Fibroblast TGF-β signaling defines spatial tumor ecosystems linked to immune checkpoint blockade resistance

Abstract

Resistance to immune checkpoint blockade (ICB) remains a major challenge in lung adenocarcinoma (LUAD), with stromal mechanisms underlying CD8⁺ T cell exhaustion still poorly understood. By integrating single-cell, bulk, and spatial transcriptomic datasets using EcoTyper, we identified a distinct immunosuppressive ecotype, EC10, enriched for TGF-β signaling and epithelial–mesenchymal transition. EC10 exhibited spatial co-localization of fibroblasts, malignant epithelial cells, and exhausted CD8⁺ T cells, and was consistently associated with immune exclusion, poor progression-free survival, and elevated TIDE scores across four ICB-treated LUAD cohorts. Cell–cell communication analyses revealed a dominant TGFB1–SERPINE1 signaling axis originating from fibroblasts, linking stromal remodeling to T cell dysfunction. In contrast, EC12 represented an inflamed, ICB-responsive state enriched in interferon signaling. These findings define EC10 as a spatially organized, fibroblast-driven immunosuppressive ecosystem predictive of ICB resistance, and highlight the therapeutic potential of targeting the TGF-β axis in LUAD.

This research identifies a fibroblast-driven immunosuppressive ecosystem in lung adenocarcinoma, linking TGF-β signaling to immune resistance.

Introduction

Lung cancer remains the most commonly diagnosed malignancy and the leading cause of cancer-related mortality worldwide, accounting for over 20% of global cancer-related deaths¹. Non-small cell lung cancer (NSCLC), which comprises approximately 60% adenocarcinoma (LUAD) and 30% squamous cell carcinoma (LUSC), dominates the disease landscape. For early stage NSCLC, such as in stage I cases, surgical resection is the cornerstone of treatment². However, 20–40% of these patients experience recurrence within two years, resulting in a 5-year post-recurrence survival rate of only 15–16.6%³. This sobering statistic underscores the urgent need to unravel the biological mechanisms that drive relapse. While genomics-based targeted therapies and immunotherapies, including immune checkpoint inhibitors (ICIs), have emerged as promising options for advanced NSCLC, the overall 5-year survival rates remain disappointing, with 64% for localized disease, 37% for regional spread, and 9% for metastatic cases, largely due to persistent recurrence⁴. These outcomes emphasize the pivotal role of the tumor immune microenvironment in shaping LUAD heterogeneity and highlight the growing importance of molecularly tailored treatments to improve patient prognosis.

Recent studies have increasingly framed tumors as complex ecosystems, in which tumor cells, immune cells, and stromal components engage in intricate interactions that collectively influence tumorigenesis, prognosis, and therapeutic responses, including ICIs. Beyond the tumor cells themselves, diverse cell types, their states, and their interplay within this ecosystem are critical determinants of cancer behavior⁵. Advances in single-cell genomics, spatial transcriptomics, and multiplexed imaging have revolutionized our ability to dissect the tumor microenvironment (TME) at unprecedented resolution, revealing the cellular populations and states that govern tumor progression and ICI efficacy⁶. However, these approaches often focus on specific tumor types or small cohorts, limiting their capacity to fully capture the heterogeneity of tumor ecosystems. To address this, tools such as EcoTyper, a machine-learning framework, integrate bulk, single-cell, and spatially resolved transcriptomic data to systematically map cellular states and multicellular communities⁷. This approach has enabled the identification of ten recurrent multicellular communities in NSCLC, with subsequent analyses linking these communities to patient survival, ICI response, and spatial organization within the TME, laying the groundwork for improved diagnostics and precision therapeutics⁸.

TGF-β signaling plays diverse roles in several cancer types. In breast cancer(https://pmc.ncbi.nlm.nih.gov/articles/PMC2390892/), it promotes epithelial–mesenchymal transition, immune evasion, and metastasis⁹. In pancreatic cancer (https://pmc.ncbi.nlm.nih.gov/articles/PMC10311106/), it activates cancer-associated fibroblasts and contributes to the fibrotic tumor microenvironment. In colorectal cancer (https://www.nature.com/articles/ng.3225), loss of SMAD4 and other pathway components is associated with metastasis and poor prognosis. These examples illustrate that TGF-β signaling is highly context-dependent and may exert both tumor-suppressive and tumor-promoting effects depending on the tumor type and microenvironment. In our study, we focus on its role in lung adenocarcinoma, where it is closely linked to fibroblast activation, CD8⁺ T cell exhaustion, and resistance to immune checkpoint blockade^10,11. While targeting TGF-β signaling has emerged as a potential strategy in cancer treatment, its pleiotropic effects complicate therapeutic approaches. Understanding the interplay between TGF-β, fibroblasts, and immune cells in non-small cell lung cancer (NSCLC), particularly in lung adenocarcinoma (LUAD), could provide new avenues for overcoming treatment resistance and improving patient outcomes, especially given the intricate cellular dynamics revealed by advanced analytical tools like EcoTyper¹⁰.

Despite advances in characterizing LUAD heterogeneity, the intercellular dynamics of the tumor microenvironment remain poorly understood¹². Using EcoTyper⁷ to integrate single-cell, bulk, and spatial transcriptomic data, we identified 12 multicellular ecotypes (E1–E12). Among these, EC10 consistently emerged as the most survival-associated ecotype, characterized by fibroblast-driven TGF-β signaling and CD8⁺ T cell exhaustion. These findings motivated us to focus our study on EC10 as a clinically relevant, immunosuppressive ecosystem.

Results

Profiling and identifying the ecosystem of lung adenocarcinoma (LUAD)

To map the cellular ecosystems of non-small cell lung cancer (NSCLC) and lung adenocarcinoma (LUAD), we utilized EcoTyper⁷, which incorporates CIBERSORTx¹³ to dissect bulk transcriptomic data into cell-type-specific profiles. Single-cell RNA sequencing (scRNA-seq) data from adenocarcinoma in situ (AIS) served as the foundational discovery dataset to define the reference signatures. These signatures were then applied to extract cellular states and ecotypes from the bulk RNA-seq and Visium spatial transcriptomic data. Our analysis began with 12 scRNA-seq datasets derived by Salcher et al.¹⁴ (Fig. 1a) and filtered to focus on primary NSCLC LUAD samples (Fig. 1b). For ecotype recovery, we analyzed five bulk RNA-seq datasets and one Visium dataset¹⁵; details are provided in Supplementary Data S1.

Fig. 1. EcoTyper-derived ecotypes from scRNA-seq data associated with LUAD outcomes.

a Reproduce UMAP using originate dataset. b Filtered NSCLC LUAD data from originate dataset. c Ecotype discover pattern and cell state abundance within Salcher et al.¹⁴ single cell RNA seq dataset. d Summary of Ecotyper subtypes (E1–E12) cell counts and patient number. e GSE11969 validation cohort of ecotype cell state abundance pattern and clinical-related survival. f GSE13213 validation cohort of ecotype cell state abundance pattern and survival. g GSE19188 validation cohort of ecotype cell state abundance pattern and survival. h GSE30219 validation cohort of ecotype cell state abundance pattern and survival. i GSE37745 validation cohort of ecotype cell state abundance pattern and survival.

Integrative analysis delineated 12 multicellular ecotypes (E1–E12) with distinct transcriptional profiles and patient distributions (Fig. 1c, d). Among them, EC10 consistently exhibited the strongest association with poor survival outcomes. EC10 was defined by malignant epithelial cells, fibroblasts, and CD8⁺ T cells, and distinguished by a dominant fibroblast-to-epithelial TGFB1–SERPINE1 signaling axis. Pseudotime analysis further indicated fibroblast-driven regulation of epithelial states, highlighting EC10 as a dynamic immunosuppressive ecosystem within LUAD.

The tumor ecosystem in NSCLC LUAD exhibits significant complexity and challenges in linking molecular mechanisms with cancer progression and patient outcomes. Using EcoTyper⁷, we identified 12 distinct ecotypes (termed NSCLC LUAD ecotypes, NEs), labeled E1–E12 (Fig. 1c). Patient and cell count distributions across these ecotypes varied significantly (Fig. 1d). For instance, E3 included 14 patients and 12,695 cells, reflecting broad patient representation. In contrast, E11 encompassed nine patients but yielded 41,772 cells, indicating a high cellular density, whereas E1 was limited to one patient and 508 cells. To explore the clinical relevance of these ecotypes, we analyzed 16 bulk RNA-seq datasets for ecotype recovery, focusing on survival differences between E10 and E12, which showed statistically significant differences (Supplementary Data S2 and S3). In the GSE11969 dataset, 9 samples were classified as E10 and 47 as E12, with a survival p-value of 0.0041 (Fig. 1e). The GSE13213 dataset assigned 10 samples to E10 and 55 to E12, yielding a p-value of 0.00011 (Fig. 1f). For GSE19188, 5 samples belonged to E10 and 34 to E12, with a p-value of 0.061 (Fig. 1g). In GSE30219, 4 samples were E10 and 37 were E12, with a p-value of 0.0025 (Fig. 1h). Finally, GSE37745 categorized 6 samples as E10 and 44 as E12, with a p-value of 0.047 (Fig. 1i). We further characterized the demographic and molecular compositions of E10 and E12 (Supplementary Data S4). Ecotype E10 included both male and female patients across early (stages I–II) and advanced (stages III–IV) disease, featuring epithelial malignant cells (state S06), fibroblasts (S08), and CD8+ T cells (S01) (Supplementary Fig. S1b). Ecotype E12 similarly spanned both sexes and stages, but was distinguished by the presence of BRAF and EGFR driver mutations, including NK cells (S01), endothelial cells (S01), and CD8+ T cells (S02).

EcoTyper analysis revealed the presence of consistent marker genes for these cellular states. In E10, epithelial malignant cells (S06) were marked by KRT6A, PITX1, CDA, S100A9, and EDIL3, with KRT6A¹⁶, PITX1¹⁷, S100A9¹⁸, and EDIL3¹⁹ implicated in invasion and progression via epithelial–mesenchymal transition (EMT). Fibroblasts (S08) expressed IL13RA2, RFLNA, TNFRSF11B, CADM3, and DKK1. TNFRSF11B²⁰ is linked to oncogenic activity through the PI3K/AKT pathway, and IL13RA2²¹ enhances lung cancer cell migration, invasion, and metastasis by activating PI3K, Akt, and TAZ signaling. CD8+ T cells (S01) were characterized by AURKB, UBE2C, ZWINT, ASPM, NUSAP1, ITGAE, and ITGA1, with UBE2C²² being associated with poor prognosis and increased proliferation and invasiveness in NSCLC.

On E12, NK cells (S01) expressed CD8A, TRGV3, TRGV4, GIMAP5, SNHG22, KLRK1, CXCR6, LAG3, TIGIT, and CTSW, with LAG3 and TIGIT indicating NK cell exhaustion. T cell CD4 (S01) was marked by IFITM1, SNHG22, GIMAP5, LINC-PINT, ITK, PDCD1, SELL, CCR7, and LEF1, where ITK²³ supports TCR signaling and T cell proliferation, and CCR7²⁴ regulates thymic structure, T-cell homing, and cancer cell migration. CD8 (S02) T cells expressed IFITM, SNHG22, NLRC5, SLAMF1, TRGV3, CXCR6, KLRK1, and CD69. Notably, IFITM1²⁵ served as an independent prognostic factor unrelated to EMT markers, NLRC5²⁶ bolstered antitumor T-cell CD8 responses, SLAMF1²⁷ promoted tumor cell survival via PI3K/Akt/mTOR signaling, CXCR6²⁸ enhanced T cell survival and tumor control, KLRK1²⁹ supported antitumor immunity, and CD69³⁰ contributed to T cell proliferation and survival.

Identification of NSCLC LUAD ecotypes (NEs) from multicellular communities

To investigate the molecular pathways underlying NSCLC LUAD ecotypes (NEs), we applied the Gene Set Variation Analysis (GSVA)³¹ algorithm using hallmark gene sets from MSigDB. This analysis revealed a distinct pathway enrichment across the 12 identified ecotypes. Ecotypes E7, E9, E11, and E12 showed consistent enrichment in proliferation and immune-related pathways, including MYC-TARGETS-V1, MYC-TARGETS-V2, and INTERFERON-GAMMA-RESPONSE. In contrast, E3, E5, and E10 were enriched in metabolic and signaling pathways such as NOTCH-SIGNALING, TGF-BETA-SIGNALING, and GLYCOLYSIS. Ecotypes E1, E2, E4, E6, and E8 were enriched in the developmental and immune pathways, including INFLAMMATORY-RESPONSE, EPITHELIAL-MESENCHYMAL-TRANSITION (EMT), and ANGIOGENESIS (Supplementary Data S5). Notably, E10 displayed strong enrichment in TGF-BETA-SIGNALING and EMT, while E12 was enriched in immune and cell cycle pathways (Fig. 2a, Supplementary Data S5). Cellular composition analysis indicated that E10 was predominantly composed of epithelial malignant cells, fibroblasts, and CD8+ T cells, whereas E12 consisted of CD4+ T cells, CD8+ T cells, and NK cells (Supplementary Fig. 2a). Dimensionality reduction via Uniform Manifold Approximation and Projection (UMAP) showed overlapping distributions of E10 and E12, with E10 encircling E12 (Fig. 2b). However, t-distributed Stochastic Neighbor Embedding (t-SNE) successfully separated the ecotypes (Fig. 2b). Differential expression analysis identified the key genes that distinguished between E10 and E12. In E10, the tumor differentiation genes S100A8 and SCGB3A2 were upregulated³², along with the downregulation of the IFN-γ-related gene SCGB3A1³³. At E12, SH2D1A³⁴, linked to cell growth and metastasis, and TIGIT ³⁵, associated with T cell suppression and immune evasion, were prominent (Supplementary Fig. S2B).

Fig. 2. Identified functional differences of NSCLC LUAD Ecotypes (E1–E12) and discovered E10 and E12 TME.

a Enrichment analysis of E1-E12 ecotypes with the Hallmark pathway using GSVA. b UMAP/t-SNE projection of E10 and E12 in the discovery dataset. c Top 10 GSEA result of E10 and E12 with Hallmark pathway. d UMAP of E10 according to cell_type_predicted. e The number of circle plots of cell–cell interaction pathways in E10 using CellChat. f Pseudotime distribution of E10 cell types using Monocle. g NicheNet result of TGF-β signaling gene set: results are shown for the Fibroblast ligands best predicting the TGF-β signaling gene set. h Top 10 L-R pair CCC score calculated the feature importance scores to predict ligand activity using Nichenet. i Pseudotime of EMT and TGF-β signaling using monocle.

To further dissect the molecular differences, we performed Gene Set Enrichment Analysis (GSEA)³⁶ using MSigDB³⁷ Hallmark gene sets, identifying the top 10 enriched pathways. On E10, MYC-TARGETS-V1, TGF-BETA-SIGNALING, and EMT were highly enriched, whereas E12 was enriched in MYC-TARGETS-V1, INTERFERON-ALPHA-RESPONSE, and INTERFERON-GAMMA-RESPONSE (Fig. 2c). These findings align with the GSVA results, confirming that TGF-BETA-SIGNALING and EMT are dominant in E10 and interferon-related pathways in E12. The statistical significance of these pathways was assessed using irGSEA³⁸, which calculated the p-values (Supplementary Fig. S2c, Supplementary Data S6). The TGF-BETA-SIGNALING pathway emerged as the most significantly enriched pathway in E10, prompting further focus on its role. UMAP visualization of E10 by cell type revealed clear separation of epithelial malignant cells and fibroblasts, though T cell CD8 formed a smaller, less distinct cluster (Fig. 2d). GSVA³¹ analysis of E10 cell types showed that epithelial malignant cells were enriched in TGF-BETA-SIGNALING and HEDGEHOG-SIGNALING, fibroblasts in EMT and NOTCH-SIGNALING, and T cell CD8 in E2F-TARGETS and G2M-CHECKPOINT. TGF-BETA-SIGNALING is a shared feature between epithelial malignant cells and fibroblasts, whereas immune pathways (INTERFERON-ALPHA-RESPONSE, INTERFERON-GAMMA-RESPONSE, and INFLAMMATORY-RESPONSE) were enriched in both fibroblasts and T cell CD8 (Supplementary Fig. S2d). Cell–cell interactions within E10, analyzed using CellChat³⁹, demonstrated strong signaling between epithelial malignant cells (S06) and fibroblasts (S08) as well as between fibroblasts and T cell CD8 (S01) (Fig. 2e). Temporal dynamics within E10 were assessed using pseudotime analysis using the Monocle⁴⁰ algorithm. Epithelial malignant cells span the entire trajectory, indicating sustained involvement in tumor progression. Fibroblast presence peaked at intermediate stages, suggesting their role in differentiation, potentially as cancer-associated fibroblasts (CAFs), while T cell CD8 were enriched at later stages, possibly reflecting delayed immune activation or exhaustion (Fig. 2f).

To identify the drivers of E10, NicheNet⁴¹ analysis predicted ligand-target interactions, designating epithelial malignant cells as receivers and fibroblasts, and CD8+ T cells as senders. TGFB1, which is highly expressed in fibroblasts, was ranked as the top ligand influencing the TGF-β gene set (Fig. 2g). The TGFB1-SERPINE1 pair exhibited a strong interaction, as quantified by CCC scores (Fig. 2h and Supplementary Data S7). Fibroblast-derived TGFB1, a key product of CAFs⁴², is associated with enhanced cancer cell migration, invasion, and EMT, with SERPINE1 serving as a prognostic marker in NSCLC due to its upregulation in TGF-β-driven EMT^43–47. Further analysis revealed an inverse relationship between TGF-β signaling and EMT in E10 cells, with EMT increasing as TGF-β signaling declined, followed by a dynamic increase in fibroblast-associated TGF-β signaling (Fig. 2i). Pseudotime analysis corroborated this, showing epithelial malignant cells persistently distributed across the trajectory, aligning with a gradual EMT shift, while fibroblast increases at intermediate stages coincided with transient TGF-β downregulation, suggesting an initial TGF-β-independent EMT phase followed by fibroblast-driven TGF-β reactivation (Fig. 2f, i).

Both E10 and E12 contained CD8+ T cells, with prior studies indicating exhaustion signatures (PDCD1, CTLA4, LAG3, and TIGIT) in states S01 and S02. At E10, these exhaustion markers were upregulated in T cell CD8, and Pearson correlation analysis showed a strong association between TGFB1 and most exhaustion markers, except CTLA4 (Supplementary Fig. S2h), supporting the role of CAFs in mediating exhausted T cell CD8.

Validation of E10 ecotype and prediction of ICB response using NSCLC LUAD ICB Cohorts

To validate the E10 ecotype and assess its association with immune checkpoint blockade (ICB) response, we analyzed four independent NSCLC LUAD cohorts treated with ICB: GSE136961, GSE162520, GSE161537, and GSE135222. Each dataset was processed using EcoTyper, which stratified tumors into 12 distinct ecotypes (EC1–EC12). We focused on E10 and E12, representing immunosuppressive and immune-active states, respectively.

Progression-free survival (PFS) analysis revealed that patients assigned to the E10 ecotype exhibited significantly poorer outcomes following ICB therapy compared to those in E12 (Fig. 3a). To understand the immunological basis for this disparity, we evaluated TIDE-based immune exclusion and dysfunction scores. E10 showed markedly higher exclusion scores and lower dysfunction scores, with most classified as non-responders to ICB. In contrast, E12 samples displayed lower exclusion, preserved dysfunction scores, and a higher likelihood of ICB benefit (Fig. 3b).

Fig. 3. Validation and prediction of ICB response using independent datasets.

a Kaplan–Meier plot showing progression-free survival (PFS) of patients stratified by Ecotype (E10 vs E12) across ICB-treated NSCLC datasets. b Distribution of TIDE dysfunction and exclusion scores in E10 and E12 samples. Red bars indicate predicted non-responders. c Correlation between TIDE-derived CAF scores and expression of immune checkpoint genes (PDCD1, LAG3, CTLA4, TIGIT) across ICB-treated samples. d Heatmap of hallmark pathway enrichment (GSVA) in E10 vs E12 groups, highlighting TGF-β signaling and immune suppression in E10.

Given the association between TGFB1⁺ fibroblasts and CD8⁺ T cell exhaustion observed in E10, we investigated whether fibroblast activity contributes to immunosuppression in the ICB context. We calculated TIDE-derived cancer-associated fibroblast (CAF) scores and performed correlation analysis with the expression of exhaustion-related immune checkpoint genes (PDCD1, LAG3, CTLA4, and TIGIT). Notably, LAG3 and TIGIT expression levels were strongly positively correlated with CAF scores, suggesting that fibroblast-mediated immunosuppression may drive T cell exhaustion and contribute to ICB resistance in the E10 ecotype (Fig. 3c).

To further characterize the transcriptional programs associated with each ecotype, we performed GSVA using Hallmark gene sets. As shown in Fig. 3d, E10 samples exhibited enrichment of TGF-β signaling, MYC targets-v1, MYC targets-v2, G2M checkpoint, and PI3K-AKT-mTOR signaling, indicating a tumor-promoting and immunosuppressive transcriptional profile. These pathways are known to contribute to tumor proliferation, immune exclusion, and checkpoint blockade resistance, aligning with the poor ICB response observed in E10. In contrast, E12 samples were enriched for interferon-α and -γ response, IL2–STAT5 signaling, and allograft rejection, consistent with a more inflamed and immunoresponsive tumor microenvironment.

Spatial transcriptome analysis of NEs

To explore the spatial organization of NSCLC LUAD ecotypes (NEs) and address their biological implications, we analyzed spatial transcriptomic (ST) data using cell2location, a Bayesian approach that integrates scRNA-seq with ST data, to estimate cell type abundance. This method employs negative binomial regression to model cell type-specific expression profiles, adjusting for batch effects and leveraging prior knowledge for precise quantification. We initially examined four NSCLC LUAD adenocarcinoma in situ (AIS) samples to uncover distinct cellular landscapes and early tumor microenvironment (TME) interactions. Spatial profiling revealed localized immune infiltration in AIS lesions, with NK and mast cells co-localized alongside tumor cells. Elevated TGF-β signaling distinguishes AIS from invasive LUAD, suggesting that AIS represents a transcriptionally unique, noninvasive state preceding immune evasion and tumor advancement. Building on prior scRNA-seq findings that identified ecotypes with heightened TGF-β signaling, increased T cell exhaustion markers (e.g., PDCD1, CTLA4, LAG3, and TIGIT), and strong TGFB1-SERPINE1 co-expression, we applied cell2location to map these ecotypes spatially. This confirmed their distribution and co-localization patterns in the ST data, with AIS samples exhibiting enhanced TGF-β signaling and a greater presence of exhausted T cells, aligning with early immune suppression. These observations are consistent with reports of progressive CD8 T-cell exhaustion in LUAD.

To refine our analysis, we used cell2location to deconvolute Visium ST data, estimating cell population abundance at each spatial location based on mRNA counts and scRNA-seq-derived reference signatures trained via negative binomial regression. Focusing on E10 and E12, we mapped their spatial fractions across four NSCLC LUAD samples: TSU-19_a-1 (Fig. 4a), TSU-19_b-1 (Fig. 4d), TSU-19_b-2 (Fig. 4g), and TSU-24 (Fig. 4j). E10 regions (orange) were enriched with epithelial malignant cells, fibroblasts, and exhausted T cell CD8, in contrast to the E12 regions (green). Co-localization analysis revealed elevated TGFB1-SERPINE1 expression (Fig. 4b, e, h, k) and increased co-expression of exhaustion markers PDCD1, CTLA4, LAG3, and TIGIT (Fig. 4b, e, h, k) in E10 across all samples, indicating a conserved mechanism of immune suppression.

Fig. 4. Deconvoluted E10 and E12 using cell2location with scRNA-reference.

a TSU-19_a-1 deconvolution result (left) and Visium clustering (right). b TSU-19_b-1 deconvolution result (left) and Visium clustering (right). c TSU-19_b-2 deconvolution result (left) and Visium clustering (right). d TSU-24 deconvolution result (left) and Visium clustering (right).

Hallmark scores for “Activating Invasion and Metastasis” and “Avoiding Immune Destruction” varied by sample. In TSU-19_a-1, these scores were prominent in clusters 6, 8, and 0 (Fig. 4c), suggesting TGF-β-driven progression along a pseudotime trajectory from early clusters (6, 8) to cluster 5. In TSU-19_b-1, enrichment occurred in clusters 2, 4, and 6 (Fig. 4c), indicating an advancement from clusters 2 and 4 toward cluster 6. For TSU-19_b-2, the scores peaked in clusters 0 and 6, with progression from cluster 6 to cluster 0. For TSU-24, clusters 2 and 7 showed high scores, with a trajectory from cluster 0 to cluster 6. These spatial patterns underscore the role of TGF-β signaling in mediating tumor progression and immune evasion across distinct ecotype regions.

Methods

LUAD Cohorts and data compilation

This study employed single-cell RNA sequencing (scRNA-seq) as the primary discovery tool, complemented by bulk RNA-seq, and spatial transcriptomic data for validation. The scRNA-seq data were sourced from the Salcher et al. study¹⁴, from which we selected 12 LUAD-specific datasets out of the 29 available NSCLC datasets. For validation using bulk RNA-seq, we utilized five datasets derived from a total of 14 GEO datasets in conjunction with the LUAD TCGA dataset, specifically GSE11969, GSE13213, GSE19188, GSE30219, and GSE37745. We utilized publicly available spatial transcriptomics data from Takano et al.¹⁵, which profiled LUAD samples across three histological stages: adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), and invasive adenocarcinoma (IA). Among the 13 available samples, we selected four representative sections (TSU-19_a-1, TSU-19_b-1, TSU-19_b-2, and TSU-24), all classified as AIS, based on tissue quality, sequencing depth, and availability of histological annotations. Our analysis focused on these early-stage lesions to explore fibroblast–immune–epithelial cell interactions at the spatial level. The details of the datasets are provided in Supplementary Data S1 and Supplementary Data S2.

Survival analysis and Log-rank test

Kaplan–Meier survival curves were generated using the ‘Surv()’ and ‘survfit()’ functions from the R packages survival (v3.7.0) and survminer (v0.5.0). To identify ecotype subtypes that were significantly associated with clinical outcomes, we performed log-rank tests across all ecotype pairs using bulk RNA-seq data and selected the pair with the smallest p-value, as detailed in Supplementary Data S3. The resulting survival analyses were visualized using the ggsurvplot function.

Ecotype discovery

EcoTyper⁷, a machine-learning-based framework, integrates bulk RNA-seq, single-cell RNA-seq (scRNA-seq), and spatial transcriptomic data to identify distinct cell states and multicellular communities. Subsequent analysis of co-occurrence patterns within these communities revealed clinically relevant multicellular ecotypes, some of which were linked to normal tissue, early tumor development, or poor survival outcomes. We applied EcoTyper to our scRNA-seq dataset following the protocols outlined in Tutorial 5: De Novo Discovery of Cell States and Ecotypes in scRNA-seq data (https://github.com/digitalcytometry/ecotyper). To determine the optimal number of clusters (k) for each cell type, we assessed the cophenetic correlation coefficients across a range of cutoffs (0.75–0.90, in increments of 0.01). A threshold of 0.98 was selected based on rank plots, where k was automatically determined within a default range of 2–20, as most cell types exhibited a steep decline in cophenetic correlation below this value (Supplementary Fig. S1a). Specifically, cell types, including alveolar type 1 cells, B cells, conventional dendritic cells (cDCs), endothelial cells, fibroblasts, mast cells, plasma cells, and CD4+ T cells, showed a pronounced drop at this threshold.

Ecotype recovery

EcoTyper⁷ provides preloaded resources to facilitate reference-guided recovery of previously defined cell states and ecotypes from user-provided bulk expression data. We utilized this functionality to recover NSCLC LUAD ecotypes from five bulk RNA-seq datasets (GSE11969, GSE13213, GSE19188, GSE30219, and GSE37745) following the protocols outlined in Tutorial 1: Recovery of cell states and ecosystems in user-provided bulk data (available at https://github.com/digitalcytometry/ecotyper). All analyses were performed using the default parameters.

Spatial mapping of Ecotype and co-localization with cell2location

To examine the spatial distribution of cell types, we used cell2location⁴⁸, a Bayesian tool that integrates scRNA-seq and spatial transcriptomic data, for high-resolution cellular mapping. Ecotypes identified from scRNA-seq were projected onto Visium spatial transcriptomic data to assess their spatial organization. We developed a custom function, plot_genes_by_ecotype_with_colocalization, adapted from the cell2location plot_genes_per_cell_type function to evaluate gene co-localization. This function calculates a colocalization score for gene pairs (e.g., TGFB1 and SERPINE1) by multiplying their normalized expression values at each spatial spot. Pearson correlation coefficients were computed for gene pairs with non-zero expressions to quantify co-expression patterns. Additionally, a plot_Tcell_colocalization function was created to analyze the spatial co-occurrence of immune-related genes (PDCD1, CTLA4, LAG3, and TIGIT) within the TME using Visium data.

Spatial hallmark score

We utilized spatial transcriptomics to dissect the tumor ecosystem by integrating a single-cell RNA-seq (scRNA-seq) reference dataset, allowing us to spatially align Ecotyper-derived cellular states, previously defined from scRNA-seq, with Visium spatial transcriptomic data. This approach facilitated the characterization and exploration of tumor architecture, highlighting the spatial interactions between cancer cells and tumor microenvironment (TME) phenotypes, and revealed intricate ecosystem interactions. Specifically, we assessed the “Activating Invasion and Metastasis” hallmark to infer activity of the TGF-β signaling pathway and the “Avoiding Immune Destruction” hallmark to examine T cell exhaustion⁴⁹.

Spatial trajectory analysis

Dimensionality reduction and clustering were performed using Seurat (v5.1.0), while spatial transcriptomic trajectories were analyzed with Monocle (v1.3.7)⁴⁰. Trajectories were constructed on a UMAP plot using Visium data, and cluster sequences were manually defined by combining Monocle outputs with spatial information. The SPATA2 package (v3.1.4)⁵⁰ createTrajectoryManually function was used to connect cluster centroids and overlay trajectories on spatial plots. The TME scores along these paths were extracted, smoothed, and analyzed for inflection points. Slopes exceeding a predefined threshold were identified as TME transition zones.

Cell–cell interaction network

To model intercellular signaling, we utilized CellChat (v1.6.1)³⁹. Overexpressed genes and interactions were detected using identifyOverExpressedGenes and identifyOverExpressedInteractions. Communication probabilities were estimated using computeCommunProb, filterCommunication, and computeCommunProbPathway, and results were summarized with aggregateNet. Circular visualizations were generated using netVisual_circle. Additionally, NicheNet (v 2.0.4)⁴¹ was used to predict ligand–target interactions by integrating expression data with signaling and regulatory knowledge. We prioritized TGF-β signaling ligands based on their ability to predict the TGF-β gene set, focusing on the key drivers of cellular communication.

Gene set enrichment analysis

Pathway enrichment was assessed using multiple approaches based on the hallmark gene sets from MSigDB³⁷. Gene Set Variation Analysis (GSVA) (v2.1.1)³¹ was applied as a non-parametric method to evaluate pathway activity across ecotypes. Gene set enrichment analysis (GSEA) (v 4.2.3)³⁶ provided a knowledge-based comparison between ecotypes E10 and E12, while irGSEA (v3.3.2)³⁸ was used to identify statistically significant pathways via single-cell rank-based enrichment.

Statistical analysis

All analyses were conducted in R (v4.3.2), with log-rank tests and Pearson’s correlation coefficients applied as needed. The cell2location analyses were performed in Python (v3.9.20). The experiments were not randomized.

Reporting summary

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

Discussion

Among the 12 LUAD ecotypes identified in our study, EC10 consistently emerged as the most clinically relevant ecosystem, defined by malignant epithelial cells, fibroblasts, and CD8⁺ T cells. EC10 was characterized by a fibroblast-driven TGFB1–SERPINE1 signaling axis that promoted CD8⁺ T cell exhaustion, thereby establishing an immunosuppressive ecosystem associated with poor survival and resistance to immune checkpoint blockade (ICB). Placing EC10 at the center of our analysis allowed us to highlight a multicellular ecosystem that integrates stromal, immune, and malignant components to explain divergent clinical outcomes in LUAD.

The precision of our data analysis is a cornerstone of this study, achieved through the systematic integration of 12 scRNA-seq datasets (161,382 cells across 141 samples), 16 bulk RNA-seq datasets, and 13 Visium spatial transcriptomic samples. Employing EcoTyper and cell2location, we not only defined NSCLC LUAD ecotypes with high fidelity but also mapped their spatial organization, as evidenced by the detailed co-localization of TGFB1-SERPINE1 and immune exhaustion markers (PDCD1, CTLA4, LAG3, and TIGIT) in E10 regions (Supplementary Fig. S2, Supplementary Fig. S3). This analytical rigor extends beyond traditional single-cell studies by leveraging scRNA-seq-derived signatures to enhance the resolution of bulk and spatial data, thus offering a novel methodological advancement in medical informatics. The conceptualization of tumors as ecosystems, increasingly recognized in the recent literature, is substantiated by the level of granularity and innovation that distinguishes our work, systematically overcoming the limitations of prior efforts focused on isolated cell types or modest cohorts.

Our findings align with and extend existing research, notably linking TGF-β signaling to T cell CD8 exhaustion in E10, consistent with Deng et al.‘s observations of immune microenvironment dynamics in early-stage LUAD⁵¹, and Castiglioni et al.’s suggestion that TGF-β signaling contributes to the maintenance of exhausted CD8 T cells in tumors⁵². However, spatial mapping of these interactions in NSCLC LUAD, supported by state-specific expression profiles (Supplementary Fig. S1b), introduces a novel dimension to these insights. Furthermore, the distinction of TGF-β signaling between adenocarcinoma in situ (AIS) and invasive LUAD, derived from precise spatial transcriptomic analysis of four AIS samples, provides fresh perspectives on early tumor evolution, underscoring the transformative potential of our information-driven approach¹⁵.

The clinical and scientific implications of this study are substantial and amplified by the precision and innovation of our analytical framework. The identification of the TGFB1–SERPINE1 pair as a key mediator in E10, validated across multiple datasets with robust statistical significance (e.g., p = 0.00011 for E10 vs. E12 survival separation in GSE13213), suggests that it is a promising therapeutic target that is potentially responsive to TGF-β inhibitors (e.g., galunisertib) or SERPINE1-directed therapies. The association of T cell CD8 exhaustion with immune checkpoint inhibitor (ICI) resistance in E10 suggests that combining TGF-β blockade with ICIs could enhance treatment efficacy, a hypothesis grounded in our spatially resolved data and ripe for clinical testing. Moreover, the spatial delineation of the E10 and E12 ecotypes, underpinned by high-resolution informatics analysis, offers a blueprint for biopsy-based molecular profiling, enabling personalized prognosis prediction and treatment stratification in LUAD. This aligns with the growing demand for ecotype-informed diagnostic tools to tackle tumor heterogeneity, showcasing the translational potential of our data-driven methodology.

Despite these strengths, our study has several limitations. The scRNA-seq and spatial transcriptomic data, sourced from Salcher et al.¹⁴ and JGAS000613¹⁵, respectively, were constrained by cohort size and diversity, although mitigated by the analytical precision of our multi-omics integration. The inferred causal link between TGFB1-SERPINE1 interactions and T cell exhaustion relies on sophisticated bioinformatic predictions yet lacks experimental validation, a gap that future studies must address. The analysis of AIS versus invasive LUAD based on four samples requires confirmation in larger cohorts despite the robust spatial insights provided. Additionally, our survival analyses, which were statistically significant across multiple datasets, were retrospective, highlighting the need for prospective clinical studies to validate these associations.

Future research should prioritize experimental validation of the TGF-β-mediated mechanisms identified herein, such as through TGFB1 inhibition in LUAD cell lines or patient-derived tissues, to assess impacts on EMT and immune responses. Expanding ecotype profiling to larger cohorts and correlating these with ICI outcomes, leveraging our informatics framework, would further enhance its clinical relevance. Extending spatial transcriptomic analyses to other NSCLC subtypes (e.g., squamous cell carcinoma) or metastatic LUAD could broaden ecotype diversity insights and capitalize on the scalability of our approach. Preclinical screening for TGFB1–SERPINE1-targeted therapies, potentially combined with ICIs, could translate these findings into clinical applications, building on the precision and innovation of our data analysis.

In conclusion, our study presents a comprehensive atlas of LUAD ecotypes and identifies EC10 as a fibroblast-driven, TGF-β–enriched immunosuppressive ecosystem associated with ICB resistance. By bridging multi-omics data with advanced informatics tools, we provide a refined perspective on LUAD heterogeneity and highlight the promise of ecotype-based patient stratification and stromal-targeted therapeutic strategies for precision oncology.

These findings suggest that TGF-β–enriched fibroblast niches orchestrate immune exclusion and CD8⁺ T cell dysfunction, ultimately contributing to resistance to immune checkpoint therapy in LUAD. By integrating spatial and bulk transcriptomic analyses with single-cell reference ecosystems, this study provides a precision immunogenomic framework for stratifying patients and rationally combining stromal-targeted therapies with ICB. Our ecotype-based profiling strategy offers a promising platform for clinical translation in tumor immunotherapy.

To directly validate these ecotype classifications, we applied EcoTyper to 12 LUAD single-cell RNA-seq datasets, encompassing over 160,000 cells. This single-cell–based analysis confirmed the recovery of key ecotypes, including E10 and E12, with conserved transcriptional profiles and cellular compositions. These findings reinforce the biological validity of the ecotypes and demonstrate their robustness across data modalities.

Author contributions

Conceptualization, S.-Y.L.; data curation, S.-Y.L.; formal analysis, S.-Y.L.; supervision, S.-Y.L. and Y.-J.L.; Writing—original draft preparation, S.-Y.L.; Writing—review & editing, Y.-J.L.

Peer review

Peer review information

Communications Biology thanks Zhiyuan Yuan and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Kaliya Georgieva. A peer review file is available.

Data availability

Bulk RNA-seq data are accessible via the TCGA portal (http://gdac.broadinstitute.org/) and GEO (accession codes GSE101929, GSE11117, GSE11969, GSE13213, GSE14814, GSE161537, GSE16250, GSE19188, GSE29013, GSE30219, GSE3141, GSE37745, GSE42127, GSE50081, and GSE81089). scRNA-seq data from Salcher et al. ¹⁴ are available at 10.5281/zenodo.6411867. Visium spatial transcriptomic data were retrieved from JGAS000613. LUAD cohorts treated with ICB: GSE136961, GSE162520, GSE161537, and GSE135222. TIDE analysis for predicting immunotherapy response. To assess the predicted response to immune checkpoint blockade (ICB), we applied the TIDE (Tumor Immune Dysfunction and Exclusion) algorithm to bulk RNA-seq data from four LUAD ICB-treated cohorts (GSE136961, GSE162520, GSE161537, and GSE135222). TIDE estimates ICB responsiveness by modeling two primary mechanisms of tumor immune evasion. The algorithm computes a composite TIDE score, in which higher values indicate a greater likelihood of immune evasion and poor response to ICB therapy. We further utilized TIDE-derived metrics, including CTL dysfunction scores, exclusion scores, and CAF (cancer-associated fibroblast) scores, and correlated them with ecotype classification (E10 vs E12) and immune checkpoint gene expression (e.g., PDCD1, LAG3, CTLA4, TIGIT) to characterize the immunological phenotype of each ecotype.

Competing interests

Y.J.L. has received consulting fees from AstraZeneca, Roche, Merck, Yuhan, Bayer, and Amgen. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-025-09087-4.