Abstract

   Immunotherapy has emerged as a promising treatment option for
   triple-negative breast cancer (TNBC); however, the pronounced
   heterogeneity of the tumor immune microenvironment significantly
   hinders the prediction of therapeutic efficacy, with effective
   biomarkers also remaining limited. This study utilizes single-cell RNA
   sequencing (scRNA-seq) and transcriptome analysis to dissect the immune
   landscapes of TNBC, emphasizing CD8 effector T cells (CD8Teff) as
   potential predictors of immune checkpoint inhibitors (ICIs) treatment
   response. Immune profiling revealed that CD8Teff cells were
   predominantly enriched in "hot" tumors strongly correlating with
   improved progression-free and overall survival. Notably, the cytokine
   C-X-C Motif Chemokine Ligand 13 (CXCL13) emerged as a key regulator,
   with its high expression marking an immune-active tumor
   microenvironment favorable to ICI efficacy. Comprehensive analysis
   uncovered critical interactions between CD8Teff cells and other immune
   and stromal components, supporting the formation of an immunologically
   active tumor microenvironment (TME) conducive to tumor control.
   Additionally, metabolic reprogramming of CD8Teff cells in responsive
   tumors highlighted pathways that may influence ICIs efficacy. CD8Teff
   cells significantly influence the TME and predict ICI success in TNBC,
   with CXCL13 as a pivotal modulator. A pathology-based artificial
   intelligence model for CD8Teff recognition was developed, achieving an
   area under the curve (AUC) of 0.823 in the training cohort and 0.805 in
   the validation cohort. This study provides an in-depth characterization
   of the TNBC immune microenvironment, identifying CD8Teff functionality,
   CXCL13 signaling, and immune-metabolic pathways as critical
   determinants of immunotherapy success. These findings support the
   development of personalized treatment strategies based on immune
   profiling to enhance therapeutic efficacy in TNBC.

Supplementary Information

   The online version contains supplementary material available at
   10.1186/s43556-025-00306-2.

   Keywords: Triple-negative breast cancer, CD8Teff cells, Single-cell RNA
   sequencing, C-X-C Motif Chemokine Ligand 13, Immunotherapy response,
   Artificial intelligence

Introduction

   Triple-negative breast cancer (TNBC) is a clinically aggressive subtype
   charscterized by poor prognosis and limited treatment options compared
   to other breast cancer subtypes [[54]1, [55]2], necessitating the
   exploration of novel therapeutic approaches. Recent advances in
   immunotherapy, particularly immune checkpoint inhibitors (ICIs)
   targeting the PD-1/PD-L1 axis, have opened new avenues for treating
   TNBC by enhancing anti-tumor immune responses [[56]3, [57]4]. The
   KEYNOTE-522 study demonstrated that adding pembrolizumab to neoadjuvant
   chemotherapy significantly improved event-free survival in early-stage
   TNBC, ICIs in combination with chemotherapy have been shown to increase
   the pathological complete response (pCR) rate by 4–17%, and improve
   3-year event-free survival rates by 8–13%. Which establishing a
   foundational framework for integrating immunotherapy into standard
   treatment regimens [[58]5]. Furthermore, immunotherapy has demonstrated
   considerable efficacy in advanced-stage TNBC, IMpassion031 and PCD4989g
   showed significant improvements in progression-free survival (PFS) and
   overall survival (OS), particularly among patients exhibiting high
   levels of tumor-infiltrating immune cells and increased immunogenicity
   [[59]6–[60]8]. However, response rates to ICIs remain highly variable,
   underscoring the urgent need for reliable biomarkers to predict which
   patients are most likely to benefit from these therapies.

   The effectiveness of immunotherapy in breast cancer is closely tied to
   the characteristics of the tumor microenvironment (TME) [[61]9]. Tumors
   are often categorized as "hot" or "cold" based on immune cell
   infiltration and activity, with hot tumors showing greater
   responsiveness to immunotherapy [[62]10–[63]12]. Currently, the
   classification of "hot" and "cold" tumors is primarily based on the
   assessment of CD8 + T cells and CD4 + T cells via immunohistochemistry
   (IHC). The IMpassion130 study assessed the efficacy of atezolizumab in
   combination with nab-paclitaxel for advanced TNBC patients, with its
   exploratory analysis revealing that CD8 + T cell infiltration in the
   tumor microenvironment could serve as a predictive biomarker for OS
   benefits from atezolizumab treatment in metastatic TNBC cases.
   Similarly, the PCD4989g trial assessed atezolizumab efficacy in various
   solid tumors, finding that patients with high CD8 + T cells expression
   had longer progression-free survival (irPFS) and OS [[64]6–[65]8].
   However, the immune landscape within TNBC is highly heterogeneous, and
   the mere presence of CD8 + T cells does not consistently correlate with
   favorable therapeutic outcomes. This suggests that additional factors,
   including the functional status of CD8 + T cells and their interactions
   with other immune and stromal components within the TME, are critical
   determinants of immunotherapy efficacy [[66]13, [67]14].

   The advent of single-cell RNA sequencing (scRNA-seq) has enabled
   high-resolution profiling of the immune landscape within tumors,
   uncovering substantial heterogeneity among immune cell subsets. This
   technology has revealed that CD8Teff cells exhibit diverse functional
   states and dynamic interactions with other immune and stromal cells,
   which may dictate tumor immunogenicity and response to ICIs [[68]15].
   However, the complex network of immune crosstalk and its role in
   modulating immune checkpoint response remains incompletely understood.
   Furthermore, current definitions of "hot" and “cold” tumors, often
   based on static immune cell staining, fail to capture the dynamic and
   functional intricacies of the TME, limiting the precision of
   immune-based classification and treatment decisions.

   This study leverages scRNA-seq and integrative multi-omics analyses to
   dissect the immune landscape of TNBC patients undergoing ICIs therapy,
   with a particular focus on CD8Teff dynamics. The functional and spatial
   characteristics of CD8Teff cells within the TME are key determinants of
   responsiveness to ICIs. By correlating single-cell immune profiles with
   clinical outcomes, we identified CD8Teff cells as central orchestrators
   of anti-tumor immunity, strongly associated with "hot" tumor phenotypes
   and favorable therapeutic responses. Furthermore, CD8Teff cells were
   found to modulate the TME through cytokine signaling via CXCL13 (C-X-C
   Motif Chemokine Ligand 13), metabolic reprogramming, and interactions
   with dendritic cells and stromal elements. To translate these findings
   into clinical practice, we developed an AI-based pathological model to
   classify "hot" and "cold" tumors based on CD8Teff distribution,
   achieving strong predictive performance. Collectively, our findings
   establish CD8Teff functionality as a critical biomarker for
   immunotherapy stratification and lay the groundwork for personalized
   immune-based strategies in TNBC.

Result

Immune cell profiles and dynamics in response to ICI treatment in TNBC

   Although CD8T IHC scores are widely used to differentiate
   between "hot" and "cold" tumors, this approach may not fully reflect
   the complexity of the immune microenvironment. In some cases, patients
   with high CD8T expression still exhibit poor responses to
   immunotherapy, suggesting significant heterogeneity in CD8T subtypes
   and differing functional roles between responders and no-responders
   (Fig. S1). This study employs single-cell RNA sequencing (scRNA-seq)
   and integrative multi-omics to achieve a more precise distinction
   between "hot" and "cold" tumors (Fig. [69]1), which included four
   patients treated with ICIs between January 2021 and September 2021.

Fig. 1.

   [70]Fig. 1
   [71]Open in a new tab

   Workflow of the study. Single-cell RNA-sequencing were applied to
   analyze the immune cell composition of TNBC tumors and peripheral blood
   from patients receiving chemotherapy combined with ICIs. This study
   characterized the CD8Teff in evaluation of tumor microenvironment hot
   and cold. Furthermore, we investigated the interactions between CD8Teff
   cells and other immune and stromal cells within the TME, which may
   guide the development of personalized treatment strategies for TNBC

   After four cycles of ICIs, two patients showed response, while the
   others did not (Fig. [72]2a, b). Single-cell RNA sequencing of two
   tumor tissues samples revealed distinct immune cell landscapes,
   including subclusters of T cells, B cells, NK cells, and stromal cells
   (Fig. [73]2c, f). The response exhibited significantly higher levels of
   Tfh and CD8Teff cells, while the no-response had more NKT and Treg
   cells (Fig. S2 a-c). These findings suggest that immune cell dynamics
   may provide insights into their functional importance in tumor immunity
   and the response to immunotherapy.

Fig. 2.

   [74]Fig. 2
   [75]Open in a new tab

   Immune cell dynamics in TNBC patients treated with ICIs. a Schematic of
   tissue and peripheral blood sample collection and (b) scRNA-seq
   workflow. c UMAP of scRNA-seq in all tumor tissue samples. d UMAP of
   scRNA-seq in all peripheral blood samples. e UMAP of scRNA-seq from 18
   TNBC patients in [76]GSE176078 and [77]GSE161892 cohorts. f Heatmap
   representing gene expression profiles of immune cells in tumor tissue
   samples. g Heatmap representing gene expression profiles of immune
   cells in peripheral blood samples. h Heatmap showing the gene
   expression patterns of immune cells across 18 TNBC patients from the
   [78]GSE176078 and [79]GSE161892 cohorts

   Further analysis of immune cell dynamics was conducted on peripheral
   blood samples collected from two additional advanced breast cancer
   patients before and after their first ICI treatment (Fig. [80]2d, g). T
   cells were classified into CD8Teff, NKT, naive T cells, Treg cells and
   GDT cells based on classical activation gene expression profiles (Fig.
   S2d-f). The results indicated dynamic shifts in T cell populations pre-
   and post-immunotherapy, with a marked increase in CD8Teff and NKT cells
   following treatment, signifying the activation of cytotoxic immune
   responses. Conversely, naive T cells decreased in proportion,
   suggesting their differentiation into more specialized subsets in
   response to therapy. These findings were consistent with broader
   analyses from 18 additional TNBC patients in the [81]GSE176078 and
   [82]GSE161892 cohorts, where similar immune cell distribution patterns
   were observed (Fig. [83]2e-h, Fig. S2g-i), further emphasizing the
   critical roles of CD8Teff and NKT cells in mediating immune responses
   to immunotherapy.

Distinct CD8Teff pathways and immune cell interactions drive ICIs response

   To explore the differences in immune response to immune checkpoint
   inhabitors, we performed Gene Set Variation Analysis (GSVA) on the
   Hallmark Pathways. Our analysis revealed significant differences in the
   pathways between T cell subgroups. In the tumor tissues samples,
   (Fig. [84]3a) T cells exhibited enhanced cell cycle signaling,
   indicating active proliferation, while CD8Teff cells in peripheral
   blood samples showed increased activation of MHC-II and exhaustion
   pathways, which are indicative of immune reactivation and improved
   tumor clearance (Fig. [85]3b). Differentiation trajectory analysis
   further revealed distinct differentiation paths for CD8Teff and Tfh
   cells, with varying distributions of proliferating T cells at the
   beginning and end of differentiation (Fig. [86]3c, d).

Fig. 3.

   [87]Fig. 3
   [88]Open in a new tab

   CD8Teff cells are the most functional T cell subsets in TNBC. a GSVA
   analysis heatmaps showing hallmark pathway enrichment in T cell subsets
   from tumor and (b) peripheral blood samples. c Pseudotime trajectory
   analysis of T cell subsets in tumor and (d) peripheral blood samples.
   e–f CellChat analysis of CD8Teff cell communication networks with other
   cells in the Response group (peripheral blood samples) before (e) and
   after (f) treatment. g–h CellChat analysis of CD8Teff cell
   communication networks with other cells in the No-Response group
   (peripheral blood samples) before (g) and after (h) treatment. i–j
   CellChat analysis of CD8Teff cell communication networks with other
   cells in the (i) Response group (tumor tissue samples) and (j)
   No-Response group (tumor tissue samples). k Heatmaps showing
   transcription factor in peripheral blood samples compared pre-treatment
   and on-treatment timepoints. l Heatmaps showing transcription factor in
   peripheral blood samples compared the Response group and No-Response
   group. m Heatmap of gene expression in tumor tissue samples from
   response and no-response groups. n–o GO enrichment analysis associated
   with CD8Teff cells in (n) peripheral blood samples and (o) 18 TNBC
   tissue samples derived from the [89]GSE176078 and [90]GSE161892
   cohorts. p-q Stacked bar plots showing the distribution of CD8Teff cell
   subpopulations before and after treatment from peripheral blood
   samples. r-s UMAP clustering of CD8Teff cells based on Seurat analysis
   from (r) peripheral blood samples and (s) 18 TNBC tissue samples
   derived from the [91]GSE176078 and [92]GSE161892 cohorts

   In response, CD8Teff cells formed strong communication networks
   comparing with other immune cells, including cytotoxic and helper
   immune cells, both before (Fig. [93]3e) and after immunotherapy
   (Fig. [94]3f). This suggests that responders maintain a coordinated and
   active immune microenvironment that is enhanced following treatment. In
   contrast, no-response exhibited weaker communication networks, with
   fewer interactions between CD8Teff cells and other immune subsets both
   before (Fig. [95]3g) and after treatment (Fig. [96]3h). Similarly,
   immune cell interactions differ markedly between the response and
   no-response groups (Fig. [97]3i-j) in tumor tissues.

   Differences in transcription factor profiles were observed between
   T-cell subsets (Fig. [98]3k, 3 m) in the response and no-response
   groups, with T cells in the response group exhibiting enhanced
   activation of CD8Teff and NKT cells. Transcription factor profiles of T
   cell subsets differed markedly (Fig. [99]3l) Pre-treatment and
   On-treatment, with On-treatment samples showing enhanced activation of
   CD8Teff cells.

   We then performed Gene Ontology (GO) enrichment analysis to further
   elucidate the biological processes associated with CD8Teff cells.
   CD8Teff cells were significantly enriched in processes related to
   immune activation, including cytotoxicity, T cell activation, and
   antigen presentation in both (Fig. [100]3n) tumor tissue and
   (Fig. [101]3o)18 TNBC tissue samples. In summary, CD8Teff cells exhibit
   distinct immune cytotoxic functions and enhanced intercellular
   interaction capabilities in the context of immunotherapy.

   To further characterize CD8Teff, we conducted a subpopulation analysis
   in both (Fig. [102]3p, q, r) peripheral blood samples and
   (Fig. [103]3s) 18 TNBC tissue samples. In peripheral blood, the
   distribution of CD8Teff subsets shifted following treatment, with
   (Fig. [104]3p-q) CD8Teff-2 predominating Pre-treatment and CD8Teff-5
   predominating On-treatment.

CD8Teff infiltration defines "Hot" tumors and predicts better clinical
outcomes

   Ro/e scoring and tissue prevalence analysis confirmed significant
   differences in Tfh, Treg, and CD8Teff cells, underscoring their roles
   in driving immunotherapy response. The results indicated a significant
   difference in CD8Teff cell proportions between patients with different
   therapeutic responses. In tumor tissues, CD8Teff cells were more
   abundant in patients who responded to immunotherapy (Fig. [105]4a),
   while no-response exhibited higher proportions of Tregs and other
   immunosuppressive cells. A similar pattern was observed in peripheral
   blood, where CD8Teff cells were enriched in responders, highlighting
   their role in systemic immune activation (Fig. [106]4b).

Fig. 4.

   [107]Fig. 4
   [108]Open in a new tab

   CD8Teff cells directly determine heat and cold of tumors. a Ro/e
   analysis of T cells in tumor tissues and (b) in peripheral blood based
   on single-cell sequencing data. c Box plot comparing CD8Teff cell
   proportions between "Hot" and "Cold" tumors by CD8Teff Ro/e in 18TNBC
   samples from [109]GSE176078 and [110]GSE161892 cohorts. d Bar plot
   showing the distribution of CD8Teff cells across different tumor
   response categories in 18TNBC samples fromGSE176078 and [111]GSE161892
   cohorts. e UMAP plots showing CD8Teff cell distribution in "Hot" and
   (f) "Cold" tumors. g Kaplan–Meier survival curves showing the
   association between CD8Teff cell levels and progression-free survival
   in TCGA cohort and (h) overall survival in SYSMH cohort. i Survival
   analysis of NSCLC immunotherapy cohort ([112]GSE12604 + [113]GSE135222)
   PFS based on CD8Teff levels. j Survival analysis of melanoma
   immunotherapy cohort (PRJEB23709) OS based on CD8Teff levels. IHC
   images showing GZMB expression in tumors with (k) high and (l) low
   CD8Teff infiltration. m Survival analysis of NSCLC immunotherapy cohort
   (POPLAR) OS based on CD8Teff levels. n Survival analysis of NSCLC
   immunotherapy cohort (OAK) OS based on CD8Teff levels. o Survival
   analysis of HCC immunotherapy cohort (Imbrave150) OS based on CD8Teff
   levels. p Survival analysis of mUC immunotherapy cohort (Immvigor210)
   OS based on CD8Teff levels. Immunofluorescence images illustrating (q)
   high and (r) low GZMB expression in tumor tissues

   Eighteen TNBC tumor samples were also classified based on Ro/e scores
   into "hot" and " cold" tumors."Hot"tumors, which displayed higher
   levels of CD8Teff cells, were more frequently found in responders,
   whereas"cold"tumors were enriched with Tregs (Fig. [114]4c-d). UMAP
   visualization further confirmed this differentiation, with "hot" tumors
   showing distinct clusters of CD8Teff cells, suggesting their active
   involvement in the anti-tumor immune response (Fig. [115]4e). In
   contrast,"cold"tumors displayed lower CD8Teff cell densities and a more
   heterogeneous immune profile, including higher proportions of Tregs and
   Tfh cells (Fig. [116]4f).

   Kaplan–Meier survival analysis demonstrated that higher levels of
   CD8Teff cells were significantly associated with improved
   progression-free survival (PFS) (Fig. [117]4g) in patients from the
   TCGA cohort and overall survival (OS) (Fig. [118]4h) in patients from
   the Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University (SYSMH)
   cohort. Patients with higher CD8Teff cell infiltration had better
   clinical outcomes, highlighting the importance of these cells in
   enhancing the effectiveness of immunotherapy. These results were
   further validated using IHC and external datasets (Fig. [119]4k-l). In
   the non-small cell lung cancer (NSCLC) ICIs cohorts ([120]GSE126044 and
   [121]GSE135222), higher CD8Teff levels were associated with improved
   PFS (Fig. [122]4i), while the melanoma ICIs (PRJEB23709) dataset
   confirmed a similar trend for OS (Fig. [123]4j). Similar results were
   observed across multiple immunotherapy cohorts, with high CD8Teff cell
   levels strongly associated with improved prognosis (Fig. [124]4. m, n,
   o, p). These findings emphasize the pivotal role of CD8Teff cells in
   fostering an immune-active tumor microenvironment and directly
   determining the "hot" or "cold" status of tumors, which is crucial for
   predicting therapeutic outcomes.

   The gene expression analysis revealed that CD8Teff cells from
   responders showed upregulation of key cytotoxic genes, such as GZMB and
   PRF1, compared to non-responders (Fig. S3a). GO enrichment analysis
   highlighted that CD8Teff cells from responders were more active in
   immune-related processes, particularly in antigen processing and
   presentation pathways (Fig. S3b). Comparing pre-treatment and
   on-treatment samples, on-treatment CD8Teff cells exhibited higher
   expression of immune activation markers, indicating increased
   tumor-killing activity (Fig. S3c). GO analysis further supported this
   by showing enhanced immune activity and protein translation in
   on-treatment CD8Teff cells (Fig. S3d). Transcription factor analysis
   revealed that CD8Teff cells in responders expressed higher levels of
   pro-inflammatory cytokines and effector molecules, consistent with
   their active role in driving anti-tumor immunity (Fig. S3e, f).

   To further validate these findings, immunohistochemical (IHC) staining
   for granzyme B (GZMB), a marker of cytotoxic activity, was performed.
   Tumors with high CD8Teff infiltration showed strong GZMB expression,
   while low CD8Teff infiltration tumors exhibited weak or absent staining
   (Fig. [125]4k, l). Immunofluorescence (IF) staining revealed similar
   results (Fig. [126]4q, r). Importantly, GZMB staining was more
   effective than CD8A staining in
   distinguishing "hot" from "cold" tumors, particularly in cases where
   CD8A staining was insufficient (Fig. [127]4q, r). This highlights
   GZMB’s utility in accurately distinguishing between hot and cold
   tumors, compensating for the limitations of traditional CD8A-based
   methods.

CXCL13 correlates with CD8Teff activity and modulates tumor immune
microenvironment

   To further investigate the role of the tumor microenvironment, we
   analyzed cytokine profiles in both blood and tissue samples across
   different groups. Significant cytokine variations were detected between
   these groups, with CXCL13 emerging as a critical cytokine that
   potentially affects multiple pathways in tumor microenvironment
   (Fig. [128]5 a, b). CXCL13 was differentially expressed in various T
   cell subpopulations and response groups, with statistically significant
   differences (Fig. [129]5 c, d). TCGA cohort confirmed that CXCL13
   influences cell survival and is closely correlated with CD8Teff
   activity, reinforcing its role as a key modulator in the tumor
   microenvironment (Fig. [130]5e,). CXCL13 expression is significantly
   positively correlated with CD8⁺ Teff cell infiltration, and samples
   with high CXCL13 expression are broadly enriched in immune-related
   genes (Fig. [131]5f). Consistent results were observed in NSCLC ICIs
   cohorts, HCC immunotherapy cohort, mUC immunotherapy cohort
   (Fig. [132]5g-l). To further validate the differences in cytokine
   expression profiles driven by CD8⁺ Teff cells, we analyzed additional
   immunotherapy cohorts. Across the melanoma immunotherapy cohort
   (PRJEB23709), the NSCLC ICIs cohorts ([133]GSE126044, [134]GSE135222,
   POPLAR, and OAK), the HCC immunotherapy cohort (Imbrave150), and the
   mUC immunotherapy cohort (IMvigor 210), CXCL13 consistently emerged as
   a cytokine whose expression differed significantly under CD8⁺ Teff
   cell–driven conditions (Fig. [135]5m-t)n addition, immunohistochemistry
   (IHC) and immunofluorescence (IF) analyses revealed distinct spatial
   distributions of CXCL13 corresponding to different GZMB expression
   states, further supporting the co-expression relationship between GZMB
   and CXCL13 (Fig. [136]5u-v). These findings highlight the significant
   role of CXCL13 in modulating immune responses and its close association
   with CD8Teff cell activity within the tumor microenvironment.

Fig. 5.

   [137]Fig. 5
   [138]Open in a new tab

   CXCL13 was the most distinct cytokine in CD8Teff cells. Heatmaps
   showing CXCL13 expression between response and no-response groups in
   tissues samples (a) and (b) peripheral blood samples. c Violin plots
   and box plots comparing CXCL13 expression between subgroups. d Box plot
   showing CXCL13 expression differences between response and no-response
   groups. e, g, i, k Kaplan–Meier survival analysis of CXCL13 expression
   in four immunotherapy cohorts: (e) TCGA, (g) POPLAR, (i) Imbrave150,
   and (k) Imvigor210. f, h, j, l Correlation of CXCL13 expression with
   CD8Teff cells in four immunotherapy cohorts: (f) TCGA, (h) POPLAR, (j)
   Imbrave150, and (l) Imvigor210. m Heatmap and (n) correlation analysis
   of CXCL13 in the melanoma ICIs dataset (PRJEB23709). o Heatmap and (p)
   correlation analysis of CXCL13 in the NSCLC ICIs datasets
   ([139]GSE12604 + [140]GSE135222). q–t Heatmaps of immune-related
   cytokine expression between CD8Teff high vs. low groups across four
   immunotherapy cohorts: (q) POPLAR, (r) OAK, (s) Imbrave150, (t) Imvigor
   210. u IHC and IF images showing CXCL13 expression in CD8Teff-low and
   CD8Teff-high scoring tumors. v IF images showing GZMB expression in
   CXCL13-low and CXCL13-high scoring tumors

CD8Teff coordinate TME remodeling and metabolic reprogramming

   Given the critical function of CD8Teff cells in antigen recognition and
   immune activation, we explored their interaction with dendritic cells.
   Single-cell analysis revealed distinct subsets of DCs, including cDC1,
   cDC2, and mature DCs, with a stronger association between CD8Teff cells
   and cDC1 in the response group from tumor tissues samples (Fig. S4,
   Fig. [141]6a, b). Trajectory analysis showed that cDC1 cells were more
   prevalent in the response group (Fig. [142]6c, d). Gene expression
   analysis revealed upregulation of antigen presentation-related genes in
   cDC1, particularly those involved in immune signaling and Fc receptor
   activity (Fig. [143]6e, f). Cell–chat analysis revealed that CD8Teff
   exhibited stronger interactions with conventional dendritic cell
   subsets cDC1 and cDC2 in the response group from peripheral blood
   samples (Fig. [144]6g-h). Pseudotime trajectory analysis revealed
   distinct temporal differentiation patterns of cDC1 and cDC2 subsets
   between immunotherapy-responsive and non-responsive groups
   (Fig. [145]6i-j). Additionally, response patients showed increased
   activity in protein translation and ribosomal processes (Fig. [146]6k,
   l). Kaplan–Meier survival analysis indicated that higher DC
   infiltration was significantly associated with better survival outcomes
   (P = 0.007) (Fig. [147]6m), and the presence of CD8Teff cells in
   DC-rich tumors further highlighted the role of cDC1 in enhancing
   anti-tumor immune responses (Fig. [148]6n). This supports the crucial
   role of cDC1 in presenting antigens to CD8Teff cells, enabling
   tumor-specific killing.

Fig. 6.

   [149]Fig. 6
   [150]Open in a new tab

   CD8Teff cells affect antigen presentation and differentiation in DC
   cells. a Cell–cell communication network in response and (b)
   no-response groups. c Proportion of DC subsets between different groups
   in tumor samples. d Pseudotime trajectory analysis showing
   differentiation paths of DCs in no-response and response groups. e
   Heatmap of gene expression in DCs from no-response and response groups.
   f GO enrichment analysis highlighting biological processes involved in
   DC function in the response group. Cell–cell communication network in
   (g) response and (h) non-response. i Proportion of DC subsets in
   different groups in blood samples. j Pseudotime trajectory analysis
   showing differentiation paths of DCs at different treatment stages. k
   Heatmap of gene expression in DCs from response and no-response. l GO
   enrichment analysis highlighting biological processes involved in DC
   function in responder groups. n Box plots comparing the expression of
   CD8Teff cells between high-DC and low-DC groups in TCGA cohort.
   Kaplan–Meier survival analysis based on CAF expression levels in (o)
   PRJEB23709 and (r) NSCLC([151]GSE12604 + [152]GSE135222); Kaplan–Meier
   survival analysis based on cDC expression levels in these cohorts: (m)
   TCGA, (q) PRJEB23709, and (t) NSCLC([153]GSE12604 + [154]GSE135222).
   Kaplan–Meier survival analysis based on DC expression levels in these
   cohorts: (p) PRJEB23709 and (s) NSCLC([155]GSE12604 + [156]GSE135222)

   Further analysis showed that high levels of CAFs were associated with
   worse survival, suggesting that CAFs contribute to a more suppressive
   tumor microenvironment (Fig. [157]6o, r). In contrast, higher
   infiltration of DC and cDC cells was consistently linked to improved
   survival, with significant results observed in both the melanoma ICIs
   cohort (P = 0.0013 for DCs, P = 0.0057 for cDCs) and the NSCLC ICIs
   datasets (P = 0.02 for DCs, P = 0.024 for cDCs) (Fig. [158]6p, q, s,
   t). These findings highlight the role of high CAF levels in promoting
   immune suppression, while elevated DC infiltration supports a more
   active immune response, thereby improving patient prognosis. In
   summary, CD8Teff cells play a central role in shaping the tumor
   microenvironment by influencing tumor cells, fibroblasts, and
   endothelial cells. Their interaction with various cell types highlights
   their importance in modulating both tumor progression and immune
   responses.

   We further investigated how CD52 + dendritic cells (DCs) regulate
   CD8Teff cell activity. CD52 + DC populations were enriched in response
   samples, indicating their active role in promoting more effective
   immune responses (Fig. S5a). Pathway analysis showed strong connections
   between CD52 + DCs and CD8Teff cells, with enhanced antigen
   presentation and immune activation in responders (Fig. S5b, c).
   CD52 + DCs exhibited upregulation of genes linked to immune activation,
   such as TNFSF10, CCL5, and CCL22, in response samples, while
   non-responders showed weaker expression patterns (Fig. S5d, e).
   Furthermore, response patients demonstrated enhanced signaling in
   immune-regulating pathways, such as TNF and CXCL, indicating that
   CD52 + DCs play a key role in modulating immune responses and shaping
   the tumor microenvironment (Fig. S5f, g). Thus, CD52 + DC cells
   regulate CD8Teff activity by modulating antigen presentation and
   directly influencing cytokine activity, ultimately shaping the tumor
   microenvironment.

   Building on these findings, we extended our analysis to other key
   components of the tumor microenvironment, such as tumor cells,
   fibroblasts, and endothelial cells (Fig.S6). CD8Teff cells exhibit
   strong communication with CancerCells_2 and CancerCells_4 (Fig. [159]7
   a, b), which are associated with the upregulation of interferon-alpha
   response, apoptosis, and epithelial-mesenchymal transition (EMT)
   pathways, further supporting their potential role in tumor suppression
   and remodeling of the tumor microenvironment (Fig. [160]7 a, b, c).
   Additionally, CD8Teff cells exhibited particularly strong interactions
   with myCAF (Fig. [161]7 d, e), activating pathways such as TNF
   signaling and cytokine-cytokine receptor interactions, which may
   contribute to immunomodulatory functions in the tumor microenvironment
   (Fig. [162]7d, e, f). In contrast, low CD8Teff expression was linked to
   stronger interactions between tumor cells and microvascular endothelial
   cells (Fig. [163]7g, h), with ECM pathway activation being more
   prominent in no-response, indicating a potential link to increased
   metastatic potential (Fig. [164]7i, j). Stronger communication has
   shown between epithelial cell, CD8Teff and NKT cells in the response
   group (Fig. [165]7k, l). The frequency and intensity of CD8Teff–NK cell
   interactions between response and no-response groups is different
   (Fig. [166]7k, l). The CD8Teff-5 subpopulation exhibited stronger
   associations with NK cells on-treatment (Fig. [167]7m, n). Notably, in
   the response group, CD8⁺ Teff-5 demonstrated particularly close
   interactions with NK cells (Fig. [168]7o, p).

Fig. 7.

   [169]Fig. 7
   [170]Open in a new tab

   CD8Teff cells affect the function of tumour cells, fibroblasts and
   endothelial cells. Network diagrams representing the interactions
   between CD8Teff cells and different tumor cell subtypes in (a) response
   and (b) no-response groups. c Heatmap of hallmark pathways affected by
   CD8Teff interactions with various cancer cell subtypes. Network
   diagrams of CD8Teff cell interactions with fibroblast subtypes in (d)
   response and (e) no-response groups. f Heatmap of hallmark pathways
   affected by CD8Teff interactions with various fibroblasts subtypes.
   Network diagrams of CD8Teff cell interactions with endothelial cell
   subtypes in (g) response and (h) no-response groups. i Heatmap of gene
   expression in the endothelial cells. j Enriched pathways impacted by
   CD8Teff cells across different groups. Network diagrams representing
   the interactions between CD8Teff cells, epithelial cell and NK cell in
   (k) response and (l) no-response groups. m, n, o, p Network diagrams
   representing the interactions between different tumor cell subtypes of
   CD8Teff cells in (m) pre-treatment, (n) on-treatment and (o) response,
   (p) no-response groups. q Immunofluorescence co-expression of GZMB and
   CD56 under high and low expression conditions

   To further investigate the relationship between CD8Teff and NK cells,
   we performed spatial co-localisation analysis using multiplex
   immunofluorescence, which demonstrated similar expression abundances
   and spatial distributions for CD8Teff and NK cells (Fig. [171]7q).

   Additionally, pathway enrichment analysis revealed that key metabolic
   processes, such as glycolysis, fatty acid oxidation (FAO), and
   glutaminolysis, were significantly upregulated in response patients,
   highlighting the importance of these metabolic pathways for effective
   anti-tumor responses (Fig. S7a, b). In the response group, CD8Teff
   cells primarily exhibited enhanced amino acid metabolism and lipid
   biosynthesis, supporting their activated and functional state. In
   contrast, CD8Teff cells from the non-responsive group displayed a
   metabolism dominated by glucose utilization, with elevated glycolytic
   activity suggesting a shift toward functional exhaustion (Fig. S7c).
   Patients with high CD8Teff infiltration also showed increased activity
   in glycolysis, FAO, and the pentose phosphate pathway (PPP), linking
   CD8Teff cells to metabolic reprogramming within the tumor (Fig. S7d-i).
   These findings suggest that metabolic shifts driven by CD8Teff cells
   are critical for shaping the tumor microenvironment and enhancing
   anti-tumor immunity.

AI-based spatial mapping of CD8Teff enables precision stratification

   To assist clinical decision-making and enhance the understanding of
   tumor immunity, we developed a network of AI algorithms using
   pathological images of TNBC patients from the TCGA database, aimed at
   identifying CD8Teff within the tumor microenvironment as key
   determinants of "cold" and "hot" tumor classification. A total of 836
   HE pathology slices from the TCGA public database were analyzed,
   quantifying immune cell distributions with high specificity
   (Fig. [172]8a). Pathologically identified CD8Teff-enriched regions from
   patients with high CD8Teff expression and CD8Teff-sparse regions from
   patients with low CD8Teff expression were highlighted. Analysis
   revealed significant spatial heterogeneity in CD8Teff distribution
   (Fig. [173]8b), underscoring the variability in immune infiltration
   within the tumor microenvironment. This spatial heterogeneity provides
   a foundation for interpreting immune cell distribution in pathology
   images and sets the stage for the future quantification of additional
   immune markers. The reliability of CD8Teff identification was
   validated, achieving an AUC of 0.823 in the training group and 0.805 in
   the validation group, confirming the accuracy of AI-driven immune cell
   classification (Fig. [174]8c, d).

Fig. 8.

   [175]Fig. 8
   [176]Open in a new tab

   Pathological AI detects CD8Teff expression and spatial heterogeneity. a
   design process of the CD8Teff pathology artificial intelligence (AI)
   algorithm, (b) visualizing CD8Teff distribution heterogeneity in
   pathology slices, the AUC of pathology based-AI identified CD8Teff is
   (c) 0.823 in the training group and (d) 0.805 in the validation group

   This AI-driven analysis plays a crucial role in clinically
   distinguishing between "hot" and "cold" tumors, aiding in personalized
   treatment strategies. Patients with "cold" tumors, characterized by low
   CD8Teff activity, may benefit from anti-angiogenic therapies that
   target the tumor-supportive microvascular environment. In contrast,
   patients with "hot" tumors, which exhibit higher CD8Teff infiltration,
   typically respond better to immunotherapy combined with chemotherapy.
   By aligning treatments with the tumor immune profile, this
   precision-driven approach not only optimizes therapeutic outcomes but
   also represents a significant advancement in personalized cancer care.

CD8Teff cells drive tumor immunity status and inform personalized therapy

   CD8Teff cells serve as a critical marker for distinguishing tumor
   immunotypes and provide essential insights for guiding personalized
   treatment. These cells modulate the tumor microenvironment by
   influencing DC antigen presentation, which enhances CD8Teff cell
   activation. In addition, CD8Teff cells are linked to the expression of
   CXCL13, a key cytokine that helps regulate immune responses and further
   amplifies CD8Teff cell activity. Together, this process drives changes
   in the tumor microenvironment, affecting the differentiation of tumor
   cells, CAFs, and endothelial cells. In "hot" tumors, CD8Teff cells are
   more active, creating an immune-responsive environment that is more
   receptive to immunotherapy (Fig.S8).

   For clinical applications, this distinction
   between "hot" and "cold" tumors is vital. Patients with "cold" tumors,
   characterized by low CD8Teff activity, may benefit from anti-angiogenic
   therapies that target the supportive microvascular environment
   promoting tumor growth. In contrast, patients with "hot" tumors, which
   have higher CD8Teff infiltration, often achieve better outcomes with
   immunotherapy combined with chemotherapy, enhancing the immune system's
   ability to fight the tumor. This personalized treatment approach can
   significantly improve therapeutic effectiveness depending on the
   tumor's immune profile.

Discussion

   In this study, single-cell RNA sequencing analyses of tumor tissues and
   peripheral blood from TNBC patients undergoing ICIs were conducted to
   investigate factors associated with the ICIs treatment efficacy. This
   study provides crucial insights into the immune landscape of TNBC,
   revealing distinct immune cell dynamics in response to chemotherapy
   combined with ICIs. Through scRNA-seq, the study identified significant
   differences in the composition and functionality of immune cell subsets
   between Response patients and No-response patients. Notably, CD8Teff
   cells emerged as a key determinant of immunotherapy efficacy and tumor
   classification as "hot" or "cold", and CD52 + DCs modulate CD8Teff cell
   activity by influencing antigen presentation. Furthermore, activated
   CD8Teff cells enhance their tumor-killing capacity by secreting
   increased levels of CXCL13, which mediates interactions with NK cells,
   tumor cells, CAFs, and vascular endothelial cells. Which interactions
   play a critical role in shaping the tumor microenvironment and
   ultimately affect the efficacy of immunotherapy. These findings offer
   clinicians enhanced tools for assessing the likely success of
   immunotherapy in breast cancer patients, ultimately contributing to
   more personalized treatment strategies. Furthermore, the study guides
   clinicians in developing treatment regimens tailored to differential
   CD8Teff expression to enhance therapeutic efficacy and employs AI-based
   pathological models to directly identify and quantify CD8Teff cells,
   thereby supporting clinical decision-making in immunotherapy. This
   study addressed the challenge of efficiently and directly identifying
   the state of the tumor immune microenvironment and elucidated its
   underlying mechanisms. By characterizing CD8Teff status, we guided the
   selection of immunotherapy strategies tailored to individual patients
   and implemented a pathological AI model capable of directly identifying
   CD8Teff cells, thereby facilitating clinical translation. This approach
   supports precision treatment and aims to enhance the efficacy of
   immunotherapy in TNBC.

   CD8Teff cells are pivotal in defining the "hot" or "cold" status of the
   TME and possess significant potential serve as clinical diagnostic
   biomarkers for guiding immunotherapy decisions. The study demonstrated
   the robust strong capability of CD8Teff cells to distinguish
   between "hot" and "cold" TME through integrated single-cell and
   multi-transcriptome data analysis. This approach challenges traditional
   methods that rely on multifactorial indicators such as CD8 + T cell
   staining, tumor mutation burden, or active cytokines, offering a more
   direct and precise means of categorizing tumor immune activity. By
   providing clinicians and pathologists with a more convenient assessment
   index, it reduces the diagnostic complexity and workload for
   pathologists, thereby improving the diagnostic efficiency for patients
   with TNBC. Furthermore, efficient AI-enabled pathological recognition
   of CD8Teff cells facilitate the classification of tumors
   as "cold" or "hot", presenting a novel approach to advancing tumor
   diagnosis and prognosis.

   Different adjuvant and neoadjuvant treatment regimens may be
   recommended for TNBC patients based on their different CD8Teff typing.
   Clinical evidence from trials such as POPLAR, OAK, IMbrave150, and
   IMvigor210 has demonstrated that "hot" tumors, characterized by higher
   CD8Teff levels, exhibited stronger immune activation and improved
   clinical outcomes [[177]10]. In contrast, the"cold"tumors, which are
   enriched with immunosuppressive cells like Tregs, show poorer responses
   to immunotherapy and are correlated with shorter PFS and OS.

   Immunotherapy-based combination regimens can be tailored according to
   CD8Teff phenotypes and their associated tumor microenvironment
   molecular signatures. The FUTURE-C-PLUS/FUTURE-SUPER study found that
   the combination of chemotherapy and immunotherapy with anti-angiogenic
   drugs improved TME and TNBC achieving objective response rates of up to
   77% [[178]16, [179]17]. For patients with cold tumors and low CD8Teff
   expression, a treatment strategy incorporating immunotherapy and
   anti-angiogenic drugs is recommended to suppress tumor invasion and
   metastasis by enhancing anti-angiogenesis [[180]18]. Additionally,
   administering active cytokines is advised to modulate the inert TME and
   increase the sensitivity of TNBC patients to immunotherapy. KEYNOTE-522
   [[181]5, [182]19], KEYNOTE-355 [[183]3, [184]20] and TORCHLIGHT
   [[185]21] have demonstrated that both early-stage and advanced TNBC can
   benefit from immunotherapy, especially when combined with
   albumin-paclitaxel. For patients with "hot" tumors and high CD8Teff
   expression, immunotherapy combined with albumin-paclitaxel-based
   chemotherapy is the preferred approach remains the preferred
   therapeutic approach.

   CD8Teff cells hold significant potential as biomarkers for predicting
   immunotherapy response, and enhanced CD8Teff activity may
   transform "cold" tumors into "hot" tumors, fostering a more
   immune-responsive microenvironment and ultimately improving patient
   prognosis. The OXEL study demonstrated that both peripheral immune cell
   scores and circulating tumor DNA status were strongly associated with
   an increased risk of relapse in patients with early TNBC [[186]22].
   Similarly, the KEYNOTE-355 study demonstrated that patients with PD-L1
   combined positive score (CPS) ≥ 10 experienced improved PFS and OS
   [[187]23]. However, the predictive capacity of PD-L1 scores remains
   limited, as they do not consistently reflect the efficacy of
   immunotherapy [[188]24]. In contrast, this study offered a more direct
   and reliable indicator of tumor immune status and therapy response by
   utilizing CD8Teff scores. Unlike with PD-L1 scores, the CD8Teff score
   integrates multiple complex factors within the TME, including
   interactions with DCs, CAFs, cancer cells, NK cells and cytokines,
   offering a more accurate characterization of tumor immune activity and
   therapy response. These elements contribute to a more accurate
   characterization of tumor immune activity and treatment response.
   Moreover, alterations in pathways such as MYC and KRAS play a critical
   role in converting tumors from a "cold" to a "hot" phenotype, further
   highlighting the importance of targeting these mechanisms to enhance
   the effectiveness of immunotherapy.

   CD8Teff cells contribute to metabolic reprogramming of the tumor
   microenvironment, enhancing antitumor immune responses and modulating
   tumor progression. The strong correlation between CD8Teff cells and
   metabolic reprogramming within the tumor microenvironment suggests that
   CD8Teff-driven metabolic shifts, including upregulation of glycolysis
   and fatty acid oxidation, which are essential for sustaining effective
   anti-tumor responses [[189]25–[190]27]. These findings highlight the
   therapeutic potential of targeting metabolic pathways to enhance
   CD8Teff functionality and improve immunotherapy outcomes. Targeting
   metabolic pathways to enhance CD8Teff activity could be a promising
   approach to boost the effectiveness of immunotherapies [[191]28]. Drugs
   or interventions that modulate key metabolic processes, such as
   glycolysis or fatty acid oxidation, could support the energy demands
   and functional persistence of CD8Teff cells within the metabolically
   challenging TME. potentially transforming the tumor microenvironment
   from a metabolically hostile one into a more supportive environment for
   immune-mediated tumor destruction. Enhancing glycolysis through
   upregulation of GLUT1 or promoting fatty acid oxidation via CPT1A
   activation can reinforce effector function and cytotoxicity. These
   interventions have the potential to remodel the TME from an
   immunosuppressive, nutrient-depleted state into one that is more
   conducive to sustained immune activation and tumor eradication.

   CD8⁺Teff function is shaped by a complex network of stimulatory and
   inhibitory interactions within TME. DCs play a central role in priming
   naïve CD8⁺ T cells by presenting antigenic peptides via MHC-I and
   delivering co-stimulatory signals. In turn, activated CD8Teff cells
   secrete IFN-γ, which enhances DC maturation and antigen-presenting
   capacity, forming a positive feedback loop that amplifies anti-tumor
   immunity. NK cells also contribute to this network by supporting CD8⁺ T
   cell–mediated tumor clearance through IL-15–mediated cross-talk and
   complementary cytotoxic mechanisms. Conversely, CAFs exert
   immunosuppressive effects by producing extracellular matrix components
   that physically hinder immune cell infiltration. They also secrete
   cytokines such as TGF-β and IL-6, which promote immune evasion and
   tumor progression. CAFs further recruit and sustain Tregs and
   myeloid-derived suppressor cells (MDSCs), both of which inhibit CD8⁺ T
   cell function. A comprehensive understanding of these intercellular
   dynamics is essential for designing synergistic and precise therapeutic
   interventions to restore and potentiate CD8⁺Teff functionality.

   Despite the valuable insights gained, the study still has the following
   shortcomings. First, there was inconsistency in the single-cell RNA
   sequencing platforms used—tissue and blood samples were sequenced using
   different technologies compared to the GEO public database datasets,
   which may introduce technical variability. Second, the study primarily
   focused on the characterization and mechanism of CD8Teff, and although
   the findings are compelling, they serve as supportive evidence for
   future basic mechanistic investigations. Third, the study did not
   include a prospective clinical cohort evaluating the effects of
   anti-angiogenic therapies. Future studies should incorporate
   prospective cohorts to explore the therapeutic efficacy and mechanistic
   contributions of combining anti-angiogenic agents with immunotherapy
   and chemotherapy, particularly in patients with'cold'tumors.

   In summary, the study identifies CD8Teff as pivotal markers for
   immunotherapy response in TNBC. Through integrated single-cell RNA
   sequencing and AI-based analysis, high CD8Teff levels correlate
   with "hot" tumors characterized by active immune responses and
   favorable treatment outcomes, whereas "cold" tumors with low CD8Teff
   infiltration exhibit reduced responsiveness to ICIs. CD8Teff cells
   actively modulate through cytokine signaling and immune activation,
   underscoring their critical role in shaping treatment response. These
   findings support the implementation of CD8Teff-based stratification as
   a foundation for personalized immunotherapy strategies in TNBC.

Materials and methods

Study design and patients

   This study prospectively enrolled four TNBC patients from Sun Yat-Sen
   Memorial Hospital, Sun Yat-Sen University (China) between January and
   September 2021. Two early-stage TNBC patients received neoadjuvant
   chemotherapy combined with immune checkpoint inhibitors (ICIs); one
   achieved a pathological complete response (pCR, defined as the
   response), and the other did not (non-pCR, defined as the no-response).
   Additionally, two advanced TNBC patients receiving first-line
   chemoimmunotherapy were included, with blood samples collected before
   and after treatment; one response and the other did not (Supplementary
   Table 1).

   Publicly available single-cell transcriptomic data from TNBC patients
   ([192]GSE176078, [193]GSE161892) were incorporated for validation
   (Supplementary Table 2). Bulk RNA-seq data from the TCGA-BRCA cohort
   and additional immunotherapy datasets were also analyzed, including:
   Melanoma: PRJEB23709 (European Nucleotide Archive); NSCLC:
   [194]GSE126044, [195]GSE135222 (GEO), POPLAR ([196]NCT01903993), and
   OAK ([197]NCT02008227); HCC: Imbrave150 ([198]NCT03434379); mUC:
   IMvigor210 ([199]NCT02951767, via IMvigor210CoreBiologies R package).
   Samples with missing clinical or survival data were excluded. All
   datasets included transcriptomic profiles, immunotherapy response
   classifications, and survival outcomes (PFS/OS) (Supplementary
   Table 2). Above mentioned datasets were accessed and downloaded in
   September 2021 and have remained publicly available since then.

Single-cell transcriptome library preparation and sequencing

   Fresh tumor tissues were preserved in sCelLive™ solution (Singleron
   Bio), processed within 30 min post-resection. Samples were minced,
   enzymatically dissociated using GEXSCOPE™ solution, filtered, and
   centrifuged. Erythrocytes were removed, and viable cells were counted
   with trypan blue. Cell suspensions (1 × 10^5 cells/mL) were loaded onto
   microfluidic chips for library construction using the GEXSCOPE™
   Single-Cell RNA Library Kit and sequenced on an Illumina NovaSeq 6000
   (150 bp paired-end reads).

   Reads were processed with CeleScope v1.1.7. Cutadapt v1.17 was used for
   trimming [[200]29], STAR v2.6.1a for genome alignment (GRCh38)
   [[201]30], and featureCounts v2.0.1 for UMI quantification [[202]31].
   Cells were excluded if they had < 200 or top 2% of gene counts, > 50%
   mitochondrial reads, or genes expressed in < 5 cells. After filtering,
   13,334 high-quality cells were retained (avg. 1,869.6 genes and 7,975.4
   UMIs/cell).

   Seurat v3.1.2 was used for normalization, scaling, PCA, and clustering
   (resolution = 2.0, 30 clusters). UMAP was used for 2D visualization.
   DEGs were identified with Seurat’s FindMarkers (Wilcoxon test,
   logFC > 0.25, expression in > 10% of cells). Cell types were annotated
   using canonical markers and the SynEcoSys database (Supplementary
   Tables 2–3). Doublets were excluded based on co-expression of multiple
   lineage markers.

   Pathway and trajectory analysis. Pathway enrichment analysis was
   conducted using clusterProfiler v4.0.2 [[203]32] for GO and KEGG,
   considering pathways with p_adj < 0.05 as significant. Gene set
   variation analysis (GSVA) was performed using the GSVA package v1.44.2
   [[204]33] to examine pathway differences across cell types.

   For cell differentiation trajectory analysis, Monocle2 v2.22.0
   [[205]34] was used with highly variable genes to order cells by
   differentiation stage. Dimension reduction was achieved with DDRTree,
   and key differentiation paths were identified.

   Transcription factor network analysis. Transcription factor regulatory
   networks were inferred using pyscenic v0.11.0 [[206]35], starting with
   scRNA expression data and transcription factors from AnimalTFDB.
   Networks were predicted with GRNBoost2, refined with CisTarget to
   identify transcription factor binding motifs, and visualized using
   pheatmap in R.

   Macrophage gene modules were detected using Hotspot [[207]36] (top 500
   genes with highest autocorrelation z-scores). Modules were scored and
   Jaccard similarity coefficients were calculated to assess
   transcriptional similarity between immune subtypes [[208]37].

   CellPhoneDB v2.1.7 was used to infer ligand–receptor interactions
   (p < 0.05, log-expression > 0.1) [[209]38]. Circlize v0.4.10 (R) was
   used for visualization.

Transcriptome sequencing data analysis

   The xCELL algorithm, based on the ssGSEA method and gene set markers,
   was used to estimate the abundance of 64 cell types across five major
   categories: adaptive and innate immune cells, hematopoietic progenitor
   cells, epithelial cells, and extracellular matrix cells (including
   non-immune cell types). Following the GitHub protocol, we performed
   calculations using the functions rawEnrichmentAnalysis,
   transformScores, and spillOver. This process yielded results for 64
   cell subtypes per sample, along with ImmuneScore, StromaScore, and
   MicroenvironmentScore. Kaplan–Meier survival curves were generated
   using the survival package in R.

   Normalization and identification of differentially expressed genes
   (DEGs) were performed using the DESeq package in R. DESeq applies a
   negative binomial distribution (NB) test to assess differential read
   counts while correcting for length-related bias. Gene expression levels
   were estimated using the baseMean value. Genes with a log₂ fold
   change > 1 or <  − 1 and p < 0.05 were considered significantly
   differentially expressed. Pathway enrichment analysis, including Gene
   Ontology (GO) and KEGG, was performed using the clusterProfiler R
   package. GO terms and KEGG pathways with p-values and false discovery
   rates (FDR) < 0.05 were considered significant. GSEA was based on 1,000
   permutations and FDR-adjusted p-values. Results were visualized using
   enrichplot. GSVA was conducted using gene sets from the MSigDB
   database.

Multiplex immunofluorescence

   FFPE tissue sections were deparaffinized twice with xylene and
   rehydrated through graded ethanol (95%, 85%, 75%), followed by rinsing
   with PBS. Sections were heated in citrate buffer (pH 6.0) at 100 °C for
   30 min, cooled, and washed with PBS. They were incubated in 1% BSA for
   30 min at room temperature to block non-specific binding. Sections were
   then incubated overnight at 4 °C with primary antibody (1:200 in BSA),
   followed by three 5-min PBS washes. HRP-conjugated secondary antibody
   (1:400, TSA kit) was applied for 60 min at room temperature, then
   washed three times with PBS. Endogenous peroxidase was blocked using
   0.3% H₂O₂ for 10 min, followed by PBS washes. TSA-tyramine substrate
   with fluorescent dye (1:200) was added for 15 min at room temperature,
   followed by PBS washes. Antigen retrieval was repeated at 95 °C for
   25 min, followed by PBS washes. These steps were repeated for
   additional antibody staining. Sections were mounted with antifade
   medium containing DAPI and sealed with coverslips. Stained sections
   were analyzed using a fluorescence microscope.

Immunohistochemistry

   Paraffin sections were baked at 70 °C for 50 min, then deparaffinized
   in xylene (15 min, twice) and rehydrated through a graded ethanol
   series (100% ethanol for 5 min, 95% for 5 min, 75% for 3 min). Sections
   were rinsed in PBS for 3 min (three times). Antigen retrieval was
   performed using sodium citrate buffer (pH 6.0) under high pressure for
   30 min, followed by natural cooling and PBS rinses. To block endogenous
   peroxidase activity, 3% hydrogen peroxide was applied for 15 min at
   room temperature, followed by PBS rinses. Sections were blocked with
   goat serum for 30 min, and primary antibody was applied for overnight
   incubation at 4 °C. After rewarming, sections were incubated with
   secondary antibody (Zhongshan Jinqiao kit) for 30 min. DAB staining
   (1:20) was monitored microscopically, followed by hematoxylin
   counterstaining, dehydration, and mounting with neutral gum.

Developing a pathological AI that recognizes CD8Teff expression and spatial
heterogeneity

   Each pathology slide was initially segmented into distinct regions to
   enable localized analysis of immune cell presence. Feature extraction
   was performed on these regions using a ResNet convolutional neural
   network (CNN), which captured intricate morphological characteristics
   associated with immune cell density. The extracted features were
   processed through a self-attention mechanism module followed by a fully
   connected layer, enabling binary classification of tumors as either
   CD8Teff-enriched or CD8Teff-sparse. This approach represents a key
   advance in quantifying immune cells from pathology images.

Statistical analysis

   All statistical analyses were performed using R software (version
   4.2.2). Appropriate statistical tests were applied based on data type.
   Chi-square, Fisher's exact, Mann–Whitney U, Kruskal–Wallis, and
   Student’s t-tests were used where applicable. A two-tailed
   P-value < 0.05 was considered statistically significant.

Supplementary Information

   [210]Supplementary Material 1.^ (4.1MB, docx)

Acknowledgements