Acute Myeloid Leukemia Skews Therapeutic WT1-specific CD8 TCR-T Cells Towards an NK-like Phenotype that Compromises Function and Persistence

T_TCR-C4 is safe, well tolerated, and produces comparable outcomes to conventional treatments for high-risk relapsed/refractory AML patients post-HCT

From December 2012 through March 2020, 15 HLA-A2-expressing patients with relapsed and/or refractory AML post-HCT were enrolled on trial NCT01640301 (Table 1, Extended data Fig. 1). The median age at diagnosis was 40. Pre-HCT, 27% of patients had secondary or treatment-related AML, at diagnosis 47% were adverse risk, and 53% favorable/intermediate risk per European LeukemiaNet (ELN) stratification (Supplementary Table 1).²⁶ Post-HCT, four patients had measurable residual disease (MRD) by ∼28 days (refractory), two relapsed by 3 months, and nine relapsed after day 100. The median time to relapse was 496 days post-HCT. Thirteen patients received salvage therapy before T_TCR-C4, including six who underwent a second HCT. Within a median of two weeks before infusion, two patients had overt disease, five were MRD-positive and the rest had no evaluable disease (NED) (Extended data Fig. 2). Ten patients received EBV-specific T_TCR-C4 and five received CMV-specific T cells (Table 1). Patients received one to four infusions depending on their place in the T_TCR-C4 dose escalation (Extended data Fig. 3).

• Download figure
• Open in new tab

Extended data Fig. 1: Flow diagram of patients enrolled on the clinical study.

Follow-up of the 67 patients screened for participation in the study and the 44 patients who signed the informed consent. Patients with no detectable disease post-HCT were assigned to the Prophylactic Arm¹, while those with evidence of disease post-HCT were treated in the Treatment Arm described here. A total of 44 pt/donor pairs were enrolled pre- HCT, TCRC4 transduced cells were initiated for 38 patients, among whom 12 disease-free (Arm 1) and 15 relapsed/refractory (Arm 2) patients received TCRC4 transduced cells. An additional patient was considered not evaluable for response outcomes as this patient received <10 % of the targeted dose and was unable to receive further infusions due to inability to generate additional T cells from the donor.

• Download figure
• Open in new tab

Extended data Fig. 2: Flow diagram of patients’ treatment history.

The cohort included 15 patients who relapsed after HCT. 7 patients received salvage chemotherapy before the infusion of T_TCR-C4, while 6 patients had two HCTs before the infusion, and 1 patient among them also received salvage chemotherapy. At the time of TCR-C4 infusion, 7 patients had detectable AML, while 8 patients were in remission (AML-negative). The time from relapse post-HCT to TCR-C4 infusion ranged from 34 to 700 days, with a median of 179 days.

• Download figure
• Open in new tab

Extended data Fig. 3. Study schema.

The first 8 patients who received T_TCR-C4 after post-HCT relapse (Treatment Arm, Stage1) received 4 escalating infusions of 10⁹-10¹⁰/m² 14 days apart, the last infusion followed by s.c. low-dose IL2 for 14 days. After safety of the highest dose was established, all subsequent patients received 2 infusions of 10¹⁰/m² T_TCR-C4 28 days apart, the second infusion followed by s.c. low-dose IL2 for 14 days.

As in Arm 1,⁹ T_TCR-C4 infusions were safe and well tolerated (Supplementary Table 2). The incidence of acute and chronic GVHD (aGVHD, cGVHD) was lower compared to DLIs or second HCTs as salvage strategies,^11,27 with only one patient (7%) developing grade 3 aGVHD and three (20%) developing cGVHD post-T_TCR-C4 infusion. Post-infusion cGVHD included one grade 3, one grade 2 and one grade 1 cGVHD. However, the onset of cGVHD occurred when T_TCR-C4 constituted <1% of the peripheral blood (PB) CD8⁺ T cells, making the infusion of T_TCR-C4 with an endogenous virus-specific TCR the unlikely cause of cGVHD.

Among the patients enrolled (Table 1), four showed no overt relapse and/or MRD following TTCR- C4 infusion(s) (Extended data Fig. 4) supporting T_TCR-C4’s potential biologic activity. However, when all 15 patients were analyzed together, the median OS was 242 days, the 2-year OS was 33%, and the 3-year OS was 20% (Extended data Fig. 5), indicating that these individual outcomes did not translate into a significant survival advantage over historical salvage treatments.^1,2,10,11 For example, DLI post-HCT relapse showed a 2-year OS of 21% (±3%) from relapse and 56% (±10%) from DLI in cases of remission or with a favorable karyotype, dropping to 15% (±3%) in aplasia or active disease.²⁷

• Download figure
• Open in new tab

Extended data Fig. 4. Swimmer Plot of Patient Responses.

The swimmer plot illustrates the response duration for patients enrolled in the trial. Time 0 represents the relapse or persistent disease after the first HCT that qualified patients for trial inclusion. The end of each bar indicates the date of death or last follow-up, with arrows marking ongoing responses. "Salvage" denotes any salvage strategy, including chemotherapy, radiotherapy, or intrathecal treatment. Salvage with a second HCT is marked by a triangle.

Patient 8: A patient who had MRD post HCT and received T_TCR-C4 with MRD. She later underwent azacitidine salvage treatment for persistent MRD (duration shown as a gray line). The patient then relapsed, underwent additional salvage chemotherapy, azacitidine, and received a second T_TCR-C4 infusion. Following this, she experienced prolonged MRD+ stable disease with detectable long-term persisting PB T_TCR-C4. Eventually, the disease progressed, and she relapsed and died (see Results for more details).

Patient 6: A patient who received salvage chemotherapy followed by a second HCT which resulted in no evaluable disease. Subsequently, she received four T_TCR-C4 infusions. The patient was alive at the last follow-up.

Patient 2: A patient who, after salvage treatment, underwent a second HCT while in overt disease. She remained MRD-positive 28 days post- second-HCT. She received four T_TCR-C4 infusions. Eventually, she achieved a MRD negative status. The patient was alive at the last follow-up.

Patient 4: A patient who achieved a complete remission after a second HCT. However, after ∼15 months, he experienced extramedullary relapse, treated with azacitidine and radiotherapy which achieved no evaluable disease. He maintained remission for ∼1 year after the first T_TCR-C4 infusion. Subsequently, he relapsed extramedullary and was refractory to an additional T_TCR-C4 infusion. Further details on this case and mechanisms of AML escape have been described previously.²

• Download figure
• Open in new tab

Extended data Fig. 5. Overall survival of AML patients treated with T_TCRC4 T cells.

Kaplan Meier estimate of overall survival (OS) of 15 AML patients with evidence of disease post-HCT, who were subsequently treated with T_TCRC4 T cells

Virus-specific substrate T cells and AML presence at the time of infusion are major determinants of T_TCR-C4 persistence

In our cohort, baseline PB WT1-specific tetramer⁺ T cells were low (Fig. 1A – red arrows, day 0) indicating limited endogenous WT1-specific T cells. By day 28 post-infusion, four patients showed TTCR- C4 levels exceeding 3% (persistence threshold) of the CD8⁺ T cells, with three maintaining these levels beyond day 100. Higher frequencies correlated with increased absolute T_TCR-C4 counts (Extended data Fig. 6A). Patients with T_TCR-C4 frequencies below 3% were eligible for additional infusions (two infusions in 6/15 patients, >2 infusions in 4/15 patients) (Table 1). However, more than two infusions did not significantly increase persistence (Fig. 1B).

• Download figure
• Open in new tab

Extended data Fig. 6. EBV specific vs. CMV specific T_TCR-C4 persistence.

(A) Line plots showing the absolute counts (cells/uL) in log scale of PB multimer+ CD8⁺ T cells collected after T_TCR-C4 infusion and stratified by virus specificity. EBV-specific cells are represented in dark orange, and CMV-specific cells are represented in dark violet. Data points correspond to individual samples collected at different timepoints post-infusion. (B) Boxplots comparing the percentage of EBV-specific T_TCR-C4 cells in Arm 1 (red) and Arm 2 (blue) with the percentage of CMV-specific T_TCR-C4 cells in Arm 2 (green) across post-infusion timepoints. The y-axis indicates the percentage (log scale) of T_TCR-C4 cells, while the x-axis represents the time post-infusion. Data were derived from flow-cytometry analysis. The horizontal dashed red line indicates the 3% threshold used to define persisting T_TCR-C4 cells. The y-axis is displayed on a log10 scale. Statistical comparison between all three groups at each time point was performed using a Kruskal-Wallis test, with p-values displayed above the corresponding boxplots. Statistical significance was defined as p < 0.05.

• Download figure
• Open in new tab

Fig. 1 Virus and time-dependent terminal differentiation skewing of T_TCRC4.

(A) Line plot showing the percentage (log scale) of T_TCRC4 in PBMCs collected after the first T_TCR-C4 infusion for all patients (n =15), with data points representing individual samples at different timepoints post-infusion. Each patient is represented by a distinct color, and red arrows indicate the percentage of WT1-specific CD8⁺ T cells already present at day 0. (B) Boxplots comparing the percentage of T_TCRC4 (log scale, y-axis) after the first infusion (x-axis), colored by the number of infusions (light blue for <= 2 infusions; dark orange > 2 infusions). The horizontal dashed red line indicates the 3% threshold used to define persisting T_TCRC4 cells. Statistical significance was determined using Kruskal-Wallis test, with p < 0.05 considered significant. (C) Line plot illustrating the percentage of T_TCRC4 (y-axis) cells derived from EBV-specific substrate cells (dark orange) or CMV-specific substrate cells (dark violet) over a 28-day period (x-axis) after the first infusion. Statistical significance was determined using Wilcoxon Rank Sum test. (* p < 0.05, ** p < 0.01) (D) Boxplots showing differential abundance analysis of CMV-specific (dark violet) and EBV-specific (dark orange) CD8⁺ T-cell subsets derived from the analysis of mass cytometry data.²⁹ CD8⁺ T-cell subsets are defined based on marker co-expression showed in Extended data Fig. 7B. A generalized linear model was used to compute differential cluster abundance between the two conditions (CMV vs. EBV), with adjusted p-values (threshold for significance of p<0.05) displayed above each comparison.

• Download figure
• Open in new tab

Extended data Fig. 7. EBV and CMV specific CD8⁺ T-cell states.

(A) Multidimensional (MDS) scaling plot illustrates the distribution of samples based on the expression of top varying markers, which are those markers showing the most variation in expression levels across the dataset. These markers are selected for their ability to highlight meaningful differences between samples or conditions. In this context, samples associated with CMV (violet dots) and EBV (orange dots) are plotted along the first two MDS dimensions (MDS dim. 1, MDS dim. 2). This unsupervised method visualizes sample clustering based on marker expression, with dissimilarities calculated using the median expression of markers across all cells in each sample. The plot provides an overview of variance and potential clustering patterns among the samples. (B) Heatmap illustrating the median expression levels of 27 markers across four PB CD8+ T cell subsets: Naïve-like (CD95-, CD45RA+, CCR7+, CD27+, CD28+), Effector Memory (Tem) (CD95+, CD45RA-, CD45RO+, CD27+, CD28+, CD127+), Central Memory (Tcm) (CD95+, CD45RA-, CD45RO+, CCR7+, CD27+, CD28+, CD127+), and Effector Memory RA+ (Temra) (CD95+, CD45RA+, CD45RO-, CD57+, KLRG1+). The data are scaled to normalize marker expression across the dataset before clustering, ensuring each marker contributes proportionally to cluster definitions while minimizing the impact of extreme values or outliers. Blue represents lower expression, and red indicates higher expression. Hierarchical clustering on the top dendrogram groups markers based on their expression patterns, revealing co-occurring markers across subsets, while the clustering on the left shows relationships between subsets based on their marker profiles, highlighting their similarities.

We next explored the contribution of substrate cell virus-specificity to post-infusion persistence. Within 28 days, EBV-specific T_TCR-C4 were more abundant than CMV-specific T_TCR-C4, which remained below the persistence threshold (Fig. 1C). These differences endured when EBV-specific T_TCR-C4 recipients from Arm 1 were included (Extended data Fig. 6B).⁹ As central-memory CD8⁺ T cells (Tcm) have shown improved persistence compared to more differentiated phenotypes post-transfer,²⁸ we sought to explore whether differences in endogenous virus-specific cells could recapitulate these findings. We analyzed the phenotypes of EBV- and CMV-specific T cells using mass cytometry data²⁹ from 143 healthy individuals and cancer patients. Multidimensional scaling (MDS) showed segregation of CMV- and EBV-specific CD8⁺ T cells along the first dimension (MDS dim.1) suggesting distinct T-cell differentiation states between these groups (Extended data Fig. 7A). Next, Flow Self-Organizing Maps (FlowSOM)³⁰ metaclustering identified four CD8⁺ T-cell subsets: naïve (CD95^-,CD45RA⁺, CCR7⁺, CD27⁺, CD28⁺), effector memory (Tem) (CD95⁺, CD45RA^-, CD45RO⁺, CD27⁺, CD28⁺, CD127⁺), Tcm (CD95⁺, CD45RA^-, CD45RO⁺, CCR7⁺, CD27⁺, CD28⁺, CD127⁺), CD45RA⁺ effector memory (Temra) (CD95⁺, CD45RA⁺, CD45RO^-, CD57⁺, KLRG1⁺) (Extended data Fig. 7B).^31,32 Differential abundance analysis revealed a significant (false discovery rate/FDR < 0.05) increase of Temra cells in CMV-specific T cells and of naïve, Tcm and Tem in EBV-specific T cells (Fig. 1D). These results suggest that the Temra state of CMV-specific substrate cells compromises post-infusion persistence, whereas EBV-specific cells, derived from less differentiated precursors, facilitate persistence.

Persistence varied among EBV-specific T_TCR-C4 recipients, (Fig. 1C, Extended data Fig. 6A), with only three of ten (4, 8 and 27) exhibiting long-term persistence (>100 days post-infusion) (Fig. 1A, Extended data Fig. 8A). We reasoned that persistence was influenced by AML high-risk factors, including post-transplant remission duration and disease burden at the time of T_TCR-C4 infusion.³ Patient 1 was excluded from this analysis due to unavailable WT1-expression data. Six of the remaining patients did not show long-term persistence: four (5,7,14, and 19) had detectable disease (MRD or overt) within two weeks of T_TCR-C4 infusion, and two (23, 26) who were disease-free at infusion, had relapsed within 3 months post- HCT, suggesting difficult-to-control disease. Statistical testing revealed a trend (Fisher’s exact test, p = 0.08) linking reduced persistence with “aggressive” disease (defined as having detectable disease pre- infusion or early relapse post-HCT). Patient 8, initially categorized as “aggressive” because MRD-positive before the first infusion, became MRD-negative through salvage therapy before the second infusion, achieving the longest observed T_TCR-C4 persistence. Reclassifying this patient as “non-aggressive”, renders the association between disease aggressiveness and persistence statistically significant (p < 0.05).These findings suggest that aggressive disease post-HCT, independently of virus-specificity, was associated with reduced T_TCR-C4 persistence in vivo.

• Download figure
• Open in new tab

Extended data Fig. 8 Persitence and functionality of T_TCR-C4 post-first infusion

(A) Line plot showing the percentage (log scale) of T_TCR-C4 in PBMCs collected after the first infusion across all patients (n =15), colored by virus-specificity (violet = CMV; orange = EBV), derived from flow- cytometry data (B) Line plot showing the TTCR_C4⁺TNFα⁺ cells following WT1-peptide stimulation. The x axis represents the days from the first infusion, while the y axis indicates the percentage (log10 scale) of Tetramer⁺ TNFα⁺ gated on CD8⁺ T cells, based on flow cytometry data

Long-term persistent T_TCR-C4 acquire NKL/terminal differentiation markers associated with progressive loss of function in vivo

To investigate the fate of T_TCR-C4 post-transfer in patients with persistent T_TCR-C4, and compare these with endogenous (TCR_C4^-) T cells, we designed a 24-color spectral flow-cytometry panel (Supplementary Table 3). PB was analyzed at ∼1 (T1), ∼7 (T2), 28 (T3) days, and ∼4 months (T4) post-transfer (Supplementary Table 4). FlowSOM clustering revealed five CD8⁺ T-cell states categorized into three higher-level clusters (Fig. 2A, top dendrogram): naïve-/cm-like cells (Fig. 2A, dark green) (CD45RA⁺, CCR7⁺, CD27⁺, CD28⁺); Tem subgroup, including endogenous CD8⁺ Tem (CD28⁺, CD27⁺, Ki67⁺, CD38⁺, TIGIT⁺, PD1⁺, Tbet⁺) (Fig. 2A, violet), and phenotypically similar T_TCR-C4_Tem (tetramer⁺) (Fig. 2A, blue); Temra subgroup of endogenous CD8⁺ T cells (CD45RA⁺, CD57⁺, KLRG1⁺, GZMB⁺) (Fig. 2A, light green) and closely clustered T_TCR-C4_Temra (Fig. 2A, red). Two-dimensional Uniform Manifold Approximation and Projection (UMAP) revealed a progressive increase in T_TCR-C4_Temra and decline in T_TCR-C4_Tem from T1 to T4 (Fig. 2B). T_TCR-C4_Temra expressed minimal Ki67 compared to T_TCR-C4_Tem (Fig. 2A), enabling manual gating of these two subsets (Fig. 2C). Over time, T_TCR-C4_Temra significantly increased, while TTCR- C4_Tem decreased (p <0.05) (Fig. 2D), suggesting differentiation from proliferative T_TCR-C4_Tem to non- proliferative T_TCR-C4_Temra. T_TCR-C4_Temra predominantly expressed cytotoxic/KLR markers (KLRG1, CD57, GNLY), previously linked to T-cell dysfunction in AML,^21–23 rather than classical T-cell exhaustion markers (Tim3, PD1, and TIGIT) (Fig. 2E).^17,33 Stratification by timepoint revealed progressive increase of the cytotoxic subset versus the exhausted subset (Fig. 2F). This phenotypic shift concided with a detectable but slow decline in T_TCR-C4 CD8⁺ T cells producing IFNγ and TNFα (Fig. 2G, Extended data Fig. 8B). Thus, long-term persisting EBV-specific T_TCR-C4 skew towards a distinct NKL phenotype linked to functional decline.

• Download figure
• Open in new tab

Fig. 2 CD8⁺ T cell subset phenotypes and functional states over time

(A) Heatmap showing fluorescence intensity values of 20 markers across five CD8⁺ PB subsets annotated as Naïve-like (CD45RA, CCR7, CD27, CD28), Tem (CD28, CD27, Ki67, CD38, TIGIT, PD1, Tbet), T_TCRC4_Tem (TETRAMER, CD28, Ki67, CD38, TIGIT, GZMB, Tbet, KLRG1, GNLY), Temra (CD45RA, CD57, KLRG1, GZMB, Tbet), T_TCRC4_Temra (TETRAMER, GNLY, KLRG1, TIM3, GZMB, Tbet). The median marker expression identifies the markers that characterize each subset with red indicating high expression and blue low. The data were scaled after aggregation to highlight population-level differences between tetramer⁺ and tetramer^- samples, ensuring a clear visualization of marker expression patterns across subsets. Hierarchical clustering on the left dendrogram groups markers based on their expression patterns, revealing co-occurring markers across subsets, while the top dendrogram shows relationships between subsets based on their marker profiles, revealing similarities and differences. The K-means algorithm was applied to further group the clusters by similarity into three groups marked by the numbers (1,2,3) at the top of the heatmap. Data were derived from spectral flow- cytometry analysis. (B) UMAP plots showing the two-dimensional distribution of the annotated CD8⁺ T-cell subsets, with the plots colored according to the subset annotation and split by timepoint. The table shows the absolute numbers of T_TCR-C4_Tem and TTCR- C4_Temra cells at each timepoint (T1, T2, T3, T4). (C) Contour plots illustrating the gating strategy for T_TCR-C4_Tem (blue) and T_TCR-C4_Temra (red). This strategy starts by identifying the CD3⁺CD8⁺ cell population, followed by selecting tetramer⁺ CD8⁺ cells. Cells are then gated based on Ki67 expression, with Ki67^- cells (red) classified as T_TCR-C4_Temra and Ki67⁺ cells (blue) as TTCR- C4_Tem. This gating strategy is based on the marker expression shown in the heatmap in Fig. 2A. (D) Boxplots depicting the percentage (y-axis) of T_TCRC4_Tem and T_TCRC4_Temra populations identified in Fig. 2C over time (x-axis, from T1 to T4). Statistical significance was assessed using Kruskal-Wallis test (threshold for significance: p< 0.05). (E) Boxplot showing the percentage (y- axis) of cytotoxic KLRG1⁺, CD57⁺, GNLY⁺ (dark red) compared to exhausted TIM3⁺, PD1⁺, TIGIT⁺ (dark green) T cells among T_TCRC4_Temra. Statistical significance was assessed using the Wilcoxon rank-sum test, with a significance threshold of p < 0.05. (F) Boxplots displaying the percentage (y-axis) of cytotoxic (dark red) compared to exhausted (dark green) T cells among T_TCRC4_Temra over time (x-axis, T1 to T4). (G) Line plot showing the TTCR_C4⁺IFNg⁺ cells following WT1-peptide stimulation. The x axis represents the days from the first infusion, while the y axis indicates the percentage of tetramer⁺IFNg⁺ cells (log10 scale) gated on CD8⁺ T cells.

T_TCR-C4 largely represent an intermediate state between effector memory and NK-like/terminally differentiated cells

To explore the relationship between CD8⁺ endogenous T-cell and T_TCR-C4 states, we performed single-cell RNA sequencing (scRNAseq) on available PB and bone marrow (BM) samples from patients with T_TCR-C4 exceeding 3% of CD8⁺ T cells at least until day 28 post-infusion. This analysis included two prophylactic⁹ and five treatment-arm patients (Supplementary Table 5).

Unsupervised clustering of PB CD8⁺ T cells (n = 24,472) identified 13 clusters, with the TCRC4 transgene expressed across several clusters, but primarily in cluster 2 (Fig. 3A). Using the marker-based purification algorithm scGate,³⁴ we identified TCR_C4⁺ cells (Fig. 3B) labeled as T_TCR-C4, while endogenous (TCRC4⁻) T cells included clusters 5 and 6,labeled as naïve-/cm-like, expressing CCR7, SELL, TCF7, LEF1, IL7R; clusters 2, 4 and 11, labeled as Tem, expressing genes associated with activation (CD69, TIGIT, GZMK); cluster 12, labeled as interferon signaling genes (ISG), expressing ISG15, ISG20, IRF7, IFI6; clusters 9 and 10, labeled as T memory/proliferative (Tmem/prolif), expressing proliferation (MKI67, MCM5, MCM7) and memory (CD27) genes; clusters 0,1,3,7, and 8, labeled as NKL/Temra, expressing genes associated with cytotoxicity/NKL (KLRF1, KIR3DL1, NKG7, FCGR3A, GZMB, KLRD1, PRF1) (Fig. 3A-C, Extended data Fig. 9A-C, Supplementary Table 6).

• Download figure
• Open in new tab

Extended data Fig. 9. Identification of transcriptional endogenous and T_TCR-C4 CD8⁺ T-cell states.

(A) UMAP plot showing the distribution of CD8⁺ T-cell states from a published independent dataset³ used as a reference atlas. (B) The UMAP projections of each T-cell state defined in Fig. 3C are overlaid onto the reference atlas as contour plots. Contours corresponding to each subset from the query are color-coded as in Fig. 3C, allowing for easy visualization of the distribution of each subset within the reference atlas. (C) Heatmap showing the differential expression of manually curated genes (Supplemental Table 7) across the five annotated CD8⁺ T-cell subsets. Blue and red indicate the relative expression levels of each marker within each subset, with blue representing lower expression and red indicating higher expression. (D) Violin plots displaying the expression levels (y-axis) of stem-like (IL7R, TCF7), activation (GZMK, CD69), and cytotoxicity/NK-like (GZMB, KLRG1, NKG7, ZEB2, S1PR5) markers across CD8⁺ T-cell transcriptional states (x-axis). Statistical significance was assessed using the Wilcoxon rank sum test. Asterisks indicate the following thresholds of significance: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

• Download figure
• Open in new tab

Fig. 3 Single cell transcriptomic analysis of endogenous, T_TCRC4 CD8⁺ T-cell states and their differentiation dynamics.

(A) Heatmap showing the top 10 differentially expressed genes identified within each cluster. The “top 10” refers to the 10 genes with the most significant differential expression across the identified clusters. The dendrogram on the left displays the similarity between the 13 clusters, which were determined through unsupervised clustering based on gene co-expression patterns. The top dendrogram shows the relationships between the genes based on their expression patterns. To further refine this clusters, a K-means algorithm was applied to group the 13 clusters into 5 main categories, labeled 1 to 5 on the left side of the heatmap. Blue represents low gene expression, while red indicates high gene expression. (B) UMAP plot illustrating the distribution of CD8⁺ T cells (endogenous and T_TCR-C4⁺) on a two-dimensional space. T_TCR-C4⁺ cells were identified using scGate,³⁴ an R package which scores cells based on TCR_C4 expression and defines thresholds to classify cells as positive or negative for the population of interest. TCR_C4⁺ cells are colored in dark red, while TCR_C4^- cells (endogenous) are represented in light grey. (C) UMAP plot displaying the two-dimensional distribution of annotated CD8⁺ T-cell transcriptional states, colored by subset. (D) Violin plot illustrating the distribution of CD8⁺ T-cell subsets identified through gene expression (Fig. 3A) and TCR_C4 score (Fig. 3B) along the principal component 1 (PC_1) axis. Each violin represents a different CD8⁺ T-cell state. The proximity of each subset along PC_1 (y-axis) indicates transcriptional similarity, with subsets closer together showing more similar gene expression profiles (E) UMAP plot illustrating the inferred developmental trajectory of CD8⁺ T-cell transcriptional states, as predicted by Monocle.⁷⁹ The color scale, ranging from blu to orange/yellow, represents the pseudotime, where blue indicates earlier developmental stages and orange/yellow indicates later stages of the trajectory. (F) Boxplot showing the distribution of the 5 CD8⁺ T-cell subsets (y-axis) along the pseudotime (x-axis). The position of the boxes along the pseudotime axis provides context for how the subsets are positioned within the developmental trajectory, with subsets at lower pseudotime values representing earlier stages of differentiation, while those at higher pseudotime values correspond to later stages. (G) Line plots displaying the smoothed gene expression of selected genes characterizing the 5 CD8⁺ T-cell subsets along the pseudotime. These plots illustrate the dynamic changes in gene expression (y-axis) as cells progress along the inferred trajectory (x-axis). The smoothed curves show how the expression of each gene varies at different pseudotime points. (H) UMAP plot colored by the annotated CD8⁺ subsets overlaid with the predicted velocity stream computed through scVelo.⁴¹ The velocity streams represent the predicted direction and magnitude of gene expression changes for each individual cell, providing insights into the dynamic transitions between cellular states. These streams highlight the likely trajectories cells follow as they evolve over time, offering a predictive view of future cellular states based on current transcriptional dynamics.

To assess transcriptional similarities among cell states, we performed principal component analysis, positioning naïve-/cm-like and NKL/Temra as differentiation spectrum extremes, with T_TCR-C4 as the intermediate state between these two extremes (Fig. 3C).

To ensure unbiased labeling of CD8⁺ T cells, we constructed a scRNAseq reference atlas comprising 109,051 CD8⁺ T cells, using a published scRNAseq dataset of tumor-infiltrating lymphocytes.³⁵ We aligned and projected our scRNAseq dataset onto the reference atlas (Extended data Fig. 9B), confirming the correspondence between the transcriptional states in our dataset and the reference. Differential gene expression (DGE) analysis using manually curated markers further supported this annotation (Supplementary Table 7, Extended data Fig. 9C).^31,32,36,37

Consistent with previous work,³² we observed a progressive reduction of stem-like markers (IL7R, TCF7) from naïve/cm-like cells through Tem and NKL/Temra cells, while activation markers (CD69, GZMK) peaked in Tem. NKL markers, including KLR-exhaustion markers (S1PR5, ZEB2),^38,39 were highest in NKL/Temra. T_TCR-C4 expressed activation and NKL markers, reflecting an intermediate state between Tem and NKL/Temra (Extended data Fig. 9D).

Monocle⁴⁰ trajectory inference confirmed T_TCR-C4 as an intermediate state between Tem and the terminal NKL/Temra state (Fig. 3E-F), and removing the TCRC4 transgene did not alter this differentiation trajectory (Extended data Fig. 10A-B). Single-gene expression over pseudotime revealed early peaks of naïve/stem-like markers (CCR7, IL7R, TCF7, SELL, LEF1) (Fig. 3G, first row), naïve/memory (CD27, CD28) and activation markers (GZMK, CD69), with the latter remaining high throughout mid-pseudotime before declining.(Fig. 3G, second row). TCRC4 expression peaked at mid-pseudotime, but remained expressed until the end of pseudotime. NKL/Temra-associated genes (PRF1, KLRG1, NKG7, ZEB2, S1PR5) peaked at the end of pseudotime (Fig. 3G, third row). RNA velocity inference⁴¹ supported the spectrum of differentiation from naïve-like to Tem and T_TCR-C4, culminating in NKL/Temra cells (Fig. 3H).

• Download figure
• Open in new tab

Extended data Fig. 10. CD8⁺ transcriptional states after TCRC4 removal.

(A) UMAP plot depicting the inferred developmental trajectory of CD8⁺ T-cell transcriptional states, as predicted by Monocle, following the removal of the TCRC4 transcript. The color scale, ranging from blu to orange/yellow, represents the pseudotime, where blue indicates earlier developmental stages and orange/yellow indicates later stages of the trajectory.

(B) UMAP plot showing the distribution of CD8+ T-cell transcriptional states after TCRC4 removal, with the same coordinates as panel A. The T- cell subsets are labeled according to the clustering defined in Fig. 3D, providing context for how the subsets are positioned within the developmental trajectory in A.

Overall, our findings suggest that T_TCR-C4 largely existed in an intermediate state between the Tem phenotype and the NKL/Temra stage.

AML drives T_TCR-C4 towards NKL/terminal differentiation instead of the dominant exhaustion pattern associated with T cells in solid tumors

To investigate AML’s impact on the differentiation of endogenous CD8⁺ T cells and T_TCR-C4, patients samples were grouped as AML- (no AML detected, including two prophylactic arm cases),⁹ or AML+ (blasts evident in BM and/or PB) (Supplementary Table 8). The increase of ISG and NKL/Temra in AML+ and T_TCR-C4 in AML- (Fig. 4A) was significant (Fig. 4B), with a wide confidence interval for ISG, making the magnitude of this difference unclear. T_TCR-C4 overexpressed NKL/Temra genes in AML+ compared to AML- (Fig. 4C), suggesting that AML blasts may induce a transcriptional shift towards NKL/Temra in T_TCR-C4 and endogenous T cells.

• Download figure
• Open in new tab

Fig. 4 AML induces T_TCRC4 NKL/Temra differentiation skewing.

(A) UMAP plot of CD8⁺ T cells colored by density and split by group (AML(-) and AML(+)). AML(+) samples are those in which leukemic cells were detected in BM or PB, while AML(-) samples lack detectable leukemic cells. The plot shows the two-dimensional distribution of CD8+ T cells, with color intensity representing cell density, where blue indicates lower density and yellow indicates higher density. Dashed lines highlight areas of highest cell density in both AML(+) and AML(-) groups. (B) Point-range plot showing the pairwise (AML(+) vs. AML(-)) proportional difference for each CD8⁺ T-cell subset. Colors indicate the statistical significance (red: FDR < 0.05, blue: FDR >= 0.05); vertical dashed lines mark the absolute value of log2FD cutoff for significance. (C) Heatmap displaying DGE of NKL genes across the CD8⁺ T-cell subsets split horizontally by AML(-) vs. AML(+), with each condition showing the gene expression patterns for CD8+ T-cell subsets. Blue indicates low expression, while red indicates high expression. Dashed boxes highlight NKL/Temra genes, emphasizing their relative expression across the two conditions. (D) Heatmap depicting significant DEGs across three conditions: T_TCR37-45_D0 (T_TCR37-45-only), K562_D13 (T_TCR37-45 T cells after 13 days of coculture with K562 AML cell line), K563_D23 (T_TCR37-45 T cells after 23 days of coculture with K562 AML cell line). Signifcance was determined using the Likelihood Ratio Test in DESeq2,⁸⁰ with genes filtered for adjusted p-value (FDR < 0.05). Color intensity represents relative gene expression, with blue indicating low expression and red indicating high expression. (E) Boxplots illustrating the z-score of Exhaustion, NK- like, Naïve-like and Tem signatures across the three conditions (co-culture timepoints). (F) Heatmap illustrating the enrichment of scRNAseq-derived (T_TCRC4, Tem, Naïve.like, NKL.Temra, ISG, Tmem.prolif) and manually curated exhaustion markers across three conditions: K562_D13 vs. T_TCR37-45_D0, K562_D23 vs. T_TCR37-45_D0, K562_D23 vs. K562_D13. The signatures shown are derived from the top 50 DEGs for each subset from scRNAseq dataset. Additionally, a manually curated exhaustion signature is included. Each row represents the relative enrichment of these gene signatures in the comparison of interest, while the color intensity indicates the degree of enrichment: red indicates high enrichment, while blue indicates low enrichment. The comparison group correspond to timepoint during coculture with the K562 cell line. T_TCR37-45 represents the baseline (T_TCR37-45-only condition) (G) UMAP plot showing the distribution of CD8⁺ T-cell states from a published independent dataset used as reference atlas.³⁵ Dashed line indicates the UMAP coordinates of Tex CD8⁺ T-cell subset. (H) UMAP plot of CD8⁺ T-cell reference atlas (light grey) overlaid with projections (dark red) of CD8⁺ T cells collected from various independent published scRNAseq datasets (queries) of AML^{20,22,23,44,45}, pancreatic cancer,⁴⁶, melanoma, ⁴⁷ and lung cancer.⁴⁸ Dashed line indicates the area corresponding to Tex.

Following previously described methodology,⁴² we exposed in vitro CD8⁺ transgenic WT1-specific T cells targeting the HLA A*0201-restricted WT137-45 epitope⁴³ (T_TCR37-45) to high-WT1-expressing HLA- A*0201-transduced K562 acutely transformed chronic myelogenous leukemia (Extended data Fig. 11A). As K562 primarily express standard proteasomes and are not readily lysed by T_TCR-C4 targeting the immunoproteasome-specific WT1126–134 peptide, we used T_TCR37-45 due to its proteasome-agnostic WT1- derived peptide recognition.⁴³ Exposing T_TCR37-45 to fresh K562 every 3-4 days at stable effector-to-target (E:T) ratios (1:1, 1:4) caused initial tumor lysis (Extended data Fig. 11B), but control was lost by day 13, marked by a declining T_TCR37-45-to-tumor ratio (Extended data Fig. 11C). Irrelevant TCR-expressing T cells exerted no control (Extended data Fig. 11D). Bulk-RNA sequencing on sorted T cells found increased (FDR <0.05) naïve/stem-like genes (IL7R, TCF7, LTB, LEF1, SELL) at day 0, followed at day 13 by increased effector (CCL5, GZMK), activation/exhaustion (CD69, LAG3, CTLA4, PDCD1), and NKL/Temra markers (KLRB1, ENTPD1, KLRC1, KLRC2, NCAM1) (Fig. 4D-E). By day 23, the activation/exhaustion markers decreased, while NKL/Temra persisted or increased (Fig. 4D-E). To integrate these findings with the scRNAseq dataset, we performed a gene-set enrichment analysis with the top 50 differentially expressed genes from each CD8⁺ subset identified in scRNAseq, including naïve-/cm- like, Tem, NKL/Temra, T_TCR-C4, Tmem/prolif and ISG from the scRNAseq dataset. We also included a manually curated exhaustion signature (Supplementary Table 9). Exhaustion and Tem signatures exhibited enrichment in later timepoints, yet there was no further increase at day 23 versus day 13 (Fig. 4F). In contrast, the NKL/Temra increased between days 13 and 23 corresponding to T_TCR37-45 loss of tumor control (Extended data Fig. 11C).

• Download figure
• Open in new tab

Extended data Fig. 11. In-vitro dysfunction coculture system.

(A) Experimental design of the in-vitro model used to study chronic tumor exposure of WT1-specific T cells (T_TCR37-45 cells). The diagram shows two conditions with Effector (T_TCR37-45) to Target (K562) (E:T) ratios of 1:1 and 1:4. Red arrows indicate tumor re-challenge every 3-4 days, while red circles on the timeline mark the time points when RNA extraction for bulk-RNA sequencing was performed. An irrelevant TCR-expressing T cell line was used as a negative control. The experiment involved exposing transgenic CD8⁺ WT1-specific T cells to HLA-A*0201-transduced K562 cells expressing high levels of WT1. This figure was realized using BioRender. (B) Line plots showing the changes in absolute cell counts (y-axis) over time (x-axis, days of coculture) for T_TCR37-45 (solid lines) and K562 tumor cells (dashed lines) in coculture at E:T ratios of 1:1 and 1:4. The lines for each condition are colored sky blue (1:4 ratio) and tan (1:1 ratio). Cell counting was performed by flow cytometry. (C) Line plots showing the T cell to tumor cell ratio (y-axis) changes at the timepoints of tumor re-challenge (x-axis). Each plot represents one condition (left, E:T 1:4; right, E:T 1:1). Cell ratios were calculated based on flow cytometry data.(D) Line plots showing the changes in absolute cell counts (y-axis) over time (x-axis, days of coculture) for T cells with irrelevant TCR (solid lines) and K562 tumor cells (dashed lines) in coculture at E:T ratios of 1:1 and 1:4. The lines for each condition are colored sky blue (1:4 ratio) and tan (1:1 ratio). Cell counting was performed by flow cytometry.

We next projected scores from published NKL dysfunction signatures^22,38,39 onto our UMAP, revealing overlaps that link the NKL/Temra CD8⁺ T cells we identified with an end-term dysfunctional differentiation subset (Extended data Fig. 12). To confirm that AML-exposed T cells skewed towards a NKL/Temra rather than exhausted (Tex) phenotype, we projected onto a scRNAseq reference atlas a BM AML CD8⁺ T-cell dataset, compiled from published sources,^{20,22,23,44,45} alongside datasets representing pancreatic,⁴⁶ melanoma,⁴⁷ and lung⁴⁸ solid tumors (Fig. 4G-H).³⁵ Unlike solid tumors, which included a Tex cluster, AML datasets aligned with our spectral flow-cytometry and scRNAseq findings confirming the absence of Tex and the presence of NKL/Temra cells (Fig. 4G, Extended data Fig. 13). Although these studies did not specifically analyze tumor-antigen reactive T cells, the findings collectively suggest that the NKL (versus exhaustion) signature is intricately associated with AML-induced T-cell dysfunction.

• Download figure
• Open in new tab

Extended data Fig. 12 UMAP projection of KLR dysfunction signatures in CD8⁺ T cells.

UMAP plots illustrate the spatial distribution of CD8⁺ T cells based on their expression of published KLR dysfunction gene signatures. ^4–6 These signatures, derived from prior studies, were applied to our scRNAseq dataset to calculate a score for each cell, reflecting the degree of expression of the KLR dysfunction features. The scores are visualized using a gradient, where blue indicates low expression and red indicates high expression. Regions with high expression (red) align with the subset we identified as Temra NK-like cells, suggesting a functional and phenotypic overlap within this subset.

• Download figure
• Open in new tab

Extended data Fig. 13 UMAP projection of NK-like module score based on gene expression in a tumor-infiltrating lymphocytes reference atlas

³ UMAP plot projection showing the module score based on the expression of the following genes: KLRG1, KLRD1, ZEB2, FCRL6, ADGRG1, S1PR5, FCGR3A, GZMB, and NKG7. Each point represents an individual cell, and the color scale ranges from light grey (low module score) to dark red (high module score), reflecting the intensity of the gene set expression across cells. The black circle marks the area with the highest NK- like score, indicating the regions where the combined expression of NK-associated genes is most prominent.

Prolonged azacitidine exposure enhances self-renewal and Tcm features of T_TCR-C4, supporting TTCR- C4-mediated AML control

Given the link between NKL differentiation and AML-induced T-cell dysfunction, we explored the state of T_TCR-C4 in patient 8, the only patient of 4 (Extended data Fig. 4) who exhibited prolonged T_TCR-C4 and MRD persistence indicative of disease control despite incomplete leukemia clearance, enabling TTCR- C4/AML interplay analysis. This female patient underwent a non-myelo-ablative (Flu/3Gy-TBI) HCT from a male donor, with MRD (0.26% blasts by flow cytometry) detected 28 days post-HCT (Fig. 5A) and received a T_TCR-C4 (10¹⁰ cells/m²) 77 days post-HCT. AML blasts expressed detectable but limited WT1 protein pre-HCT (Extended data Fig. 14). scRNAseq at day 49 post-infusion revealed circulating CD34⁺ female-origin blasts (XIST⁺) expressing WT1 (Fig. 5E-F). Persistent MRD and limited T-cell persistence at day 63 post-infusion (T cells declined to only 0.1 multimer⁺ CD8⁺ T cells/µl on day 60), (Fig. 5B, 5C- red arrow, 5D) prompted azacitidine treatment (five days every 28 days). Post-azacitidine, T_TCR-C4 counts increased (∼50 multimer⁺ CD8⁺ T cells/ul; >3% CD8 multimer⁺ T cells) without additional T_TCR-C4 (Fig. 5D), leading to azacitidine discontinuation after seven cycles when no AML was detected. However, AML recurrence at day 414 post-infusion triggered chemotherapy (mitoxantrone, etoposide and cytarabine), additional azacitidine cycles followed by a second T_TCR-C4 (10¹⁰ cells/m²) infusion (Fig. 5A). Azacitidine was held post-infusion but restarted upon MRD detection 162 days post-second T_TCR-C4 infusion, continuing for over 20 monthly cycles, maintaining marrow MRD below 1% for 21 months (641 days, from day 868 to day 1509) before progression was detected on day 1546 (Fig. 5A-B). During this 21-month period, blasts were detectable via scRNAseq in BM but not in PB (Fig. 5F, Extended data Fig. 15A-B). Throughout the azacitidine cycles, platelet counts decreased (∼100 x 10³ platelets/µl) mid-cycle before rising (∼300 x 10³ platelets/µl) immediately before the next cycle, consistent with transient azacitidine-mediated myelosuppression. Neutrophil counts, however, paradoxically increased mid-cycle (∼4.5 x 10³ neutrophils/µl) and decreased (∼1.6 x 10³ neutrophils/µl) just before the next cycle, suggesting that treatment-related AML contraction facilitated enhanced mid-cycle neutrophil production despite azacitidine’s myelosuppressive effect (Extended data Fig. 15C). This 21-month response exceeds the typical 4-month median for post-HCT azacitidine treatment,⁴⁹ suggesting that persisting T_TCR-C4 plus azacitidine contributed to long-term disease control.

• Download figure
• Open in new tab

Extended data Fig. 14. Immunohistochemistry of bone marrow biopsy stained for WT1.

This immunohistochemistry image shows a bone marrow biopsy stained for WT1. Black arrows indicate WT1-positive nuclei, visible as distinct dark-stained areas within the cells, which signify expression of the WT1 protein.

• Download figure
• Open in new tab

Extended data Fig. 15 Patient 8 blasts and T cells characterization

(A) UMAP plot showing a blast score (dark red) calculated based on the co- expression of CD34 and XIST (a female-specific gene) in a patient previously transplanted with a sex-mismatched donor (female patient, male donor). The score was used to identify cells expressing a high blast score. Each UMAP plot represents a different timepoint (day 49, 256, 405, 1322, 1343) after the first infusion, with lightgrey indicating low expression and dark red indicating high expression of the blast score. (B) UMAP plot showing a blast score calculated based on the co-expression of XIST, CD34, CD33, and KIT. This score was used to visualize cells with high blast scores in bone marrow samples. Each UMAP plot represents a specific timepoint (day 1322 and 1344) post-first T_TCR-C4 infusion, (C) Line plot showing platelets and neutrophil absolute counts over time after the first infusion. Platelets (left y-axis, dark blue line plot, measured as platelets/μL × 1000) and neutrophils (right y-axis, red line plot, measured as ANC/μL × 1000) are shown in relation to the days after the first infusion (x-axis). Vertical dark green dashed lines indicate the timing of azacytidine cycles, while solid orange lines represent the timepoints when BM scRNA-seq was performed.

• Download figure
• Open in new tab

Extended data Fig. 16 UMAP visualization of gene expression in AML patient samples.

(A) UMAP plots at day 49 post-first T_TCR-C4 infusion depict the expression patterns of leukemia-associated markers WT1 and XIST (a female-specific gene) in patient 8, previously transplanted with a sex-mismatched donor (female patient, male donor). Additionally, CD34 expression across all cell populations is shown. Cells with high expression of each marker are highlighted in red, while other cells appear in grey.

(B) UMAP plots illustrate a longitudinal analysis of leukemia-associated gene expression in patient 26, who was transplanted with a sex-mismatched donor (female patient, male donor), at days 7 (d7) and 28 (d28) post-first T_TCR-C4 infusion. XIST indicates leukemic cells, while RPS4Y1 (a male- specific gene) marks non-leukemic cells. Cells with high gene expression are represented in dark red, and those with lower expression appear in light grey.

• Download figure
• Open in new tab

• Download figure
• Open in new tab

Fig. 5 Azacitidine as a potential boost for T_TCR-C4 in vivo function and persistence

(A) Timeline of patient’s treatment regimen. Dark red highlights timeframes with detectable BM AML (B) Percent of BM AML blasts (y-axis, log scale) by multiparametric flow cytometry (MFC) at specific timepoints (black dots, dark red shaded area). The horizontal dashed line represents the sensitivity limit of MFC. (C) Multimer⁺ cells/µl and (D) Percent multimer⁺ of CD8⁺ T cells in PB (black dots) and BM (red dots) collected before and after T_TCR-C4 infusions. The red arrow indicates the lack of T_TCR-C4 persistence before the start of Azacitidine (E) UMAP plots showing a blast score (dark red) calculated based on the co-expression of CD34 and XIST (a female-specific gene) in a patient previously transplanted with a sex-mismatched donor (female patient, male donor). The score was used to identify cells expressing a high blast score. Additionally, WT1 expression across all cell populations is shown. Cells with high expression of each marker are highlighted in red, while other cells appear in grey. (F) Violin plot of the WT1 expression (y-axis) across the timepoints analyzed after first infusion of T_TCR-C4 (x-axis, d265, d405, d1322, d1343). Violin plots are colored by timepoint. (G) UMAP plots display the distribution of total PB CD8⁺ T cells from the scRNAseq dataset, with the T_TCR-C4 subset from patient 8 highlighted in dark red. Single-cell points represent individual cells, while kernel density contours (also in dark red) depict the density of T_TCR-C4 cells within the CD8⁺ T cell landscape. The plots are stratified by timepoints post-infusion: (day 49, day 256, day 405, day 1322, and day 1343). This series illustrates the spatiotemporal dynamics of the T_TCR-C4 subset, showing its distribution in relation to the broader PB CD8⁺ T cell population and highlighting potential transcriptional changes over time. Red arrows were added manually to indicate the different skewing of T_TCR-C4 across the timepoints examined. (H) UMAP plots showing the distribution of PB CD8⁺ T cells, with the density contours highlighting cells with high self-renewing (dark green) and NK-like (dark red) scores. The self- renewing score was calculated based on a set of stem-like/survival genes (TCF7, LEF1, SELL, CCR7, BCL2, IL7R, CD27, CD28) and the NK-like score was based on NK-associated genes (ZEB2, S1PR5, CX3CR1, KLRG1, NKG7, FCRL6, KLRD1, ADGRG1). The plots are color-coded by score (light grey for the background, darker shades indicating higher scores). Cells with scores above the 75th percentile for each gene set are highlighted with density in dark green for high self-renewing score cells and dark red for high NK-like score cells. (I) The plot shows the temporal changes in the self-renew (left) and NK-like (right) scores of T_TCR-C4 from patient 8. The smoothed trends of the scores were calculated using LOESS regression at multiple timepoints. The dark green (self-renew) and dark red (NK-like) lines represent the smoothed score trends, with shaded areas indicating the confidence interval. The y-axis reflects the score expression values. (J) Dot plot displaying the expression of the self-renew and NK-like scores across timepoints post-infusion. The size of each dot corresponds to the expression level of the respective score and the color indicates the relative intensity of the expression. (K) Stacked plot showing cell count of T_TCRC4 in specific clonal frequency ranges over time (d256, d405, d1322, d1343). Colors indicate the clonal frequencies. (L) Stacked plot showing cell count for BM T_TCRC4 within specific clonal frequency ranges at d1322 and d1343. Colors indicate the clonal frequencies. (M) Cartoon (created with Biorender) showing the influence of blasts and azacitidine on T_TCR-C4. The red area under the curve represents the blasts percentage (y-axis) over time (x-axis). Blasts induce T_TCR-C4 skewing towards NKL and cell death. However, prolonged exposure to azacitidine allows T_TCR-C4 to maintain self-renewal, facilitating persistence and long-term disease control.

T_TCR-C4 transcriptomics were analyzed to identify associations with disease control. We scored genes associated with self-renewal and NK-like profiles and found that at earlier timepoints (d49, d256, and d405 post-first infusion), T_TCR-C4 were predominantly skewed towards an NKL/Temra transcriptional state. However, at later timepoints (d1322, d1343) although T_TCR-C4 remained along the same differentiation trajectory (from less differentiated states to NKL/Temra T_TCR-C4), they exhibited fewer NKL features, and showed an enrichment of self-renewal markers (Fig. 5G-I). scTCRseq analysis of PB at days 265, 405, 1322, 1343 post-first T_TCR-C4 infusion revealed progressive clonal expansion of T_TCR-C4 (Fig. 5K). T_TCR-C4 clonal expansion peaked during the 21 months of continued azacitidine treatment, suggesting that during this phase, T_TCR-C4 were capable of both self-renewal, which in turn facilitated their long-term persistence, and antigen recognition, which triggered further expansion and disease control. Similarly, BM scRNAseq revealed clonal expansion of T_TCR-C4 between d1322 and d1344 post-first infusion (Fig. 5L) despite no additional T_TCR-C4. These findings indicate that azacitidine-exposed T_TCR-C4, retained a self-renewing (versus NKL) features, which enabled their long-term persistence and enhanced AML-targeting efficacy that may have contributed to long-term disease control (Fig. 5M).

Clinical protocol

The trial was approved by the Fred Hutchinson Cancer Center (FHCC) Institutional Review Board, the US Food and Drug Administration and the National Institutes of Health Recombinant DNA Advisory Committee. It was registered at ClinicalTrials.org under the identifier NCT01640301. Eligible participants included ‘high-risk’ AML patients with relapsed or refractory disease (overt or MRD) post-HCT along with their fully HLA-matched (10 of 10) related or unrelated donors expressing HLA A*0201 (HLA-A2).

Patient selection

HLA-A2 genotype was confirmed by high-resolution typing before enrollment. Exclusion criteria included: refractory central nervous system disease, HIV seropositivity, grade ≥ 3 GVHD and no available CMV/EBV-seropositive matched donor. The sample size for this study was not based on formal power calculations, but on feasibility, the potential to provide descriptive information, determine whether further study was warranted and evaluate toxicity.

Treatment plan

Isolation of TCRC4, lentiviral vector construction, and T_TCR-C4 generation were performed as previously described.⁹ Patients were eligible to receive a first infusion of T_TCR-C4 after demonstrating relapse (overt or MRD) at any time post allogeneic HCT. Patients 26,27,28 received lymphodepleting treatment before the first T_TCR-C4 infusion with cyclophosphamide (300 mg/m² IV) and fludarabine phosphate (30 mg/m² IV) daily on days -4 to -2. A second T_TCR-C4 infusion was administered only if the frequency of T_TCR-C4 was <3% of total peripheral CD8⁺ T cells. Patients 5,7,19,23,27 received 1 infusion, patients 4,8,15,26,28 received 2 infusions, patient 1 received 1 infusion, the remainder received 4 infusions (Table 1). Patients were monitored for toxicities, based on Common Toxicity Criteria v.4.0.

Assessment of disease status

Morphology, multiparameter flow cytometry, standard cytogenetics or genomic technologies were routinely performed on bone marrow aspirates and peripheral blood samples that were obtained from all patients. Any level of residual disease was considered to indicate positivity for MRD.

T cell tracking by WT1 peptide/HLA (pHLA) tetramers

WT1 pHLA-specific tetramers (produced by the FHCC Immune Monitoring Core Facility) were used to detect T_TCR-C4 in PBMCs collected after infusions, with a staining sensitivity of 0.01% of total CD8⁺ T cells, as previously described.⁹ T_TCR-C4 percentages were calculated using FlowJo v.10 (Treestar).

Patient outcomes and survival analysis

Kaplan-Meier OS curves were estimated using the survminer (v0.4.9) and survival (v3.3-1) packages in R. OS was calculated from the date of first T_TCR-C4 infusion to the date of death or censoring. Outcomes of responding patients were represented using a swimmerplot. Day 0 was defined as the post-HCT relapse date or, for HCT-refractory patients, as day 28 post-HCT, at which timepoint persistent disease was observed. R package swimplot (v1.2.0) was used for visualization.

Analysis of T_TCR-C4 persistence

Persistence of TCR-T cells (%T_TCR-C4) over time was visualized using ggplot2 (v3.4.4) R package. Wilcoxon rank-sum tests were used for pairwise comparisons between T_TCR-C4 with different virus- specificity (EBV vs. CMV) over time. Kruskal-Wallis tests were applied for multi-group comparisons.

Patients were categorized based on disease status prior to infusion: those with detectable MRD or overt disease before T_TCR-C4 infusion, or those who relapsed within 3 months post-HCT, were classified as having "aggressive" disease. Patients without these risk factors were considered "non-aggressive." One patient was excluded from the analysis due to the unavailability of WT1-expression data, which could confound interpretation of the persistence results. Fisher’s exact test was used to evaluate the association between disease aggressiveness and T_TCR-C4 persistence. The test was chosen due to the small sample size and binary classification of persistence (long-term vs. short-term) and disease aggressiveness (aggressive vs. non- aggressive). Statistical significance was assessed with a p-value threshold of 0.05.

Flow-cytometry

Cryopreserved PBMCs were thawed and allowed to rest overnight in RPMI medium supplemented with 10% fetal bovine serum (R10). The cells underwent stimulation with a cocktail containing the WT1126-134 peptide at a final concentration of 1 μg/ml in R10 and intracellular cytokine staining using a 15-color staining panel, as previously described.⁹ Flow cytometry was conducted on an LSRII instrument (Becton Dickinson) with data acquisition using FACS-Diva software v.8.0.1. Flow cytometry data were subsequently analyzed using FlowJo v.10 (Treestar). Finally, the percentage of Tetramer⁺cytokine⁺ cells gated on the CD8⁺ T cell population was determined.

Spectral flow cytometry

Post-infusion PB samples were analyzed using a 5-laser Cytek Aurora. Antibodies used are listed in Supplementary Table 1. T_TCR-C4 were identified by binding to the APC dye-labeled HLA-A2:WT1126-134 tetramer. Spectral flow-cytometry data were biexponentially transformed, compensated and preprocessed (aggregates and dead cell removal) in FlowJo V10 (TreeStar). Pregated CD8⁺ T cells were exported from FlowJo and loaded in R (v4.3.2). First we created a flowSet using flowCore (v2.12.2)⁶⁶and subsequently analyzed the data using CATALYST (v1.24.0).⁶⁷

Ki67 was then used to manually gate subsets of interest using the FlowJo software. Statistical analysis included Kruskal-Wallis to compare T_TCR-C4_Temra and T_TCR-C4_Tem percentages over time, and Wilcoxon test for comparing KLRG1⁺CD57⁺GNLY⁺ and Tim3⁺PD1⁺TIGIT⁺ T_TCR-C4_Temra.

In-silico mass cytometry validation

Mass cytometry files used to generate an atlas of virus-specific CD8⁺ T cells²⁹ were downloaded from https://zenodo.org/records/8330231. Data were analyzed using flowCore and CATALYST R packages as specified above.

Single Cell RNA Sequencing

Patients with persisting PB T_TCR-C4 (> 3% of the total CD8⁺ T cells) at least until day 28 after infusion were selected for scRNAseq analysis. Available PBMCs or BMMCs were thawed, washed and loaded on a 10x Chromium Controller based on the 3’ Chromium or 5’ Chromium Single Cell V(D)J Reagent Kit manual (10x Genomics). Library preparation was performed as per manufacturer’s protocol with no modifications. Library quality was confirmed by TapeStation High Sensitivity (Agilent, evaluates library size), Qubit (Thermo Fisher, evaluates dsDNA quantity), and KAPA qPCR analysis (KAPA Biosystems, evaluates quantity of amplifiable transcript). Samples were mixed in equimolar fashion and sequenced on an Illumina HiSeq 2500 rapid run mode according to the standard 10X Genomics protocol. TCR target enrichment, 5’ gene expression library, and TCR library were carried out to the 5’ Chromium Single Cell V(D)J Reagent Kit manual (10x Genomics). The 10X Genomics software Cell Ranger (v2.0.0) was used to process the raw data FASTAQ files with default parameters. The EmptyDrops method⁶⁸ was used to identify cells with low RNA contents. The “count” function was used to perform alignment, filtering, barcode counting and UMI counting. Reads were aligned to the hg38 human reference genome (Ensembl) and the known transgene codon-optimized sequence using Spliced Transcripts Alignment to a Reference (STAR).⁶⁹

For V(D)J sequencing assembly and paired clonotype calling, we used CellRanger “vdj” function. This function leverages Chromium cellular barcodes and UMIs to assemble V(D)J transcripts for each cell. CellRanger V(D)J calling produces an output named “filtered_contig_annotations.csv” for each sample, which lists CDR3 amino acid and nucleotide sequences for single cells identified by their barcodes.

scRNAseq quality control and subsetting

The filtered feature matrices generated by the CellRanger pipeline were used for downstream quality control (QC) and analyses. We used the function read10xCounts from the R package DropletUtils (version 1.14.2) to load the CellRanger output in R as a SingleCellExperiment⁷⁰ object. Doublet cells filtering was performed on each sample using the scds package (v1.10.0).⁷¹ QC and filtering were conducted using the scater R package (v1.22.0)⁷².⁷¹ Genes not detected across all the cells were removed and cells were filtered based on feature counts, the percentage of mitochondrial and ribosomal genes, and the number of expressed features. Cells with values beyond a specific threshold, between 1 and 3.5 median absolute deviations (MAD) from the median, were excluded. These MAD thresholds were established according to the quality of each sample. Features with a count greater than 1 in at least 3 cells were retained for downstream analysis. We then split cells by sample in 15 datasets, normalized, found the 2000 most variable genes and scaled for each dataset using Seurat (v4.3.0.9001) SplitObject, NormalizeData, FindVariableFeatures and ScaleData respectively.^73,74 For batch correction, we used the FindIntegrationAnchors and IntegrateData functions from Seurat. The integrated dataset was then used for scaling (ScaleData) and dimensionality reduction (PCA and UMAP) using RunPCA and RunUMAP, respectively. UMAP dimension reduction and clustering were computed using the first 20 principal components (PCs). The number of PCs capturing most of the variation in our data was selected using Seurat function ElbowPlot which visualizes the standard deviation of each PC. Clusters were identified via shared-nearest-neighbor-based (SNN) clustering and further analyzed at a resolution of 0.6. CD8⁺ cell subset was identified using scGate (v1.0.1)³⁴ and subsequently extracted with Seurat subset function.

scRNAseq CD8⁺ endogenous and T_TCR-C4 analysis

The CD8⁺ subset was used to find the 2000 most variable features (FindVariableFeatures), scale the data (ScaleData), and run dimensionality reduction (runPCA and runUMAP). SNN was used for clustering with a resolution of 0.5 for downstream analysis. Differential gene expression across clusters or condition (e.g. AML(+) vs AML(-)) was computed using Seurat FindAllMarkers with default parameters. TCRC4- transgene⁺ cells were identified using scGate (v1.0.1). For visualization purposes, we used Seurat built-in functions alongside ComplexHeatmap (v2.15.1)⁷⁵, scCustomize (v1.1.0.9001) (https://github.com/samuel-marsh/scCustomize) and SCP (https://github.com/zhanghao-njmu/SCP). The pan-cancer CD8⁺ single-cell reference atlas was built using the CD8.thisStudy_10X.seu.rds file³⁵ downloaded from Zenodo and processed using the make.reference function from the ProjecTILs R package (v3.2.0).⁷⁶ We then used Run.Projectils function, from the same package, to project the cell states from the query dataset onto the reference.

Cell trajectory inference was computed using the Monocle R package (v1.3.4), with UMAP used as dimensionality reduction. Single-gene expression patterns along the pseudotime were visualized using the plot_genes_in_pseudotime function in Monocle. RNA velocity was conducted by exporting the CD8⁺ T- cell Seurat object as an h5ad file using the R package seurat-disk (v0.0.0.9020) (https://github.com/mojaveazure/seurat-disk), loading this file in Python (v3.8.14) as AnnData object,⁷⁷ and estimating velocities with the scvelo Python package (v0.2.4)⁴¹ using the deterministic model.⁷⁷

We classified patients as AML(+) or AML(-) based on the presence of detectable blasts in BM or PB. Three patients lacked BM or PB evaluations at one timepoint each. For two of these patients, we used the donor- recipient (male-female) sex mismatch. Patient 8 at day 49 exhibited cells expressing the female-specific gene XIST alongside the AML-associated genes CD34 and WT1 (Extended data Fig. 15A); similarly, patient 25 expressed XIST in a RPS4Y1 (male-specific gene) negative region at days 7 and 28, thus classified as AML(+) at these timepoints (Extended data Fig. 15B). Patient 4 was categorized as AML(-) at day 100 and AML(+) at day 581, as previously reported.⁴³

To assess the significance of differences in cell proportions per CD8⁺ cell state between groups (AML(+) vs AML(-)), a permutation test was applied. Specifically, the permutation_test and permutation_plot functions from the R package scProportionTest (version 0.0.0.9000) (https://github.com/rpolicastro/scProportionTest) were used with default parameters.

For scRNAseq in-silico validation, we compiled scRNAseq datasets from independent studies on AML^{20,22,23,44,45}, lung cancer⁴⁸, pancreatic cancer⁴⁶ and melanoma.⁴⁷ CD8⁺ T cells were identified and extracted using the scGate³⁴ R package. These cells were then projected onto the pan-cancer CD8⁺ T cell reference atlas³⁵ using ProjecTILs R package.

The TCR repertoire was analyzed using the R package scRepertoire (v1.10.1).⁷⁸ Initially the degree of single cells clonal expansion was defined based on the number of cells sharing the same clonotype: Hyperexpanded (100 < X <= 500), Large (20 < X <= 100), Medium (5 < X <= 20), Small (1 < X <= 5), Single (0 < X <= 1) The function occupiedscRepertoire was used to visualize the degree of clonal expansion by cell-state over time.

In-vitro chronic antigen stimulation model

The in-vitro dysfunction model was established as previously described.⁴² The WT1⁺ K562 tumor cell line was transduced with lentiviral constructs to express HLA-A*02:01 and GFP, then sorted for purity using the Sony MA900 cell sorter. Cells where then cultured in media consisting of IMDM with GlutaMAX (Gibco, Life Technologies, #31980030) with 10% FBS, and 1% of penicillin/streptomycin.

T_TCR37-45 cells were obtained following previously published protocols.⁴³ We selected TCR-T cells targeting WT137-45 as K562 cell line primarily express the standard proteasome and is not lysed by T_TCR-C4 targeting WT1126-134.⁴³ T cells expressing an irrelevant virus-specific TCR were used as negative control. Co-cultures of T cells with K562 cells were established at a 1:1 and 1:4 E:T ratio (2.5 x 10⁵ T cells, 2.5 x 10⁵ or 1 x 10⁶ tumor cells, respectively). After 3-4 days of coculture, 250 μl of the cell suspension was used for T cell counting and flow-cytometry staining. The remaining cell suspension was spun down, and cells were resuspended in fresh media. A Novocyte 3 lasers flow cytometer was used to quantify GFP⁺ tumor cells and T cells and to maintain constant E:T ratios by reseeding K562 cells. Notably, during the peak of T-cell expansion (day 9-13), the volumes of the cocultures were reduced to ensure the reseeding of an adequate number of tumor cells, thereby maintaining constant E:T ratios. This protocol was followed for 23 days.

Bulk RNA sequencing

Bulk RNA sequencing was performed using BGISEQ-500 platform at BGI Genomics. Briefly, total RNA was extracted using the Qiagen RNeasy Micro Kit according to the manufacturer’s instructions. For the construction of low input polyA mRNA-seq libraries, the SMARTseq (v4) Package was used. Sequencing was performed on a DNBseq T7 machine (MGI) with paired-end 150 bp reads, generating 30M raw reads per sample. Raw sequencing data were filtered and trimmed using the software Soapnuke developed by BGI Genomics. The filtered reads were then aligned to the reference transcriptome using Bowtie2 (v2.2.5). Gene read counts were subsequently generated from the alignment results using RSEM (v1.2.8).

To assess differences over time across conditions (T_TCR37-45_D0, K562_D14, and K562_D23), we used the likelihood ratio test (LRT) with a full model of ∼Condition and a reduced model including only the intercept (reduced = ∼ 1). Normalized counts were obtained using the rlog transformation, and differential expression analysis was performed using the R package DESeq2 (v1.42.0). We manually curated a list of genes of interest (Supplementary Table 10) and filtered the results to retain only those genes that were significant (padj < 0.05) with an absolute log2 fold-change > 1). For visualization purposes, average normalized counts of biological replicates were calculated using the avereps function from the R package limma (v3.56.2) and visualized using the R package pheatmap (v1.0.12). To visualize the expression patterns of manually curated gene signatures, we calculated z-scores for the averaged normalized counts across conditions (T_TCR37-45_D0, K562_D14, and K562_D23). Finally, to compare the enrichment of scRNAseq-derived gene signatures (top 50 differentially expressed genes) and of a manually curated exhaustion signature across conditions, we used the hciR (v1.7) function fgsea_all with default parameters and plot_fgsea for plotting.