliferation (Fig. [105]S7a). Overall, most CD8^+ CAR T cells were TBX21^+ EM or TE, which is consistent with the established link between CD28 costimulation and effector memory (rather than central memory) differentiation^[106]14,[107]15. Phenotypic compositions of CD8^+ CAR T cells at T[exp] and T[per] were compared. CAR T cells at T[exp] were predominantly exhausted-like EM (64%) whereas CAR T cells at T[per] were predominantly TE (63% for T[per1], 77% for T[per2]) (Fig. [108]2c). These findings were statistically significant and consistent across all seven patients (Fig. [109]2d, Fig. [110]S5c). Moreover, the large clone sizes within the T[per]-specific TE cluster (Fig. [111]S7a) suggest active TE proliferation at T[per]. CAR T cells at T[per1] were enriched for EM. From T[per1] to T[per2], the EM proportion decreased while the TE proportion increased, suggesting progressive differentiation from EM to TE. However, the overall phenotypic compositions at T[per1] and T[per2] were more similar than different, which is concordant with findings from repertoire overlap analysis (Fig. [112]1e, f). Moreover, we observed decreasing proliferating proportions and increasing CM proportions over time (Fig. [113]2d), though this was not always statistically significant. These changing proportions may suggest some CAR T cells were returning from an activated to a resting phenotype. In conclusion, CD8^+ CAR T cells phenotypically shifted from exhausted-like EM to TE between T[exp] and T[per]. CD4^+ CAR T cells maintain a memory phenotype with CAR Treg persistence We next performed clustering, annotation, and longitudinal analyses of CD4^+ CAR T cells to investigate phenotypes at T[exp] and T[per]. We identified five T-cell clusters, all expressing CAR transgene and CD4, based on gene and protein markers (Fig. [114]2e, f; expanded marker set in Fig. [115]S6a). No cluster was patient-specific (Fig. [116]S6b). All clusters expressed CXCR3, indicating T-cell activation and type 1 helper polarization. One of the five clusters represented proliferating T cells (MKI67^+TOP2A^+). The remaining four non-proliferative clusters were annotated as memory (Mem, TBX21^–FOXP3^–), type 1 helper (Th1, TBX21^+FOXP3^–), or regulatory (Treg, TBX21^–FOXP3^+) T cells. The two TBX21^–FOXP3^– Mem clusters at the center of the UMAP comprised most of the cells. One of the Mem clusters upregulated IRF7 and interferon-stimulated genes (MX1, OAS1, ISG15), indicating response to type I interferon signaling. This signature suggests dynamic interferon secretion in vivo and is in concordance with type I interferon’s role in memory CD4^+ T-cell differentiation^[117]30. The TBX21^+FOXP3^– Th1 cluster along the lower half of the UMAP upregulated cytolytic genes (PRF1, GNLY, GZMK, NKG7) and tissue-homing chemokine receptors (CX3CR1, CXCR6), resembling the cytotoxic CD4^+ CAR T cells described by Melenhorst et al.^[118]19. Lastly, the TBX21^–FOXP3^+ Treg cluster expressed classic Treg markers (IL2RA, CD25, low IL7R, low CD127) and upregulated IKZF2, indicating a suppressive phenotype^[119]31. CAR Tregs inhibit conventional CAR T cells^[120]27,[121]32 and portend progressive disease^[122]12,[123]27. Phenotypic compositions of CD4^+ CAR T cells at T[exp] and T[per] were compared. CD4^+ CAR T cells predominantly exhibited a Mem phenotype (49–82%) across all timepoints (Fig. [124]2g). This observation was consistent across all patients (Fig. [125]S6c). Between T[exp] and T[per1], the Mem proportion significantly decreased while the Th1 proportion significantly increased, indicating that CD4^+ CAR T cells may be increasingly polarized during the early contraction period (Fig. [126]2h). Moreover, across all patients, the Treg proportion steadily increased from T[exp] (5%) to T[per1] (11%) to T[per2] (21%), indicating that CAR Tregs were maintained in peripheral blood after peak expansion. The CAR Treg cluster was predominantly non-clonal (Fig. [127]S7b), suggesting CAR Tregs were maintained through persistence rather than proliferation. While previous studies have only investigated CAR Tregs at day 7^[128]12,[129]27, this current study affirms and extends the presence of CAR Tregs until at least day 28. Importantly, persistence of CAR Tregs points towards their possible involvement in decreasing acute inflammation and restoring immune homeostasis after peak expansion. In conclusion, we discovered that CD4^+ CAR T cells did not exhibit an abrupt T[exp]-to-T[per] phenotypic shift. Rather, they consistently exhibited a memory phenotype with CAR Tregs persisting over the course of therapy. Integration of clonotypic and phenotypic shifts in CD8^+ CAR T cells supports a two-stage differentiation model Having observed shifts in both CD8^+ CAR T-cell clonotypes and phenotypes between T[exp] and T[per], we hypothesized the existence of two distinct waves of in vivo clonal expansion. To test this hypothesis, we linked the clonotype and phenotype of single CD8^+ CAR T cells in our dataset using their cell barcodes as unique indices. After this linking process, we confirmed that the proportions of T cells with exhausted-like EM (EM-exh) and terminal effector (TE) phenotypes (Fig. [130]3a) are consistent with prior results (Fig. [131]2c). A phenotype label was subsequently assigned to each CD8^+ CAR T-cell clone based on its predominant phenotype at each timepoint. For each patient, we tracked the total abundances of clones sharing a common phenotype label across timepoints. Notably, the abundance of EM-exh clones at T[exp] was significantly reduced at T[per1] and T[per2] (Fig. [132]3b, left), while the abundance of TE clones at T[per1] and T[per2] was significantly reduced at T[exp] (Fig. [133]3b, middle and right). In addition to redemonstrating the shifts in CD8^+ CAR T-cell clonotypes and phenotypes, these findings support the existence of two distinct waves of in vivo clonal expansion. (Fig. [134]3c). Fig. 3. Integration of clonotypic and phenotypic shifts supports a two-stage differentiation model. [135]Fig. 3 [136]Open in a new tab All figure panels are based on the clonotype-phenotype linked dataset (n = 32432 cells). a Proportion of CAR T cells with an exhausted-like effector memory (EM-exh), terminal effector (TE), or other phenotypes at each timepoint. b Total abundance of all clones that were predominantly EM-exh at T[exp] (left), TE at T[per1] (middle), or TE at T[per2] (right), measured across timepoints. A clone’s predominant phenotype was defined as the phenotype with the greatest representation. Each point represents clones from a patient (n = 6). Abundance across timepoints were compared by repeated measures ANOVA with paired two-sided post hoc t tests, whereby **** indicates p < 0.0001, *** indicates p < 0.001, ** indicates p < 0.01, * indicates p < 0.05, and ns indicates not significant. c Overlay of clonal abundance dynamics with 95% confidence intervals for clones annotated as EM-exh from T[exp] (red) or TE from T[per1]/T[per2] (blue). d Pie chart depicting proportion of the top 500 clones classified as Wave 1, Wave 2, or Other. e Heatmap depicting normalized CAR abundance across timepoints for the top 500 largest clones. f Clonal dynamics and phenotypic distribution for the top 500 clones, grouped into Wave 1 and Wave 2. Left panels depict the change in clonal abundance across timepoints, while right panels show the predominant phenotype distribution at each timepoint. g Cartoon summarizing the two-stage model for CAR T-cell differentiation. Bulk CAR T-cell expansion and contraction (black line) masks the dynamics of Wave 1 (EM-exh, expansion phase timeframe, red) and Wave 2 (TE, persistence phase timeframe, blue) clones. Created in BioRender. Hu, Y. (2025) [137]https://BioRender.com/rwp4v6g. Source data are provided as a Source Data file. Next, we evaluated the explanatory power of our hypothesis for two distinct clonal waves. To minimize noise, we filtered for the top 500 largest clones (mean size 10.8, minimum size 4). Resulting clones were annotated with the following three definitions, based on our hypothesis for two waves: “Wave 1” (phenotypically EM-exh with highest abundance at T[exp]), “Wave 2” (phenotypically TE with highest abundance at T[per1]/T[per2]), or “Other” (not Wave 1 or Wave 2). Wave 1 and Wave 2 clones together accounted for a significant majority (76%) of the top 500 clones (Fig. [138]3d). In accordance with our definitions, Wave 1 clones were maximally abundant at T[exp], while Wave 2 clones were maximally abundant at T[per] (Fig. [139]3e). Wave 1 clones predominantly exhibited an EM-exh phenotype at T[exp] and sharply decreased in abundance at T[per] (Fig. [140]3f, left). In contrast, Wave 2 clones predominantly exhibited a TE phenotype at T[per] and sharply decreased in abundance at T[exp] (Fig. [141]3f, right). To investigate whether these findings are dependent on clone size, we expanded our analysis to the top 3000 largest clones (mean size 3.7, minimum size 2). Under this filter, Wave 1 and Wave 2 clones together continued to account for a significant majority (79%) of the top 3000 clones (Fig. [142]S8a) and reaffirm the expected clonal and phenotypic dynamics (Fig. [143]S8b, c). Therefore, our hypothesis for two distinct waves of in vivo clonal expansion has high explanatory power and is robust to clone size. In conclusion, the two distinct waves of in vivo clonal expansion strongly substantiate a two-stage CD8^+ CAR T-cell differentiation model (Fig. [144]3g). Under this model, some CAR T cells (Wave 1) with an exhausted-like EM phenotype expand earlier to dominate the peak expansion timeframe (T[exp]), while other CAR T cells (Wave 2) with a TE phenotype expand later to dominate the post-peak persistence timeframe (T[per]). Of note, the two-stage differentiation model uncouples CD8^+ CAR T cells from peak expansion and post-peak persistence by designating these two waves as separate lineages. Moreover, these findings provide evidence against the intuitive idea that the post-peak contraction in CAR abundance is solely apoptosis or extravasation of short-lived CAR T cells from peak expansion. Rather, even as total CAR abundance contracts after peak expansion, a distinct subset of CAR T-cell clones simultaneously expands to eventually dominate the peripheral blood CAR T-cell repertoire. T[exp]- and T[per]-specific transcriptional signatures and regulatory networks We set out to identify the molecular determinants underlying CD8^+ CAR T cells at T[exp] and T[per] using gene set enrichment analysis (GSEA), differential gene expression analysis (DGEA), and regulatory network analysis. For internal validation, the average expression of timepoint-specific genes (see Methods) for each sample (patient by timepoint) were hierarchically clustered and compared via correlation (Fig. [145]4a). Samples largely clustered by timepoint, validating the existence of patient-independent molecular signatures. We then validated our dataset against external data from Maus et al., which consisted of CD8^+ CAR T cells from patients with large B-cell lymphoma (Fig. [146]4a, b, colored in tan and brown, six in total after filtering)^[147]27. Based on their sample collection timing (day 7), external data from Maus et al. should resemble our samples at T[exp], rather than at T[per]. Consistent with expectations, external data resembled T[exp] transcriptomes on both pseudo-bulk (Fig. [148]4a, via correlation) and single-cell (Fig. [149]4b, via label transfer) levels, which increases the external validity of our findings. Fig. 4. Transcriptional signatures and regulatory networks of CD8^+ CAR T cells at T[exp] and T[per]. [150]Fig. 4 [151]Open in a new tab a Heatmap depicting correlation between pseudo-bulk transcriptome of each sample (patient by timepoint). Samples were ordered along columns and rows by hierarchical clustering. Transcriptomes of day 7 samples from Maus et al. were added for external validation. b Stacked bar graph depicting label transfer of timepoint (T[exp] or T[per]) from this study’s dataset onto single-cell transcriptomes of samples from Maus et al. c Gene set enrichment analysis comparing CAR T cells between T[exp] and T[per]. Gene sets were ordered by direction of upregulation and magnitude of enrichment. d Tile map depicting normalized expression of genes (columns) among different cell clusters at T[exp] and T[per] (rows). Genes were manually grouped into modules according to known functions. e Schematic for regulon construction and T[exp] versus T[per] classification. After transcriptomes were transformed into regulomes, regulon scores were calculated to train a machine-learning model to classify CAR T cells from T[exp] and T[per]. Key regulons were identified based on importance for the model’s predictions. Created in BioRender. Hu, Y. (2025) [152]https://BioRender.com/3ltrr3g. f Bar graph depicting the SHapley Additive exPlanation (SHAP) values for the top eight regulons underlying T[exp] and T[per] predictions. g Tile map depicting normalized signature scores of regulons (columns) among different cell clusters at T[exp] and T[per] (rows). Regulons were grouped as T[exp]-determining (left) and T[per]-determining (right). Target networks for the top eight T[exp]-determining (h) and T[per]-determining regulons (i). In each regulatory network, only the top differentially expressed genes are depicted. Each gene is colored according to log[2] fold-change between expression in T[exp] (red) and T[per] (blue). Source data are provided as a Source Data file. To elucidate the immunological processes at each timepoint, CD8^+ CAR T-cell transcriptomes at T[exp] and T[per] from our seven patients were compared via GSEA (Fig. [153]4c and data file [154]S1 for single-cell method, Fig. [155]S9a and data file [156]S2 for pseudo-bulk method) and DGEA (Fig. [157]4d, Fig. [158]S9b, data file [159]S3-[160]4). CAR T cells at T[exp] upregulated gene sets for cell cycling and apoptosis, consistent with a short-lived phenotype. Upregulation of the cell cycling gene set remained true even when only comparing the proliferating cell cluster at T[exp] and T[per] (Fig. [161]S9d). This indicates that, while proliferating cells are present at both T[exp] and T[per], those at T[exp] are more proliferative than those at T[per]. We also observed T[exp]-specific and cluster-independent upregulation of genes related to T-cell activation (including CD69, CD44, CD74) and exhaustion (including PRDM1, CTLA4, NR4A2), implicating antigen-engagement and CAR-mediated signaling during T[exp]. On the other hand, CAR T cells at T[per] upregulated gene sets for cytotoxicity and immune regulation. Expression of cytotoxicity-related genes (including KLRB1, FCGR3A, PRF1, GZMB, NKG7) was restricted to TE and exhausted-like TE clusters, indicating that this gene signature arises from T[per]-specific preponderance of the terminal effector phenotype. We also observed T[per]-specific upregulation of regulatory genes, including GTPase immune-associated proteins (GIMAP5, GIMAP7) which predict long-term persistence^[162]33, as well as sphingosine-1-phosphate receptors (S1PR1, S1PR5) which indicate blood localization^[163]34. These signatures suggest that CAR T cells at T[per] are functional and capable of long-term persistence in peripheral blood. Lastly, downregulation of genes related to T-cell activation and exhaustion implicate decreased in vivo antigen load and/or CAR-mediated signaling during T[per]. Interestingly, some differentially expressed gene signatures between CD8^+ CAR T cells at T[exp] and T[per] pertained to cytokine signaling (Fig. [164]4c, d, see Fig. [165]S9c). CAR T cells at T[exp] upregulated TNF response genes (including NFKBIA, NFKBIE, DUSP4, RELB). TNF can be secreted by CAR T cells for autocrine signaling^[166]20, predicts efficacious CAR T cells^[167]20, and contributes to CRS^[168]35. On the other hand, CAR T cells at T[per] upregulated type I interferon (IFN-I) response genes (including STAT1, IFITM1, IFITM2, IRF7). Many IFN-I response genes have known antiviral functions. The temporally specific and cluster-independent upregulation of TNF and IFN-I response genes indicates a dynamic in vivo cytokine environment during the CAR T-cell immune response. Next, we investigated T[exp]- and T[per]-specific regulatory networks using a machine-learning model (Fig. [169]4e, see Methods). Single-cell transcriptomes were transformed into regulomes to calculate regulon scores. A machine-learning model was trained using the regulon scores to classify CD8^+ CAR T cells at T[exp] and T[per]. This strategy resulted in high (~90%) classification accuracy across patients (Fig. [170]S10a). Top T[exp]- and T[per]-determining regulons were identified based on each regulon’s importance (SHAP value) for the model’s prediction (Fig. [171]4f, data file [172]S5). Sizes of timepoint-determining regulons span approximately 10–1000 genes (Fig. [173]S10b). Regulon expression was compared between cell clusters and timepoints (Fig. [174]4g). T[exp]- and T[per]-determining regulons were broadly upregulated at T[exp] and T[per], respectively. However, expression of some regulons was cluster-specific (e.g., FLI1 regulon among EM clusters). The top three T[exp]-determining regulons (for JUND, RELB, BHLHE40) exhibited higher expression among non-proliferating clusters and included co-regulated inflammation-associated genes (including NFKBIA, TNFAIP3, DUSP5) (Fig. [175]4h, see Fig. [176]S11a)^[177]36. The RELB and BHLHE40 genes themselves are included in the JUND regulon, indicating they may be directly upregulated by JUND. These three regulons were upregulated at T[exp] and suggest response to pro-inflammatory cytokines (such as TNF, discussed above) or CAR signaling. BHLHE40 rewires mitochondrial metabolism for tissue residency^[178]37, which may contribute to CAR T-cell fitness after entering the lymphoma. Shared genes between the JUND and BHLHE40 regulons included AP-1 transcription factors (JUN, JUNB, FOSL2), which have been implicated in T-cell and CAR T-cell exhaustion^[179]38,[180]39, and may contribute to the exhausted-like EM phenotype that predominates the first clonal wave at T[exp]. Among remaining T[exp]-determining regulons, FLI1 is notable for antagonizing effector T-cell differentiation^[181]40. Consistent with known biology, expression of the FLI1 regulon was upregulated among EM clusters, downregulated among TE clusters, and did not significantly change between T[exp] and T[per] within TE clusters. On the other hand, T[per]-determining regulons were more obscure (Fig. [182]4i, see Fig. [183]S11b). Among top T[per]-determining regulons, KLF13 regulates T-cell apoptosis^[184]41, BPTF activates a stemness gene expression program^[185]42 and maintains peripheral T-cell homeostasis^[186]43, and ELF4 promotes memory CD8^+ T cells by regulating quiescence^[187]44. Altogether, the putative functions of the KLF13, BPTF, and ELF4 regulons may underlie the delayed expansion and persistence of the second clonal wave at T[per]. Moreover, these functions indicate restoration of immune homeostasis during T[per]. The TBX21 regulon was notable for effector-associated genes (including NKG7, ZEB2) and was upregulated within non-CM clusters, consistent with preponderance of the TE phenotype at T[per]. The TBX21 regulon also includes S1PR5, which is a known T-BET target^[188]45 and retains T cells in peripheral blood^[189]34. Lastly, the STAT1 regulon was notable for interferon-associated antiviral genes (including OAS2, ISG20), in concordance with upregulation of IFN-I signaling gene sets at T[per]. CD8^+ CAR T cells at T[exp] and T[per] originate from distinct precursors in the infusion product The distinct clonotypes, phenotypes, transcriptional signatures, and regulatory networks underlying the two waves of CD8^+ CAR T cells led us to hypothesize that they originate from distinct precursors in the infusion product. To test this hypothesis, we first investigated infusion product phenotypes. After filtering for infusion product CD8^+ T cells, we identified four clusters based on gene expression density (Fig. [190]5a, b, Fig. [191]S12b) and levels (Fig. [192]5c, Fig. [193]S12a): proliferating (high MKI67), activated naïve-like (naïve-act, high TCF7), type 1 effector (EFF-Tc1, high TBX21), and type 2 effector (EFF-Tc2, high GATA3) T cells. The TCF7^hi naïve-act cluster upregulated markers of naiveness (IL7R, LEF1, CCR7, CD45RA) and T-cell activation (CD38, CD95), consistent with an early activated or stem cell-like memory phenotype. The two EFF clusters upregulated cytolytic molecules (GZMB, PRF1, GNLY) and downregulated markers of naiveness (IL7R, LEF1). EFF-Tc1 and EFF-Tc2 were distinguished by expression density of lineage-specific transcription factors (TBX21 for Tc1, GATA3 for Tc2) and receptors (KLRG1 for Tc1, CCR4 for Tc2). The EFF-Tc1 cluster exhibited lower CAR transgene expression than the EFF-Tc2 cluster (Fig. [194]S12b, c), which may impact their functional phenotypes in vivo^[195]46,[196]47. The three non-proliferating clusters upregulated IRF7 and anti-viral genes (ISG20, IFITM1, IFITM2), indicating type I interferon signaling during ex vivo transduction and/or expansion. Cluster proportions varied between patients (Fig. [197]S12d). In general, proliferating T cells (54%) were the most prevalent, while proportions of non-proliferating phenotypes (naïve-act, 14%; EFF-Tc1, 11%; EFF-Tc2, 21%) were mutually comparable (Fig. [198]5e, top row). Fig. 5. Infusion product precursors of peripheral blood CD8^+ CAR T cells. [199]Fig. 5 [200]Open in a new tab a UMAP depicting single-cell transcriptomes of infusion product CD8^+ T cells colored by cell cluster. Density maps (b) and violin plots (c) depicting expression levels of key genes and proteins for annotation and phenotyping. For extended version, see Fig. [201]S11a, b. d Cartoon depicting identification of Pre-T[exp] and Pre-T[per] using endogenous TCR clonotypes as unique indices. Created in BioRender. Hu, Y. (2025) [202]https://BioRender.com/49klewt. e Stacked bar graph depicting proportion distribution of all infusion product (top row) or precursors of peripheral blood CD8^+ CAR T cells (bottom two rows) among the four cell clusters. f Colored UMAPs depicting distribution of Pre-T[exp], Pre-T[per], and non-linked infusion product cells (“IP only”) on the overall UMAP. g Tile map depicting normalized expression of genes (columns) among Pre-T[exp], Pre-T[per], and IP only infusion product cells (rows). Genes were manually grouped into modules according to known functions. h Violin plots depicting expression levels of select differentially expressed gene sets between Pre-T[exp] (n = 1813 cells), Pre-T[per] (n = 2502 cells), and IP only (n = 8464 cells) infusion product cells. Expression levels were compared by two-sided Wilcoxon Rank-Sum test with p values adjusted for multiple hypotheses testing using the Benjamini-Hochberg method, whereby **** indicates p < 0.0001 and ** indicates p < 0.01. The central line indicates the median. The bounds of the box indicate the 25th–75th percentile. The leng^th of the whiskers indicates 1.5 times the inter-quartile range from the first and third quartiles. i Gene set enrichment analysis comparing Pre-T[exp]and Pre-T[per]. Gene sets were ordered according to direction of upregulation and magnitude of enrichment. j Enrichment plots for select gene sets differentially expressed between Pre-T[exp] and Pre-T[per] infusion product cells. Enrichment scores are computed by the Kolmogorov-Smirnov test with false discovery rate (FDR) adjusted q-values. k Cartoon depicting fates of CD8^+ CAR T cells over the entire course of therapy, from infusion product precursors to peripheral blood CAR T cells at T[exp] and T[per]. Created in BioRender. Hu, Y. (2025) [203]https://BioRender.com/3dtq7c4. NES normalized enrichment score. Source data are provided as a Source Data file. Having established infusion product phenotypes, we next performed TCR lineage tracing analysis to developmentally link infusion product CAR T cells with peripheral blood CAR T cells using their endogenous TCR clonotypes as unique indices. This method delineates infusion product precursors of CD8^+ CAR T cells at T[exp] (“Pre-T[exp]”) and at T[per] (“Pre-T[per]”) (Fig. [204]5d). Both Pre-T[exp] and Pre-T[per] comprised all four infusion product clusters (Fig. [205]5e, bottom rows) and largely overlapped on the UMAP (Fig. [206]5f). Hence, we conclude that coarse cluster phenotype alone cannot accurately predict in vivo differentiation. Compared to T cells not linked to peripheral blood (“IP only”), Pre-T[exp] and Pre-T[per] cells upregulated expression of effectorness (including GZMB, TNF, GNLY) and type 1 polarization (including IFNG, KLRG1) genes, as well as lower expression of stemness (including SELL, IL7R) and type 2 polarization (including GATA3, IL4R) genes (Fig. [207]5g). The significance of type 1 versus type 2 polarized CD8^+ CAR T cells is not well-understood in the literature and warrants future studies. Pathway analysis indicated that Pre-T[exp] and Pre-T[per] upregulated gene sets for T-cell effector function (“PD-1 signaling”, “T-cell-mediated cytotoxicity”), response to interferon γ (“interferon γ signaling”, “antigen processing and presentation”), and DNA replication, whereas non-linked CAR T cells upregulated gene sets for stemness (Fig. [208]5h, data file [209]S6). Our discovery of CD8^+ CAR T-cell precursors with an effector-like phenotype is corroborated by findings from Thomas et al., who also identified effector-like precursors from CAR T-cell patients treated for B-cell acute lymphoblastic leukemia^[210]48. Possible fates for the non-linked, naïve-like, IP-only CAR T cells include extravasation into the lymphoma, expansion in peripheral blood outside the T[exp]-T[per] window (day 8–28), or failure to persist in vivo. Lastly, we directly compared transcriptomic signatures of Pre-T[exp] and Pre-T[per] via GSEA. Although both Pre-T[exp] and Pre-T[per] broadly exhibited an effector phenotype (Fig. [211]5g, h), Pre-T[exp] upregulated effectorness-associated gene sets (including TCR, PD-1, interferon γ, and IL-1 signaling), whereas Pre-T[per] upregulated stemness-associated gene sets (Fig. [212]5i). In concordance, DGEA showed that Pre-T[exp] upregulated effector molecules (GZMB, GZMK, FCGR3A, NKG7, KLRG1), AP-1 transcription factors (JUN, JUND, FOS), and MHC class II expression (HLA-DRB1, HLA-DRB5, HLA-DRA), whereas Pre-T[per] upregulated naïve-like markers (SELL, IL7R, S1PR1) (Fig. [213]S12e, data file [214]S7). Upregulation of JUN in Pre-T[exp] confers exhaustion resistance^[215]38, potentially underlying their eventual efficacy at T[exp]. Interestingly, Pre-T[per] exhibited greater CAR transgene expression (Fig. [216]S12e), which may reflect transduction differences among apheresis precursors or a link between CAR expression and in vivo differentiation^[217]47. Next, we compared Pre-T[exp] and Pre-T[per] using transcriptomic signatures from a human CD8^+ differentiation atlas from Giles et al.^[218]49. Pre-T[exp] more closely resembled an effector memory state, whereas Pre-T[per] more closely resembled either a stem-cell memory or central memory state (Fig. [219]5j). Altogether, these patterns indicate that, while both Pre-T[exp] and Pre-T[per] exhibited an effector phenotype in the infusion product, Pre-T[exp] were more differentiated with greater effectorness, whereas Pre-T[per] were less differentiated with greater stemness. In conclusion, TCR lineage tracing analysis supports our hypothesis that CD8^+ CAR T cells at T[exp] and T[per] originate from different infusion product precursors. Integrating these findings into the two-stage differentiation model paints a more complete picture of CD8^+ CAR T cells over the month following CAR T-cell administration (Fig. [220]5k). Effector CD8^+ CAR T cells exist along a gradient of effectorness and stemness in the infusion product. Following infusion into the patient, effector CAR T cells with greater effectorness rapidly expand until peak expansion (T[exp], days 8–12), adopting a functional and cytotoxic EM phenotype with exhaustion-like characteristics upon antigen stimulation. Lymphoma infiltration and in vivo killing^[221]50, CRS^[222]5, and ICANS^[223]5 coincide with the T[exp] timeframe, suggesting that these CAR T cells mediate tumor clearance and side effects. Subsequently, this first wave of expanded CAR T cells diminishes through apoptosis or extravasation. Simultaneously, the remaining effector CAR T cells with greater stemness from the infusion product expand during the post-peak persistence timeframe (T[per], day 21–28). These newer and longer-lived CAR T cells adopt effector characteristics and persist in vivo, where they may ensure a durable response through long-term immunosurveillance. CD8^+ Exhausted-like EM CAR T cells at T[exp] exhibit characteristics of early exhaustion Discovery of the predominant exhausted-like EM phenotype in peripheral blood at T[exp] may be surprising since exhaustion was conventionally defined in lymphoid organs or tumor-infiltrating lymphocytes^[224]51. To verify this annotation, we further characterized the molecular signatures of the exhausted-like EM cluster. The exhausted-like EM cluster was densely situated at the lower half of the UMAP (Fig. [225]6a, b). We used RNA-seq/CITE-seq to examine gene and protein expression density, respectively (Fig. [226]6c, Fig. [227]S13a for additional markers). The exhausted-like EM cluster highly expressed exhaustion-associated transcription factors (NR4A2, TOX, IRF4) and inhibitory receptors (ENTPD1, PDCD1, TIGIT, LAG3, CTLA4). Intermediate expression of memory (TCF7, LEF1, CD27, IL7R) and effector (TBX21, GZMB, PRF1, GNLY, IFNG, NKG7) genes was also observed. Additionally, the cluster exhibited low expression of CX3CR1 (effector lineage marker) and B3GAT1 (senescence marker). CITE-seq measurements of protein expression for cell surface receptors (including CD39, PD-1, TIGIT, LAG3, CD27, CD57, CD127, TIM-3, CXCR3) were largely consistent with corresponding gene expression. Sequencing data indicated that high expression of CD39 (protein for ENTPD1) and low expression of CD57 (protein for B3GAT1) most clearly differentiated the exhausted-like EM cluster from other clusters. To validate these characteristics, flow cytometry was employed to analyze a set of longitudinal patient PBMCs (Fig. [228]S13b for gating). Consistent with sequencing data, flow cytometry data demonstrated that the CD39^+CD57^– phenotype was highest among CD8^+ CAR T cells at T[exp] (75%) compared to at T[per1] (33%) and T[per2] (18%) (Fig. [229]6d, top row). Furthermore, flow cytometry showed decreased CX3CR1 expression at T[exp], and increased expression at T[per1] and T[per2], in agreement with sequencing data (Fig. [230]6d, bottom row). This mixed expression pattern of exhaustion, memory, and effector markers resembles the circulating PD1^+CD39^+ T[ex] cells described by the Wherry group’s human T-cell differentiation atlas^[231]49. Fig. 6. Transcriptomic signatures of exhausted-like EM CD8^+ CAR T cells. [232]Fig. 6 [233]Open in a new tab UMAP depicting single-cell transcriptomes of peripheral blood CD8^+ T cells colored by cell cluster (a) or density contours of each cluster (b). The exhausted-like EM cluster is located on the lower half of the UMAP. c Density maps depicting expression levels of major T-cell genes and proteins, divided into categories. In the “receptor” category, proteins are placed directly beneath the corresponding gene. d Flow plots for validating transcriptomic data. Plots depict expression of CD39 and CD57 (top row) or CX3CR1 (bottom row) in CD8^+ CAR T cells at each timepoint from P2. Exhausted-like EM CAR T cells were expected to be T[exp]-specific with high CD39 and low CD57/CX3CR1 expression. e, f Heat map depicting normalized expression of major gene sets from Wherry et al. (e). Expression of two of the gene sets were depicted as violin plots (f), ordered by decreasing expression level per cluster (EM, exh-like: n = 7163 cells; CM: n = 1883 cells, EM: n = 4937 cells; TE, exh-like: n = 552 cells; TE: n = 21019 cells). The central line indicates the median. The bounds of the box indicate the 25th–75th percentile. The length of the whiskers indicates 1.5 times the inter-quartile range from the first and third quartiles. Expression levels were compared to that of the cluster with highest expression via two-sided Wilcoxon Rank-Sum test, with p values adjusted for multiple hypotheses testing using the Benjamini-Hochberg method, whereby **** indicates p < 0.0001, *** indicates p < 0.001. g Volcano plot depicting differentially expressed genes found via two-sided Wilcoxon Rank-Sum test with Bonferroni adjustment between EM and exhausted-like EM CD8^+ CAR T cells. Genes were colored according to direction of upregulation. h Enrichment plots for select gene sets differentially expressed between EM and exhausted-like EM CD8^+ CAR T cells. Enrichment scores are computed by the Kolmogorov-Smirnov test with false discovery rate (FDR) adjusted q-values. Source data are provided as a Source Data file. Next, we conducted GSEA using gene sets from Wherry et al. that identified T-cell differentiation states in the Armstrong/clone 13 LCMV model (Fig. [234]6e).^[235]52 These gene sets constitute comprehensive references for defining T-cell exhaustion because they (1) incorporate