Single-cell profiling of natural killer/T-cell lymphoma reveals stratified immune features and potential therapeutic implications

Yi Cao; Jun Cai; Danling Dai; Jinni Wang; Yuchen Zhang; Jie Xiong; Huiqiang Huang; Yan Gao; Yi Xia; Fang Liu; Xinna Gao; Shenrui Bai; Qihua Zou; Jiahui Wang; Shuyun Ma; Junfeng Jiang; Weili Zhao; Wenlong Zhong; Qingqing Cai

doi:10.1038/s41467-026-76152-9

. 2026 Jul 29;17:9184. doi: 10.1038/s41467-026-76152-9

Single-cell profiling of natural killer/T-cell lymphoma reveals stratified immune features and potential therapeutic implications

Yi Cao ^1,^2,^#, Jun Cai ^1,^3,^#, Danling Dai ^1,^#, Jinni Wang ^1,^#, Yuchen Zhang ^1,^2,^#, Jie Xiong ^4,^#, Huiqiang Huang ^1,², Yan Gao ^1,², Yi Xia ^1,², Fang Liu ¹, Xinna Gao ^1,², Shenrui Bai ^1,², Qihua Zou ¹, Jiahui Wang ^1,², Shuyun Ma ⁵, Junfeng Jiang ⁶, Weili Zhao ^4,^✉, Wenlong Zhong ^7,^✉, Qingqing Cai ^1,^2,^6,^✉

PMCID: PMC13522500 PMID: 42660902

Abstract

Tumor cell heterogeneity and interactions with the immune microenvironment play a key role in the progression and therapeutic efficacy of natural killer/T-cell lymphoma (NKTCL). We perform single-cell RNA sequencing analysis of 63 samples, integrating spatial transcriptomics, bulk transcriptomics, proteomics, and metabolomics to dissect inter- and intra-tumoral heterogeneity. Four meta-programs (MP) are identified, including MP1 (immune-responsive), MP2 (proliferative), MP3 (inflammatory), and MP4 (metabolic), each linked to distinct molecular and immune features. MP1 exhibits an immune-exhausted tumor microenvironment and high programmed death-ligand 1 expression, suggesting a potential response to immune checkpoint blockade. MP2 shows an immune-desert phenotype with elevated HDAC2 and MKI67 expression, indicating epigenetic regulation in tumor proliferation. MP3 is characterized by a myeloid-dominant tumor microenvironment, JAK/STAT pathway activation, and an aggressive clinical course. MP4 exhibits a distinct amino acid metabolic profile and enriched tertiary lymphoid structures. Collectively, our study provides a high-resolution molecular atlas of NKTCL heterogeneity, offering insights into patient stratification and potential avenues for future therapeutic development.

Subject terms: T-cell lymphoma, Cancer microenvironment, Immune evasion, Cancer genomics, Tumour heterogeneity

Tumor cell heterogeneity and interactions with the immune microenvironment play a key role in natural killer/T-cell lymphoma (NKTCL). Here, the authors characterize 63 NKTCL samples using single-cell, spatial, and bulk multiomics; they identify four gene expression meta-programs that are associated with tumor proliferation and response to therapy.

Introduction

Natural killer/T-cell lymphoma (NKTCL) is a rare subtype of non-Hodgkin lymphoma (NHL) characterized by rapid progression and poor prognosis^1,2. The cornerstone of treatment for NKTCL patients is L-asparaginase-based chemotherapy, which exploits the metabolic vulnerability of NKTCL resulting from its dependence on extracellular asparagine for tumor growth³. Similar to many other hematologic malignancies, immune checkpoint inhibitors (ICIs) are increasingly used to treat advanced NKTCL in the first-line setting^3–5. However, approximately 40% of patients are resistant to ICIs⁵. Given the molecular heterogeneity among patients, stratified targeted therapies are essential to improve treatment efficacy and clinical outcomes.

NKTCL is highly heterogeneous and displays unique pathological complexity. First, it could originate from either NK or cytotoxic T cells⁶. Second, the morphology of the tumor cells ranges from small to large transformed cells⁷. Intensive lymphocyte infiltration is a hallmark of NKTCL, but cellular composition and its relationship with disease progression remain unclear⁸. This diversity, coupled with the intricate crosstalk between tumor cells and immune components within the tumor microenvironment (TME), represents significant challenges to designing effective treatment approaches. Three molecular subtypes with distinct genetic mutations have been identified using bulk transcriptomic and genomic analyses⁶, but lack the resolution to dissect cellular heterogeneity and decode intercellular communication networks⁹.

Advancements in single-cell RNA sequencing (scRNA-seq) have identified specific tumor subpopulations, such as the tumorigenic LMP1⁺ malignant NK cells¹⁰, as well as TME heterogeneities, i.e., three TME phenotypes (immune-deficient, immune-inflamed and immune-desert) in NKTCL¹¹. Nevertheless, these studies are constrained by small sample sizes and insufficient patient stratification, hindering the generalization and reproducibility of findings across diverse cohorts. Interpatient heterogeneity significantly influences clinical outcomes^12,13, but patient stratification of NKTCL patients based on scRNA-seq remains lacking.

In this work, we perform the scRNA-seq analysis on 63 samples and identify four subgroups in NKTCL, each associated with distinct immune-related TME characteristics, e.g., immune exhaustion, immune desert, immune exclusion, and tertiary lymphoid structures (TLSs) enrichment. The findings are further validated through complementary multi-omics profiling, including high-resolution spatial transcriptomics, bulk transcriptomics, metabolomics, and proteomics. This study provides valuable insights into understanding tumor heterogeneity and the immune microenvironment in NKTCL, suggesting potential prognostic markers and biologically plausible therapeutic implications.

Results

A single-cell landscape of NKTCL

To establish a comprehensive single-cell atlas of NKTCL, we conducted scRNA-seq analysis on 63 samples. These included 48 in-house collected samples from 38 patients enrolled at Sun Yat-sen University Cancer Center (SYSUCC), consisting of 31 patients with previously untreated NKTCL (of which 10 patients had matched post-treatment samples after the first treatment cycle) and seven patients who had experienced relapse following either radiotherapy alone or anthracycline-based chemotherapy as first-line treatment (Supplementary Data 1). scRNA-seq data from 10 additional NKTCL patients from a publicly available dataset¹⁰ were also included in the analysis (Fig. 1a). Across the 48 NKTCL patients, early-stage disease (Ann Arbor stage I–II) accounted for 54.2% of cases, whereas stage IV disease accounted for 31.3%. All tumor specimens profiled by scRNA-seq were collected from lesions in the nasal cavity or nasopharynx. scRNA-seq data from five healthy nasal mucosa samples from a publicly available dataset were included as controls¹⁴. Xenium, a high-resolution multiplexed in situ spatial profiling platform¹⁵, was used to resolve the spatial distribution of cell clusters in samples from 14 patients (Supplementary Data 1, Supplementary Fig. 1). In addition to the single-cell discovery cohort, complementary multi-omics datasets were included. The bulk RNA-seq dataset comprised a previously published asparaginase-based chemotherapy cohort (n = 97)⁶, an immunochemotherapy cohort from a clinical trial (n = 21)⁵, and 141 newly generated cases from three institutions, including 93 from SYSUCC and 48 from Beijing Tongren Hospital and Fudan University Shanghai Cancer Center. Corresponding proteomic (n = 93) and metabolomic (n = 21) datasets were matched to the SYSUCC cohort and the clinical trial cohort⁵, respectively.

Fig. 1 — a A schematic overview of the study design. b The uniform manifold approximation and projection (UMAP) plot showing the annotation and color codes of the cell clusters. Each dot represents an individual cell; colors indicate distinct cell populations. c The expression patterns of marker genes in major cell clusters. For clarity, clusters are grouped into five panels: NK & T cells, B & plasma cells, myeloid cells, stromal cells, and epithelial cells. d Heatmap showing the copy number variation (CNV) profiles of malignant NK&T cells, as inferred by inferCNV. NK and T cells from healthy controls were used as reference populations. Copy number gains and losses are shown in red and blue, respectively. e Box plots showing CNV score in non-malignant (n = 114,615) and malignant NK&T cells (n = 127,248). Center lines denote median values, box limits denote interquartile ranges, and whiskers denote 1.5 × the interquartile range. P value was calculated using a two-sided Wilcoxon rank-sum test. f UMAP plots of NK&T cells grouped by non-malignant (n = 114,615) and malignant (n = 127,248) populations. A schematic was created in BioRender. Cao, Y. (2026) https://BioRender.com/ylz8st7. Source data are provided as a Source Data file.

After quality control of the 63 scRNA-seq samples, a total of 448,073 cells were classified based on distinct gene expression patterns into 62 subsets, including 30 NK and T cell subsets, nine B and plasma cell subsets, 13 myeloid cell subsets, eight stromal cell subsets and two others (ciliated/epithelial cells) (Fig. 1b, Supplementary Fig. 2a, b). Cell types were verified by known marker genes (Fig. 1c). To more precisely identify the malignant cell populations, copy number variations (CNVs)¹⁶ were used to distinguish tumor cells from non-malignant cells (Fig. 1d–f). T-cell receptor (TCR) rearrangement data were available in 28 samples, showing 80% tumor cells from NK-cell origin and 20% tumor cells from T-cell origin. EBV-related genes were highly expressed in the malignant cell populations (Supplementary Fig. 2c), aligning with the known EBV association and the characteristic pathology of NKTCL². We next assessed the batch effects and observed that cells grouped predominantly by cell types, rather than by the origin of the datasets (Supplementary Fig. 2d–f).

NMF-derived tumor meta-programs reveal distinct patient stratifications

Non-negative matrix factorization (NMF)¹⁷ of the scRNA-seq data revealed four distinct meta-programs (MP1-MP4) in tumor cells (Fig. 2a, b, Supplementary Fig. 3a–c and Supplementary Data 2). MP1 was enriched in immune response pathways and showed upregulation of immune-related transcription factors (TFs) (e.g., JUN and BCL3^18,19) (Fig. 2c, Supplementary Fig. 3d, e). MP2 was enriched in cell cycle transition and protein deacetylation pathways, along with a high level of expression of cell cycle-related TFs (e.g., E2F1 and HDAC2^20,21). MP3 was enriched in inflammatory response and JAK/STAT signaling pathways, with a high level of expression of inflammatory TFs (e.g., STAT3 and NFKB1²²). MP4 was enriched in amino acid metabolism pathways. Next, we computed the functional module score of each MP based on its signature genes and stratified 48 untreated and relapsed patients into four subgroups (MP1, MP2, MP3, and MP4) (Fig. 2d).

Principal component analysis (PCA) and distance calculations based on cell subtype proportions revealed clear distinctions among the MP subgroups (Supplementary Fig. 3f). Ro/e (observed-to-expected ratio) analysis²³, which evaluates the relative abundance and distribution of each cell cluster across sample groups (Fig. 2e, Supplementary Fig. 3g), revealed substantial T cell infiltration in patients of the MP1 subgroup, minimal immune cell infiltration in MP2 subgroup, dominance of myeloid and stromal cell populations in MP3 subgroup, and enrichment of B and plasma cells in MP4 subgroup. Jaccard similarity index²⁴, which is a measure of co-existence between cell subsets, was calculated to confirm distinct cellular co-occurrence patterns within each MP subgroup. Strong co-existence of CD4⁺ and CD8⁺ T cells was observed in MP1, a relatively sparse network involving NK cells in MP2, predominant co-occurrence of macrophages and dendritic cells in MP3, and abundant presence of B cells and IgG⁺ plasma cells in MP4 (Fig. 2f). Correlation analysis between TME cell subset frequencies revealed positive correlation between macrophage and stromal subsets (Supplementary Fig. 3h). Consistent with the role of fibroblasts in hindering immune cell infiltration to tumors²⁵, the NK/T cell subsets correlated negatively with the stromal subsets.

We next explored the relationship of the four MP subgroups with the molecular subtypes of NKTCL reported in a previous study⁶. Notably, MP1 closely matched the immune-activating TSIM subtype (Fig. 2g), both characterized by high T-cell infiltration. In a patient cohort treated with asparaginase-based chemotherapy⁶, the MP4 subgroup had significantly superior progression-free survival (PFS) compared to other three subgroups (Fig. 2h). In a patient cohort treated with immunochemotherapy⁵, longer PFS was observed in both the MP4 and MP1 subgroups. The MP2 and MP3 subgroups had poorer PFS in the immunochemotherapy cohort, indicating more aggressive disease phenotypes and treatment resistance. In our newly generated SYSUCC cohort (n = 93), multivariate analysis further confirmed that the MP2 and MP3 subgroups were independent adverse predictors of PFS after adjustment for the International Prognostic Index (IPI) or the Prognostic Index of Natural Killer lymphoma with Epstein-Barr virus (PINK-E) (Supplementary Fig. 3i, j). These findings illustrated the clinical relevance of MP-based classification.

Characterizing T cell-dominant microenvironment in MP1

In the scRNA-seq cohort, the MP1 subgroup (n = 26) was characterized by a T cell-dominant microenvironment (Supplementary Fig. 4a). Survival analysis in the immunochemotherapy cohort confirmed a positive association between MP1 signature genes and superior PFS (Fig. 3a), suggesting a potential synergy between T cell infiltration and immunotherapy. Multicolor immunohistochemistry (mIHC) analysis and single-cell resolution Xenium data confirmed a higher infiltration of CD8⁺ T cells in the MP1 subgroup (Fig. 3b, c). Analysis of TCR repertoire characteristics revealed a significantly lower Shannon index and higher TCR clonality within CD8⁺ T cells in the MP1 subgroup compared to the other three subgroups (Fig. 3d), indicating reduced TCR diversity. Notably, this pattern has been associated with improved treatment responses to ICIs in previous studies^26,27.

Fig. 3 — a Kaplan-Meier plot comparing PFS between patients stratified by MP1 signature expression in the asparaginase-based chemotherapy cohort and the immunochemotherapy cohort. P-values were derived from the log-rank test. b Representative CD8⁺ T cell mIHC images from an MP1 sample. Scale bars, 50 μm (left) and 20 μm (right). The bar plot shows quantification results (n = 5 samples per group). c Xenium spatial plots of sample P42 (MP1 subgroup) and sample P9 (MP3 subgroup). d Comparison of Shannon index and TCR clonality between the MP1 (n = 13) and the pooled MP2–MP4 subgroups (n = 9). e Heatmap illustrating the mean scaled expression of co-inhibitory and co-stimulatory molecules in CD8⁺ T cells, and *PD-L1* in tumor cells. f UMAP of CD8⁺ T cell subsets. g Density plots showing CD8⁺ T cell subset distribution along pseudotime (left) and temporal expression changes of key genes across MP subgroups (right). h Prognostic association of CD8⁺ T cell subsets in the immunochemotherapy cohort. Points indicate hazard ratios, and error bars indicate 95% confidence intervals (CIs). P-values were derived from the Cox proportional hazards model. i, Boxplots showing temporal changes in the CD8-LAYN subset. j Experimental scheme. Anti-PD-1 antibody and/or anti-LAG-3 antibody were administered to huPBMC-NOG mice following YT-wt cell injection. Anti-IgG isotype was used as a control. k Tumor growth was measured every three days using a digital caliper. l, m Flow cytometry analysis of GZMB (l) and TOX (m) expression in CD8⁺ T cells. MFI, mean fluorescence intensity. Data are presented as mean ± SD in (b, **k–m**). Box plots in (d, i) show the median, interquartile range (IQR), and 1.5 × IQR whiskers. For **j–m**, n = 4 mice per group. Statistical significance was determined using two-sided Student’s t-tests in b, a two-sided Wilcoxon rank-sum test in (d), a two-sided Wilcoxon test in (i), two-way ANOVA in (k), and one-way ANOVA in (l, m). P-values are indicated in the corresponding panels. Schematic in (j) was created in BioRender. Cao, Y. (2026) https://BioRender.com/tfvq39u. Source data are provided as a Source Data file.

The expression level of co-inhibitory molecules in CD8⁺ T cells, such as PDCD1, LAG3, and CTLA4, was highest in the MP1 subgroup, reflecting a higher degree of T cell exhaustion (Fig. 3e)²⁸. Concurrently, CD274, the ligand for programmed cell death 1 (PD-1), was notably upregulated in malignant cells within the MP1 subgroup (Fig. 3b, e). The presence of inhibitory signals in both malignant cells and CD8⁺ T cells suggests a highly immunosuppressive TME. Subclustering analyses identified 10 CD8⁺ T cell states and 11 CD4⁺ T cell states (Fig. 3f, Supplementary Fig. 4b–d and Supplementary Data 3). Pseudotime trajectory and RNA velocity analyses^29–31 identified two major differentiation branches: one leading to tissue-resident memory T cells (Trm) and the other towards an exhausted T cell (Tex) phenotype, characterized by CD8-LAYN (Fig. 3g, Supplementary Fig. 4e). Notably, CD8-LAYN cells showed a higher distribution preference in the MP1 subgroup (Supplementary Fig. 4f). Exhaustion-related genes, such as PDCD1, LAG3, and TOX, were progressively upregulated along the pseudotime trajectory and showed higher expression in MP1, correlating with CD8-LAYN accumulation (Fig. 3g). Patients with a relatively higher expression of the CD8-LAYN signature had prolonged PFS in the immunochemotherapy cohort (Fig. 3h), suggesting its potential role as a favorable prognostic marker for immunochemotherapy response.

We next evaluated the proportion of the CD8-LAYN subset before and after chemotherapy and immunochemotherapy to analyze its dynamic changes following treatment. The fraction of the CD8-LAYN subset was decreased after treatment with immunochemotherapy but not after chemotherapy (Fig. 3i). Immunochemotherapy significantly decreased the exhaustion score and increased the effector memory score of the CD8-LAYN subset (Supplementary Fig. 4g). This phenotypic shift was similar to that reported after immunochemotherapy in breast cancer³². Notably, the CD8-LAYN subset within the MP1 subgroup displayed increased LAG-3 expression after immunochemotherapy (Supplementary Fig. 4h). To this end, we next examined whether inhibition of both PD-1 and LAG-3 could enhance immunotherapeutic efficacy. Briefly, RNA sequencing data from six NKTCL cell lines (Supplementary Data 4) identified the YT cell line as representative of MP1-like characteristics, based on the expression of MP1-signature genes. In a humanized peripheral blood mononuclear cell (hu-PBMC) NOG mouse model (Fig. 3j), combination treatment with anti-PD-1 and anti-LAG-3 antibodies resulted in significantly greater tumor growth inhibition compared with either anti-PD-1 or anti-LAG-3 single-agent treatment, as well as anti-IgG antibody controls (Fig. 3k, Supplementary Fig. 4i). Combined treatment further enhanced CD8⁺ T cell function, as evidenced by elevated GZMB and reduced TOX expression (Fig. 3l, m). Collectively, these results suggest that dual immune checkpoint inhibition may offer a biologically plausible therapeutic avenue for NKTCL with MP1 characteristics, warranting further investigation in future studies.

The immune desert microenvironment with HDAC2 involvement in MP2

In the scRNA-seq cohort, the MP2 subgroup (n = 6) exhibited a distinct immune-desert microenvironment³³, characterized by low T cell infiltration (Supplementary Fig. 5a). Gene set enrichment analysis (GSEA) demonstrated downregulation of key immune-related pathways, including T cell activation, leukocyte chemotaxis, and immune effector processes, indicating a suppressed immune response in tumor cells of the MP2 subgroup (Fig. 4a). Collectively, these findings highlight the impaired immune activity associated with MP2.

Fig. 4 — a A circular plot of GSEA results for MP2 tumor cells. b Spatial expression of *HDAC2*, *CD8A* and *NCAM1* in Xenium data from sample P10 (MP2 subgroup). c Box plot showing HDAC2 protein expression in MP2 (n = 46) and other subgroups (n = 47). d Representative HDAC2 IHC images in MP2 (n = 10) and the other three subgroups (n = 10). Scale bars, 500 μm (left), 50 μm (right). e The Anti-PD-1-resistant model establishment, with three mice per group in each selection round. IP, intraperitoneal injection. f Immunoblot analysis of HDAC2 and H3K27 acetylation (H3K27ac) in YT and YT-R cells treated with or without chidamide (0.4 μM). GAPDH served as a loading control. g The half maximal inhibitory concentration (IC50) of chidamide in YT and YT-R cells (n = 3 biological replicates per group). h Experimental scheme. Anti-PD-1 antibody and/or chidamide were administered to huPBMC-NOG mice bearing YT-R tumors. Anti-IgG and PBS were used as controls. i Tumor growth was measured every three days. j Endpoint tumor weight. k Representative H3K27ac IHC images in YT-R xenograft tumors with or without chidamide treatment. Scale bars, 500 μm (left) and 50 μm (right). AOD, average optical density. l, m Flow cytometry analysis of tumor-infiltrating CD8⁺ T cells (l) and PD-1 expression (m). MFI, mean fluorescence intensity. Data are presented as mean ± SD in (d, g, **i–m**). Box plots in (c) show the median, interquartile range (IQR), and 1.5 × IQR whiskers. For **h–m**, n = 4 mice per group. Statistical significance was determined using two-sided Wilcoxon rank-sum test in c, two-sided Student’s t-tests in (d, k), two-way ANOVA in (g, i) and one-way ANOVA in (j, l, m). Schematics in (e, h) were created in BioRender. Cao, Y. (2026) https://BioRender.com/m6ca3wi and https://BioRender.com/tfvq39u. Source data are provided as a Source Data file.

Analysis of transcription factors indicated that HDAC2 was a key regulator of tumor cells in the MP2 subgroup (Supplementary Fig. 3d). Xenium spatial profiling revealed high expression of HDAC2 in tumor regions, alongside reduced infiltration of CD8⁺ T cells in a patient from the MP2 subgroup (Fig. 4b, Supplementary Fig. 5b, c). Proteomic differential analysis identified a distinct set of upregulated proteins in the MP2 subgroup, which are primarily involved in deacetylation processes and cell cycle regulation, including HDAC2, HDAC5, and MKI67 (Fig. 4c, Supplementary Fig. 5d and Supplementary Data 5). IHC analysis confirmed significantly higher HDAC2 expression and Ki-67 expression in MP2 compared to other subgroups (Fig. 4d, Supplementary Fig. 5e). Given the previous findings of reduced CD8⁺ T cell infiltration in anti-PD-1-resistant mouse models^34,35, along with the poor PFS of MP2 subgroup in the immunochemotherapy cohort (Fig. 2h), we sought to establish an NKTCL cell line with acquired resistance to anti-PD-1 antibody. Through a stepwise in vivo selection process³⁴ using hu-PBMC mice, we successfully generated the YT-R cell line (Fig. 4e, Supplementary Fig. 5f). Transcriptomic analysis revealed that YT-R cells exhibited an MP2-like phenotype (Supplementary Fig. 5g, Supplementary Data 4).

To determine whether HDAC2 promotes the growth of MP2-like tumor cells, we generated control (scramble) and HDAC2-knockdown (shHDAC2) YT-R cells using short hairpin RNA (shRNA) (Supplementary Fig. 5h) and compared their growth in vitro and in vivo. HDAC2 silencing markedly suppressed YT-R cell proliferation and tumor growth (Supplementary Fig. 5i, j). We next compared the expression of HDAC2 and global histone acetylation patterns between parental YT and YT-R cells. YT-R cells exhibited markedly elevated HDAC2 expression, and treatment with chidamide, a class I-selective histone deacetylase inhibitor (HDACi)³⁶, led to a marked increase in global H3K27 acetylation (H3K27ac) (Fig. 4f). In viability assays, YT-R cells also showed greater sensitivity to chidamide than parental YT cells (Fig. 4g).

Next, we examined the effects of chidamide, anti-PD-1 antibody, or their combination in YT-R tumor-bearing mice (Fig. 4h). The combination of chidamide and anti-PD-1 antibody further enhanced tumor inhibition in immune-resistant models (Fig. 4i, j). Chidamide significantly inhibited tumor growth, whereas anti-PD-1 treatment alone had minimal effect. Consistent with its pharmacologic activity, chidamide treatment increased H3K27ac levels in YT-R tumors (Fig. 4k). Interestingly, chidamide treatment alone markedly increased CD8⁺ T cell infiltration, whereas addition of an anti-PD-1 antibody did not further enhance this effect (Fig. 4l, Supplementary Fig. 5k). While chidamide alone did not alter PD-1 expression in CD8⁺ T cells, the combination treatment led to a notable reduction, indicating decreased T cell exhaustion (Fig. 4m). This aligned with enhanced T cell function and tumor suppression observed in a sarcoma model treated with the same combination³⁷. These results suggest that this regimen may represent a biologically plausible therapeutic avenue for NKTCL with MP2 characteristics.

Tumor-driven inflammation and myeloid-enriched immune microenvironment in MP3

In the scRNA-seq cohort, the MP3 subgroup (n = 12) was characterized by a predominance of myeloid cells. Survival analyses in both the chemotherapy and immunochemotherapy cohorts showed a positive association between MP3 signature genes and poor PFS (Fig. 5a, Supplementary Fig. 6a). Notably, all three patients with hemophagocytic lymphohistiocytosis (HLH) in our scRNA-seq data belonged to the MP3 subgroup (Fig. 2e). HLH, a hyperinflammatory condition marked by uncontrolled cytokine release³⁸, correlates with poor prognosis in NKTCL patients^39,40. Proteomic analysis revealed elevated inflammatory mediators in this subgroup, including IL-18, CCL5, and TGFB1 (Fig. 5b, Supplementary Fig. 6b and Supplementary Data 6)^41–44. Among these, IL-18 is well-established as a key mediator of HLH pathogenesis^45,46, with its overproduction contributing to the inflammatory cascade.

Next, we stratified myeloid cells into 12 subsets based on representative gene signatures (Fig. 5c, Supplementary Fig. 6c). The M.Sig score, a metric for assessing immunotherapy response⁴⁷, was highest in the Macro-TREM2 subset (Fig. 5d), underscoring its strong association with immunotherapy resistance. In the immunochemotherapy cohort, elevated Macro-TREM2 signature score correlated with poor PFS (Fig. 5e). Intercellular communication analysis revealed the highest interaction frequencies between tumor cells and tumor associated macrophages (TAMs) in the MP3 subgroup (Fig. 5f). Notably, tumor cells in MP3 demonstrated strong interactions with Macro-TREM2 via inflammatory ligand-receptor pairs (Fig. 5g). These results suggest the critical role of Macro-TREM2 in shaping the inflammatory and immunosuppressive TME of MP3.

Bulk RNA-seq analysis showed significantly higher Macro-TREM2 signature score in patients with HLH compared to those without HLH (Fig. 5h). Meanwhile, the Macro-TREM2 subset was enriched in functional activation and an inflammatory response (Supplementary Fig. 6d). Since macrophage activation is a hallmark feature of HLH^38,48, these observations suggest a potential role of Macro-TREM2 in the development of HLH. Given the reported encouraging efficacy of JAK inhibitors in treating HLH⁴⁹ and the activation of JAK/STAT signaling in MP3, we conducted a co-culture experiment to investigate the role of tumor-intrinsic JAK/STAT signaling in driving Macro-TREM2 differentiation (Fig. 5i). NKYS cells exhibited the highest MP3 signature score by transcriptome sequencing and showed increased pSTAT3 expression, consistent with an MP3-like profile, whereas SNK-6 cells displayed lower pSTAT3 expression (Supplementary Data 4, Supplementary Fig. 6e). Co-culture of macrophages with NKYS cells led to a marked increase in CD68⁺TREM2⁺ macrophages compared to SNK-6 cells and controls, suggesting that MP3-like tumor cells promote TREM2⁺ macrophage differentiation. Pretreatment with the JAK inhibitor ruxolitinib prior to co-culture reduced the proportion of CD68⁺TREM2⁺ macrophages (Fig. 5j), highlighting the role of JAK/STAT signaling in this process.

To further explore the role of Macro-TREM2 in immune evasion within the MP3 subgroup, we analyzed its intercellular interactions and identified the strongest ligand-receptor interactions with stromal cells (Supplementary Fig. 6f). Among the eight stromal cell clusters (Supplementary Fig. 6g, h), Macro-TREM2 showed the most frequent interactions with CXCL1⁺ fibroblasts (Fig. 5k), which were enriched in extracellular matrix (ECM) remodeling pathways (Fig. 5l). Bulk transcriptomics further confirmed a positive correlation between the Macro-TREM2 signature score and the CXCL1⁺ fibroblast signature score (Supplementary Fig. 6i). Proteomic analysis revealed elevated collagen-related proteins, including COL1A1, COL14A1, and COL16A1⁵⁰, highlighting ECM remodeling in MP3 (Supplementary Fig. 6b). This was further supported by Masson trichrome staining, which revealed a greater extent of fibrosis in the MP3 subgroup compared to other subgroups (Fig. 5m). mIHC analysis demonstrated abundant TREM2⁺ macrophages in the MP3 subgroup, accompanied by the exclusion of CD8⁺ T cells at the tumor-stroma interface (Supplementary Fig. 6j). Taken together, the MP3 subgroup is defined by tumor-intrinsic JAK/STAT signaling, which may drive a TAM-enriched TME and likely underlies its unfavorable prognosis. These observations suggest a biologically plausible rationale for exploring JAK inhibition as a potential avenue to modulate MP3-associated immunosuppression in future studies.

Distinct amino acid metabolic profile and TLS-enriched microenvironment in MP4

In the scRNA-seq cohort, the MP4 subgroup (n = 4) was characterized by enrichment of amino-acid metabolic pathways in malignant cells. Survival analyses in both the asparaginase-based chemotherapy and immunochemotherapy cohorts showed a positive association between MP4 signature genes and superior PFS (Fig. 6a). To explore the metabolic characteristics underlying sensitivity to asparaginase-based anti-metabolic treatment, we performed untargeted serum metabolomic analysis in 21 NKTCL patients. Metabolomic analysis revealed a distinct differential metabolite profile in the MP4 subgroup, with significant enrichment of the alanine, aspartate, and glutamate metabolism pathway (Supplementary Fig. 7a, Fig. 6b and Supplementary Data 7). This aligns with a previous study reporting that enrichment of this pathway correlates with increased sensitivity to asparaginase-based chemotherapy⁵¹.

Fig. 6 — a Kaplan-Meier PFS analysis by MP4 signature. b MetaboAnalyst pathway enrichment of MP4-associated differential metabolites (n = 21). Enrichment P-values were calculated using the global test. c Violin plots showing *SLC1A5* expression in malignant cells across MP subgroups (n = 44,054, 37,923, 25,367 and 19,904 cells for MP1–MP4). d Correlation between *SLC1A5* expression and asparagine metabolism score in the SYSUCC cohort (n = 93). e Representative SLC1A5 IHC images and AOD quantification in MP4 (n = 10) and other subgroups (n = 10). Scale bars, 500 μm (left), 50 μm (right). AOD, average optical density. f Kaplan-Meier PFS analysis by *SLC1A5* expression in the asparaginase-based chemotherapy cohort (n = 97). g SLC1A5 mRNA and protein levels in SNK-6 cells transfected with scramble control or *SLC1A5* shRNA. h Asparaginase IC50 in SNK-6 cells with scramble control or *SLC1A5* shRNA. i Differential amino acid profiles between scramble and shSLC1A5 groups; metabolites were selected by variable importance in projection (VIP) ≥ 1 and fold change ≥ 1.20 or ≤ 0.83. j Xenium map of the TLS boundary in sample P28 (MP4 subgroup); each dot represents an mRNA molecule. k Neighborhood enrichment of Leiden clusters. l Representative TLS mIHC images and TLS density in MP4 (n = 11) and other subgroups (n = 11). Scale bars, 50 μm. m Kaplan-Meier PFS analysis of TLS⁺ (n = 14) and TLS⁻ (n = 7) patients in the immunochemotherapy cohort. Data are presented as mean ± SD in (e, g, h, l); n = 3 biological replicates in (g, h, i). In (d), shading indicates the 95% CI; box plots in (c) show median, IQR and 1.5 × IQR whiskers. Statistical significance was determined using log-rank tests in (a, f, m), a two-sided Wilcoxon rank-sum test in (c), a two-sided Spearman correlation in (d), two-sided Student’s t-tests in (e, l), one-way ANOVA in (g), two-way ANOVA in (h), and a two-sided Welch’s t-test with Benjamini–Hochberg correction in (i). Source data are provided as a Source Data file.

Given the role of solute carrier (SLC) transporters in metabolic reprogramming⁵², we examined 27 amino acid transporters and identified SLC1A5—a Na⁺-dependent exchanger for glutamine, alanine, and asparagine^53,54, as the most highly expressed in the MP4 subgroup (Supplementary Fig. 7b, Fig. 6c). Bulk-RNA analyses showed positive correlations between SLC1A5 expression and aspartate/asparagine metabolism scores (Fig. 6d), and IHC further confirmed elevated SLC1A5 in the MP4 subgroup (Fig. 6e). In the asparaginase-based chemotherapy cohort, patients with high SLC1A5 expression had a prolonged PFS compared with those with low expression (Fig. 6f). In the MP4-like SNK-6 cell line (Supplementary Data 4), SLC1A5 knockdown reduced asparaginase sensitivity relative to scramble control (Fig. 6g, h). Metabolomic analysis of the culture supernatant revealed increased extracellular asparagine following SLC1A5 knockdown, suggestive of impaired cellular uptake (Fig. 6i). These findings suggest that SLC1A5 may contribute to drug sensitivity to anti-metabolic therapies and serve as a potential biomarker for the MP4 subgroup.

In addition to metabolic characteristics, the MP4 subgroup exhibited unique immune features, characterized by enriched B cells and plasma cells (Supplementary Fig. 7c). Subclustering analyses identified nine B and plasma cell subsets (Supplementary Fig. 7d, e), with IgG⁺ plasma cells and cycling memory B cells enriched in MP4 (Supplementary Fig. 7f). Given the reported predominance of IgG-producing plasma cells in TLS-positive tumors⁵⁵ and the established role of B cell-driven immunity in TLS formation⁵⁶, we evaluated TLS presence in the MP4 subgroup. Spatial transcriptomic analysis on a sample of the MP4 subgroup revealed well-organized TLS structures, characterized by distinct immune aggregates of co-localized B cells and T cells, with clear spatial segregation from surrounding tumor cells (Fig. 6j, Supplementary Fig. 7g). Neighborhood-enrichment analysis demonstrated a high degree of spatial proximity between B cells and T cells within TLS regions, underscoring their coordinated immune activity (Fig. 6k)⁵⁷. Additionally, mIHC staining further showed higher TLS density in the MP4 subgroup (Fig. 6l).

To evaluate the clinical relevance of TLS, we first analyzed our newly generated SYSUCC cohort (n = 93), most of whom received asparaginase-based chemotherapy. TLS positivity was defined by IHC detection of at least one organized CD20⁺ B-cell follicular aggregate, as described in previous studies^55,58. Multivariate analysis revealed that TLS positivity was independently associated with improved PFS (Supplementary Data 8). We also evaluated TLS maturation by assessing germinal-center (GC) structures, defined by CD21⁺/CD23⁺ follicular dendritic cell (FDC) networks⁵⁸. GC-positive TLS were enriched for immune activation-related pathways and were associated with more favorable PFS than GC-negative TLS (Supplementary Fig. 7h–j). We further examined our immunochemotherapy cohort, where TLS-positive patients had significantly better PFS than TLS-negative patients (Fig. 6m), supporting its predictive value for immunotherapy response. Intriguingly, the amino acid metabolism score was positively correlated with TLS signature (Supplementary Fig. 7k), suggesting crosstalk between metabolic and immune programs in the MP4 subgroup. Collectively, these findings suggest that the MP4 subgroup was defined by distinct metabolic features and a TLS-enriched immune microenvironment, potentially contributing to its favorable outcomes in both asparaginase-based chemotherapy and immunochemotherapy cohorts, although further functional validation is warranted.

Independent validation and prognostic relevance of MP-based classification

To independently validate this classification framework, we analyzed an external immunochemotherapy cohort comprising 48 NKTCL patients profiled by bulk RNA sequencing. The distribution of baseline clinical characteristics and immune cell compositions across the four subgroups is presented in Fig. 7a. No significant differences were observed in age, sex, or Ann Arbor stage (Supplementary Fig. 8a–c). By contrast, the incidence of HLH was significantly higher in the MP3 subgroup compared with the other subgroups (Fig. 7b). Consistent with our discovery cohort, representative signature genes displayed distinct expression patterns, with higher CD274 and PDCD1 expression in MP1, HDAC2 and MKI67 in MP2, TREM2 and COL1A1 in MP3, and SLC1A5 and MS4A1 in MP4 (Supplementary Fig. 8d). Immune infiltration analysis revealed distinct tumor microenvironmental landscapes across the four subgroups (Fig. 7c), with CD8⁺ T cells enriched in MP1 but markedly reduced in MP2, M2 macrophages predominant in MP3, and B cells enriched in MP4. Importantly, survival analysis demonstrated that MP1 and MP4 subgroups demonstrated improved PFS compared to MP2 and MP3 subgroups (Fig. 7d). These findings are consistent with those observed in our discovery cohort, highlighting the reproducibility and clinical relevance of MP-based classification in independent datasets.

Fig. 7 — a Heatmap showing the distribution of MP1–MP4 subgroups in the external immunochemotherapy cohort (n = 48). Each column represents an individual patient, and stacked bar plots indicate the relative proportions of immune cell subsets estimated by CIBERSORT. PINKE, prognostic index of natural killer lymphoma with Epstein-Barr virus; HLH, hemophagocytic lymphohistiocytosis; PFS, progression-free survival; OS, overall survival. b Stacked bar plot showing the proportion of HLH cases in MP3 versus other subgroups. P-value was calculated using a two-sided Fisher’s exact test. c Box plots showing the proportions of CD8⁺ T cells, M2 macrophages, and memory B cells across MP1-MP4 subgroups (n = 13, 12, 13 and 10 patients for MP1–MP4, respectively). Each point denotes an independent patient; boxes show medians and interquartile ranges, with whiskers extending to 1.5 × IQR. P-values were calculated using two-sided Wilcoxon rank-sum tests. d A Kaplan-Meier survival analysis of PFS stratified by MP subgroup classification in the external immunochemotherapy cohort (n = 48). P-value was calculated using the log-rank test. e Summary of the characteristics for each MP subgroup. Schematic in e was created in BioRender. Cao, Y. (2026) https://BioRender.com/xvfkk69. Source data are provided as a Source Data file.

Discussion

In this study, we performed a comprehensive analysis of the molecular and immune landscapes of NKTCL at the single-cell level. We identified four MP subgroups distinguished by specific tumor-intrinsic and immune features, with varying clinical outcomes and associations with potential biomarkers and therapeutic implications. These findings were further validated by spatial transcriptomics, bulk transcriptomics, metabolomics, proteomics, and by in vitro and in vivo experiments. The distinct features of the four MP subgroups are summarized in Fig. 7e.

ICIs have significantly altered the landscape of cancer treatment⁵⁹. Previous studies have demonstrated the effectiveness of ICIs in NKTCL, propelling it into the era of immunotherapy^3,5. However, durable responses are observed in only a subset of NKTCL patients⁵. In our study, patients in the MP1 subgroup with high baseline levels of the CD8-LAYN subset showed a favorable response to immunotherapy. The CD8-LAYN subset expresses exhaustion-related genes at high levels and resembles the CD8-CXCL13 subset in breast and colorectal cancers^32,60, which is also associated with favorable immunotherapy responses. We hypothesize that tumors in the MP1 subgroup are characterized by a ‘hot’ immune microenvironment, where PD-1 blockade reinvigorates T cell function and elicits potent anti-tumor responses. Additionally, dual blockade of PD-1 and LAG-3 has shown synergy in melanoma⁶¹, though its effect in NKTCL remains unexplored. Experiments in the hu-PBMC mouse model in the current study showed synergistic reinvigoration of CD8⁺ T cell-mediated immunity in MP1-like tumors. Consistent with this rationale, a phase I study (NCT05894421) evaluating dual blockade of LAG-3 and PD-1 in relapsed/refractory (r/r) lymphoma, including NKTCL, reported encouraging antitumor activity, with an overall response rate (ORR) of 52.38% and a disease control rate (DCR) of 80.95%⁶².

The MP2 subgroup was characterized by cell cycle activation, high expression of HDAC2, and limited immune cell infiltration, potentially limiting ICI efficacy. HDACi has been shown to enhance anti-tumor immunity by improving antigen presentation, activating NK and cytotoxic T cells, and suppressing regulatory T cells (Tregs)^63,64. In the current study, chidamide reversed the immune desert phenotype in anti-PD-1 resistant tumors by improving CD8⁺ T cell infiltration. Combined chidamide and PD-1 inhibitor alleviated T cell exhaustion and inhibited tumor growth in vivo. This aligns with encouraging results (59.5% ORR and 48.6% CR rate) reported in a phase 1b/2 trial of chidamide plus anti-PD-1 therapy in patients with r/r NKTCL⁶⁵. In that study, patients with low CD3⁺CD8⁺ T cell levels benefited more from this combination, which is consistent with the immune-desert features observed in the MP2 subgroup and supports further clinical evaluation of this strategy in NKTCL with MP2-like characteristics. A phase 2 clinical trial (NCT04994210) is ongoing to evaluate anti-PD-1 antibody plus chidamide as a first-line treatment for NKTCL.

Tumor-promoting inflammation drives immune dysregulation and accelerates tumor progression⁶⁶. HLH is characterized by hyperinflammation, with uncontrolled activation of lymphocytes and macrophages. A retrospective study showed that NKTCL patients with HLH had a 2-year overall survival (OS) of 17.4%, compared to 72.1% in those without HLH⁶⁷. Notably, all HLH patients in our single-cell data belonged to the MP3 subgroup, which exhibited the poorest prognosis and was characterized by tumor-driven inflammatory activation and elevated IL-18 levels—a key mediator linked to hemophagocytosis and poor survival in NKTCL⁶⁸. TAMs are crucial regulators in tumor immunity⁶⁹ and HLH pathogenesis⁷⁰. Consistent with a previous study showing upregulated TREM2 expression in macrophages under inflammatory conditions⁷¹, HLH patients in the current study had a higher Macro-TREM2 signature score. Furthermore, TREM2⁺ macrophages have been associated with immunotherapy resistance⁷². JAK inhibitors are being investigated as potential therapies for lymphoma-associated HLH and have shown promising results⁷³. Co-culture experiments in the current study demonstrated that inhibiting the JAK/STAT pathway reduced TREM2⁺ macrophage differentiation. These findings support the hypothesis that targeting this pathway may help mitigate immune evasion in the MP3 subgroup and may improve responsiveness to immunotherapy. A phase 2 trial (NCT06733051) is ongoing to evaluate the combination of ICIs and JAK inhibitors in r/r NKTCL.

Tumors with low asparagine synthetase (ASNS) expression, including lymphomas and leukemias, rely on exogenous asparagine for growth⁷⁴. Asparaginase-based regimens^75,76 exploit this vulnerability and improve survival outcomes in NKTCL. Dysregulation of SLC transporters, responsible for amino acid acquisition, is a critical driver of metabolic reprogramming in tumor cells⁷⁷. SLC1A1 has been identified as a key mediator of glutamine metabolism reprogramming in NKTCL, promoting glutamine addiction and enhancing susceptibility to asparaginase-based therapies⁷⁸. Previous research has indicated that SLC1A5 functions as a key transporter for extracellular asparagine uptake⁵³. Here, we identified SLC1A5 as a key transporter in the MP4 subgroup, contributing to heightened sensitivity to asparaginase-based therapies. Moreover, the MP4 subgroup was enriched in B cells and TLSs, which are well-documented to correlate with improved immunotherapy outcomes across multiple malignancies^79,80. Intriguingly, we observed a positive correlation between asparagine metabolism and TLS signatures in MP4, suggesting that metabolic activity may synergize with TLS-mediated immune activation to augment treatment efficacy. Nevertheless, the mechanistic link between tumor metabolism and TLS formation in NKTCL warrants further investigation. A phase 2 trial (NCT06583083) is ongoing to compare PD-1 blockade plus P-GEMOX versus P-GEMOX alone in NKTCL.

While this study provides valuable insights into the molecular stratification of NKTCL and its therapeutic vulnerabilities, several limitations should be acknowledged. First, the rarity of NKTCL imposes constraints on sample size and statistical power, particularly in analyses linked to treatment response and survival outcomes, and the predominance of cohorts from China may further limit the generalizability of our findings to broader populations. Second, although bulk RNA-seq provides an independent means of validating MP classification in larger cohorts, such validation is inherently subject to signal averaging across cell populations, which may affect the robustness of MP-based stratification. Third, the evaluation of TLS is constrained by its focal and heterogeneous distribution⁸¹, as limited tissue samples may not adequately capture their presence or organization, leading to underestimation and interpretative uncertainty. Fourth, although we identified key molecular features associated with therapeutic vulnerabilities, experimental validation remains limited, as mechanistic studies are incomplete and current in vitro and in vivo models may not fully capture the native disease context. Fifth, baseline–relapse longitudinal pairs are not available, and therefore, dynamic changes in cellular composition or subtype transitions over the disease course cannot be assessed. These aspects may see further advancements in the future. Despite these limitations, our findings highlight the potential of molecular stratification to inform biologically plausible therapeutic implications in NKTCL.

Methods

Ethics statement

This study complies with all relevant ethical regulations. The study protocol was approved by the Institutional Review Board of Sun Yat-sen University Cancer Center (SYSUCC). Written informed consent was obtained from all patients. Written informed consent for publication of potentially identifiable participant information included in this paper was obtained from all participants.

Patients and specimens

For the scRNA-seq analysis, we included a total of 63 samples from 53 individuals, comprising five normal nasal mucosa samples from healthy controls and 58 NKTCL patient samples. Among the 58 NKTCL samples, 10 were obtained from publicly available datasets, while 48 were in-house collected from our institution. The in-house samples included 31 treatment-naive cases, 10 matched post-treatment samples collected after the first course of therapy, and seven independent relapse samples. The relapsed samples were from patients who had experienced relapse following either radiotherapy alone or anthracycline-based chemotherapy as first-line treatment. For the 10 paired post-treatment samples, all were collected after the first course of treatment. The remaining two datasets were obtained from the GEO and GSA database under accession numbers GSE203663¹⁰ and HRA000772¹⁴, respectively. Xenium in situ expression analysis was conducted in 14 treatment-naive samples from 14 patients. Detailed clinicopathologic features are summarized in Supplementary Data 1.

The bulk RNA-seq analysis included a total of 259 patients: 97 patients who received asparaginase-based chemotherapy from a previously published dataset⁶; 21 patients from a previously published clinical trial of immunochemotherapy (ClinicalTrials.gov identifier NCT04127227; https://clinicaltrials.gov/study/NCT04127227; first posted on October 14, 2019)⁵; and 141 newly generated cases, including 93 from Sun Yat-sen University Cancer Center (SYSUCC) and 48 from Beijing Tongren Hospital and Fudan University Shanghai Cancer Center. The 48 external validation cases were treated with immunochemotherapy. The proteomic analysis included 93 samples, which were derived from the same 93 patients diagnosed at SYSUCC in the transcriptomic analysis. The metabolomics analysis included 21 serum samples with matched bulk RNA-seq data from the previously published immunochemotherapy trial cohort (NCT04127227)⁵. The existing data from this cohort were reused for retrospective, exploratory secondary analyses in the present study.

For the newly generated patient cohorts included in this study, all individuals were diagnosed and treated at the respective institutions. The histological diagnosis of NKTCL was independently reviewed and confirmed by two experienced hematopathologists, in accordance with the WHO classification criteria. In cases where there was any discrepancy between the initial assessments, a third experienced hematopathologist was consulted to reach a final consensus on the diagnosis. This multi-tiered diagnostic review process helped ensure the accuracy and consistency of the diagnoses prior to inclusion in the study. Specifically, most cases demonstrated surface CD3 negativity, cytoplasmic CD3ε positivity, CD56 positivity, and Epstein-Barr virus-encoded small RNA (EBER)-in situ hybridization (ISH) positivity. For CD56-negative cases, the expression of cytotoxic molecules such as granzyme B, together with EBER-ISH positivity, was required. Histologically, NKTCL was characterized by a diffuse, permeative lymphoid infiltrate, often associated with coagulative necrosis and admixed apoptotic bodies.

We did not report gender. Participant sex was reported as stated by the participants themselves. Sex was not used as an eligibility or sample-selection criterion. We did not perform analyses stratified by sex because the sample size at the participant level was insufficient to support meaningful sex-based comparisons.

Tissue portions used for single-cell dissociation and molecular analyses were consumed during sample processing. Remaining archived tissue blocks or slides, where available, are retained at the corresponding participating institutions and are not publicly shared because of institutional ethics approval, participant consent and relevant human genetic resources regulations. Requests for access to remaining archived materials for research purposes should be submitted to the corresponding author via email at caiqq@sysucc.org.cn and are subject to institutional approval.

Sample processing and single-cell preparation for scRNA-seq

Fresh tumor specimens were transported on ice in sterile culture dishes containing 10 mL of 1 × Dulbecco’s phosphate-buffered saline (DPBS; Thermo Fisher, Cat. no. 14190144) to remove residual tissue storage solution. The samples were minced and enzymatically dissociated with 0.25% trypsin (Thermo Fisher, Cat. no. 25200-072) and 10 µg/mL DNase I (Sigma, Cat. no. 11284932001) in PBS, supplemented with 5% fetal bovine serum (FBS; Thermo Fisher, Cat. no. SV30087.02). Dissociation was performed at 37 °C with gentle agitation at 50 rpm for approximately 40 min, with cell suspensions collected every 20 min to optimize yield and viability. The resulting cell suspensions were filtered through a 40-µm nylon cell strainer to remove debris and treated with 1 × red blood cell lysis solution (Thermo Fisher, Cat. no. 00-4333-57) to eliminate red blood cells. Cells were then washed with DPBS containing 2% FBS and assessed for viability using 0.4% trypan blue (Thermo Fisher, Cat. no. 15250061) on a Countess II Automated Cell Counter (Thermo Fisher).

scRNA-seq and TCR V(D)J library generation and sequencing

scRNA-seq and TCR V(D)J libraries were prepared using the 10x Genomics Single Cell Immune Profiling Solution Kit v2.0 (10x Genomics, 1000263). Cells were resuspended in PBS containing 0.04% BSA at a concentration of 100–800 cells/μL and processed to capture approximately 500–7000 cells per sample. After reverse transcription and cell barcoding in droplets, cDNA was purified using Dynabeads MyOne SILANE and amplified by PCR. The amplified cDNA was used to generate both 5’ gene expression libraries and V(D)J-enriched libraries. Library preparation involved cDNA fragmentation, end repair, size selection using SPRIselect beads, and sample indexing PCR. Sequencing was conducted on the Illumina NovaSeq 6000 platform. For the single-cell RNA libraries, sequencing was performed to achieve a minimum depth of 25,000 reads per cell, utilizing read lengths of 26 bp for Read1, 8 bp for the i7 Index, and 98 bp for Read2. Similarly, the single-cell TCR V(D)J libraries were sequenced to a minimum depth of 5000 reads per cell, with read lengths of 150 bp for Read1, 8 bp for the i7 Index, and 150 bp for Read2.

scRNA-seq data processing

The raw scRNA-seq data were analyzed using the Cell Ranger 3.0.1 pipeline from 10× Genomics. The FASTQ files obtained from Illumina sequencing were aligned to the human reference genome (GRCh38) and EBV reference sequence (Akata). Next, gene-barcode matrices were constructed by counting unique molecular identifiers (UMIs) and excluding non-cell-associated barcodes. Finally, a gene-barcode matrix, containing gene expression counts for the barcoded cells, was generated. The resulting matrix was imported into the Seurat (version 4.3.0) R package⁸² for downstream analysis of the single-cell RNA-seq data. For quality control, cells with fewer than 300 detected genes, fewer than 1000 UMIs, or more than 20% mitochondrial gene expression were excluded. Doublets were identified and removed using the DoubletFinder package (version 2.0.3)⁸³, with an expected doublet rate of 8%. After filtering low-quality cells, a total of 448,073 cells were retained for downstream analysis. The filtered matrix underwent normalization and scaling, incorporating the top 2000 highly variable genes.

Unsupervised clustering and cell type annotation

Principal component analysis (PCA) was conducted on the matrix of variable genes, with the top 20 principal components (PCs) selected for downstream analysis. To mitigate inter-sample batch effects arising from independent sample processing and dataset origin, the Harmony algorithm⁸⁴ was applied to the PCA embeddings with the per-sample identifier (orig.ident) specified as the batch covariate. Harmony was run with default parameters (theta = 2, lambda = 1, max.iter.cluster = 20). The Harmony-corrected embeddings were subsequently used to construct the k-nearest-neighbor graph and to perform unsupervised graph-based clustering (Louvain; resolution = 0.5), and were also used for UMAP visualization of cell neighborhoods. For the initial round of clustering that delineated the five major cell types, the resolution of 0.5 was selected based on Clustree analysis; each major cell type was further clustered at an optimal resolution, likewise determined by Clustree, yielding a total of 62 subsets. To annotate the first-round cell clusters, we classified them into 5 major types based on the expression of well-established marker genes (NK&T cells: CD3D, CD3E, GZMB, NCAM1, NKG7; B&Plasma cells: MZB1, CD19, MS4A1, CD79A, CD79B; Myeloid cells: CD14, LYZ, CD36, CD68, C1QC; Stromal cells: COL1A1, COL11A1, COL15A1, DCN, PECAM1; Epithelial cells: EPCAM, KRT5, KRT8, KRT19, CLDN4).

To refine cell subset annotation, the Seurat pipeline was applied to each major cell type, with batch effects corrected using Harmony. Cluster-specific marker genes were identified using the ‘FindAllMarkers’ function in Seurat, with the Wilcoxon rank-sum test applied by default (min.pct = 0.25, logfc.threshold = 0.25). For normal NK&T cells, we identified 8 NK cell clusters, 11 CD4⁺ clusters, and 10 CD8⁺ clusters, based on published signature genes. For myeloid cells, 13 clusters were identified, including monocyte-like cells, macrophages, neutrophils, conventional dendritic cells (cDC), mast cells, and plasmacytoid dendritic cells (pDC). B and plasma cells were classified into 9 clusters and further annotated based on marker genes. For stromal cells, we identified 8 clusters, including fibroblasts, pericytes, and endothelial cells, all of which were annotated using marker genes. Detailed marker gene lists are provided in Supplementary Data 3.

Single-cell copy-number variation analysis

To distinguish malignant from non-malignant NK/T cells, copy number variation (CNV) analysis was conducted using the inferCNV R package (version 1.16.0)¹⁶. NK and T cells from healthy controls were used as reference populations. The expression matrix was extracted, and genes were mapped to chromosomal locations. CNV inference was conducted with noise reduction enabled, applying a hidden Markov model and setting a threshold of 0.1 to detect copy number alterations. Cells were classified based on CNV profiles, and a total CNV score was calculated to quantify genomic instability across individual cells.

Characterizing intratumor heterogeneity and meta-programs of NKTCL

To explore intratumor transcriptional heterogeneity in NKTCL, we applied a non-negative matrix factorization (NMF) algorithm¹⁷ to the relative expression matrix of malignant cells. For each tumor sample, variable genes (n = 3000) were identified after excluding mitochondrial, ribosomal, and hemoglobin-related genes. Expression matrices were normalized using counts-per-million (CPM), and NMF decomposition was performed with ranks ranging from 3 to 7. The top 50 genes from each rank were selected as candidate programs, and robust programs were identified based on consistency across ranks. Pairwise Jaccard similarities were calculated to cluster programs into meta-programs (MPs), with each MP defined by genes shared in at least 25% of its constituent programs. This analysis identified four distinct MPs, with the top 5 highest-scoring genes selected from each MP as its signature, ranked according to their contribution to each program. The complete list of the top 50 genes for each MP is provided in Supplementary Data 2.

Differential expression and gene set enrichment analysis

Differentially expressed genes (DEGs) between clusters were determined using the FindMarkers function in Seurat, employing a Wilcoxon rank-sum test with thresholds of min.pct = 0.25 and logfc.threshold = 0.25. Functional enrichment analysis was then conducted on the identified DEGs with the clusterProfiler package (version 4.8.3), using hallmark and Gene Ontology (GO) gene sets from the Molecular Signatures Database (MSigDB). Enrichment scores across clusters were compared to identify biological pathways with significant associations. Furthermore, to assess gene set enrichment for individual MPs, Gene Set Variation Analysis (GSVA) was performed with default settings from the GSVA package (version 1.50.2).

Definition and calculation of signature score

To define the transcriptional signatures of distinct cell subsets, we identified the top 20 upregulated genes for each subset based on average log₂(fold change) values and used them as cell cluster signature genes. Signature scores were computed at the single-cell level using the AddModuleScore function in Seurat. For the analysis of functional states of T cells, we applied a previously defined gene signature and scoring method³². The exhaustion score was calculated using PDCD1, LAG3, LAYN, HAVCR2, CXCL13, TIGIT, CTLA4, GZMB, TOX, RBPJ, VCAM1, and MYO7A, while the effector memory score was determined based on IFNG, CCL4, HLA-DQA1, PRF1, GZMA, CD44, DUSP2, GZMK, KLRB1, KLRD1, GZMH, and CTSW. Additionally, a previously defined TLS signature⁵⁷, composed of CD79B, CCR6, CD1D, LAT, CETP, SKAP1, RBP5, EIF1AY, and PTGDS, was used to evaluate TLS score across samples. To assign the aspartate and asparagine metabolism score, gene signatures were derived from the GSEA MSigDB gene set “REACTOME_ASPARTATE_AND_ASPARAGINE_METABOLISM”. For bulk transcriptomic analysis, GSVA was applied to assess the enrichment of defined gene signatures at the patient level. The detailed gene signatures are summarized in Supplementary Data 9.

Single-cell transcription factor regulon analysis

Single-cell regulatory network analysis was performed using the R package SCENIC (version 1.3.1)⁸⁵. Regulon activity was assessed using the AUCell package, and transcription factors with significant activity across tumor MPs were identified. Wilcoxon rank-sum tests were used to compare the activity of transcription factors between distinct MPs.

Matching with reported NKTCL molecular subtypes

To compare our classification with previously reported molecular subtypes, we identified marker genes for each subgroup by performing differential expression analysis on transcriptomic data using the limma package. For each subgroup, the top 50 upregulated genes were selected based on adjusted P-values and log₂(fold change) values as subgroup-specific signature genes. Next, pseudo-bulk expression data were generated by aggregating single-cell RNA-seq counts at the patient level. Signature scores for each patient were calculated using the AddModuleScore function in Seurat, normalized against control feature sets. Patient-level subtype assignments were then determined based on the highest z-score for the respective signature genes. Finally, subtype distributions were compared between MP classifications and existing subtypes using cross-tabulation and visualized via Sankey diagrams to highlight concordance patterns.

Pseudotime analysis and RNA velocity analysis

To investigate the differentiation trajectory of CD8⁺ T cell subsets, pseudotime analysis was conducted using the Monocle3 (version 1.3.4) algorithm^29,30. The raw gene expression data were first preprocessed and normalized using the preprocess_cds function, followed by batch effect correction with the align_cds function. For dimensionality reduction, PCA was applied, and clustering was performed using the cluster_cells function. The cells were ordered along a trajectory using the learn_graph and order_cells functions. The CD8_C02_CCR7 (naive T) cell subset was designated as the root for trajectory construction. Trajectories were visualized using UMAP, and pseudotime scores were assigned to each cell. To assess the expression patterns of specific genes along pseudotime, plots were generated to visualize changes in gene expression across the trajectory. For RNA velocity analysis, spliced and unspliced transcript counts were extracted using the velocyto pipeline³¹. A k-nearest neighbor (k = 25) graph was constructed to model transcriptomic relationships, and RNA velocity vectors were estimated and projected onto the reconstructed trajectory to infer lineage transitions within CD8⁺ T cells. Default parameters were used for velocity estimation and visualization.

Cell subsets distribution analysis

To assess the distribution preference of cell subsets, we computed the ratio of observed to expected cell counts (Ro/e) for each cluster across various subgroups. The expected cell counts were calculated using a chi-square test, based on a previously established methodology²³. An Ro/e value greater than 1 suggests that a specific cell subset is observed more frequently than expected by chance (enriched), while an Ro/e value less than 1 suggests depletion in that tissue.

Correlation analysis

For single-cell level analysis, we calculated the proportions of immune cell subsets for each sample and assessed their correlations using Spearman’s rank correlation. Proportions were determined by normalizing the number of cells in each subset relative to the total number of cells in the corresponding sample. For bulk transcriptomic analysis, gene signature scores were log-transformed to mitigate the impact of extreme values. Spearman’s correlation test was applied to assess associations between gene signatures.

Cellular abundance calculation and cellular co-existence analysis

To calculate the relative abundance of different cell subsets in each sample, we initially determined the proportions of each subset by normalizing the number of cells within each subset to the total count of non-epithelial and malignant cells. Subsequently, the sample-cell proportion matrix was scaled by cell type to normalize the abundance values across the subsets. Jaccard similarity index was calculated to assess the co-existence of cell subsets²⁴. The Jaccard index for each pair of subsets was computed as the ratio of their intersection to their union, where subsets with a relative cellular abundance above a certain threshold were considered present.

scTCR-seq data processing and analysis

scTCR-seq data were processed using the Chromium Single-Cell V(D)J Enrichment Kit (10x Genomics) according to the manufacturer’s instructions. The Cell Ranger vdj pipeline (version 5.0.0; 10× Genomics) was used to align raw TCR-seq reads to the human reference genome (GRCh38) and identify productive TCR alpha and beta chain sequences. The assembly was restricted to full-length, productive chains that were associated with valid cell barcodes. Each cell was assigned the pair of TCR alpha and beta chains with the highest UMI counts. Cells with identical TCR alpha and beta pairs were defined as clonally related, originating from the same ancestral lineage. Cells with at least two productive TCR chains (alpha and beta) were retained for downstream analysis. For clonality analysis in CD8⁺ T cells, we utilized the scRepertoire R package. Cells with clonally expanded TCRs were defined as those with shared TCR alpha and beta chains across at least three cells.

Cell-cell communication analysis

Cell-cell communication analysis was analyzed using the LIANA R package⁸⁶, which integrates multiple computational methods to infer ligand-receptor interactions at the single-cell level. For each MP subgroup, interactions were identified based on predefined ligand-receptor databases, with interaction strength assessed using the built-in CellPhoneDB scoring metric. To refine the characterization of key interactions, we focused on ligand-receptor pairs involved in cell communication. Specific interactions were further investigated to identify those preferentially enriched in certain subgroups.

Xenium in situ analysis

Xenium in situ transcriptomic profiling was conducted using the Xenium Prime 5 K Human Pan Tissue and Pathways Assay Kit (2 rxns) PN-1000671. Formalin-fixed paraffin-embedded (FFPE) tissue sections (5 μm thick) were prepared, mounted onto Xenium slides, deparaffinized and de-crosslinked before hybridization with probes at 50 °C overnight (~20 h). Hybridized sections were subjected to post-hybridization washes, followed by ligation (37 °C for 2 h) and amplification (30 °C for 2 h). Autofluorescence quenching and nuclei staining were performed before imaging on the Xenium Analyzer. The resulting fluorescence signals were captured at high resolution to ensure accurate spatial localization of transcripts. Image segmentation and transcript assignment were performed using Xenium’s built-in processing pipeline.

Raw Xenium output files, including transcript count matrices, cell segmentation masks, and spatial coordinates, were exported from Xenium Explorer for further analysis. Cell segmentation and transcript localization were performed using the default Xenium Explorer pipeline. For downstream processing, data were imported into Python and converted to Zarr format using the spatialdata package. Scanpy was used for quality control, normalization, and clustering. Quality control (QC) metrics included total transcript counts per cell, the number of unique genes per cell, cell segmentation area distribution, and nucleus-to-cytoplasm area ratio. Low-quality cells were removed based on a minimum transcript count threshold, and genes were retained if detected in at least five cells. Following QC, transcript counts were log-normalized, and PCA was performed using the top 30 principal components. UMAP embedding was applied to visualize the spatial distribution of cells, and Leiden clustering was used to identify transcriptionally distinct populations. To assess spatial organization, a spatial connectivity graph was computed using squidpy.gr.spatial_neighbors, defining neighbors based on centroid-based spatial coordinates. Neighborhood enrichment analysis was performed using squidpy.gr.nhood_enrichment to evaluate cell-type interactions within local spatial contexts. All analyses were conducted using Python 3.9.21, Scanpy (version 1.10.3), Squidpy (version 1.6.1), and Spatialdata (version 0.2.5).

Bulk-RNA sequencing and survival analysis

Total RNA was extracted from FFPE tissues using the RNeasy mini kit (Qiagen, Germany). Strand-specific RNA libraries were prepared with the TruSeq Stranded Total RNA Sample Preparation Kit (Illumina, USA) following the manufacturer’s instructions. mRNA was enriched with oligo(dT) beads, purified, and fragmented using divalent cations at 86 °C for 6 min. First-strand cDNA synthesis was performed using random primers and reverse transcriptase, followed by second-strand synthesis with DNA Polymerase I and RNase H. The cDNA fragments underwent end repair, A-tailing, and adapter ligation prior to PCR amplification. Libraries were purified, quantified using a fluorometric method, and their fragment distribution was assessed on an Agilent 2100 Bioanalyzer (Agilent Technologies, USA). Sequencing was conducted on an Illumina NovaSeq 6000 platform.

Raw sequencing reads underwent quality control using Seqtk, which included adapter trimming and removal of low-quality sequences. The filtered reads were aligned to the reference genome with HISAT2 (v2.0.4), enabling spliced alignment to map reads across exon-exon junctions. Gene expression was quantified using StringTie (v1.3.3b), generating transcript abundance estimates for each sample. To adjust for variations in sequencing depth and RNA composition, expression values were normalized using the Trimmed Mean of M values (TMM) method. The normalized values were then converted to Fragments Per Kilobase of transcript per Million mapped reads (FPKM) for subsequent analyses.

Survival analysis was conducted using the R packages survival (version 3.5.5) and survminer (version 0.4.9). Patients were classified into one of four MP groups based on their gene signature score, with each patient assigned to the group with the highest score. While patients were further stratified into high and low expression groups, the optimal threshold for each gene set was determined using the surv_cutpoint function. All survival analyses were visualized using the ggsurvplot function, and statistical comparisons between groups were performed using the log-rank test.

Proteomic analysis

Protein extraction, denaturation, reduction, alkylation, tryptic digestion, and peptide cleanup were performed using the iST Sample Preparation Kit (PreOmics, Germany) following the manufacturer’s protocol. Briefly, FFPE tissue sections were lysed in LYSE buffer, treated with a beatbox for 10 min, and incubated at 95 °C for 1 h with agitation (1000 rpm). After cooling, supernatants were collected, and trypsin digestion was carried out at 37 °C for 3 h (500 rpm). Digestion was stopped with a stop buffer, and peptides were cleaned and desalted using the iST cartridge. Peptides were eluted (2 × 100 µL) and lyophilized via SpeedVac.

For high-pH fractionation, peptides were dissolved in buffer A (H₂O, pH 10.0, ammonium hydroxide) and separated using a nanoACQUITY UPLC system (Waters, USA) with a C18 column (2.5 mm × 250 mm, 1.9 µm, Omicsolution). A non-linear gradient was applied: 4% to 53% buffer B (80% ACN, pH 10.0) in 40 min, 53% to 68% in 15 min, 68% to 95% in 5 min (held for 3 min), and re-equilibrated (17 min, 2 µL/min, 30°C). Fractions (n = 24) were collected (EasyPept Frac nano, Omicsolution) and dried. For LC-MS/MS, peptides (200 ng) were resuspended in 0.1% FA and separated on an AUR3-15075C18 column (15 cm × 75 µm, 1.7 µm, 120 Å, IonOpticks) using an UltiMate 3000 system (Thermo Fisher) coupled to a timsTOF Pro 2 (Bruker). The 80-min gradient was: 4% to 28% buffer B (80% ACN, 0.1% FA) in 45 min, 28% to 44% in 10 min, 44% to 90% in 10 min (held for 7 min), and re-equilibrated at 4% for 8 min (400 nL/min, 50°C). DIA data were acquired in diaPASEF mode with 22 × 40 Th precursor isolation windows (m/z 349–1229). The MS1 cycle time was optimized with variable repetitions (2–5) in a 13-scan diaPASEF scheme. PASEF MS/MS scans applied a collision energy ramp from 59 eV at 1/K0 = 1.6 Vs/cm² to 20 eV at 1/K0 = 0.6 Vs/cm².

DIA raw data were processed in Spectronaut 19 (Biognosys) with default settings. The database used was Homo sapiens (UniProt, version 2022, 20,610 entries). Trypsin digestion (specific cleavage) was carried out with fixed carbamidomethylation (Cys) and variable modifications, including oxidation (Met) and N-terminal acetylation. Retention time prediction utilized dynamic iRT, with extraction windows adjusted according to mass calibration. Q-value (FDR) cutoffs were set at 1% for precursor, peptide, and protein levels. Decoy sequences were generated through a mutation-based approach. Normalization was done using local normalization, and protein quantification followed the MaxLFQ method.

Protein expression levels were normalized before statistical analysis. Patients were grouped based on transcriptomic classifications. Differentially expressed proteins were identified using the limma package, with multiple testing correction applied via the Benjamini-Hochberg FDR procedure. Proteins with an FDR-adjusted p-value < 0.05 and a fold change greater than 1.5 or less than 1/1.5 were considered significant. Significantly upregulated proteins in the MP2 and MP3 subgroups are listed in Supplementary Data 5, Supplementary Data 6, respectively.

Serum metabolomics analysis

Serum metabolomics analysis was conducted as previously described⁵. Briefly, the analysis included 21 independent serum samples; samples were processed within 2 hours of collection and stored at −80 °C. For each sample, hydrophilic and hydrophobic compounds were extracted from 50 μL of serum using internal-standard-containing extraction solvents, including acetonitrile/methanol for hydrophilic compounds and methyl tert-butyl ether/methanol for hydrophobic compounds. Chromatographic separation was performed using a Waters ACQUITY UPLC HSS T3 C18 column for hydrophilic compounds and a Thermo Accucore C30 column for hydrophobic compounds, followed by LC-ESI-MS/MS on a UPLC ExionLC AD system coupled to a QTRAP® triple quadrupole-linear ion trap mass spectrometer (Sciex), operated in both positive- and negative-ion modes. Pooled quality-control samples and blank samples were included in the analytical sequence, with QC samples inserted approximately every ten injections to monitor instrument stability, reproducibility and carryover. Data acquisition used scheduled multiple reaction monitoring (MRM) with linear ion trap (LIT) and triple quadrupole (QQQ) scans; instrument control and data processing were performed using Analyst version 1.6.3 (Sciex).

Metabolite identification and quantification were carried out in MRM mode using a stepwise multiple ion monitoring-enhanced product ions (MIM–EPI) strategy. MS2 spectral-tag data were interpreted by comparing the accurate precursor spectral ion values selected by the first quadrupole (Q1; precursor-ion selector) and product spectral ion values selected by the third quadrupole (Q3; product-ion selector), together with retention time (RT) and fragmentation patterns. Primary matching used the Metware Database (MWDB, custom target standard-product library), with additional reference to established resources (MassBank, KNApSAcK, Human Metabolome Database, and METLIN). The reproducibility of extraction and detection was evaluated by monitoring total ion current (TIC) and multiple-peak MRM quantifications. Peak areas were exported, log₂-transformed, and normalized across samples prior to downstream statistical analysis.

Differential analysis was performed using Metware Cloud, an online data analysis platform (https://cloud.metware.cn). Metabolites with a p-value < 0.05 and a fold change ≥ 1.20 or ≤ 0.83 were considered statistically significant. Pathway enrichment analysis was conducted using MetaboAnalyst6.0⁸⁷. Metabolic pathways enriched in the MP4 subgroup are listed in Supplementary Data 7.

Cell lines and cell culture

This study used six human NKTCL cell lines: YT, NKYS, KHYG-1, SNK-1, SNK-6, and NK-92. KHYG-1 cells were obtained from the Japanese Collection of Research Bioresources Cell Bank (Cat. no. JCRB0156), and NK-92 cells were obtained from the American Type Culture Collection (ATCC, Cat. no. CRL-2407). SNK-1, SNK-6 and YT cells were purchased from Otwo Biotech (China; Cat. nos. HTX2588C, SHC2318 and HTX3529, respectively), and NKYS cells were kindly provided by Dr. Bai (Sun Yat-sen University Cancer Center, Guangzhou). For culture conditions, NKYS and KHYG-1 cells were grown in RPMI-1640 medium (Gibco) supplemented with 10% fetal bovine serum (FBS, Gibco) and 100 IU/mL recombinant human IL-2 (PeproTech), while YT cells were maintained in IMDM (Gibco) with 20% FBS. NK-92, SNK-1, and SNK-6 cells were cultured in RPMI-1640 medium with 10% FBS, 10% horse serum (HS, Gibco), and 100 IU/mL IL-2. All culture media contained 1% penicillin-streptomycin (Gibco), and mycoplasma contamination was regularly monitored. Cells were cultured at 37 °C in a humidified incubator with 5% CO₂.

Cell transfection

Two short hairpin RNAs (shRNAs) targeting HDAC2 were designed, with the following sequences: shHDAC2-1, forward sequence GACGGTATCATTCCATAAATA; and shHDAC2-2, forward sequence CCCAATGAGTTGCCATATAAT. Two shRNAs targeting SLC1A5 were also generated. For shSLC1A5-1, the forward sequence was GCCTGAGTTGATACAAGTGAA, and the reverse sequence was TTCACTTGTATCAACTCAGGC. For shSLC1A5-2, the forward and reverse sequences were GCAGTCCTTGGACTTCGTAAA and TTTACGAAGTCCAAGGACTGC, respectively. All shRNA sequences targeting HDAC2 and SLC1A5 were cloned into the pLKO.1-puro lentiviral vector. The vector was obtained from IGE (Guangzhou, China). Lentiviral particles were produced by co-transfecting HEK293T cells with the pLKO.1-shRNA plasmid, the packaging plasmid psPAX2, and the envelope plasmid pMD2.G at a ratio of 4:3:1, using polyethyleneimine (PEI; Sigma-Aldrich) according to the manufacturer’s instructions. Supernatants were harvested 48 h after transfection, filtered through a 0.45-μm membrane, and concentrated by ultracentrifugation.

For transduction, YT-R and SNK-6 cells were seeded at a density of 5 × 10⁵ cells per well in 6-well plates and infected with lentiviral supernatants in the presence of 8 μg/mL polybrene. After 48 h, SNK-6 cells were refreshed with complete RPMI-1640 medium, whereas YT-R cells were replaced with fresh complete IMDM to remove residual viral particles.

Quantitative Real-Time PCR (qRT-PCR)

Total RNA was extracted from cells using TRIzol reagent (Thermo Fisher Scientific) following the manufacturer’s instructions. RNA concentration and purity were measured using a NanoDrop 2000 spectrophotometer (Thermo Fisher Scientific). cDNA was synthesized from total RNA using the HiScript III RT SuperMix (Vazyme, China). Quantitative real-time PCR (qRT-PCR) was performed with ChamQ Universal SYBR qPCR Master Mix (Vazyme) on a LightCycler 480 system (Roche, Germany). Relative gene expression levels were calculated using the 2^−ΔΔCt method and normalized to GAPDH expression. The primers for human SLC1A5 and human GAPDH were purchased from IGE (Guangzhou, China). The SLC1A5 primer sequences were as follows: forward primer, 5′-GAGCTGCTTATCCGCTTCTTC-3′; reverse primer, 5′-GGGGCGTACCACATGATCC-3′. The GAPDH primer sequences were as follows: forward primer, 5′-GGAGCGAGATCCCTCCAAAAT-3′; reverse primer, 5′-GGCTGTTGTCATACTTCTCATGG-3′.

Cell viability

Cells (5 × 10³) were plated into 96-well plates and treated with the indicated concentrations of reagents. Cell viability was determined with the CCK-8 assay (1:10 dilution; Biosharp), and absorbance at 450 nm was recorded using a microplate reader.

In vivo experiments

The humanized PBMC mice (female, HuPBMC-NOG-DKO) aged 3–4 weeks were sourced from Beijing Vital River Laboratory Animal Technology Co., Ltd (China). The NOG-DKO (Strain NO. 411) mice were humanized using cryopreserved PBMCs derived from a single healthy donor. Briefly, PBMCs were isolated from fresh leukopak via Ficoll density gradient centrifugation, assessed for viability and immune cell composition by flow cytometry, and cryopreserved in CryoStor® CS10 medium using a controlled-rate freezing protocol. Prior to injection, PBMCs were thawed and used immediately without ex vivo expansion. Each NOG-DKO mouse was intravenously injected with 5 × 10⁶ human PBMCs via the tail vein.

These mice were randomly divided into experimental groups and housed in standard filter-top cages under specific pathogen-free conditions at 22–26 °C and 40–60% relative humidity, with ad libitum access to sterile water and food and a regular 12-h light/dark cycle. All procedures involving animals were performed in compliance with the protocol approved by the Institutional Animal Care and Use Committee (IACUC) at SYSUCC (L102012025020D). The maximal tumor volume permitted by the institutional animal ethics committee was 2000 mm³, and this limit was not exceeded in any animal during the study.

YT wild-type (YT-wt) cells (3 × 10⁷) suspended in 200 μL of matrigel (17.34 mg/ml, 0827245, ABW) were subcutaneously injected into the dorsal flank of each mouse. Six days after inoculation, the anti-PD-1 antibody sintilimab (10 mg/kg, Selleck Chemical) and the LAG-3 inhibitor relatlimab (5 mg/kg, Selleck Chemical) were administered intraperitoneally every 3 days. Anti-IgG4 isotype (Selleck Chemical) was used as a control for both anti-PD-1 and anti-LAG-3 treatments.

To establish the in vivo anti-PD-1 resistant model, YT-wt cells stably expressing green fluorescent protein (GFP) were injected (1 × 10⁷ cells) subcutaneously into the dorsal flank of each Hu-PBMC mouse. Mice were treated with anti-PD-1 antibody sintilimab (10 mg/kg, Selleck Chemical) intraperitoneally four times at 3-day intervals. Tumors were dissociated for in vitro culturing, and the cycle was repeated two more times to obtain YT-R cells. For PD-1 blockade combined with chidamide (HDAC2 inhibitor) treatment, YT-R tumor-bearing mice were treated with anti-PD-1 antibody and/or chidamide (20 mg/kg orally daily, Chipscreen Biosciences). Anti-IgG control (Selleck Chemical) was used as the control for anti-PD-1; PBS was used as the control for chidamide.

Macrophage differentiation and construction of in vitro co-culture system

Human monocytes were isolated from the peripheral blood mononuclear cells (PBMCs) of healthy donors using the method described in a previous study⁸⁸. Briefly, PBMCs were separated via Ficoll density gradient centrifugation, and CD14⁺ monocytes were enriched with magnetic beads (Miltenyi Biotec) following the manufacturer’s guidelines. The isolated CD14⁺ monocytes were then plated into 6-well plates (5 × 10⁵ cells per well) and cultured in RPMI-1640 medium containing 10% FBS and 20 ng/mL granulocyte-macrophage colony-stimulating factor (GM-CSF) for 5–7 days to promote macrophage differentiation. Next, an in vitro co-culture system was established using 0.4 µm Transwell inserts (Corning Costar). A total of 1 × 10⁵ tumor cells were seeded into the upper compartment of the Transwell and co-cultured with differentiated macrophages for 24 h. Cells in the lower compartment were then collected and analyzed by flow cytometry. For experiments involving JAK inhibition, tumor cells were pre-treated with 10 μM ruxolitinib (Selleck) for 48 h before co-culture.

Flow cytometry analysis

For flow cytometry analysis of primary tumor samples, tumors were dissociated using 0.5 mg/mL collagenase II (Worthington Biochemical) and 0.25 mg/mL DNase (Servicebio) at 37 °C for 30 min. The resulting cell suspension was filtered through 70 μm nylon strainers to remove large fragments and washed with PBS. Cells for surface staining were incubated with fluorophore-conjugated antibodies targeting specific markers for 30 min at 4 °C in the dark. For intracellular staining, cells were fixed and permeabilized with the Foxp3/Transcription Factor Staining Buffer Set (Thermo Fisher Scientific) before incubation with antibodies against intracellular targets for 30 min at 4 °C. Following staining, cells were washed, fixed with 1% paraformaldehyde, and analyzed using flow cytometry. The antibodies used included PD-1 (Cat. no. 379209, Clone A17188A, BioLegend, 1:200), CD45 (Cat. no. 368515, Clone 2D1, BioLegend, 1:200), CD3 (Cat. no. 317329, Clone OKT3, BioLegend, 1:200), CD8 (Cat. no. 344741, Clone SK1, BioLegend, 1:200), GZMB (Cat. no. 372225, Clone QA16A02, BioLegend, 1:200), and TOX (Cat. no. 50-6502-82, Clone TXRX10, Thermo Fisher Scientific, 1:200). Dead cells were excluded using the Zombie Aqua fixable viability kit (Cat. no. 423101, BioLegend, 1:200). GFP-positive cells were excluded during gating, as the YT and YT-R cells express GFP.

For the co-culture experiments, cells from the lower compartment of the Transwell system were stained with anti-human CD68 (Cat. no. 333809, Clone Y1/82 A, BioLegend, 1:200) and TREM2 (Cat. no. 91068 T, Clone D8I4C, Cell Signaling Technology, 1:400) antibodies. The antibody binding assays were performed as previously described⁸⁹. Briefly, cells were incubated with TREM2 monoclonal antibody (mAb) for 30 min at 4 °C, followed by washing and staining with Alexa Fluor 488 Donkey anti-rabbit IgG (BioLegend, 1:200) for an additional 30 minutes at 4 °C.

Flow cytometry data were collected using a CytoFLEX flow cytometer (Beckman Coulter) and analyzed with FlowJo software (version 10.6.2).

IHC

FFPE were cut into 4 μm-thick slides for immunohistochemical analysis. Tissue sections were deparaffinized and rehydrated through a gradient of xylene and ethanol. After a 30-min incubation with 0.3% hydrogen peroxide, antigen retrieval was conducted using citrate buffer at a sub-boiling temperature for 20 min. Subsequently, the tissue was blocked with 5% bovine serum albumin (BSA) for 30 min. The sections were incubated overnight at 4 °C with the primary antibody, followed by incubation with HRP-conjugated secondary antibodies at 37 °C for 30 min. Slices were then incubated with a 3, 3’-diaminobenzidine tetrahydrochloride (DAB) kit (Biosharp) for color development and incubated with hematoxylin for nuclear counterstaining. Primary antibodies included anti-SLC1A5 (Cat. no. 20350-1-AP, Proteintech, 1:5000), anti-CD20 (Cat. no. 48750 T, Clone E7B7T, Cell Signaling Technology, 1:400), anti-CD4 (Cat. no. 48274S, Clone EP204, Cell Signaling Technology, 1:400), anti-CD8 (Cat. no. 85336 T, Clone D8A8Y, Cell Signaling Technology, 1:400), anti-MIB1 (Cat. no. AF0198, Clone MIB1, Affinity, 1:200, for Ki-67 detection), anti-H3K27ac (Cat. no. 82902-1-RR, Clone 1M16, Proteintech, 1:3000), and anti-HDAC2 (Cat. no. 12922-3-AP, Proteintech, 1:5000). Images were obtained using the NanoZoomer S360 (Hamamatsu Photonics, Japan).

mIHC

Tissue sections were deparaffinized with xylene and rehydrated through graded ethanol solutions. Antigen retrieval was performed using citrate buffer at a sub-boiling temperature for 20 min. To reduce nonspecific binding, the sections were blocked with 5% bovine serum albumin (BSA) for 30 min at room temperature. Primary antibodies, diluted in the blocking buffer, were incubated overnight at 4 °C in a humidified environment. Afterward, sections were treated with HRP-conjugated secondary antibodies for 30 min at 37 °C. Signal amplification was achieved using tyramide signal amplification (TSA), followed by microwave heat treatment. The procedure for applying primary antibodies, secondary antibodies, and TSA was repeated in sequence. Nuclei were stained with DAPI after the antigen labeling process was complete. Images were captured using the Akoya Vectra Polaris Automated Quantitative Pathology Imaging System. All experiments included at least three biological replicates. The following primary antibodies were used: anti-CD20 (Cat. no. 48750 T, Clone E7B7T, Cell Signaling Technology, 1:400), anti-CD4 (Cat. no. 48274S, Clone EP204, Cell Signaling Technology, 1:400), anti-CD8 (Cat. no. 85336 T, Clone D8A8Y, Cell Signaling Technology, 1:400), anti-CD3 (Cat. no. ZA-0503, Clone EP41, ZSGB-BIO, 1:300), anti-CD21 (Cat. no. 76069 T, Clone 2G9, Cell Signaling Technology, 1:800), anti-CD23 (Cat. no. 87960 T, Clone E1R4G, Cell Signaling Technology, 1:2000), anti-CD68 (Cat. no. 76437 T, Clone D4B9C, Cell Signaling Technology, 1:1500), anti-COL1A1 (Cell Signaling Technology, 1:500), and anti-TREM2 (Cat. no. EPR26209-22, Clone EPR26209-22, Abcam, 1:100).

Haematoxylin and eosin (H&E) staining

Tissue sections were deparaffinized using xylene and subsequently rehydrated through a graded ethanol series. Afterward, the sections were stained with hematoxylin for 3 min, followed by differentiation in 1% hydrochloric acid in 70% ethanol. The sections were then stained with eosin Y for 1 min, dehydrated through graded ethanol, and cleared in xylene. Images were captured using the Automated Quantitative Pathology Imaging System (Akoya Vectra Polaris).

Masson trichrome staining

Formalin-fixed paraffin-embedded (FFPE) tissue sections were cut at 4 μm thickness and processed using the Masson Stain Kit (Yeasen Biotechnology, Shanghai, China) following the manufacturer’s instructions. Briefly, sections were pre-moistened with distilled water for 30–60 s, then stained with hematoxylin solution for approximately 60 s and rinsed for 30 s. Acid fuchsin solution was applied for 30–60 s, followed by another rinse. Sections were then differentiated in phosphomolybdic acid for 6–8 min and counterstained with aniline blue for 5 min. After dehydration with absolute ethanol, the sections were dried and mounted for microscopic examination.

Western blotting assay

Cell lysates were prepared in RIPA buffer (CWBIO), supplemented with protease and phosphatase inhibitors. The samples were separated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) and transferred to polyvinylidene difluoride (PVDF) membranes (Merck Millipore). Membranes were blocked with 5% skimmed milk at room temperature for 1 h, followed by overnight incubation at 4 °C with the primary antibody. The next day, membranes were incubated with secondary antibodies, and protein signals were detected using the UltraSignal ECL reagent (4 AW011, 4A Biotech). The primary antibodies used included anti-HDAC2 (Cat. no. 12922-3-AP, Proteintech, 1:5000), anti-STAT3 (Cat. no. 9139 T, Clone 124H6, Cell Signaling Technology, 1:1000), anti-Phospho-STAT3 (Cat. no. 9145 T, Clone D3A7, Cell Signaling Technology, 1:2000), anti-H3K27ac (Cat. no. 8173 T, Clone D5E4, Cell Signaling Technology, 1:1000), anti-SLC1A5 (Cat. no. 20350-1-AP, Proteintech, 1:4000), and anti-GAPDH (Cat. no. 10494-1-AP, Proteintech, 1:50000). Secondary antibodies used were anti-rabbit IgG (Proteintech, 1:5000) and anti-mouse IgG (Proteintech, 1:5000). Unprocessed scans of immunoblot membrane strips are provided in the Source Data file.

Statistics and reproducibility

Statistical analyses were performed using GraphPad Prism (version 9.5.1) for experimental data. Spearman’s correlation coefficient was utilized to evaluate correlations, and survival analysis was conducted using the log-rank test. For assessing statistical significance, tests such as Student’s t-test, Kruskal-Wallis test, Wilcoxon rank-sum test, ANOVA, and the Limma moderated t-test were applied. Differential expression analyses and group-wise comparisons were adjusted for multiple testing using the Benjamini-Hochberg FDR method. ANOVA was followed by Tukey’s multiple comparisons test. Univariate Cox analyses were adjusted using the Benjamini-Hochberg FDR method, and variables significant in univariate analysis were entered into multivariate models. A significance threshold of P < 0.05 was considered. Additional statistical procedures were carried out using R (version 4.3.3). No statistical method was used to predetermine sample size; sample sizes were determined based on the availability of eligible samples. No data were excluded from the analyses except through standard quality-control filtering. Randomization and blinding were not applicable to bioinformatic analyses due to the observational design of the study. For validation assays, blinding was not formally performed because sample identities or treatment information were required for handling and processing. Animals in in vivo experiments were randomly assigned to treatment groups.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Supplementary information

Supplementary Information^{(26.1MB, pdf)}

Peer Review File^{(5MB, pdf)}

41467_2026_76152_MOESM3_ESM.pdf^{(190.7KB, pdf)}

Description of Additional Supplementary Files

Supplementary Data 1^{(16.9KB, xlsx)}

Supplementary Data 2^{(12KB, xlsx)}

Supplementary Data 3^{(60.3KB, xlsx)}

Supplementary Data 4^{(10KB, xlsx)}

Supplementary Data 5^{(219.8KB, xlsx)}

Supplementary Data 6^{(158.4KB, xlsx)}

Supplementary Data 7^{(11KB, xlsx)}

Supplementary Data 8^{(10.2KB, xlsx)}

Supplementary Data 9^{(10.1KB, xlsx)}

Reporting Summary^{(102.7KB, pdf)}

Source data

Source Data^{(549.7KB, xlsx)}

Acknowledgments

The authors thank Professor Liang Wang (Beijing Tongren Hospital) and Professor Rong Tao (Fudan University Shanghai Cancer Center) for their assistance with clinical sample collection and data support. The authors thank Kehong Zhang, M.D., Ph.D., from Ivy Medical Editing (Shanghai, China) for revising the final manuscript.

Author contributions

Q.-Q.C. and W.-L.Zhong. conceived and designed the study; Y.C., J.C., J.-N.W., Y.-C.Z., J.X., H.-Q.H., Y.G., Y.X., J.-H.W., S.-Y.M., J.-F.J. and W.-L.Zhao. performed sample preparation and data generation; Y.C. and J.C. performed bioinformatics analyses; Y.C., J.C., D.-L.D., F.L., X.-N.G., S.-R.B. and Q.-H.Z. performed functional validations; W.-L.Zhong. and Q.-Q.C. supervised the study. Y.C. and J.C. wrote the manuscript. Q.-Q.C. and W.-L.Zhong. revised the manuscript. All authors reviewed and approved the final version of the manuscript.

Peer review

Peer review information

Nature Communications thanks Rong Fan, Ankur Sharma and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by grants from National Natural Science Foundation of China (82230001, 82270199, 82400228, and U24A20680), National Key Research and Development Program (2022YFC2502602), Guangzhou Science and Technology Program (2024B03J1291), the Sun Yat-Sen University Clinical Research 5010 Program (2020009), the Clinical Oncology Foundation of Chinese Society of Clinical Oncology (Y-SY2021ZD-0110), and the China National Postdoctoral Program for Innovative Talents (BX20240445).

Data availability

The raw scRNA-seq, bulk RNA-seq and spatial transcriptomics data generated in this study have been deposited in the Genome Sequence Archive (GSA) database under accession codes HRA012856, HRA012928, and OMIX011063. The mass spectrometry proteomics raw data have been deposited to the ProteomeXchange Consortium via the iProX partner repository with the dataset identifier IPX0014865000. The mass spectrometry-based metabolomics raw data have been deposited in the Metabolomics Workbench under Study ID ST004944. The human scRNA-seq, bulk RNA-seq and spatial transcriptomics data are available under restricted access because of data privacy, institutional ethics approval and relevant human genetic resources regulations in China. Researchers could submit data access requests for scientific and non-commercial use via the National Genomics Data Center (NGDC) after registration, using BioProject accession PRJCA041960. Data access requests will be processed within one month according to the guidelines of the Genome Sequence Archive for noncommercial use. Once access has been approved, the data will be available for one month. Publicly available sequencing data analyzed in this study were obtained from the Gene Expression Omnibus (GEO) under accession code GSE203663, and from the Genome Sequence Archive of the BIG Data Center at the Beijing Institute of Genomics, Chinese Academy of Sciences, under accession code HRA000772. The remaining data are available within the Article, Supplementary Information, or Source Data file. Source data are provided with this paper.

Code availability

Code for data analysis and visualization in this study is available at: https://github.com/cy358/Single-cell-profiling-of-NKTCL. The code is released under an MIT License. The codes used in this study have been uploaded to Zenodo and are freely available at: https://doi.org/10.5281/zenodo.20780871⁹⁰.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

These authors contributed equally: Yi Cao, Jun Cai, Danling Dai, Jinni Wang, Yuchen Zhang, Jie Xiong.

Contributor Information

Weili Zhao, Email: zhao.weili@yahoo.com.

Wenlong Zhong, Email: zhongwlong3@mail.sysu.edu.cn.

Qingqing Cai, Email: caiqq@sysucc.org.cn.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76152-9.

References

1.Shankland, K. R., Armitage, J. O. & Hancock, B. W. Non-Hodgkin lymphoma. Lancet380, 848–857 (2012). [DOI] [PubMed] [Google Scholar]
2.Tian, X. P. et al. Novel target and treatment agents for natural killer/T-cell lymphoma. J. Hematol. Oncol.16, 78 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
3.Xiong, J. et al. Anti-metabolic agent pegaspargase plus PD-1 antibody sintilimab for first-line treatment in advanced natural killer T cell lymphoma. Signal Transduct. Target Ther.9, 62 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Cai, J. et al. Combination of anti-PD-1 antibody with P-GEMOX as a potentially effective immunochemotherapy for advanced natural killer/T cell lymphoma. Signal Transduct. Target Ther.5, 289 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Tian, X. P. et al. First-line sintilimab with pegaspargase, gemcitabine, and oxaliplatin in advanced extranodal natural killer/T cell lymphoma (SPIRIT): a multicentre, single-arm, phase 2 trial. Lancet Haematol.11, e336–e344 (2024). [DOI] [PubMed] [Google Scholar]
6.Xiong, J. et al. Genomic and transcriptomic characterization of natural killer T cell lymphoma. Cancer Cell37, 403–419.e406 (2020). [DOI] [PubMed] [Google Scholar]
7.Quintanilla-Martinez, L. et al. Histological and immunophenotypic profile of nasal NK/T cell lymphomas from Peru: high prevalence of p53 overexpression. Hum. Pathol.30, 849–855 (1999). [DOI] [PubMed] [Google Scholar]
8.Ishii, H. et al. Monocytes enhance cell proliferation and LMP1 expression of nasal natural killer/T-cell lymphoma cells by cell contact-dependent interaction through membrane-bound IL-15. Int J. Cancer130, 48–58 (2012). [DOI] [PubMed] [Google Scholar]
9.Suvà, M. L. & Tirosh, I. Single-cell RNA sequencing in cancer: lessons learned and emerging challenges. Mol. Cell75, 7–12 (2019). [DOI] [PubMed] [Google Scholar]
10.Li, Y. Q. et al. Single-cell analysis reveals that malignant cells reshape the cellular landscape and foster an immunosuppressive microenvironment of extranodal NK/T-cell lymphoma. Adv. Sci.10, e2303913 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Xiong, J. et al. GPCR signaling contributes to the immune characteristics of the microenvironment and the process of EBV-induced lymphomagenesis. Sci. Bull.68, 2607–2619 (2023). [DOI] [PubMed] [Google Scholar]
12.Chu, X. et al. Integrative single-cell analysis of human colorectal cancer reveals patient stratification with distinct immune evasion mechanisms. Nat. Cancer5, 1409–1426 (2024). [DOI] [PubMed] [Google Scholar]
13.Guo, D. Z. et al. Single-cell tumor heterogeneity landscape of hepatocellular carcinoma: unraveling the pro-metastatic subtype and its interaction loop with fibroblasts. Mol. Cancer23, 157 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Wang, W. et al. Single-cell profiling identifies mechanisms of inflammatory heterogeneity in chronic rhinosinusitis. Nat. Immunol.23, 1484–1494 (2022). [DOI] [PubMed] [Google Scholar]
15.Marco Salas, S. et al. Optimizing Xenium In Situ data utility by quality assessment and best-practice analysis workflows. Nat. Methods22, 813–823 (2025). [DOI] [PMC free article] [PubMed]
16.Patel, A. P. et al. Single-cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma. Science344, 1396–1401 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Puram, S. V. et al. Single-cell transcriptomic analysis of primary and metastatic tumor ecosystems in head and neck cancer. Cell171, 1611–1624.e1624 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Herrington, F. D. & Nibbs, R. J. Regulation of the adaptive immune response by the IκB family protein Bcl-3. Cells5, 14 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Redmer, T. et al. JUN mediates the senescence-associated secretory phenotype and immune cell recruitment to prevent prostate cancer progression. Mol. Cancer23, 114 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Ertosun, M. G., Hapil, F. Z. & Osman Nidai, O. E2F1 transcription factor and its impact on growth factor and cytokine signaling. Cytokine Growth Factor Rev.31, 17–25 (2016). [DOI] [PubMed] [Google Scholar]
21.Medina Rangel, P. X. et al. Cell cycle and senescence regulation by podocyte histone deacetylase 1 and 2. J. Am. Soc. Nephrol.34, 433–450 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Fan, Y., Mao, R. & Yang, J. NF-κB and STAT3 signaling pathways collaboratively link inflammation to cancer. Protein Cell4, 176–185 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Zhang, L. et al. Lineage tracking reveals dynamic relationships of T cells in colorectal cancer. Nature564, 268–272 (2018). [DOI] [PubMed] [Google Scholar]
24.Wang, R. et al. Evolution of immune and stromal cell states and ecotypes during gastric adenocarcinoma progression. Cancer Cell41, 1407–1426.e1409 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Jenkins, L. et al. Cancer-associated fibroblasts suppress CD8+ T-cell infiltration and confer resistance to immune-checkpoint blockade. Cancer Res.82, 2904–2917 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Reuben, A. et al. Comprehensive T cell repertoire characterization of non-small cell lung cancer. Nat. Commun.11, 603 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Cheng, B. et al. The activity and immune dynamics of PD-1 inhibition on high-risk pulmonary ground glass opacity lesions: insights from a single-arm, phase II trial. Signal Transduct. Target Ther.9, 93 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Blank, C. U. et al. Defining ‘T cell exhaustion. Nat. Rev. Immunol.19, 665–674 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Qiu, X. et al. Reversed graph embedding resolves complex single-cell trajectories. Nat. Methods14, 979–982 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
30.Cao, J. et al. The single-cell transcriptional landscape of mammalian organogenesis. Nature566, 496–502 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Bergen, V., Lange, M., Peidli, S., Wolf, F. A. & Theis, F. J. Generalizing RNA velocity to transient cell states through dynamical modeling. Nat. Biotechnol.38, 1408–1414 (2020). [DOI] [PubMed] [Google Scholar]
32.Zhang, Y. et al. Single-cell analyses reveal key immune cell subsets associated with response to PD-L1 blockade in triple-negative breast cancer. Cancer Cell39, 1578–1593.e1578 (2021). [DOI] [PubMed] [Google Scholar]
33.Bagaev, A. et al. Conserved pan-cancer microenvironment subtypes predict response to immunotherapy. Cancer Cell39, 845–865.e847 (2021). [DOI] [PubMed] [Google Scholar]
34.Xie, P. et al. CRKL dictates anti-PD-1 resistance by mediating tumor-associated neutrophil infiltration in hepatocellular carcinoma. J. Hepatol.81, 93–107 (2024). [DOI] [PubMed] [Google Scholar]
35.Benguigui, M. et al. Interferon-stimulated neutrophils as a predictor of immunotherapy response. Cancer Cell42, 253–265.e212 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Cao, H. Y., Li, L., Xue, S. L. & Dai, H. P. Chidamide: targeting epigenetic regulation in the treatment of hematological malignancy. Hematol. Oncol.41, 301–309 (2023). [DOI] [PubMed] [Google Scholar]
37.Que, Y. et al. Frequent amplification of HDAC genes and efficacy of HDAC inhibitor chidamide and PD-1 blockade combination in soft tissue sarcoma. J. Immunother. Cancer9, e001696 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Ramos-Casals, M., Brito-Zerón, P., López-Guillermo, A., Khamashta, M. A. & Bosch, X. Adult haemophagocytic syndrome. Lancet383, 1503–1516 (2014). [DOI] [PubMed] [Google Scholar]
39.Takahashi, N., Miura, I., Chubachi, A., Miura, A. B. & Nakamura, S. A clinicopathological study of 20 patients with T/natural killer (NK)-cell lymphoma-associated hemophagocytic syndrome with special reference to nasal and nasal-type NK/T-cell lymphoma. Int J. Hematol.74, 303–308 (2001). [DOI] [PubMed] [Google Scholar]
40.Wen, H. et al. Recurrent ECSIT mutation encoding V140A triggers hyperinflammation and promotes hemophagocytic syndrome in extranodal NK/T cell lymphoma. Nat. Med.24, 154–164 (2018). [DOI] [PubMed] [Google Scholar]
41.Soria, G. & Ben-Baruch, A. The inflammatory chemokines CCL2 and CCL5 in breast cancer. Cancer Lett.267, 271–285 (2008). [DOI] [PubMed] [Google Scholar]
42.Landy, E., Carol, H., Ring, A. & Canna, S. Biological and clinical roles of IL-18 in inflammatory diseases. Nat. Rev. Rheumatol.20, 33–47 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
43.Batlle, E. & Massagué, J. Transforming growth Factor-β signaling in immunity and cancer. Immunity50, 924–940 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
44.Dostert, C., Grusdat, M., Letellier, E. & Brenner, D. The TNF family of ligands and receptors: communication modules in the immune system and beyond. Physiol. Rev.99, 115–160 (2019). [DOI] [PubMed] [Google Scholar]
45.Henter, J. I. Hemophagocytic lymphohistiocytosis. N. Engl. J. Med.392, 584–598 (2025). [DOI] [PubMed] [Google Scholar]
46.Tang, Y. et al. Excessive IL-10 and IL-18 trigger hemophagocytic lymphohistiocytosis-like hyperinflammation and enhanced myelopoiesis. J. Allergy Clin. Immunol.150, 1154–1167 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
47.Liu, C. et al. Pan-Cancer single-cell and spatial-resolved profiling reveals the immunosuppressive role of APOE+ macrophages in immune checkpoint inhibitor therapy. Adv. Sci.11, e2401061 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
48.Mazodier, K. et al. Severe imbalance of IL-18/IL-18BP in patients with secondary hemophagocytic syndrome. Blood106, 3483–3489 (2005). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Zhang, Q. et al. A study of ruxolitinib response-based stratified treatment for pediatric hemophagocytic lymphohistiocytosis. Blood139, 3493–3504 (2022). [DOI] [PubMed] [Google Scholar]
50.Thorlacius-Ussing, J. et al. The collagen landscape in cancer: profiling collagens in tumors and in circulation reveals novel markers of cancer-associated fibroblast subtypes. J. Pathol.262, 22–36 (2024). [DOI] [PubMed] [Google Scholar]
51.Xu, P. P. et al. A phase II study of methotrexate, etoposide, dexamethasone and pegaspargase sandwiched with radiotherapy in the treatment of newly diagnosed, stage IE to IIE extranodal natural-killer/T-cell lymphoma, nasal-type. EBioMedicine25, 41–49 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Lin, L., Yee, S. W., Kim, R. B. & Giacomini, K. M. SLC transporters as therapeutic targets: emerging opportunities. Nat. Rev. Drug Discov.14, 543–560 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
53.Wu, J. et al. Asparagine enhances LCK signalling to potentiate CD8(+) T-cell activation and anti-tumour responses. Nat. Cell Biol.23, 75–86 (2021). [DOI] [PubMed] [Google Scholar]
54.Sharma, D., Yu, Y., Shen, L., Zhang, G. F. & Karner, C. M. SLC1A5 provides glutamine and asparagine necessary for bone development in mice. Elife10, e71595 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
55.Zhang, Y. et al. CCL19-producing fibroblasts promote tertiary lymphoid structure formation, enhancing anti-tumor IgG response in colorectal cancer liver metastasis. Cancer Cell42, 1370–1385.e1379 (2024). [DOI] [PubMed] [Google Scholar]
56.Schumacher, T. N. & Thommen, D. S. Tertiary lymphoid structures in cancer. Science375, eabf9419 (2022). [DOI] [PubMed] [Google Scholar]
57.Cabrita, R. et al. Tertiary lymphoid structures improve immunotherapy and survival in melanoma. Nature577, 561–565 (2020). [DOI] [PubMed] [Google Scholar]
58.Helmink, B. A. et al. B cells and tertiary lymphoid structures promote immunotherapy response. Nature577, 549–555 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
59.Kubli, S. P., Berger, T., Araujo, D. V., Siu, L. L. & Mak, T. W. Beyond immune checkpoint blockade: emerging immunological strategies. Nat. Rev. Drug Discov.20, 899–919 (2021). [DOI] [PubMed] [Google Scholar]
60.Chen, Y. et al. Spatiotemporal single-cell analysis decodes cellular dynamics underlying different responses to immunotherapy in colorectal cancer. Cancer Cell42, 1268–1285.e1267 (2024). [DOI] [PubMed] [Google Scholar]
61.Andrews, L. P. et al. LAG-3 and PD-1 synergize on CD8(+) T cells to drive T cell exhaustion and hinder autocrine IFN-γ-dependent anti-tumor immunity. Cell187, 4355–4372.e4322 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
62.Cai, Q. et al. TQB2223 in combination with penpulimab in patients with relapsed or refractory lymphoma: primary analysis of an open-label phase I study. Blood144, 3047–3047 (2024). [Google Scholar]
63.Orillion, A. et al. Entinostat neutralizes myeloid-derived suppressor cells and enhances the antitumor effect of PD-1 inhibition in murine models of lung and renal cell carcinoma. Clin. Cancer Res.23, 5187–5201 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
64.Llopiz, D. et al. Enhanced anti-tumor efficacy of checkpoint inhibitors in combination with the histone deacetylase inhibitor Belinostat in a murine hepatocellular carcinoma model. Cancer Immunol. Immunother.68, 379–393 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
65.Gao, Y. et al. Sintilimab (anti-PD-1 antibody) plus chidamide (histone deacetylase inhibitor) in relapsed or refractory extranodal natural killer T-cell lymphoma (SCENT): a phase Ib/II study. Signal Transduct. Target Ther.9, 121 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
66.Hanahan, D. Hallmarks of cancer: new dimensions. Cancer Discov.12, 31–46 (2022). [DOI] [PubMed] [Google Scholar]
67.Jia, J. et al. Clinical features and survival of extranodal natural killer/T cell lymphoma with and without hemophagocytic syndrome. Ann. Hematol.95, 2023–2031 (2016). [DOI] [PubMed] [Google Scholar]
68.Lim, S. W. et al. Serum IL18 is associated with hemophagocytosis and poor survival in extranodal natural killer/T-cell lymphoma. Leuk. Lymphoma60, 317–325 (2019). [DOI] [PubMed] [Google Scholar]
69.DeNardo, D. G. & Ruffell, B. Macrophages as regulators of tumour immunity and immunotherapy. Nat. Rev. Immunol.19, 369–382 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
70.Setiadi, A., Zoref-Lorenz, A., Lee, C. Y., Jordan, M. B. & Chen, L. Y. C. Malignancy-associated haemophagocytic lymphohistiocytosis. Lancet Haematol.9, e217–e227 (2022). [DOI] [PubMed] [Google Scholar]
71.Deczkowska, A., Weiner, A. & Amit, I. The physiology, pathology, and potential therapeutic applications of the TREM2 signaling pathway. Cell181, 1207–1217 (2020). [DOI] [PubMed] [Google Scholar]
72.Xiong, D., Wang, Y. & You, M. A gene expression signature of TREM2(hi) macrophages and γδ T cells predicts immunotherapy response. Nat. Commun.11, 5084 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
73.Zhou, L. et al. Ruxolitinib combined with doxorubicin, etoposide, and dexamethasone for the treatment of the lymphoma-associated hemophagocytic syndrome. J. Cancer Res. Clin. Oncol.146, 3063–3074 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
74.Grima-Reyes, M. et al. Tumoral microenvironment prevents de novo asparagine biosynthesis in B cell lymphoma, regardless of ASNS expression. Sci. Adv.8, eabn6491 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
75.Zhang, Y. et al. Sequential P-GEMOX and radiotherapy for early-stage extranodal natural killer/T-cell lymphoma: A multicenter study. Am. J. Hematol.96, 1481–1490 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
76.Li, X. et al. DDGP versus SMILE in newly diagnosed advanced natural killer/T-cell lymphoma: a randomized controlled, multicenter, open-label study in China. Clin. Cancer Res.22, 5223–5228 (2016). [DOI] [PubMed] [Google Scholar]
77.Kandasamy, P., Gyimesi, G., Kanai, Y. & Hediger, M. A. Amino acid transporters revisited: New views in health and disease. Trends Biochem Sci.43, 752–789 (2018). [DOI] [PubMed] [Google Scholar]
78.Xiong, J. et al. SLC1A1 mediated glutamine addiction and contributed to natural killer T-cell lymphoma progression with immunotherapeutic potential. EBioMedicine72, 103614 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
79.Hu, C. et al. Tertiary lymphoid structure-associated B cells enhance CXCL13(+)CD103(+)CD8(+) tissue-resident memory T-cell response to programmed cell death protein 1 blockade in cancer immunotherapy. Gastroenterology166, 1069–1084 (2024). [DOI] [PubMed] [Google Scholar]
80.Meylan, M. et al. Tertiary lymphoid structures generate and propagate anti-tumor antibody-producing plasma cells in renal cell cancer. Immunity55, 527–541.e525 (2022). [DOI] [PubMed] [Google Scholar]
81.Munoz-Erazo, L., Rhodes, J. L., Marion, V. C. & Kemp, R. A. Tertiary lymphoid structures in cancer - considerations for patient prognosis. Cell Mol. Immunol.17, 570–575 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
82.Hao, Y. et al. Integrated analysis of multimodal single-cell data. Cell184, 3573–3587.e3529 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
83.McGinnis, C. S., Murrow, L. M. & Gartner, Z. J. DoubletFinder: doublet detection in single-cell RNA sequencing data using artificial nearest neighbors. Cell Syst.8, 329–337.e324 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
84.Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nat. Methods16, 1289–1296 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
85.Aibar, S. et al. SCENIC: single-cell regulatory network inference and clustering. Nat. Methods14, 1083–1086 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
86.Dimitrov, D. et al. Comparison of methods and resources for cell-cell communication inference from single-cell RNA-Seq data. Nat. Commun.13, 3224 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
87.Pang, Z. et al. MetaboAnalyst 6.0: towards a unified platform for metabolomics data processing, analysis and interpretation. Nucleic Acids Res.52, W398–w406 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
88.Kuang, D. M. et al. Activated monocytes in the peritumoral stroma of hepatocellular carcinoma foster immune privilege and disease progression through PD-L1. J. Exp. Med.206, 1327–1337 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
89.Ng, K. W. et al. Antibodies against endogenous retroviruses promote lung cancer immunotherapy. Nature616, 563–573 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar] [Retracted]
90.Cao, Y. Single-cell profiling of natural killer T-cell lymphoma reveals stratified immune features and potential therapeutic implications. Zenodo 10.5281/zenodo.20780871 (2026). [DOI] [PubMed]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

[CR1] 1.Shankland, K. R., Armitage, J. O. & Hancock, B. W. Non-Hodgkin lymphoma. Lancet380, 848–857 (2012). [DOI] [PubMed] [Google Scholar]

[CR2] 2.Tian, X. P. et al. Novel target and treatment agents for natural killer/T-cell lymphoma. J. Hematol. Oncol.16, 78 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR3] 3.Xiong, J. et al. Anti-metabolic agent pegaspargase plus PD-1 antibody sintilimab for first-line treatment in advanced natural killer T cell lymphoma. Signal Transduct. Target Ther.9, 62 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR4] 4.Cai, J. et al. Combination of anti-PD-1 antibody with P-GEMOX as a potentially effective immunochemotherapy for advanced natural killer/T cell lymphoma. Signal Transduct. Target Ther.5, 289 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR5] 5.Tian, X. P. et al. First-line sintilimab with pegaspargase, gemcitabine, and oxaliplatin in advanced extranodal natural killer/T cell lymphoma (SPIRIT): a multicentre, single-arm, phase 2 trial. Lancet Haematol.11, e336–e344 (2024). [DOI] [PubMed] [Google Scholar]

[CR6] 6.Xiong, J. et al. Genomic and transcriptomic characterization of natural killer T cell lymphoma. Cancer Cell37, 403–419.e406 (2020). [DOI] [PubMed] [Google Scholar]

[CR7] 7.Quintanilla-Martinez, L. et al. Histological and immunophenotypic profile of nasal NK/T cell lymphomas from Peru: high prevalence of p53 overexpression. Hum. Pathol.30, 849–855 (1999). [DOI] [PubMed] [Google Scholar]

[CR8] 8.Ishii, H. et al. Monocytes enhance cell proliferation and LMP1 expression of nasal natural killer/T-cell lymphoma cells by cell contact-dependent interaction through membrane-bound IL-15. Int J. Cancer130, 48–58 (2012). [DOI] [PubMed] [Google Scholar]

[CR9] 9.Suvà, M. L. & Tirosh, I. Single-cell RNA sequencing in cancer: lessons learned and emerging challenges. Mol. Cell75, 7–12 (2019). [DOI] [PubMed] [Google Scholar]

[CR10] 10.Li, Y. Q. et al. Single-cell analysis reveals that malignant cells reshape the cellular landscape and foster an immunosuppressive microenvironment of extranodal NK/T-cell lymphoma. Adv. Sci.10, e2303913 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR11] 11.Xiong, J. et al. GPCR signaling contributes to the immune characteristics of the microenvironment and the process of EBV-induced lymphomagenesis. Sci. Bull.68, 2607–2619 (2023). [DOI] [PubMed] [Google Scholar]

[CR12] 12.Chu, X. et al. Integrative single-cell analysis of human colorectal cancer reveals patient stratification with distinct immune evasion mechanisms. Nat. Cancer5, 1409–1426 (2024). [DOI] [PubMed] [Google Scholar]

[CR13] 13.Guo, D. Z. et al. Single-cell tumor heterogeneity landscape of hepatocellular carcinoma: unraveling the pro-metastatic subtype and its interaction loop with fibroblasts. Mol. Cancer23, 157 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR14] 14.Wang, W. et al. Single-cell profiling identifies mechanisms of inflammatory heterogeneity in chronic rhinosinusitis. Nat. Immunol.23, 1484–1494 (2022). [DOI] [PubMed] [Google Scholar]

[CR15] 15.Marco Salas, S. et al. Optimizing Xenium In Situ data utility by quality assessment and best-practice analysis workflows. Nat. Methods22, 813–823 (2025). [DOI] [PMC free article] [PubMed]

[CR16] 16.Patel, A. P. et al. Single-cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma. Science344, 1396–1401 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR17] 17.Puram, S. V. et al. Single-cell transcriptomic analysis of primary and metastatic tumor ecosystems in head and neck cancer. Cell171, 1611–1624.e1624 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR18] 18.Herrington, F. D. & Nibbs, R. J. Regulation of the adaptive immune response by the IκB family protein Bcl-3. Cells5, 14 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR19] 19.Redmer, T. et al. JUN mediates the senescence-associated secretory phenotype and immune cell recruitment to prevent prostate cancer progression. Mol. Cancer23, 114 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR20] 20.Ertosun, M. G., Hapil, F. Z. & Osman Nidai, O. E2F1 transcription factor and its impact on growth factor and cytokine signaling. Cytokine Growth Factor Rev.31, 17–25 (2016). [DOI] [PubMed] [Google Scholar]

[CR21] 21.Medina Rangel, P. X. et al. Cell cycle and senescence regulation by podocyte histone deacetylase 1 and 2. J. Am. Soc. Nephrol.34, 433–450 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR22] 22.Fan, Y., Mao, R. & Yang, J. NF-κB and STAT3 signaling pathways collaboratively link inflammation to cancer. Protein Cell4, 176–185 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR23] 23.Zhang, L. et al. Lineage tracking reveals dynamic relationships of T cells in colorectal cancer. Nature564, 268–272 (2018). [DOI] [PubMed] [Google Scholar]

[CR24] 24.Wang, R. et al. Evolution of immune and stromal cell states and ecotypes during gastric adenocarcinoma progression. Cancer Cell41, 1407–1426.e1409 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR25] 25.Jenkins, L. et al. Cancer-associated fibroblasts suppress CD8+ T-cell infiltration and confer resistance to immune-checkpoint blockade. Cancer Res.82, 2904–2917 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR26] 26.Reuben, A. et al. Comprehensive T cell repertoire characterization of non-small cell lung cancer. Nat. Commun.11, 603 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR27] 27.Cheng, B. et al. The activity and immune dynamics of PD-1 inhibition on high-risk pulmonary ground glass opacity lesions: insights from a single-arm, phase II trial. Signal Transduct. Target Ther.9, 93 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR28] 28.Blank, C. U. et al. Defining ‘T cell exhaustion. Nat. Rev. Immunol.19, 665–674 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR29] 29.Qiu, X. et al. Reversed graph embedding resolves complex single-cell trajectories. Nat. Methods14, 979–982 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR30] 30.Cao, J. et al. The single-cell transcriptional landscape of mammalian organogenesis. Nature566, 496–502 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR31] 31.Bergen, V., Lange, M., Peidli, S., Wolf, F. A. & Theis, F. J. Generalizing RNA velocity to transient cell states through dynamical modeling. Nat. Biotechnol.38, 1408–1414 (2020). [DOI] [PubMed] [Google Scholar]

[CR32] 32.Zhang, Y. et al. Single-cell analyses reveal key immune cell subsets associated with response to PD-L1 blockade in triple-negative breast cancer. Cancer Cell39, 1578–1593.e1578 (2021). [DOI] [PubMed] [Google Scholar]

[CR33] 33.Bagaev, A. et al. Conserved pan-cancer microenvironment subtypes predict response to immunotherapy. Cancer Cell39, 845–865.e847 (2021). [DOI] [PubMed] [Google Scholar]

[CR34] 34.Xie, P. et al. CRKL dictates anti-PD-1 resistance by mediating tumor-associated neutrophil infiltration in hepatocellular carcinoma. J. Hepatol.81, 93–107 (2024). [DOI] [PubMed] [Google Scholar]

[CR35] 35.Benguigui, M. et al. Interferon-stimulated neutrophils as a predictor of immunotherapy response. Cancer Cell42, 253–265.e212 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR36] 36.Cao, H. Y., Li, L., Xue, S. L. & Dai, H. P. Chidamide: targeting epigenetic regulation in the treatment of hematological malignancy. Hematol. Oncol.41, 301–309 (2023). [DOI] [PubMed] [Google Scholar]

[CR37] 37.Que, Y. et al. Frequent amplification of HDAC genes and efficacy of HDAC inhibitor chidamide and PD-1 blockade combination in soft tissue sarcoma. J. Immunother. Cancer9, e001696 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR38] 38.Ramos-Casals, M., Brito-Zerón, P., López-Guillermo, A., Khamashta, M. A. & Bosch, X. Adult haemophagocytic syndrome. Lancet383, 1503–1516 (2014). [DOI] [PubMed] [Google Scholar]

[CR39] 39.Takahashi, N., Miura, I., Chubachi, A., Miura, A. B. & Nakamura, S. A clinicopathological study of 20 patients with T/natural killer (NK)-cell lymphoma-associated hemophagocytic syndrome with special reference to nasal and nasal-type NK/T-cell lymphoma. Int J. Hematol.74, 303–308 (2001). [DOI] [PubMed] [Google Scholar]

[CR40] 40.Wen, H. et al. Recurrent ECSIT mutation encoding V140A triggers hyperinflammation and promotes hemophagocytic syndrome in extranodal NK/T cell lymphoma. Nat. Med.24, 154–164 (2018). [DOI] [PubMed] [Google Scholar]

[CR41] 41.Soria, G. & Ben-Baruch, A. The inflammatory chemokines CCL2 and CCL5 in breast cancer. Cancer Lett.267, 271–285 (2008). [DOI] [PubMed] [Google Scholar]

[CR42] 42.Landy, E., Carol, H., Ring, A. & Canna, S. Biological and clinical roles of IL-18 in inflammatory diseases. Nat. Rev. Rheumatol.20, 33–47 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR43] 43.Batlle, E. & Massagué, J. Transforming growth Factor-β signaling in immunity and cancer. Immunity50, 924–940 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR44] 44.Dostert, C., Grusdat, M., Letellier, E. & Brenner, D. The TNF family of ligands and receptors: communication modules in the immune system and beyond. Physiol. Rev.99, 115–160 (2019). [DOI] [PubMed] [Google Scholar]

[CR45] 45.Henter, J. I. Hemophagocytic lymphohistiocytosis. N. Engl. J. Med.392, 584–598 (2025). [DOI] [PubMed] [Google Scholar]

[CR46] 46.Tang, Y. et al. Excessive IL-10 and IL-18 trigger hemophagocytic lymphohistiocytosis-like hyperinflammation and enhanced myelopoiesis. J. Allergy Clin. Immunol.150, 1154–1167 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR47] 47.Liu, C. et al. Pan-Cancer single-cell and spatial-resolved profiling reveals the immunosuppressive role of APOE+ macrophages in immune checkpoint inhibitor therapy. Adv. Sci.11, e2401061 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR48] 48.Mazodier, K. et al. Severe imbalance of IL-18/IL-18BP in patients with secondary hemophagocytic syndrome. Blood106, 3483–3489 (2005). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR49] 49.Zhang, Q. et al. A study of ruxolitinib response-based stratified treatment for pediatric hemophagocytic lymphohistiocytosis. Blood139, 3493–3504 (2022). [DOI] [PubMed] [Google Scholar]

[CR50] 50.Thorlacius-Ussing, J. et al. The collagen landscape in cancer: profiling collagens in tumors and in circulation reveals novel markers of cancer-associated fibroblast subtypes. J. Pathol.262, 22–36 (2024). [DOI] [PubMed] [Google Scholar]

[CR51] 51.Xu, P. P. et al. A phase II study of methotrexate, etoposide, dexamethasone and pegaspargase sandwiched with radiotherapy in the treatment of newly diagnosed, stage IE to IIE extranodal natural-killer/T-cell lymphoma, nasal-type. EBioMedicine25, 41–49 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR52] 52.Lin, L., Yee, S. W., Kim, R. B. & Giacomini, K. M. SLC transporters as therapeutic targets: emerging opportunities. Nat. Rev. Drug Discov.14, 543–560 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR53] 53.Wu, J. et al. Asparagine enhances LCK signalling to potentiate CD8(+) T-cell activation and anti-tumour responses. Nat. Cell Biol.23, 75–86 (2021). [DOI] [PubMed] [Google Scholar]

[CR54] 54.Sharma, D., Yu, Y., Shen, L., Zhang, G. F. & Karner, C. M. SLC1A5 provides glutamine and asparagine necessary for bone development in mice. Elife10, e71595 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR55] 55.Zhang, Y. et al. CCL19-producing fibroblasts promote tertiary lymphoid structure formation, enhancing anti-tumor IgG response in colorectal cancer liver metastasis. Cancer Cell42, 1370–1385.e1379 (2024). [DOI] [PubMed] [Google Scholar]

[CR56] 56.Schumacher, T. N. & Thommen, D. S. Tertiary lymphoid structures in cancer. Science375, eabf9419 (2022). [DOI] [PubMed] [Google Scholar]

[CR57] 57.Cabrita, R. et al. Tertiary lymphoid structures improve immunotherapy and survival in melanoma. Nature577, 561–565 (2020). [DOI] [PubMed] [Google Scholar]

[CR58] 58.Helmink, B. A. et al. B cells and tertiary lymphoid structures promote immunotherapy response. Nature577, 549–555 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR59] 59.Kubli, S. P., Berger, T., Araujo, D. V., Siu, L. L. & Mak, T. W. Beyond immune checkpoint blockade: emerging immunological strategies. Nat. Rev. Drug Discov.20, 899–919 (2021). [DOI] [PubMed] [Google Scholar]

[CR60] 60.Chen, Y. et al. Spatiotemporal single-cell analysis decodes cellular dynamics underlying different responses to immunotherapy in colorectal cancer. Cancer Cell42, 1268–1285.e1267 (2024). [DOI] [PubMed] [Google Scholar]

[CR61] 61.Andrews, L. P. et al. LAG-3 and PD-1 synergize on CD8(+) T cells to drive T cell exhaustion and hinder autocrine IFN-γ-dependent anti-tumor immunity. Cell187, 4355–4372.e4322 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR62] 62.Cai, Q. et al. TQB2223 in combination with penpulimab in patients with relapsed or refractory lymphoma: primary analysis of an open-label phase I study. Blood144, 3047–3047 (2024). [Google Scholar]

[CR63] 63.Orillion, A. et al. Entinostat neutralizes myeloid-derived suppressor cells and enhances the antitumor effect of PD-1 inhibition in murine models of lung and renal cell carcinoma. Clin. Cancer Res.23, 5187–5201 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR64] 64.Llopiz, D. et al. Enhanced anti-tumor efficacy of checkpoint inhibitors in combination with the histone deacetylase inhibitor Belinostat in a murine hepatocellular carcinoma model. Cancer Immunol. Immunother.68, 379–393 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR65] 65.Gao, Y. et al. Sintilimab (anti-PD-1 antibody) plus chidamide (histone deacetylase inhibitor) in relapsed or refractory extranodal natural killer T-cell lymphoma (SCENT): a phase Ib/II study. Signal Transduct. Target Ther.9, 121 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR66] 66.Hanahan, D. Hallmarks of cancer: new dimensions. Cancer Discov.12, 31–46 (2022). [DOI] [PubMed] [Google Scholar]

[CR67] 67.Jia, J. et al. Clinical features and survival of extranodal natural killer/T cell lymphoma with and without hemophagocytic syndrome. Ann. Hematol.95, 2023–2031 (2016). [DOI] [PubMed] [Google Scholar]

[CR68] 68.Lim, S. W. et al. Serum IL18 is associated with hemophagocytosis and poor survival in extranodal natural killer/T-cell lymphoma. Leuk. Lymphoma60, 317–325 (2019). [DOI] [PubMed] [Google Scholar]

[CR69] 69.DeNardo, D. G. & Ruffell, B. Macrophages as regulators of tumour immunity and immunotherapy. Nat. Rev. Immunol.19, 369–382 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR70] 70.Setiadi, A., Zoref-Lorenz, A., Lee, C. Y., Jordan, M. B. & Chen, L. Y. C. Malignancy-associated haemophagocytic lymphohistiocytosis. Lancet Haematol.9, e217–e227 (2022). [DOI] [PubMed] [Google Scholar]

[CR71] 71.Deczkowska, A., Weiner, A. & Amit, I. The physiology, pathology, and potential therapeutic applications of the TREM2 signaling pathway. Cell181, 1207–1217 (2020). [DOI] [PubMed] [Google Scholar]

[CR72] 72.Xiong, D., Wang, Y. & You, M. A gene expression signature of TREM2(hi) macrophages and γδ T cells predicts immunotherapy response. Nat. Commun.11, 5084 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR73] 73.Zhou, L. et al. Ruxolitinib combined with doxorubicin, etoposide, and dexamethasone for the treatment of the lymphoma-associated hemophagocytic syndrome. J. Cancer Res. Clin. Oncol.146, 3063–3074 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR74] 74.Grima-Reyes, M. et al. Tumoral microenvironment prevents de novo asparagine biosynthesis in B cell lymphoma, regardless of ASNS expression. Sci. Adv.8, eabn6491 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR75] 75.Zhang, Y. et al. Sequential P-GEMOX and radiotherapy for early-stage extranodal natural killer/T-cell lymphoma: A multicenter study. Am. J. Hematol.96, 1481–1490 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR76] 76.Li, X. et al. DDGP versus SMILE in newly diagnosed advanced natural killer/T-cell lymphoma: a randomized controlled, multicenter, open-label study in China. Clin. Cancer Res.22, 5223–5228 (2016). [DOI] [PubMed] [Google Scholar]

[CR77] 77.Kandasamy, P., Gyimesi, G., Kanai, Y. & Hediger, M. A. Amino acid transporters revisited: New views in health and disease. Trends Biochem Sci.43, 752–789 (2018). [DOI] [PubMed] [Google Scholar]

[CR78] 78.Xiong, J. et al. SLC1A1 mediated glutamine addiction and contributed to natural killer T-cell lymphoma progression with immunotherapeutic potential. EBioMedicine72, 103614 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR79] 79.Hu, C. et al. Tertiary lymphoid structure-associated B cells enhance CXCL13(+)CD103(+)CD8(+) tissue-resident memory T-cell response to programmed cell death protein 1 blockade in cancer immunotherapy. Gastroenterology166, 1069–1084 (2024). [DOI] [PubMed] [Google Scholar]

[CR80] 80.Meylan, M. et al. Tertiary lymphoid structures generate and propagate anti-tumor antibody-producing plasma cells in renal cell cancer. Immunity55, 527–541.e525 (2022). [DOI] [PubMed] [Google Scholar]

[CR81] 81.Munoz-Erazo, L., Rhodes, J. L., Marion, V. C. & Kemp, R. A. Tertiary lymphoid structures in cancer - considerations for patient prognosis. Cell Mol. Immunol.17, 570–575 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR82] 82.Hao, Y. et al. Integrated analysis of multimodal single-cell data. Cell184, 3573–3587.e3529 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR83] 83.McGinnis, C. S., Murrow, L. M. & Gartner, Z. J. DoubletFinder: doublet detection in single-cell RNA sequencing data using artificial nearest neighbors. Cell Syst.8, 329–337.e324 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR84] 84.Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nat. Methods16, 1289–1296 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR85] 85.Aibar, S. et al. SCENIC: single-cell regulatory network inference and clustering. Nat. Methods14, 1083–1086 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR86] 86.Dimitrov, D. et al. Comparison of methods and resources for cell-cell communication inference from single-cell RNA-Seq data. Nat. Commun.13, 3224 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR87] 87.Pang, Z. et al. MetaboAnalyst 6.0: towards a unified platform for metabolomics data processing, analysis and interpretation. Nucleic Acids Res.52, W398–w406 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR88] 88.Kuang, D. M. et al. Activated monocytes in the peritumoral stroma of hepatocellular carcinoma foster immune privilege and disease progression through PD-L1. J. Exp. Med.206, 1327–1337 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR89] 89.Ng, K. W. et al. Antibodies against endogenous retroviruses promote lung cancer immunotherapy. Nature616, 563–573 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar] [Retracted]

[CR90] 90.Cao, Y. Single-cell profiling of natural killer T-cell lymphoma reveals stratified immune features and potential therapeutic implications. Zenodo 10.5281/zenodo.20780871 (2026). [DOI] [PubMed]

PERMALINK

Single-cell profiling of natural killer/T-cell lymphoma reveals stratified immune features and potential therapeutic implications

Yi Cao

Jun Cai

Danling Dai

Jinni Wang

Yuchen Zhang

Jie Xiong

Huiqiang Huang

Yan Gao

Yi Xia

Fang Liu

Xinna Gao

Shenrui Bai

Qihua Zou

Jiahui Wang

Shuyun Ma

Junfeng Jiang

Weili Zhao

Wenlong Zhong

Qingqing Cai

Abstract

Introduction

Results

A single-cell landscape of NKTCL

Fig. 1. Landscape profiling of tumor and immune microenvironment in NKTCL by scRNA-seq.

NMF-derived tumor meta-programs reveal distinct patient stratifications

Fig. 2. Molecular features of tumor meta-programs defined by NMF and patients classification.

Characterizing T cell-dominant microenvironment in MP1

Fig. 3. Characterization of T cell subsets in the MP1 subgroup.

The immune desert microenvironment with HDAC2 involvement in MP2

Fig. 4. Characterization of HDAC2-involved immune suppression in the MP2 subgroup.

Tumor-driven inflammation and myeloid-enriched immune microenvironment in MP3

Fig. 5. Characterization of myeloid and stromal cell subsets in the MP3 subgroup.

Distinct amino acid metabolic profile and TLS-enriched microenvironment in MP4

Fig. 6. Characterization of metabolic and immune features in the MP4 subgroup.

Independent validation and prognostic relevance of MP-based classification

Fig. 7. Clinical validation and summarized characteristics of MP subgroups.

Discussion

Methods

Ethics statement

Patients and specimens

Sample processing and single-cell preparation for scRNA-seq

scRNA-seq and TCR V(D)J library generation and sequencing

scRNA-seq data processing

Unsupervised clustering and cell type annotation

Single-cell copy-number variation analysis

Characterizing intratumor heterogeneity and meta-programs of NKTCL

Differential expression and gene set enrichment analysis

Definition and calculation of signature score

Single-cell transcription factor regulon analysis

Matching with reported NKTCL molecular subtypes

Pseudotime analysis and RNA velocity analysis

Cell subsets distribution analysis

Correlation analysis

Cellular abundance calculation and cellular co-existence analysis

scTCR-seq data processing and analysis

Cell-cell communication analysis

Xenium in situ analysis

Bulk-RNA sequencing and survival analysis

Proteomic analysis

Serum metabolomics analysis

Cell lines and cell culture

Cell transfection

Quantitative Real-Time PCR (qRT-PCR)

Cell viability

In vivo experiments

Macrophage differentiation and construction of in vitro co-culture system

Flow cytometry analysis

IHC

mIHC

Haematoxylin and eosin (H&E) staining

Masson trichrome staining

Western blotting assay

Statistics and reproducibility

Reporting summary

Supplementary information

Source data

Acknowledgments

Author contributions