Combination anti-PD-1 and anti-CTLA-4 therapy generates waves of clonal responses that include progenitor-exhausted CD8⁺ T cells

Summary

Combination checkpoint blockade with anti-PD-1 and anti-CTLA-4 antibodies has shown promising efficacy in melanoma. However, the underlying mechanism in humans remains unclear. Here, we perform paired single-cell RNA and TCR sequencing across time in 36 patients with stage IV melanoma treated with anti-PD-1, anti-CTLA-4, or combination therapy. We develop the algorithm Cyclone to track temporal clonal dynamics and underlying cell states. Checkpoint blockade induces waves of clonal T cell responses that peak at distinct timepoints. Combination therapy results in greater magnitude of clonal responses at 6 and 9 weeks compared to single-agent therapies, including melanoma-specific CD8⁺ T cells and exhausted CD8⁺ T cell (T_EX) clones. Focused analyses of T_EX identify that anti-CTLA-4 induces robust expansion and proliferation of progenitor T_EX, which synergizes with anti-PD-1 to reinvigorate T_EX during combination therapy. These next generation immune profiling approaches can guide the selection of drugs, schedule, and dosing for novel combination strategies.

eTOC blurb

Wang et al. analyze longitudinal blood from melanoma patients on checkpoint blockade using paired single-cell RNA and TCR sequencing and uncover waves of clonal responses that peak at different timepoints. By studying anti-CTLA-4, anti-PD-1, and in combination, they show that anti-CTLA-4 reinvigorates progenitor exhausted CD8⁺ T cells while anti-PD-1 drives their differentiation.

Introduction

Immune checkpoint inhibitors, such as anti-PD-1 therapy (αPD-1) and anti-CTLA-4 (αCTLA-4) therapy, have transformed clinical oncology by inducing long-term remissions in a variety of histologies, even in the metastatic setting. Combination checkpoint blockade with anti-PD-1 and anti-CTLA-4 antibodies (combination therapy) is more effective than single-agent therapy in metastatic melanoma, with a 52% overall survival at 5 years, compared to 44% and 26% in αPD-1 and αCTLA-4, respectively¹. A better understanding of the cellular and molecular mechanisms of αPD-1 and αCTLA-4 individually, and in combination, will guide the development of safer and more effective combination immunotherapy strategies.

CTLA-4 and PD-1 blockade each have distinct mechanisms of tumor control that have been defined using pre-clinical models². CTLA-4 is upregulated early upon T cell activation^3,4 and is a competitive inhibitor of the costimulatory molecule CD28^5-7, attenuating T cell activation in both a T cell intrinsic^3,4,8 and extrinsic manner^9-12. Thus, CTLA-4 blockade enhances the priming and early activation of effector T cells, resulting in improved tumor control¹³. Indeed, the human αCTLA-4 antibody ipilimumab results in increased CD8⁺ tumor infiltrating lymphocytes (TIL)¹⁴ as well as broadening of the T cell receptor repertoire^15,16 and tumor-specific CD8⁺ T cell responses¹⁷. Ipilimumab also has major pharmacodynamic effects on CD4⁺ T cells, including the expansion of Th1-like ICOS⁺Tbet⁺CD4⁺ T cells^18-21 and regulatory T cells (Tregs)^14,22,23, and the generation of CD4⁺ phenotypes, or archetypes, that are normally absent in physiological settings²⁴.

The major ligand for PD-1 is PD-L1, which is expressed by myeloid cells and cancer cells in the tumor microenvironment. PD-L1 is induced by IFN-γ²⁵, and represents an adaptive resistance mechanism of tumors in response to immune pressure^26-29. Exhausted CD8⁺ T cells (T_EX), in particular, are sensitive to PD-1 inhibition as they have the highest expression of PD-1 among CD8⁺ T cells³⁰. PD-1 blockade reinvigorates T_EX, allowing for a temporary burst of proliferation and recovery of effector function^31-35, resulting in improved viral and tumor control. In addition, PD-1 blockade has also been shown to modulate effector CD8⁺ T cells^36,37 and Treg populations^38,39. Thus CTLA-4 represents an early proximal checkpoint of T cell activation, while PD-1 more subtly modulates T cell differentiation upon T cell activation.

Nevertheless, the cellular and molecular mechanisms of combination checkpoint blockade in humans are still unclear, including the individual contributions of αPD-1, and αCTLA-4. Moreover, the immune responses of patients receiving αPD-1, αCTLA-4, and combination therapy have not been compared directly. Herein, we interrogated the immune responses of cohorts of patients with melanoma receiving αPD-1, αCTLA-4, or combination therapy. We used flow cytometry and single cell transcriptomics paired with clonotypic analyses to study the pharmacodynamic immune properties of checkpoint blockade in several dimensions - in terms of magnitude, breadth, and durability of responses. Using clonotypic trajectory analyses, we determined that these pharmacodynamic responses of CD8⁺ T cells were composed of waves of clonotypic responses with distinct temporal kinetics and cellular composition. Combination therapy induced large clonal responses at 6 and 9 weeks after treatment, which enriched for exhausted and effector CD8⁺ T cell clones. Finally, we identified that combination checkpoint blockade resulted in enhanced generation of progenitor T_EX, which was largely due to the effect of αCTLA-4.

Results

Combination checkpoint blockade results in exhausted and effector CD8⁺ T cell responses

By tracking proliferating CD8⁺ T cells, we and others have identified that PD-1 blockade induces systemic and intra-tumoral immune responses that peak after a single dose of αPD-1 by 3 weeks^40-43. We extended these analyses to cohorts of patients treated with αCTLA-4 and αPD-1 plus αCTLA-4 (combination therapy). Similar to αPD-1, αCTLA-4 and combination therapy had maximal pharmacodynamic effects on PD-1 expressing CD8⁺ T cells (Figures S1A and S1B). Interestingly, treatment-naïve patients who received αPD-1 had a more durable immunologic response than those who had prior αCTLA-4. Combination therapy resulted in a greater magnitude of immunologic response compared to either single-agent therapy alone and had additional effect on PD-1 negative cells, consistent with previous studies⁴⁴.

To understand the underlying cellular and molecular mechanisms of combination therapy on CD8⁺ T cells, we performed single-cell analyses of transcriptome and T cell receptor (TCR) clonotypes on 36 patients with melanoma receiving αCTLA-4, αPD-1, or combination therapy across 4 timepoints. We focused on non-naïve CD8⁺ T cells (Star methods: CD8⁺ T cell Atlas, Figure 1A), as naïve CD8⁺ T cells were almost exclusively singlet T cell clones and did not expand in response to therapy. We constructed a batch-corrected non-naïve CD8⁺ T cell atlas using Harmony⁴⁵ and Symphony⁴⁶, with a total of 273,508 non-naïve CD8⁺ T cells (Figure 1B, Figure S2A, Star methods: CD8⁺ T cell atlas). Baseline clinical characteristics are summarized in Tables 1 and S1.

Figure 1. Combination therapy induces T_EX and effector CD8⁺ T cell responses.

A. Study schema. Peripheral blood collected from 36 patients at baseline and every 3 weeks after checkpoint blockade. Clinical metadata. Median age. Response includes complete response and partial response.

B. Uniform Manifold Approximation and Projection (UMAP) representation of non-naive CD8⁺ T cells (n=273,508). CM indicates central memory, SCM indicates stem cell-like memory.

C. Auto-correlation of top 3000 genes ranked by Hotspot and 4 identified modules by agglomerative clustering.

D. Pathway enrichment analysis of modules, with selected enriched pathways colored by −log 10 of the adjusted p-value (p-adj). Abbreviations: Effector – Eff, Memory – Mem, AP-1/IFN – AP1, Mito/ATP – ATP.

E. UMAPs colored by gene module signature scores.

F. Average module expression by metacluster.

G. Fold change from baseline of frequencies for each metacluster. Significance as indicated by p-value or star is the change in frequency from baseline. One patient in the αPD-1 group with an extremely large exhausted CD8⁺ T cell population at baseline (15.8%), confirmed by flow cytometry, was excluded from the heatmap. Flow cytometry data is in Figure S1. αCTLA-4 (n=10), αPD-1 (n=15), combination (αPD-1 and αCTLA-4) (n=9). Wilcoxon signed-rank test.

H. Frequency of exhausted and effector CD8⁺ T cells for combination therapy at indicated times. Transparent lines indicate individual patients (n=9). Opaque lines indicate mean. Wilcoxon signed-rank test.

NS non-significance; *p<0.05.

See also Figures S1 and S2, Tables 1 and S2.

Table 1.

Baseline characteristics of single cell sequencing cohort, related to Figure 1 and Star Methods. See also Table S1.

	Pembrolizumab (n=16)	Ipilimumab (n=11)	Combination (Ipi+nivo) (n=9)
Age (Median)	71 (39-83)	59 (50-79)	52 (29-71)
Men	8 (50%)	6 (55%)	7 (78%)
Women	8 (50%)	5 (45%)	2 (22%)
Race
White	12 (75%)	9 (82%)	9 (100%)
Other	-	2 (18%)	-
Unknown	4 (25%)	-	-
ECOG performance status
0-1	15 (94%)	11 (100%)	8 (89%)
>=2	1 (6%)		1 (11%)
LDH
Normal	10 (63%)	3 (27%)	5 (56%)
Elevated	5 (31%)	6 (55%)	2 (22%)
Unknown	1 (6%)	2 (18%)	2 (22%)
Median number of cycles	10 (3-31)	4 (4-4)	4 (2-4) / 32 (2-55)a
Clinical Response
Complete response	5 (31%)	3 (27%)	4 (44%)
Partial response	3 (19%)	3 (27%)	3 (33%)
Stable disease	2 (12%)	-	-
Progressive disease	6 (38%)	3 (27%)	1 (11%)
Not applicableb	-	2 (18%)	1 (11%)
Number of previous lines of therapy
0	4 (25%)	7 (64%)	5 (56%)
1	7 (44%)	1 (9%)	2 (22%)
2	4 (25%)	3 (27%)	1 (11%)
>3	1 (6%)	-	1 (11%)
Previous therapyc
Immunotherapy	12 (75%)	3 (27%)	3 (33%)
Ipilimumab	12 (75%)	1 (9%)	-
Pembrolizumab	-	1 (9%)	2 (22%)
Interferon	-	1 (9%)	1 (11%)
IL-2	-	-	1 (11%)
Chemotherapy	5 (31%)	1 (9%)	-
BRAF or MEK inhibitor	-	2 (18%)	2 (22%)
Kinase inhibitor	-	1 (9%)	-

All patients included in the study had stage IIIC, recurrent, or IV disease. Combination therapy is ipilimumab and nivolumab.

^aTwo patients in combination therapy group lost to follow up while on the nivolumab maintenance phase of the therapy

^bUnable to assign clinical response to patients if they received radiation therapy while on treatment

^cImmunotherapy – ipilimumab, pembrolizumab, interferon, IL-2; Chemotherapy – dacarbazine, temozolomide

BRAF or MEK inhibitor – vemurafenib, dabrafenib, encorafenib, binimetinib; Kinase inhibitor – temsirolimus.

We first sought to understand the underlying transcriptional machinery of human non-naïve CD8⁺ T cells using gene module analysis. We used the algorithm Hotspot⁴⁷ to identify modules of co-expressed genes through an unsupervised approach. Hotspot identified 4 distinct modules which included co-regulated genes and pathways known to be important in effector, memory, AP-1 and interferon, and mitochondrial/ATPase biology (Figures 1C and 1D, Table S2). The effector module included genes and pathways associated with T cell activation, TCR signaling, migration, and cytotoxicity^48,49. The memory module included metabolic pathways known to be used in memory cells such as oxidative phosphorylation, as well as pathways important for translation machinery, cytokine signaling, and survival^48,49. We also identified a gene module that enriched for pathways downstream of cellular perturbations, including AP-1⁵⁰, interferon⁵¹, and HIF-1α pathways⁵², which are modulated downstream of T cell activation and cytokine signaling. Finally, a module was identified that included genes related to the mitochondrial electron transport chain and ATP production (ATP5F1E, ATP6V0E1) in conjunction with annexins (ANXA2, ANXA1) and S100 family members (S100A4, S100A10) (Figures 1C and 1D, Table S2).

We then defined non-naïve CD8⁺ T cell subsets using differential expression of these modules, which allowed us to fine-tune the cluster resolution and generate clusters with distinct molecular identities (Figures 1E and 1F, Star methods: CD8⁺ T cell atlas). We identified a spectrum of CD8⁺ T cell differentiation states. Both stem-cell memory (SCM) and memory CD8⁺ T cells had enrichment of the memory module, but SCM cells were also enriched for a naïve transcriptional signature (Figure S2B); activated CD8⁺ T cells had high expression of the AP-1/IFN module, and effector and NK-like CD8⁺ T cells were identified by high expression of the effector module as well as an EMRA gene signature (Figures 1B, 1E, 1F, and S2B).

Exhausted CD8⁺ T cells in the blood did not cluster independently in unsupervised analysis. This was not surprising due to the low frequency of T_EX in the blood and the known heterogeneity of T_EX which span a spectrum from progenitor to more differentiated T_EX⁵⁴. Therefore, we defined T_EX using a human exhausted gene signature from sorted PD1⁺CD39⁺ CD8⁺ T cells⁵³ (Figure S2C, Table S2, Star methods: Gene lists and exhausted CD8⁺ T cells). Effector CD8⁺ T cells had high effector and mitochondrial ATPase activity, while T_EX had decreased expression of effector, AP-1, and mitochondrial modules in comparison (Figure 1F). Thus, these analyses capture exhausted CD8⁺ T cell biology in which partial effector function loss is driven by mitochondrial dysfunction^55,56 as well as partnerless NFAT in the absence of AP-1⁵⁷. Altogether, we identified CD8⁺ T cell states that were consistent with published gene signatures of memory, effector, and exhausted CD8⁺ T cells (Figures S2B-F) as well as their underlying biology.

Having constructed a non-naïve CD8⁺ T cell atlas, we sought to identify the cellular changes induced by single-agent and combination checkpoint blockade. αPD-1 induced an expansion of T_EX at week 3 as previously reported⁴¹, while combination checkpoint blockade significantly increased the frequency of both exhausted and effector CD8⁺ T cells, which peaked at 6 and 9 weeks, respectively (Figures 1G and 1H). Thus, combination checkpoint blockade therapy had not only a larger magnitude of CD8⁺ T cell response but also a broad and durable pharmacodynamic effect on CD8⁺ T cell subsets. These data are consistent with previous studies in patients showing that combination therapy induced a greater magnitude of transcriptional and cytokine changes⁵⁸ as well as expansion of Tbet⁺Eomes⁺CD8⁺ T cells⁴⁴ compared to single-agent therapy.

Checkpoint blockade induces waves of clonal CD8⁺ T cell responses with distinct kinetics and cell state compositions

To understand more precisely how CD8⁺ T cells respond to checkpoint blockade, we used each cell’s TCR to track clonal dynamics across time. First, to control for normal clonal fluctuations over time, we performed paired single-cell RNA and TCR sequencing on sorted CD8⁺ T cells from peripheral blood mononuclear cells (PBMC) collected three weeks apart from 2 healthy donors (HD) and assessed the clonal dynamics of non-naïve CD8⁺ T cells. Overall, healthy donors exhibited minimal clonal variation over time. On the contrary, checkpoint blockade therapy induced significant changes in the frequency of CD8⁺ T cell clones (Figures S3A-C). Thus, checkpoint blockade therapy induced a greater magnitude of clonal perturbation, above that of normal physiologic variation, consistent with previous reports^15,16.

We wanted to focus our analyses on robust clonal responses that were meaningful and therapeutically induced, both in terms of fold change and absolute change in frequency. Only cells with both productive TCR alpha and beta chains were included. We additionally excluded patients with prior immunotherapy to eliminate possible pretreatment effects, except for a large cohort of patients who received αPD-1 and had prior αCTLA-4 therapy with ipilimumab, which we separated into a distinct treatment group (ipi pretreated patients). This resulted in a subset of 83,287 cells constituting a total of 32,080 unique clones across eligible patients and timepoints (Star methods: Cyclone and TCR trajectory). Finally, only significantly changing clones were included in downstream analysis, based on the distribution of clonal changes in the HD reference, which yielded 508 clones, or 1.58% of the total pool (Figures S3D and S3E). Strikingly, each treatment had different frequencies of clones that met the threshold: αCTLA-4 (0.95%), αPD-1 (prior ipi) (1.15%), αPD-1 (2.49%), combination (2.6%), and HD (0.61%) (Figure S3F). αPD-1 and combination therapy had greater dynamic clonal changes than αCTLA-4 and αPD-1 in ipi pretreated patients, suggesting that αCTLA-4 pretreatment affects immune response to subsequent therapies. Thus, immune checkpoint blockade dynamically modulates 1-3% of non-naïve CD8⁺ T cell clones in the blood, and different treatments induced different magnitudes of clonal responses, consistent with the changes seen in Ki67 (Figure S1).

To understand the pharmacodynamic patterns of T cell clones over time, we developed the algorithm Cyclone⁵⁹, which treats TCR clonotypic frequencies across time as time-series data and models every clonotype as individual trajectories. Using insights from the field of trajectory clustering, we explored the distinct patterns of TCR expansion and contraction by binning clones through an unsupervised approach based on pairwise frequency correlation (Star methods: Cyclone and TCR trajectory). Using Cyclone, we tracked 508 unique TCR clonotypes in 18 patients across 9 weeks of treatment and identified 6 patterns of clonotypic trajectories (Figures 2A and 2B). Notably, the trajectories generated by immune checkpoint blockade were composed of unique clones that peaked at different timepoints, at weeks 3, 6, and 9, indicating waves of clonal T cell responses. Four out of the six clonotypic patterns consisted of expanding trajectories, including those peaking at week 3 (Traj 3), week 6 (Traj 4 and 5), and week 9 (Traj 6). Traj 1 and 2 were both contracting trajectories, representing clonotypes that decreased in frequency during treatment, albeit with different kinetics.

Figure 2. Checkpoint blockade induces distinct clonal trajectories over time.

A. Pairwise correlation of clonotypic trajectories using Cyclone.

B. Cumulative frequencies for 6 clonotypic patterns. Each line denotes a single clone, and height between adjacent lines at each timepoint denotes frequency.

C. Cumulative clonotypic frequencies split by treatment. Asterisk indicates significantly greater expansion per clone for the following comparisons: Traj 4: αPD-1 vs αPD-1 (prior ipi) ***, αPD-1 vs αCTLA-4 **, Combination vs αPD-1 (prior ipi) ***; Traj 5: Combination vs αPD-1 *, Combination vs αCTLA-4 **; Traj 6: Combination vs αPD-1 **, Combination vs αPD-1 (prior ipi) **, Combination vs αCTLA-4 ***. *p<0.05, **p<0.01, ***p<0.001, Wilcoxon rank-sum test. See Figure S4 for all comparisons and statistics. Combination therapy indicates αPD-1 and áCTLA-4.

D. Trajectory composition of each treatment, scaled by absolute number of clones.

E. Fold change of cell state composition stratified by pattern, averaged by patient. Percent of cells that have Ki67 gene expression, capped at 10% for visualization. CM indicates central memory, SCM indicates stem cell-like memory.

For all: αCTLA-4 (n=4), αPD-1 (prior ipi) (n=7), αPD-1 (n=4), combination (n=3). Large outlier clones greater than 2.5 standard deviations above mean frequency were excluded from visualization in B-E.

See also Figures S1, S3, S4, and S5.

To understand whether single-agent and combination checkpoint blockade strategies induced distinct clonal dynamics, we separated each clonotypic pattern by treatment (Figures 2C, 2D, S4A, and S4B). While αCTLA-4 induced smaller changes in CD8⁺ T cell clonal frequency, it has previously been shown to increase clonal diversity^15,16, which may be below the sensitivity of detection by single-cell TCR sequencing. αPD-1 in ipi pretreated patients induced Traj 3 that peaked at week 3, consistent with the early kinetics of αPD-1 previously described^40-43, although major clonal responses were also seen at week 6 and to a lesser extent at week 9. Treatment-naïve patients, however, had significantly larger expansion of clones at week 6 (Traj 4) (Figures 2C, 2D, and S4C) and generated large novel clones at each timepoint, while ipi pretreated patients had few or no novel clones (Figures S4D-F). Combination therapy, however, induced large clonal responses peaking at weeks 6 and 9 (Traj 4-6) and generated significantly larger novel clones compared to αCTLA-4 or αPD-1 alone. All three treatments resulted in substantial clonal contraction (Traj 1), consistent with the TCR turnover and evolution previously described^15,16.

We then analyzed the underlying cell state compositions of these clonal responses using cell state assignments from the non-naïve CD8⁺ T cell atlas (Figure 1B). Compared to ipi pretreated patients, treatment-naïve patients receiving αPD-1 had clonal expansion across a number of cell states spanning early activated, activated, and early effector at week 3, and early effector and NK-like responses at week 6 (Figures 2E, S5A, and S5B). These phenotypes were not captured amongst the proliferation-based responses that had been described previously^40-42, likely due to the fact that these other cell states had a lower Ki67 expression as compared to exhausted CD8⁺ T cells (Figure 2E: %Ki67 overlay). αCTLA-4 resulted in broad responses with lower magnitude, while combination therapy induced major T_EX and effector responses that peaked at 6 and 9 weeks, respectively (Figures 2E, S5A, and S5B).

Combination therapy generates melanoma-specific responses that peak at 3 and 6 weeks.

We next sought to understand the antigen specificity associated with each clonal trajectory to uncover possible clinical relevance. We used a combinatorial tetramer approach to identify antigen-specific responses in serial blood samples from patients expressing A1, A2, A3, or B7 alleles across three treatments: αCTLA-4 (n=2), αPD-1 (n=3), and combination therapy (n=6) (Star methods: Combinatorial Tetramer). In total, we identified 13 melanoma-specific (MART-1, gp100, Tyrosinase, LAGE-1, NY-ESO1, TAG, and MAGE-A1) and 11 viral-specific (CMV, EBV, Flu) CD8⁺ T cell populations in 9 out of 11 patients (Figure 3A). We detected melanoma-specific CD8⁺ T cells in 0/2 patients on αCTLA-4, 2/3 patients on αPD-1, and 5/6 patients on combination therapy. Therefore, we primarily focused on combination therapy, which had the majority of detected melanoma-specific responses and had statistically significant expansions at 3 and 6 weeks post-treatment. In contrast, viral-specific CD8⁺ T cells expanded at 9+ weeks (Figures 3B and 3C).

Figure 3. Combination checkpoint blockade induces melanoma-specific CD8⁺ T cell responses at 3 and 6 weeks.

A. Representative flow plots showing combinatorial tetramer gating for HLA-A2 restricted (indicated by A2) melanoma and viral specific CD8⁺ T cells in one sample.

B. Representative flow plots showing melanoma and viral specific CD8⁺ T cells at indicated timepoints. Week 9+ indicates weeks 9-15. A2 and A3 indicate HLA-A2 and HLA-A3 restricted epitopes, respectively.

C. Frequencies of HLA-A1, HLA-A2, HLA-A3, and HLA-B7 restricted melanoma and viral specific CD8⁺ T cells in combination therapy (αPD-1 and αCTLA-4) at indicated timepoints. Each line indicates one patient-specific antigen-specificity. Melanoma (n=6 patients, n=13 patient antigen-specificities); Viral (n=4 patients, n=7 patient antigen-specificities). NS non-significance; *p<0.05, **p<0.01, Wilcoxon signed-rank test.

D. Cumulative frequencies of antigen specific CD8⁺ T cells. Each line denotes an antigen specificity for individual patients in combination therapy.

E. Phenotypic compositions of antigen specific CD8⁺ T cells. Each line denotes a phenotype within an antigen specificity for individual patients in combination therapy.

C-E. One large viral clone (CMV pp65 NLV) in patient 16-2189 with frequencies 0.99%, 1.4%, 1.37%, and 1.25% at weeks 0, 3, 6, and 9, respectively, was excluded for clarity.

See also Figure S6 and Table S3.

Trajectory clustering of these antigen-specific responses demonstrated that melanoma-specific CD8⁺ T cells occupied early trajectories that peaked at 3 or 6 weeks, while viral-specific responses, largely driven by EBV-specific CD8⁺ T cells, peaked at 9+ weeks (Figure 3D). Finally, CD8⁺ T cells responding to combination therapy were almost exclusively exhausted or EM1 (CD45RA⁻CCR7⁻CD27⁺) CD8⁺ T cells, whether melanoma- or viral-specific (Figure 3E).

We also performed bulk TCR sequencing on FFPE tumor specimens from 10 patients in our main sequencing study (4 αCTLA-4, 5 αPD-1, 1 combination) to generate a library of TIL clonotypes, which were mapped back to our main single-cell RNA sequencing dataset. This generated a set of 79 circulating clonotypes that were also present in the tumor. We then ran Cyclone analysis on these 79 clonotypes, yielding 4 trajectories, including clones that peaked at 3, 6, and 9 weeks. For αPD-1 therapy, the majority of tumor-associated clonal responses were in Traj 2, peaking at week 3. In contrast, after combination therapy, tumor-associated clones expanded primarily at week 6 in Traj 3 (Figure S6 and Table S3).

Altogether, combination therapy expanded T_EX and effector clones at 6 and 9 weeks after treatment, while antigen-specific and tumor-associated data suggest that the early trajectories, which peak at 3 to 6 weeks, may be relevant for clinical efficacy.

Progenitor exhausted CD8⁺ T cells identified in peripheral blood

Exhausted CD8⁺ T cells, as identified by exhaustion signature, had features of memory, effector, and NK-like clusters across the non-naïve CD8⁺ T cell atlas (Figure 1B), reminiscent of T_EX subsets with a parallel differentiation trajectory^35,60-67. These T_EX are enriched for tumor-specific CD8⁺ T cells^68,69, respond robustly to αPD-1^{31,33,35,41,60,64}, and are a major mediator of the clinical efficacy of checkpoint blockade in human cancers⁷⁰. Moreover, recent studies have revealed the heterogeneity of T_EX, including progenitor^61,62,65, intermediate/NK-like^66,67,71-73, and terminal^33,61,62,65 T_EX that play complementary roles in tumor control. We therefore subset on T_EX and fit a pseudotime trajectory, choosing the more central memory-like T_EX as the starting point (Figure 4A). To characterize exhausted CD8⁺ T cells, we evaluated gene expression across the pseudotime gradient (Figures 4B and 4C, Star methods: Pseudotime), identifying three main subsets with expression patterns consistent with known T_EX subsets, including progenitor (CTLA-4, TCF7, SPRY1)^61,62,65,74, terminal^33,61,62,65 (TOX, LAG3, EOMES), and intermediate/NK-like (ZEB2, BATF, KLRC)^66,67,71-73. We additionally analyzed surface protein (antibody-derived tags, ADT) data for a subset of the patients and identified associated protein expression profiles for each subset including progenitor (CD62L, ICOS, CD28), terminal (CD38, CD11a, CD244), and NK-like (CD16, CX3CR1, CD56) exhausted CD8⁺ T cells (Figure 4D).

Figure 4. Combination therapy reinvigorates progenitor and differentiated exhausted CD8⁺ T cells.

A. Monocle 3 tree fitted onto exhausted CD8⁺ T cells (T_EX), colored by pseudotime.

B. Relative expression of top differentially expressed genes for each cluster, across pseudotime.

C. Gene expression for individual cells along pseudotime with fitted spline curve, for selected markers. Cells with 0 expression of a gene are excluded from the respective plot.

D. Relative expression of selected protein markers across pseudotime.

E. Fold change from baseline of frequencies for T_EX subsets. Significance indicates change in frequency from baseline. One patient in the αPD-1 group with an extremely large T_EX population at baseline (15.8%), confirmed by flow, was excluded from the heatmap. αCTLA-4 (n=10), áPD-1 (n=15), combination (αPD-1 and αCTLA-4) (n=9). Wilcoxon signed-rank test.

F. Frequency of progenitor (Prog) and differentiated (Diff) T_EX subsets after combination therapy. Transparent lines indicate individual patients (n=9). Opaque line indicates mean.

*p<0.05, **p<0.01.

See also Figures S7, S8, and S9.

T_EX in tumor have a terminally differentiated exhausted signature that clearly separates them from memory and effector CD8⁺ T cells^68,75. However, these circulating T_EX did not have the same degree of exhaustion signature. To understand how these blood T_EX compared to tumor T_EX, we performed single-cell RNA sequencing on paired blood and tumor specimens from 3 patients with melanoma and generated an atlas consisting of CD8⁺ T cells from both blood and tumor (Figures S7A and S7B, Star methods: Paired blood-tumor atlas). UMAP dimension 1 revealed the path of differentiation as we had seen in the blood, with naïve, memory, effector, and NK-like CD8⁺ T cells, while UMAP dimension 2 separated blood from tumor. Indeed, the addition of tumor data revealed an additional arm of T_EX composed of terminal and progenitor T_EX clusters (Figures S7A and S7B). Terminal T_EX expressed high levels of exhaustion genes such as TOX, PDCD1, LAG3, and CXCL13. Progenitor T_EX had comparatively higher expression of progenitor-associated genes such as TCF7, CCR7, SELL, and TNF (Figure S7C). At the same time, Progenitor T_EX also expressed TOX and other exhaustion genes, albeit at a lower amount as compared to terminal T_EX but higher than other cell states, consistent with the known biology of Progenitor T_EX^60,62,67,76. As the distribution of T_EX subsets differ based on tissue location^60-62,77, we suspected that circulating T_EX were enriched for late progenitor and NK/intermediate T_EX as described in preclinical models⁶⁰. We defined circulating and tumor T_EX based on the same threshold of human exhaustion signature (Figures S2C, S7D). By this criterion, 2.8% of circulating CD8⁺ T cells were T_EX; in contrast 43% of CD8⁺ T cells in the tumor were T_EX. Tumor T_EX also had a greater exhaustion score compared to circulating T_EX (Figure S7D). We then mapped these circulating and tumor T_EX back onto the blood-tumor atlas, as defined by exhaustion score. Tumor T_EX were largely restricted to terminal T_EX as previously reported. However, blood T_EX largely mapped to progenitor and effector clusters, consistent with the phenotype of progenitor and intermediate/NK T_EX (Figure S7E). Given the small numbers of circulating T_EX in the paired blood-tumor dataset, we also mapped circulating T_EX from the main study, which also fell primarily in progenitor and effector clusters (Figure S7F). Altogether, these data suggest that the T_EX identified in circulation are largely composed of progenitor and intermediate T_EX while terminal T_EX are tissue restricted.

Combination therapy generates progenitor exhausted CD8⁺ T cell responses that are induced by αCTLA-4

Progenitor T_EX (Prog_EX) had high expression of CTLA-4 (Figures 4B and 4C) and thus might be modulated by CTLA-4 blockade. To determine the impact of checkpoint blockade on T_EX subsets, we defined Prog_EX and differentiated T_EX (Diff_EX), which included terminal and intermediate/NK-like subsets, using a central memory score (Star methods: Defining Progenitor T_EX). This revealed expansion of DiffEx by both αCTLA-4 and αPD-1, although changes in Prog_EX were underpowered to reach statistical significance (Figure 4E). However, combination therapy was associated with significant expansion of both Prog_EX and Diff_EX which peaked at weeks 3 or 6 (Figure 4F).

To confirm whether combination checkpoint blockade modulated Prog_EX, we analyzed flow cytometry data generated from patients with melanoma treated with combination therapy (Figure S1). We interrogated the proliferating Ki67⁺ responding cells after combination therapy to understand the cell states that were involved in the pharmacodynamic immune responses. UMAP visualization of these responding cells revealed an enrichment of PD-1, CTLA-4, and TCF-1 expression in responding cells after treatment (Figure S8A), suggesting that combination therapy modulated Prog_EX. As CTLA-4 is expressed in both Prog_EX and Diff_EX⁶², we defined T_EX as PD-1⁺CTLA-4⁺CD8⁺ T cells, consistent with our previously published data (Figure S8B)^40,41. Both TCF-1⁺PD-1⁺CTLA-4⁺ (Prog_EX) and TCF-1⁻PD1⁺CTLA-4⁺ (Diff_EX) CD8⁺ T cells had high expression of classic exhaustion markers, including TIGIT, CD39 and TOX (Figure 5A). Combination therapy generated the largest magnitude of proliferative response in Prog_EX and Diff_EX compared to other canonical CD8⁺ T cell differentiation states (Figure 5B), which was associated with expansion of these T_EX subsets that peaked at week 6 (Figure S8C). Thus, consistent with single-cell sequencing data, combination therapy robustly reinvigorated both Prog_EX and Diff_EX subsets of T_EX.

Figure 5. CTLA-4 blockade reinvigorates progenitor exhausted CD8⁺ T cells.

A. Expression of selected markers for indicated CD8⁺ T cell subsets 3 weeks after combination therapy (αPD-1 and αCTLA-4). Flow cytometry was performed on PBMC from 8 patients. Central memory (CM), effector memory (EM), terminal effector memory (EMRA).

B. Ki67 expression over time for indicated CD8⁺ T cell subsets after combination therapy (n=8). Differentiated T_EX (Diff), Progenitor T_EX (Prog).

C. Study design for sequential treatment cohort. Flow cytometry was performed on blood collected every 3 weeks from a cohort of 8 patients who received ipilimumab followed by pembrolizumab.

D. Ki67 expression in CD8⁺ T cells over time after ipilimumab (red) and pembrolizumab (blue).

E. Frequency and Ki67 expression for progenitor exhausted CD8⁺ T cells (n=4).

F. Study design for Checkmate 238 study. Flow cytometry was performed on blood collected before treatment and two weeks post-treatment from patients on ipilimumab (n=24) or nivolumab (n=21).

G. Ki67 expression in progenitor T_EX for ipilimumab (red, n=24) and nivolumab (blue, n=21).

H. Frequency of progenitor T_EX.

I. Ki67 expression in progenitor T_EX at Week 2, comparing ipilimumab and nivolumab. Error bars represent mean and 95th percentile confidence interval.

B, D-I. Transparent lines indicate individual patients. Opaque line indicates mean.

NS non-significance; *p<0.05, **p<0.01, ***p<0.001, ****p<0.0001, Wilcoxon signed-rank test.

See also Figures S8 and S9 and Table S4.

To characterize how each monotherapy contributed to the Prog_EX response in combination therapy, we performed flow cytometry on a cohort of 8 patients who received sequential treatment with αCTLA-4 (ipilimumab) followed by αPD-1 (pembrolizumab) to control for patient-specific effects (Figure 5C, 5D). αCTLA-4 induced a greater magnitude of Prog_EX expansion and proliferation compared to αPD-1 (Figure 5E), which was also associated with Diff_EX responses (Figure S9A). Finally, to further validate these findings and to control for pretreatment effects, flow cytometry was independently performed on peripheral blood at baseline and 2 weeks post treatment from patients with Stage III/IV resectable melanoma treated with adjuvant ipilimumab or nivolumab from a randomized clinical trial (Checkmate 238)⁷⁸ (Table S4, Figure 5F). αPD-1 resulted in proliferation but no expansion of Prog_EX and Diff_EX (Figures 5G, 5H and S9B), In fact, there was a trend towards a decrease in the frequency of Prog_EX, suggesting differentiation and/or migration events. To understand the relevance of these immune dynamics in the tumor, we performed flow cytometry on baseline and post-treatment tumor samples from a cohort of 17 patients with Stage III melanoma who received neoadjuvant αPD-1 (Figure S9C). Prog_EX in the tumor were defined as PD1⁺TCF-1⁺Tim3⁻CD8⁺ T cells, while terminally exhausted CD8⁺ T cells (Term_EX) were defined as PD1⁺TCF-1⁻Tim3^+35,62 (Figure S9C). αPD-1 resulted in proliferation of Prog_EX in the tumor microenvironment, which was coupled with a decrease in frequency of Prog_EX and increase in Term_EX (Figure S9D), consistent with the paradigm that PD-1 blockade drives the differentiation of Prog_EX to Term_EX⁶².

In contrast, αCTLA-4 resulted in the expansion of Prog_EX in the blood and induced significantly greater proliferation of Prog_EX compared to αPD-1 (Figures 5G-I) which was coupled to both expansion and proliferation of Diff_EX (Figure S9B). Altogether these data demonstrate that αCTLA-4 results in expansion of Prog_EX while αPD-1 drives its differentiation. These immune mechanisms may explain some of the immune and clinical effects of combination therapy.

Discussion

We used single-cell multi-modal analyses to directly compare the pharmacodynamic effects of αPD-1, αCTLA-4, and combination checkpoint blockade therapy in humans being treated for melanoma. By using single-agent therapy cohorts as comparisons, we were able to deconvolute the cellular and molecular effects of combination checkpoint blockade, revealing novel insights.

First, αPD-1 and αCTLA-4 therapy each have their own distinct pharmacodynamic properties, including kinetics of immune responses. In combination, these therapies resulted in immune responses that were both durable and greater in magnitude which may contribute to both enhanced efficacy as well as toxicity.

Second, despite clear pharmacodynamic responses by proliferation (Ki67) and frequency of cell states, these measurements represent snapshots in time and did not inform if cellular responses at different timepoints were clonally related or distinct. Therefore, we developed a novel algorithm Cyclone to characterize patterns of clonotypic trajectories in an unsupervised manner. Cyclone identified distinct waves of clonal CD8⁺ T cell responses that peaked at different post-treatment timepoints. Thus, in many cases, the sustained immunologic responses seen in flow cytometry reflected the sum of different clonotypes with distinct temporal trajectories. While there was sustained expansion for some clones (e.g. Traj 5), the majority of clones had trajectories characterized by brief expansion at a specific timepoint followed by a return to baseline. Each dose of immune checkpoint blockade induced the activation and expansion of distinct groups of clones which likely recognize distinct groups of epitopes. Moreover, tumor-specific CD8⁺ T cell clones appear to be primarily modulated within the first 3-6 weeks, consistent with the early pathologic responses seen in neoadjuvant checkpoint blockade trials in melanoma^79-83. There also may be a point at which tumor-specific CD8⁺ T cell clones are depleted by PD-1 blockade, and therapies to promote epitope spreading^84,85 or immune rest^86-88 may be required to rejuvenate tumor-specific immunity. These observations are consistent with preclinical data demonstrating the importance of sustained and coordinated systemic cellular responses for effective immunotherapy⁸⁹, as well as the emergence of new exhausted CD8⁺ T cell clones after PD-1 blockade⁹⁰. With the recent success of neoadjuvant therapy⁹¹, understanding the pharmacodynamics and antigen-specificity of these waves of clonal responses will be critical for defining the optimal time for surgery after neoadjuvant therapy.

Third, immune monitoring through clonal trajectories identified a dimension of immune response distinct from cellular proliferation or Ki67. While clonal responses to αPD-1 peaked at 3 and 6 weeks, consistent with what was seen using Ki67^41,42, Cyclone identified αPD-1-responsive cell states (e.g. effector, NK-like)^36,37,92 in addition to the T_EX responses identified by Ki67^40-42. Thus, as proliferation, differentiation, and clonal expansion may have distinct temporal kinetics, clonal monitoring provides the ability to capture pharmacodynamic effects independent of proliferation. Interestingly, prior αCTLA-4 treatment appeared to blunt αPD-1 responses in terms of the magnitude of clonal responses and the diversity of cell states recruited. These data may explain why sequential therapy with nivolumab followed by ipilimumab was superior to ipilimumab followed by nivolumab⁹³, and are consistent with published data demonstrating the presence of distinct predictive features for αPD-1 depending on prior αCTLA-4 treatment history^94-96.

Finally, one of the advantages of this large dataset is the ability to study exhausted CD8⁺ T cell responses, despite their low frequency in the blood. Focused analyses of T_EX identified subsets previously defined in preclinical models, including progenitor, terminal, and intermediate/NK-like^{33,35,60-62,64-67,71-73,77}. Combination therapy induced major T_EX responses that peaked at 6 weeks, which included progenitor and likely intermediate/NK T_EX, which are known to be in circulation⁶⁰. We demonstrate that αCTLA-4 resulted in the expansion of Prog_EX. Thus, in the context of combination therapy, αCTLA-4 may sustain and expand the pool of Prog_EX while αPD-1 drives the differentiation of Prog_EX into Diff_EX.

Altogether, we performed large-scale human immune profiling to study the pharmacodynamic immune responses of checkpoint blockade in several dimensions, by assessing proliferation, changes in cellular frequency, and clonotypic responses. This multidimensional approach to immune monitoring has direct translational implications and can be used to quantify and modulate the magnitude, breadth, and durability of immune responses to novel therapeutic strategies, as well as optimize timings and dosages.

Limitations of the study

There are several limitations to keep in mind for this study. First, because we chose to deeply interrogate three treatment types across several timepoints, the numbers within each treatment type are small, and we were underpowered to find associations with clinical outcomes. This was especially true for the Cyclone analyses because of the requirement for Cyclone for a complete sample set across every timepoint at the required threshold of cell number. These observations need to be validated in larger clinical studies. Second, as many of the clinical progressors were too sick to provide serial samples, some of the cohorts were skewed towards clinical responders. This was especially true for patients treated with αCTLA-4. Finally, the lack of paired tumor specimens before and after treatment limited our ability to study the pharmacodynamics of intra-tumoral T cell clones in relation to the clonal trajectories in the blood. The application of Cyclone to neoadjuvant studies would be valuable.

STAR Methods:

Resource Availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Alexander C. Huang (alexander.huang@pennmedicine.upenn.edu)

Materials availability

This study did not generate new unique reagents.

Data and code availability

Raw and processed single-cell sequencing data generated in this study have been deposited in the National Center for Biotechnology Information Gene Expression Omnibus and are publicly available: (Main sequencing study: GSE272993, treatment-naïve aPD-1: GSE272734, healthy donor: GSE272735, paired blood tumor: GSE273718).

All original code for Cyclone has been deposited on GitHub: (https://github.com/jiwen90/cyclone) and Zenodo: (https://doi.org/10.5281/zenodo.12754871).

Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Experimental model and study participant details

Patients/Specimen Collection

Patients were consented for blood and tissue collection under the University of Pennsylvania Abramson Cancer Center’s melanoma research program tissue collection protocol UPCC 08607 and neoadjuvant clinical trial UPCC02619, in accordance with the Institutional Review Board. Peripheral blood was obtained in sodium heparin tubes at pre-treatment and before each infusion and on-treatment every 3 weeks. Peripheral blood mononuclear cells (PBMC) were isolated using Ficoll gradient and stored in liquid nitrogen. For specimens from Checkmate 238, PBMC were obtained from a randomly selected subset of patients following informed consent under an IRB-approved protocol at NYU.

All relevant clinical characteristics, including age and sex, are listed in Tables 1, S1, and S4.

Method details

Single-cell RNA, TCR, and ADT sequencing

For the 32-patient cohort, following processing of frozen aliquots, cells were stained and sorted via FACS (BD Aria II). For PBMC samples, 300,000 cells from each sample were sorted on a DAPI negative gate for live cells. CD3⁺ T cell samples were additionally gated on CD3⁺. Non-naïve T cell sorted samples were additionally gated as cells that were not CD45RA⁺CCR7⁺. Non-naïve CD4⁺ and CD8⁺ sorted samples were gated on CD4⁺ and CD8⁺, respectively, in addition to CD3⁺ and not CD45RA⁺CCR7⁺. Cells were then stained for 30 minutes at room temperature with either a panel of 60 or 101 Total-Seq-C antibodies that also labeled samples individually through sample-specific barcodes, and washed 2x using the HT1000 laminar wash system (Curiox). Cells were then counted using the Cellaca MX High-throughput Automated Cell Counter as described in the manufacturer’s protocol (Nexcelom), pooled and loaded on the 10x Chromium Next GEM Single Cell 5' Kit (v2) using a superloading strategy, mixing cells from different samples in each 10x lane to control for lane-to-lane batch effect. TCR sequences were enriched using the human V(D)J B/T cell enrichment kit. Libraries were prepared according to manufacturer’s protocol (10x Genomics) and sequenced on a NovaSeq 6000 System using the S4 2x150 kit (Illumina). Each sample/lane was sequenced to the average depth of 30,000 to 35,000 reads for gene expression and 10,000 to 15,000 reads for CITE-seq per cell.

For the treatment-naïve patients on αPD-1 and healthy donors, single-cell sequencing was performed on sorted CD3⁺CD8⁺ T cells from peripheral blood using the 10x Chromium Next GEM Single Cell 5’ Kit (v2). TCR sequences were enriched using the human V(D)J T cell enrichment kit. Sequencing was performed as described above.

For the samples used to construct the paired blood-tumor atlas, single-cell sequencing was performed on sorted CD4⁺ and CD8⁺ T cells from blood and CD45⁺ cells from tumor biopsies using the 10x Chromium Next GEM Single Cell 5’ Kit (v2). TCR sequences were enriched using the human V(D)J T cell enrichment kit. Sequencing was performed as described above.

Pre-processing of scRNA/TCR-seq libraries

Raw reads were aligned to the human transcriptome (refdata-cellranger-GRCh38-2020-A) to produce cell-by-gene matrices using Cell Ranger count (10x Genomics, version 7.0.0) with default parameters and a Feature Reference CSV file containing the feature barcode library if applicable. For VDJ, raw reads were aligned to the VDJ reference (refdata-cellranger-vdj-GRCh38-alts-ensembl-5.0.0) using Cell Ranger vdj. For the 32-patient cohort, single cells were dehashed into their respective samples by their sample-specific barcodes.

CD8⁺ T cell atlas construction

Out of the 36 patients in the single-cell sequencing cohort, the 32 patients on αCTLA-4, αPD-1 (prior ipi), and combination therapy were sequenced with one or more of the following 5 sorting strategies: PBMC, CD3⁺ T cells (CD3⁺), non-naïve T cells (not CD45RA⁺CCR7⁺), non-naïve CD4⁺ (CD4⁺ and not CD45RA⁺CCR7⁺), and non-naïve CD8⁺ T cells (CD8⁺ and not CD45RA⁺CCR7⁺). The number of patients for each sorting strategy combination were as follows: PBMC, non-naïve CD4⁺, and non-naïve CD8⁺ (2); PBMC and non-naïve T cells (13); CD3⁺ T cells (17). RNA, TCR, and ADT data were generated for all samples. The CD3⁺ T cell sort is referred to as PanT in the Seurat object metadata. The non-naïve T cell sort is referred to as Tcells in the Seurat object metadata.

Four additional treatment-naïve patients on αPD-1 and two healthy donors were sorted on CD3⁺CD8⁺ T cells. RNA and TCR data were generated for these samples.

Cells with greater than 200 detected genes, less than 2500 detected genes, and less than 10 percent mitochondrial reads were retained.

The 32-patient cohort was used for initial analysis. CD8⁺ T cell subsets were defined in silico based on protein expression differently for each sorting strategy. PBMC samples were subset on CD3⁺CD8⁺CD4⁻, CD3⁺ T cell sorted samples were subset on CD8⁺CD4⁻, and non-naïve T cell sorted samples were subset on CD8⁺CD4⁻. Non-naïve CD8⁺ T cell sorted samples were included as-is. Naïve cells were removed from PBMC and CD3⁺ T cell sorts by Azimuth reference mapping.

The standard Seurat workflow was performed separately for the CD3⁺ T cell sort and all other sorts: NormalizeData, FindVariableFeatures, ScaleData, RunPCA, RunHarmony, FindNeighbors, FindClusters, and RunUMAP. Leftover NK, MAIT, and monocytes were removed by unsupervised clustering and verification with Azimuth. Cells with high TCR-Va7.2 protein expression were excluded. Unsupervised clustering was used to identify clusters with low quality scores, which were removed.

A reference map was created with clean CD3⁺ T cell sort samples (Figure S2), using Harmony batch correction on patient, timepoint, and sort variables. For dimension reduction, ribosomal genes were excluded. The top 30 PCs were used for Harmony batch correction, and the first 10 Harmony dimensions were used for downstream analysis.

Unsupervised clustering was iteratively performed on the CD3⁺ T cell sort samples at different resolutions. For each resolution, the object was downsampled to 2000 cells per cluster. The resulting UMI matrix was fed into Hotspot⁴⁷, excluding genes expressed in less than 3 cells. The Hotspot algorithm was run using the top 10 Harmony embeddings as distance metric and subsequently on the top 3000 informative genes. Gene modules were created using agglomerative clustering with a minimum of 200 genes per module and FDR less than 0.05. Overrepresentation analysis was performed on the top 250 genes of each module, ranked by Z autocorrelation score using a hypergeometric test against the ConsensusPathDB database (http://cpdb.molgen.mpg.de/) with Bonferroni correction. Metaclusters were biologically defined as described in the text. All other samples, including the treatment-naïve αPD-1 patients and healthy donors, were mapped onto the reference map with Symphony⁴⁶ to create the final non-naïve CD8⁺ T cell atlas. Cluster labels were transferred to the query using the k-nearest neighbors prediction function in Symphony.

Gene module expression was calculated for each cell using the AddModuleScore function in Seurat, and a final clustering resolution of 0.5 was chosen. Differentially expressed geens were calculated using the FindAllMarkers function in Seurat using default values: a log2 fold change threshold of 0.25 and an adjusted p-value of less than 0.01.

Gene lists and defining exhausted CD8⁺ T cells

Bulk RNA-seq data from sorted CD8⁺ T cell subsets were obtained from a previously published paper⁵³. Differential expression analysis one-vs-all for each subset was performed. Patient ID and batch was added as a covariate to the design formula along with its intercept in the linear model in DESeq2. The top 250 genes, ranked by p-value, were chosen for each cell subset.

Seurat’s AddModuleScore was used to calculate an expression score for every cell. The top 3% of cells by exhaustion score were defined as T_EX.

Defining Progenitor T_EX in scRNA-seq data

Imputation was performed using ALRA⁹⁷ on the T_EX subset to smooth the central memory score mapped onto cells due to dropouts, and progenitor T_EX was defined as the top 1/3 of cells by central memory score.

Cyclone and TCR trajectory analyses

Only immunotherapy-naïve patients were included for analysis, in addition to the cohort of αPD-1 patients who received prior αCTLA-4 therapy with ipilimumab. One subject (14-1189_αCTLA-4) on αCTLA-4 was excluded from this analysis because of massive clonal expansion after checkpoint blockade that resulted in a distinctive clonal trajectory.

Only cells with both productive TCR alpha (TRA) and beta (TRB) chains were included (58.7% of CD8⁺ T cells), with clonotypes defined by the combination of patient ID, TRA, and TRB chain nucleotide sequences. Only TCR clonotypes with an absolute change of magnitude at least 0.25% and a fold change of greater than 2 or less than 0.5 in a 3-week interval were included. Clonotypes with less than 5 cells at every timepoints were filtered out because frequency variance is high with low counts due to Poisson sampling effects.

Increasing clonotypes with fold change greater than 2 were also required to have a peak frequency at a timepoint other than at baseline for inclusion. On the other hand, decreasing clonotypes with fold change of less than 0.5 were required to have a peak frequency at baseline. Pairwise Pearson correlation between each clonotype was calculated based on frequency at each timepoint, and trajectories were created using hierarchical clustering of pairwise Euclidean distances of correlation vectors.

The cell state composition of each trajectory was calculated by breaking down the underlying cell states contained within the cells in a trajectory at each timepoint for each patient, where the frequency is out of all cells in each respective patient sample. In other words, the sum of cell state frequencies for a trajectory at a timepoint equals the total frequency of that trajectory at that timepoint for each patient. Fold change of cell state frequencies were calculated with a 2% regularization term to prevent infinite fold changes. The average fold change across all patients was calculated for plotting.

HLA typing

HLA typing was performed by the University of Pennsylvania Histocompatibility and Immunogenetics Laboratory using one or more of the following molecular methods: SBT, SSP, SSO, RT-PCR, NGS.

Antigen-specific CD8⁺ T cell trajectory clustering

Unique antigen-specificities for each patient were treated as individual trajectories, which were binned into 4 patterns by the timepoint of maximum frequency (eg Traj 1: week 0, Traj 2: week 3, Traj 3: week 6, Traj 4: week 9+). In order to discover antigen-specific T cell populations that do not change over time, no filtering was performed based on fold change or absolute change as done in Cyclone; however, antigen-specificities with a fold change of less than 1.5 between week 0 and week 3 were binned into Traj 1.

Bulk TCR sequencing

Manual macrodissection was performed on FFPE slides, if necessary, using a scalpel and a slide stained with haematoxylin and eosin (H&E) as a guide. Tissue deparaffinization and DNA extraction were performed using standard methods. DNA amplification, library preparation, sequencing, and preliminary bioinformatics analysis was performed by Adaptive Biotechnologies. Amplification and sequencing of TCRB CDR3 was performed at single-level (ultra-deep) resolution using the immunoSEQ Platform (Adaptive Biotechnologies).

Tumor TCR trajectory clustering

Only productive TCRs were included. TCR beta chain CDR3 amino acid sequences were mapped onto the main single-cell sequencing study by exact match, for each of the 10 patients with bulk TCR data. Clones were binned by the timepoint of maximum frequency.

Pseudotime analysis

Pseudotime analysis was performed using Monocle 3 on the T_EX subset, retaining the original UMAP embeddings. Default parameters were used, except for use_partition set to FALSE, close_loop set to FALSE, and a minimal branch length of 8. Central memory-like cells were chosen as the root. A spline curve with 3 degrees of freedom was fit on gene and protein expression across pseudotime. Only the CD3⁺ sort was used for the protein expression analysis to reduce batch effects because the other sorts used a different ADT panel.

Blood-tumor atlas construction

Cell types were assigned by Azimuth reference mapping, and CD8⁺ T cells were subset in silico. The standard Seurat workflow was performed: NormalizeData, FindVariableFeatures, ScaleData, RunPCA, RunHarmony, FindNeighbors, FindClusters, and RunUMAP. Samples were batch corrected using Harmony by patient. Circulating exhausted cells from the main study were mapped onto the blood-tumor atlas using Symphony.

Flow cytometry

For samples processed at Penn, cryopreserved PBMC samples were thawed and stained with master mix of antibodies for surface stains including CD4 (Biolegend, SK3), CD8 (ebioscience, RPA-T8), CD45RA (Biolegend, HI100), Tim3 (F38-2E2), IgG4 (Southern Biotech, HP6025), CD39 (Biolegend, A1), CD3 (BD, UCHT1), CD127 (Biolegend, A019D5), CCR7 (Biologend, G043H7), CD28 (BD, CD28.2), CD226 (BD, DX11), CD73 (Biolegend, AD2), TIGIT (Biolegend, VSTM3) and CD27 (BD, L128) and intracellular stains for FoxP3 (Invitrogen, PCH101), CTLA4 (BD, BNI3), Eomes (Invitrogen, WD1928), Tbet (Biolegend, 4B10), Tox (Miltenyi, REA473), TCF-1 (Cell Signaling and Technology, C63D9), and Ki67(BD, B56). Permeabilization was performed using the Foxp3 Fixation/Permeabilization Concentrate and Diluent kit (eBioscience). Cells were resuspended in 1% para-formaldehyde until acquisition on a BD Biosciences LSR II cytometer or Symphony A5.

For samples stained with combinatorial tetramer panels at Penn, cryopreserved PBMC samples were thawed and stained with fluorophore-conjugated tetramers, followed by master mix of antibodies for surface stains, including CD45RA (BD Biosciences, HI100), Blue-Fluorescent Reactive Dye (Invitrogen), CD8 (BD Biosciences, RPA-T8), CD69 (BD Biosciences, FN50), CD226 (BD Biosciences, DX11), CXCR6 (BD Biosciences, 13B 1E5), CD28 (BD Biosciences, CD28.2), PD-1 (BioLegend, EH12.2H7), CD161 (BD Biosciences. DX12), CD14 (BD Biosciences, M5E2), CD19 (BD Biosciences, HIB19), CD41a (BD Biosciences, HIP8), CD4 (BioLegend, SK3), CD127 (BioLegend, A019D5), CCR7 (BioLegend, G043H7), CXCR5 (BD Biosciences, RF8B2), TIGIT (Invitrogen, MBSA43), CX3CR1 (BD Biosciences, 2A9-1), CD103 (BioLegend, Ber-ACT8), TIM-3 (BioLegend, F38-2E2), CD39 (BioLegend, A1), and CD27 (BioLegend, QA17A18) and intracellular stains for LAG-3 (BD Biosciences, T47-530), Ki-67 (BioLegend, Ki-67), CD3 (BioLegend, SK7), CTLA-4 (BioLegend, BNI3), FOXP3 (Invitrogen, PCH101), TBET (BioLegend, 4B10), TOX (Miltenyl Biotec, REA473), and TCF-7/TCF-1 (BD Biosciences, S33-966). Permeabilized was performed using the Foxp3 / Transcription Factor Staining Buffer Set (eBioscience). Cells were resuspension in 1% para-formaldehyde (Electron Microscopy Sciences) until acquisition on a five-laser Aurora cytometer (Cytek Biosciences) and unmixed using non-staining controls and single-color reference libraries.

For samples processed at NYU (Checkmate 238 trial), following thaw, PBMC underwent Fc blockade with Human TruStain FcX (BioLegend) and NovaBlock (Thermo Fisher Scientific) for 10 min at room temperature, followed by surface antibody staining for 20 min at room temperature in the dark. Cells were then permeabilized using the eBioscience Foxp3/Transcription Factor Staining Buffer Set (ThermoFisher) for 20 minutes at room temperature in the dark. Following permeabilization, cells were intracellularly stained for 1 hour at room temperature in the dark, followed by resuspension in 1% para-formaldehyde (Electron Microscopy Sciences). All samples were acquired on a five-laser Aurora cytometer (Cytek Biosciences).

All flow cytometry was analyzed using FlowJo (Becton Dickinson).

Quantification and statistical analysis

All statistical analyses were carried out using base R. Wilcoxon signed-rank test was used for paired comparisons. Wilcoxon rank-sum test was used for unpaired comparisons. Two-sided tests were used for all statistical comparisons. Statistical significance was as follows: p < 0.05 (*), 0.01 (**), 0.001 (***), 0.0001 (***). Error bars represent the 95^th percentile confidence interval.

Supplementary Material

2

Table S2: Hotspot Module Genes, Pathways, and Human CD8⁺ T cells Gene Lists. Related to Figure 1.

Hotspot module genes: List of all genes for each of the 4 gene modules generated by Hotspot.

Hotspot module pathways: List of top enriched pathways in the 4 gene modules generated by Hotspot.

Human CD8⁺ T cell Gene Lists: List of top 250 differentially expressed genes for each CD8⁺ T cell subset defined by sorted cells from human healthy donors, as described in Star Methods.

3

Table S3: Tumor TCR sequences. Related to Figures 3 and S6.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
PE/Fire™ 640 anti-human CD8	BioLegend	Cat# 344761; RRID: AB_2860887; clone SK1
PerCP anti-human CD11c	BioLegend	Cat# 337234; RRID: AB_2566656; clone Bu15
BD Horizon™ BUV805 Mouse Anti-Human CD14	BD Biosciences	Cat# 612902; clone M5E2
BD Horizon™ BUV496 Mouse Anti-Human CD19	BD Biosciences	Cat# 612939; clone 612939
PE/Dazzle™ 594 anti-human CD20	BioLegend	Cat# 302348; RRID: AB_2564387; clone 2H7
BD Pharmingen™ PE-Cy™5 Mouse Anti-Human CD21	BD Biosciences	Cat# 551064; RRID: AB_394028; clone B-ly4
BD OptiBuild™ BUV615 Mouse Anti-Human CD23	BD Biosciences	Cat# 751104; RRID: AB_2875134; clone M-L233
BD Horizon™ BUV563 Mouse Anti-Human CD25	BD Biosciences	Cat# 612919; RRID: AB_2870204; clone 2A3
Brilliant Violet 711™ anti-human CD27 Antibody	BioLegend	Cat# 302834; RRID: AB_2563809; clone O323
BD Horizon™ BV480 Mouse Anti-Human CD28	BD Biosciences	Cat# 566110; RRID: AB_2739512; clone CD28.2
APC/Fire™ 810 anti-human CD38	BioLegend	Cat# 303549; RRID: AB_2860783; clone HIT2
Brilliant Violet 605™ anti-human CD39	BioLegend	Cat# 328236; RRID: AB_2750430; clone A1
Pacific Blue™ anti-human CD40	BioLegend	Cat# 334319; RRID: AB_10612573; clone 5C3
Brilliant Violet 570™ anti-human CD45RA	BioLegend	Cat# 304132; RRID: AB_2563813; clone HI100
Super Bright™ 780 CD71 (Transferrin Receptor)	Invitrogen	Cat# 78-0719-42; RRID: AB_2784898; clone OKT9 (OKT-9)
Spark YG™ 581 anti-human CD127 (IL-7Rα)	BioLegend	Cat# 351367; RRID: AB_2890778; clone A019D5
Brilliant Violet 510™ anti-human CD138 (Syndecan-1)	BioLegend	Cat# 356517; RRID: AB_2562661; clone MI15
Brilliant Violet 421™ anti-human CD152 (CTLA-4)	BioLegend	Cat# 369606; RRID: AB_2616795; clone BNI3
BD OptiBuild™ BV750 Mouse Anti-Human CD183 (CXCR3)	BD Biosciences	Cat# 746895; RRID: AB_2871692; clone 1C6/CXCR3
BD Horizon™ APC-R700 Rat Anti-Human CXCR5 (CD185)	BD Biosciences	Cat# 565191; RRID: AB_2739103; clone RF8B2
Brilliant Violet 650™ anti-human CD197 (CCR7)	BioLegend	Cat# 353234; RRID: AB_2563867; clone G043H7
APC/Fire™ 750 anti-human/mouse/rat CD278 (ICOS)	BioLegend	Cat# 313536; RRID: AB_2632923; clone C398.4A
BD Horizon™ BB700 Mouse Anti-Human CD279 (PD-1)	BD Biosciences	Cat# 566460; RRID: AB_2744348; clone EH12.1
BD Horizon™ BUV661 Mouse Anti-Human HLA-DR	BD Biosciences	Cat# 612980; RRID: AB_2870252; clone G46-6
BD Horizon™ BUV737 Mouse Anti-Human IgD	BD Biosciences	Cat# 612798; RRID: AB_2870125; clone IA6-2
REAfinity™ PE anti-human/mouse TOX	Miltenyi Biotec	Cat# 130-120-716; RRID: AB_2801780; clone REA473
TCF1/TCF7 (C63D9) Rabbit mAb (Alexa Fluor® 647 Conjugate)	Cell Signaling	Cat# 6709S; RRID: AB_2797631; clone C63D9
BD Horizon™ BUV395 Mouse Anti-Ki-67	BD Biosciences	Cat# 564071; RRID: AB_2738577; clone B56
Anti-human IgG	Miltenyi Biotec	Cat# 130-119-880; RRID: AB_2784374; clone IS11-3B2.2.3
Mouse Anti-Human IgG4 pFc'-FITC	SouthernBiotech	Cat# 9190-02; RRID: AB_2796682; clone HP6023
BD Horizon™ BUV395 Mouse Anti-Human CD45RA	BD Biosciences	Cat# 568712; clone HI100
LIVE/DEAD™ Fixable Blue Dead Cell Stain Kit, for UV excitation	Invitrogen	Cat# L23105; RRID: AB_2796682
BD Horizon™ BUV496 Mouse Anti-Human CD8	BD Biosciences	Cat# 612942; clone RPA-T8
BD OptiBuild™ BUV563 Mouse Anti-Human CD69	BD Biosciences	Cat# 748764; RRID: AB_2873167; clone FN50
BD OptiBuild™ BUV661 Mouse Anti-Human CD226	BD Biosciences	Cat# 749934; RRID: AB_2874171; clone DX11
BD OptiBuild™ BUV737 Mouse Anti-Human CXCR6 (CD186)	BD Biosciences	Cat# 748449; RRID: AB_2872865; clone 13B 1E5
BD OptiBuild™ BUV805 Mouse Anti-Human CD28	BD Biosciences	Cat# 742037; RRID: AB_2871331; clone CD28.2
BD Horizon™ BV421 Mouse Anti-Human CD279 (PD-1)	BD Biosciences	Cat# 565935; RRID: AB_11153482; clone EH12.1
BD OptiBuild™ BV480 Mouse Anti-Human CD161	BD Biosciences	Cat# 746305; RRID: AB_2743630; clone DX12
BD Horizon™ V500 Mouse Anti-Human CD14	BD Biosciences	Cat# 562693; RRID: AB_2737727; clone MφP9
BD Horizon™ V500 Mouse anti-Human CD19	BD Biosciences	Cat# 561125; RRID: AB_10562391; clone HIB19
D Horizon™ BV510 Mouse Anti-Human CD41a	BD Biosciences	Cat# 563250; RRID: AB_2738096; clone HIP8
Spark Violet™ 538 anti-human CD4 Antibody	BioLegend	Cat# 344673; RRID: AB_2890774; clone SK3
Brilliant Violet 570™ anti-human CD127 (IL-7Rα) Antibody	BioLegend	Cat# 351308; RRID: AB_2832685 ; clone A019D5
Brilliant Violet 750™ anti-human CD197 (CCR7) Antibody	BioLegend	Cat# 353253; RRID: AB_2800945; clone G043H7
BD Horizon™ BB700 Rat Anti-Human CXCR5 (CD185)	BD Biosciences	Cat# 566469; RRID: AB_2869769; clone RF8B2
TIGIT Monoclonal Antibody (MBSA43), PerCP-eFluor™ 710, eBioscience™	Invitrogen	Cat# 46-9500-42; RRID: AB_10853679; clone MBSA43
BD OptiBuild™ BB790 Rat Anti-Human CX3CR1	BD Biosciences	Custom design; clone 2A9-1
PE/Fire™ 640 anti-human CD103 (Integrin αE) Antibody	BioLegend	Cat# 350243; RRID: AB_2924550 ; clone Ber-ACT8
PE/Cyanine5 anti-human CD366 (Tim-3) Antibody	BioLegend	Cat# 345052; RRID: AB_2819987 ; clone F38-2E2
APC/Fire™ 750 anti-human CD39 Antibody	BioLegend	Cat# 328229; RRID: AB_2650838 ; clone A1
APC/Fire™ 810 anti-human CD27 Recombinant Antibody	BioLegend	Cat# 393213; RRID: AB_2860961; clone QA17A18
BD OptiBuild™ BUV615 Mouse Anti-Human LAG-3 (CD223)	BD Biosciences	Cat# 752362; RRID: AB_2875879; clone T47-530
Brilliant Violet 711™ anti-human Ki-67 Antibody	BioLegend	Cat# 350516; RRID: AB_2563861; clone Ki-67
Spark Blue™ 550 anti-human CD3 Antibody	BioLegend	Cat# 344852; RRID: AB_2819985; clone SK7
PE/Dazzle™ 594 anti-human CD152 (CTLA-4) Antibody	BioLegend	Cat# 369616; RRID: AB_2632878; clone BNI3
TOX Antibody, anti-human/mouse, APC, REAfinity™	Miltenyi Biotec	Cat# 130-118-335; clone REA473
BD Horizon™ R718 Mouse Anti-TCF-7/TCF-1	BD Biosciences	Cat# 567587; clone S33-966
Brilliant Violet 570™ anti-human CD3 Antibody	BioLegend	Cat# 300435; RRID: AB_10898117; clone UCHT1
PE anti-mouse TIGIT (Vstm3) Antibody	BioLegend	Cat# 142103; RRID: AB_10895760; clone 1G9
BD Horizon™ BUV737 Mouse Anti-Human CD27	BD Biosciences	Cat# 612829; RRID: AB_2870151; clone L128
EOMES Monoclonal Antibody (WD1928), PE-eFluor™ 610, eBioscience™	Invitrogen	Cat# 61-4877-42; clone WD1928
BD Pharmingen™ Alexa Fluor® 700 Mouse anti-Ki-67	BD Biosciences	Cat# 561277; RRID: AB_10611571; clone B56
Spark Violet™ 538 anti-human CD4	BioLegend	Cat# 344673; RRID: AB_2890774; clone SK3
PE/Fire™ 640 anti-human CD8	BioLegend	Cat# 344761; RRID: AB_2860887; clone SK1
PerCP anti-human CD11c	BioLegend	Cat# 337234; RRID: AB_2566656; clone Bu15
BD Horizon™ BUV805 Mouse Anti-Human CD14	BD Biosciences	Cat# 612902; clone M5E2
BD Horizon™ BUV496 Mouse Anti-Human CD19	BD Biosciences	Cat# 612939; clone 612939
PE/Dazzle™ 594 anti-human CD20	BioLegend	Cat# 302348; RRID: AB_2564387; clone 2H7
BD Pharmingen™ PE-Cy™5 Mouse Anti-Human CD21	BD Biosciences	Cat# 551064; RRID: AB_394028; clone B-ly4
BD OptiBuild™ BUV615 Mouse Anti-Human CD23	BD Biosciences	Cat# 751104; RRID: AB_2875134; clone M-L233
BD Horizon™ BUV563 Mouse Anti-Human CD25	BD Biosciences	Cat# 612919; RRID: AB_2870204; clone 2A3
Brilliant Violet 711™ anti-human CD27 Antibody	BioLegend	Cat# 302834; RRID: AB_2563809; clone O323
BD Horizon™ BV480 Mouse Anti-Human CD28	BD Biosciences	Cat# 566110; RRID: AB_2739512; clone CD28.2
APC/Fire™ 810 anti-human CD38	BioLegend	Cat# 303549; RRID: AB_2860783; clone HIT2
Brilliant Violet 605™ anti-human CD39	BioLegend	Cat# 328236; RRID: AB_2750430; clone A1
Pacific Blue™ anti-human CD40	BioLegend	Cat# 334319; RRID: AB_10612573; clone 5C3
Brilliant Violet 570™ anti-human CD45RA	BioLegend	Cat# 304132; RRID: AB_2563813; clone HI100
Super Bright™ 780 CD71 (Transferrin Receptor)	Invitrogen	Cat# 78-0719-42; RRID: AB_2784898; clone OKT9 (OKT-9)
Spark YG™ 581 anti-human CD127 (IL-7Rα)	BioLegend	Cat# 351367; RRID: AB_2890778; clone A019D5
Brilliant Violet 510™ anti-human CD138 (Syndecan-1)	BioLegend	Cat# 356517; RRID: AB_2562661; clone MI15
Brilliant Violet 421™ anti-human CD152 (CTLA-4)	BioLegend	Cat# 369606; RRID: AB_2616795; clone BNI3
BD OptiBuild™ BV750 Mouse Anti-Human CD183 (CXCR3)	BD Biosciences	Cat# 746895; RRID: AB_2871692; clone 1C6/CXCR3
BD Horizon™ APC-R700 Rat Anti-Human CXCR5 (CD185)	BD Biosciences	Cat# 565191; RRID: AB_2739103; clone RF8B2
Brilliant Violet 650™ anti-human CD197 (CCR7)	BioLegend	Cat# 353234; RRID: AB_2563867; clone G043H7
APC/Fire™ 750 anti-human/mouse/rat CD278 (ICOS)	BioLegend	Cat# 313536; RRID: AB_2632923; clone C398.4A
BD Horizon™ BB700 Mouse Anti-Human CD279 (PD-1)	BD Biosciences	Cat# 566460; RRID: AB_2744348; clone EH12.1
BD Horizon™ BUV661 Mouse Anti-Human HLA-DR	BD Biosciences	Cat# 612980; RRID: AB_2870252; clone G46-6
BD Horizon™ BUV737 Mouse Anti-Human IgD	BD Biosciences	Cat# 612798; RRID: AB_2870125; clone IA6-2
PE-Cyanine5.5 FOXP3	Invitrogen	Cat# 35-4776-42; RRID: AB_11218682; clone PCH101
REAfinity™ PE anti-human/mouse TOX	Miltenyi Biotec	Cat# 130-120-716; RRID: AB_2801780; clone REA473
TCF1/TCF7 (C63D9) Rabbit mAb (Alexa Fluor® 647 Conjugate)	Cell Signaling	Cat# 6709S; RRID: AB_2797631; clone C63D9
BD Horizon™ BUV395 Mouse Anti-Ki-67	BD Biosciences	Cat# 564071; RRID: AB_2738577; clone B56
Anti-human IgG	Miltenyi Biotec	Cat# 130-119-880; RRID: AB_2784374; clone IS11-3B2.2.3
Mouse Anti-Human IgG4 pFc'-FITC	SouthernBiotech	Cat# 9190-02; RRID: AB_2796682; clone HP6023
BD Horizon™ BUV395 Mouse Anti-Human CD45RA	BD Biosciences	Cat# 568712; clone HI100
LIVE/DEAD™ Fixable Blue Dead Cell Stain Kit, for UV excitation	Invitrogen	Cat# L23105; RRID: AB_2796682
BD Horizon™ BUV496 Mouse Anti-Human CD8	BD Biosciences	Cat# 612942; clone RPA-T8
BD OptiBuild™ BUV563 Mouse Anti-Human CD69	BD Biosciences	Cat# 748764; RRID: AB_2873167; clone FN50
BD OptiBuild™ BUV661 Mouse Anti-Human CD226	BD Biosciences	Cat# 749934; RRID: AB_2874171; clone DX11
BD OptiBuild™ BUV737 Mouse Anti-Human CXCR6 (CD186)	BD Biosciences	Cat# 748449; RRID: AB_2872865; clone 13B 1E5
BD OptiBuild™ BUV805 Mouse Anti-Human CD28	BD Biosciences	Cat# 742037; RRID: AB_2871331; clone CD28.2
BD Horizon™ BV421 Mouse Anti-Human CD279 (PD-1)	BD Biosciences	Cat# 565935; RRID: AB_11153482; clone EH12.1
BD OptiBuild™ BV480 Mouse Anti-Human CD161	BD Biosciences	Cat# 746305; RRID: AB_2743630; clone DX12
BD Horizon™ V500 Mouse Anti-Human CD14	BD Biosciences	Cat# 562693; RRID: AB_2737727; clone MφP9
BD Horizon™ V500 Mouse anti-Human CD19	BD Biosciences	Cat# 561125; RRID: AB_10562391; clone HIB19
D Horizon™ BV510 Mouse Anti-Human CD41a	BD Biosciences	Cat# 563250; RRID: AB_2738096; clone HIP8
Spark Violet™ 538 anti-human CD4 Antibody	BioLegend	Cat# 344673; RRID: AB_2890774; clone SK3
Brilliant Violet 570™ anti-human CD127 (IL-7Rα) Antibody	BioLegend	Cat# 351308; RRID: AB_2832685 ; clone A019D5
Brilliant Violet 750™ anti-human CD197 (CCR7) Antibody	BioLegend	Cat# 353253; RRID: AB_2800945; clone G043H7
BD Horizon™ BB700 Rat Anti-Human CXCR5 (CD185)	BD Biosciences	Cat# 566469; RRID: AB_2869769; clone RF8B2
TIGIT Monoclonal Antibody (MBSA43), PerCP-eFluor™ 710, eBioscience™	Invitrogen	Cat# 46-9500-42; RRID: AB_10853679; clone MBSA43
BD OptiBuild™ BB790 Rat Anti-Human CX3CR1	BD Biosciences	Custom design; clone 2A9-1
PE/Fire™ 640 anti-human CD103 (Integrin αE) Antibody	BioLegend	Cat# 350243; RRID: AB_2924550 ; clone Ber-ACT8
PE/Cyanine5 anti-human CD366 (Tim-3) Antibody	BioLegend	Cat# 345052; RRID: AB_2819987 ; clone F38-2E2
APC/Fire™ 750 anti-human CD39 Antibody	BioLegend	Cat# 328229; RRID: AB_2650838 ; clone A1
APC/Fire™ 810 anti-human CD27 Recombinant Antibody	BioLegend	Cat# 393213; RRID: AB_2860961; clone QA17A18
BD OptiBuild™ BUV615 Mouse Anti-Human LAG-3 (CD223)	BD Biosciences	Cat# 752362; RRID: AB_2875879; clone T47-530
Brilliant Violet 711™ anti-human Ki-67 Antibody	BioLegend	Cat# 350516; RRID: AB_2563861; clone Ki-67
Spark Blue™ 550 anti-human CD3 Antibody	BioLegend	Cat# 344852; RRID: AB_2819985; clone SK7
PE/Dazzle™ 594 anti-human CD152 (CTLA-4) Antibody	BioLegend	Cat# 369616; RRID: AB_2632878; clone BNI3
FOXP3 Monoclonal Antibody (PCH101), PE-Cyanine5.5, eBioscience™	Invitrogen	Cat# 35-4776-42; RRID: AB_11218682; clone PCH101
PE/Cyanine7 anti-T-bet Antibody	BioLegend	Cat# 644823; RRID: AB_2561760; clone 4B10
TOX Antibody, anti-human/mouse, APC, REAfinity™	Miltenyi Biotec	Cat# 130-118-335; clone REA473
BD Horizon™ R718 Mouse Anti-TCF-7/TCF-1	BD Biosciences	Cat# 567587; clone S33-966
Brilliant Violet 570™ anti-human CD3 Antibody	BioLegend	Cat# 300435; RRID: AB_10898117; clone UCHT1
PE anti-mouse TIGIT (Vstm3) Antibody	BioLegend	Cat# 142103; RRID: AB_10895760; clone 1G9
BD Horizon™ BUV737 Mouse Anti-Human CD27	BD Biosciences	Cat# 612829; RRID: AB_2870151; clone L128
EOMES Monoclonal Antibody (WD1928), PE-eFluor™ 610, eBioscience™	Invitrogen	Cat# 61-4877-42; clone WD1928
TCF1/TCF7 (C63D9) Rabbit mAb (Alexa Fluor® 647 Conjugate)	Cell Signaling and Technology	Cat# 6709S; clone C63D9
BD Pharmingen™ Alexa Fluor® 700 Mouse anti-Ki-67	BD Biosciences	Cat# 561277; RRID: AB_10611571; clone B56
APC/Cyanine7 anti-human CD3 Antibody	BioLegend	Cat# 344817; RRID: AB_10644011; clone SK7
PerCP anti-human CD4 Antibody	BioLegend	Cat# 344623; RRID: AB_2563325; clone SK3
Alexa Fluor® 700 anti-human CD8 Antibody	BioLegend	Cat# 344723; RRID: AB_2562789; clone SK1
FITC anti-human CD45RA Antibody	BioLegend	Cat# 304105; RRID: AB_314409; clone HI100
PE anti-human CD197 (CCR7) Antibody	BioLegend	Cat# 353203; RRID: AB_10916391; clone G043H7
Custom TotalSeq™-C Human Cocktail	BioLegend	Custom
Biological Samples
Human healthy donor PBMC	Human Immunology Core, University of Pennsylvania	N/A
Patient PBMC and Tumor	University of Pennsylvania	IRB
Chemicals, Peptides, and Recombinant Proteins
CMV pp65 (YSE) YSEHPTFTSQY	GenScript	Custom peptide synthesis
FluNP (CTE) CTELKLSDY	GenScript	Custom peptide synthesis
CMV pp50 (VTE) VTEHDTLLY	GenScript	Custom peptide synthesis
Tyrosinase (SSD) SSDYVIPIGTY	GenScript	Custom peptide synthesis
MAGE A1 (EAD) EADPTGHSY	GenScript	Custom peptide synthesis
MAGE A3 (EVD) EVDPIGHLY	GenScript	Custom peptide synthesis
AIM2 (RSD) RSDSGQQARY	GenScript	Custom peptide synthesis
N-RAS Q61R (ILD) ILDTAGREEY	GenScript	Custom peptide synthesis
CMV pp65 (NLV) NLVPMVATV	GenScript	Custom peptide synthesis
MART-1 (ELA) ELAGIGILTV	GenScript	Custom peptide synthesis
gp100 280-9V (YLE) YLEPGPVTV	GenScript	Custom peptide synthesis
gp100 209-2M (IMD) IMDQVPFSV	GenScript	Custom peptide synthesis
gp100 (KTW) KTWGQYWQV	GenScript	Custom peptide synthesis
GnTV (VLP) VLPDVFIRCV	GenScript	Custom peptide synthesis
LAGE-1 (MLM) MLMAQEALAFL	GenScript	Custom peptide synthesis
SSX-2 (KAS) KASEKIFYV	GenScript	Custom peptide synthesis
NY-ESO-1 (SLL) SLLMWITQA	GenScript	Custom peptide synthesis
PRDX5 (AMA) AMAPIKVRL	GenScript	Custom peptide synthesis
CMV pp50 (TVR) TVRSHCVSK	GenScript	Custom peptide synthesis
Flu M1 (SII) SIIPSGPLK	GenScript	Custom peptide synthesis
EBV BRLF1 (RVR) RVRAYTYSK	GenScript	Custom peptide synthesis
EBV EBNA 3A (RLR) RLRAEAQVK	GenScript	Custom peptide synthesis
gp100 (ALL) ALLAVGATK	GenScript	Custom peptide synthesis
gp100 (LIY) LIYRRRLMK	GenScript	Custom peptide synthesis
gp100 (ALN) ALNFPGSQK	GenScript	Custom peptide synthesis
TAG (RLS) RLSNRLLLR	GenScript	Custom peptide synthesis
MAGE A1 (SLF) SLFRAVIIK	GenScript	Custom peptide synthesis
Flex-T™ HLA-A*01:01 Monomer UVX	BioLegend	Cat# 280001
Flex-T™ HLA-A*02:01 Monomer UVX	BioLegend	Cat# 280003
Flex-T™ HLA-A*03:01 Monomer UVX	BioLegend	Cat# 280005
Flex-T™ HLA-B*07:02 Monomer UVX	BioLegend	Cat# 280009
Critical Commercial Assays
eBioscience™ Foxp3 / Transcription Factor Staining Buffer Set	Invitrogen	Cat# 00-5523-00
Next GEM Chip K	10x Genomics	Cat# PN-1000286
Next GEM Single Cell 5' Kit v2	10x Genomics	Cat# PN-1000263
Chromium Single Cell V(D)J Enrichment Kit, Human T Cell	10x Genomics	Cat# PN-1000005
Chromium Single Cell 5' Feature Barcode Library Kit	10x Genomics	Cat# PN-1000080
NovaSeq 6000 S4 Reagent Kit v1.5	Illumina	Cat# 20028312
Deposited Data
This work (scRNA/TCR/ADT-seq of main 32-patient study)	NIH GEO	GSE272993
This work (scRNA/TCR-seq of treatment-naive patients on aPD-1)	NIH GEO	GSE272734
This work (scRNA/TCR-seq of healthy donors)	NIH GEO	GSE272735
This work (scRNA-seq of paired blood-tumor atlas)	NIH GEO	GSE273718
Software and Algorithms
R	R	Version 4.3.1
Python	Python	Version 3.9.13
FlowJo	BD Biosciences	Version 10.10.0
CellRanger	10x Genomics	Version 7.0.0
Cyclone	This paper	10.5281/zenodo.12754871
Seurat	https://github.com/jiwen90/seurat	Version 4.3.0
Other
CD8 Human Gene Lists, see Table S2	This paper	N/A
Bulk TCR-seq of TILs, see Table S3	This paper	N/A
Multi-database gene pathways	http://cpdb.molgen.mpg.de/	N/A

Highlights.

• A novel algorithm Cyclone uncovers patterns of clonal trajectories across time

• Immune checkpoint blockade induces distinct waves of clonal responses after each dose

• Melanoma-specific CD8⁺ T cell responses are present at 3 and 6 weeks after treatment

• CTLA-4 blockade induces robust progenitor-exhausted CD8⁺ T cell responses