Targeting ADGRG1 in cancer-associated adipocytes reprograms the immunosuppressive microenvironment and enhances immunotherapy responsiveness in triple-negative breast cancer models

Abstract

Background

The adipocyte-rich microenvironment of the breast plays a crucial role in tumor progression, yet the functional heterogeneity of cancer-associated adipocytes (CAAs) and their specific mechanisms in driving immunosuppressive microenvironment and therapy resistance remain inadequately characterized.

Methods

We performed an integrated analysis of single-nucleus RNA sequencing (snRNA-seq) data from human adipose tissue and bulk transcriptomic data from the TCGA-BRCA and METABRIC cohorts. A prognostic signature was developed based on transcriptional features of CAA subpopulations. Functional roles of the key gene ADGRG1 were investigated through in vitro co-culture systems and in vivo breast cancer models using genetic knockdown/overexpression and pharmacological inhibition.

Results

snRNA-seq revealed a distinct CAA subpopulation enriched in aggressive contexts, labeled as HAD7. The HAD7-based signature stratified patients into high- and low-score groups, with the high-score group exhibiting poor survival and an immunosuppressive microenvironment. ADGRG1 was identified as a core driver. In vitro, ADGRG1 in CAAs promoted cancer cell migration, invasion, and epithelial mesenchymal transition. In vivo, adipocyte-specific ADGRG1 enhanced tumor growth, metastasis, and impaired anti-tumor immunity. Mechanistically, ADGRG1 activated a CXCL8-AKT/mTOR signaling axis. Crucially, pharmacological inhibition of ADGRG1 synergized with anti-PD-1 therapy, effectively suppressing tumor growth and metastasis by reversing immunosuppressive microenvironment.

Conclusions

Our study identifies ADGRG1 as a master regulator of a pro-tumorigenic CAA subpopulation. Targeting the ADGRG1-CXCL8-AKT/mTOR axis represents a promising stromal-targeting strategy to overcome immunotherapy resistance in breast cancer.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12916-026-05049-7.

Background

Breast cancer (BC) persists as a formidable global health challenge, ranking as the most commonly diagnosed cancer worldwide. Statistical projections indicate a concerning rise in its incidence, underscoring the ongoing urgency to address this disease [1, 2]. The integration of immune checkpoint inhibitors (ICIs), specifically anti-PD-1 therapy, into the treatment paradigm for early-stage Triple-Negative Breast Cancer (TNBC) has demonstrated improved pathological responses and event-free survival outcomes in high-risk patients, as established by the landmark KEYNOTE-522 trial [3]. However, significant challenges persist, including heterogeneous patient responses, the lack of universally reliable predictive biomarkers, and the critical need to identify novel cellular targets within the tumor microenvironment to broaden therapeutic efficacy and overcome mechanisms of resistance. This is largely attributable to profound tumor heterogeneity and the development of therapy resistance, highlighting the critical need to decipher the complex biology of the tumor microenvironment (TME), which is now recognized as a key orchestrator of cancer progression and treatment failure [4].

The breast TME represents a highly organized ecosystem comprising various cellular components, including immune cells, fibroblasts, endothelial cells, and adipocytes. As the most abundant stromal element in breast tissue, adipocytes play a particularly important role in shaping the TME through their dynamic interactions with cancer cells [5, 6]. When exposed to tumor-derived signals, normal adipocytes undergo significant phenotypic and functional alterations, transforming into cancer-associated adipocytes (CAAs). This transition is characterized by delipidation, acquisition of a fibroblast-like morphology, and profound changes in secretory activity [7, 8].

CAAs actively contribute to tumor progression through multiple interconnected mechanisms. They secrete a diverse array of bioactive molecules including pro-inflammatory cytokines (IL-6, IL-8, TNF-α), chemokines (CCL2, CCL5), adipokines (leptin, adiponectin), and extracellular matrix remodeling enzymes [9, 10]. This altered secretome creates a permissive environment for cancer cell invasion, epithelial-mesenchymal transition, and metastatic dissemination [11, 12]. Importantly, CAAs play a crucial role in modulating immune responses within the TME. Through the release of specific cytokines and expression of immunoregulatory molecules, CAAs promote the recruitment and polarization of immunosuppressive cells such as tumor-associated macrophages and myeloid-derived suppressor cells, while simultaneously inhibiting the function and infiltration of cytotoxic T lymphocytes [13, 14]. This immunomodulatory capacity contributes significantly to the limited efficacy of immune checkpoint inhibitors in breast cancer patients [15–17].

Beyond their immunoregulatory functions, CAAs engage in metabolic crosstalk with cancer cells, providing energy substrates and supporting the bioenergetic demands of rapidly proliferating tumors [18]. The metabolic symbiosis between CAAs and cancer cells, particularly through fatty acid transfer and oxidation, represents another mechanism by which CAAs promote treatment resistance and tumor progression [19]. Recent evidence further suggests that CAAs contribute to the maintenance of cancer stem cell populations, adding another layer of complexity to their pro-tumorigenic activities [20].

Although the importance of CAAs in breast cancer pathogenesis is increasingly recognized, several key aspects remain poorly understood. The heterogeneity within the CAAs population, the molecular drivers of their functional diversification, and their specific contributions to immunotherapy resistance represent significant knowledge gaps [12, 21–23]. Recent advances in single-cell technologies have begun to reveal the cellular complexity of the breast TME, yet the translation of these findings into clinically relevant stratification systems and therapeutic strategies remains limited [24].

In this study, we employed an integrative approach combining single nuclear RNA sequencing with bulk transcriptomic data from multiple clinical cohorts to comprehensively characterize CAAs heterogeneity in breast cancer. We identified a distinct pro-inflammatory CAA subpopulation associated with aggressive disease features and poor clinical outcomes. Through subsequent mechanistic investigations, we revealed that ADGRG1 signaling in CAAs drives CXCL8 secretion, which in turn activates the AKT/mTOR pathway in tumor cells to promote EMT and fosters an immunosuppressive microenvironment amenable to reprogramming. This specific CAAs subset orchestrates an immunosuppressive microenvironment and contributes to therapy resistance. Our findings not only provide new insights into CAAs biology but also identify potential therapeutic targets for overcoming resistance to current treatment modalities in breast cancer patients.

Methods

Data source

The single-nucleus transcriptome profile of human subcutaneous adipose tissue was downloaded from NCBI The Gene Expression Omnibus (GEO) with an accession GSE176171, GSE184687, GSE225700 and European Bioinformatics Institute (EBI) under an accession number E-MTAB-9199 [25–32]. These subcutaneous adipose tissue specimens obtained from abdominal and neck regions of patients undergoing bariatric or plastic surgery. Dataset GSE176171: Subcutaneous adipose tissue samples from the abdominal region; Dataset GSE184687: Subcutaneous adipose tissue samples from the neck region; Dataset GSE225700: Subcutaneous adipose tissue samples from the abdominal region; Dataset E-MTAB-9199: Subcutaneous adipose tissue samples from the neck region.

All donor samples for these reference datasets were obtained from individuals without diagnosed breast cancer, providing a baseline adipocyte atlas. The mRNA expression profile and clinical information of multi-patient BC samples were downloaded from The Cancer Genome Atlas (TCGA) (http://portal.gdc.cancer.gov/) database and cBioPortal (https://www.cbioportal.org/) database.

Quality control of single-nucleus and bulk tissue transcriptomic samples

The inclusion criteria for the single-nucleus transcriptomic samples includes: (1) From the datasets ≥ 4 samples and availability of raw data or gene expression data; (2) Extraction protocol: Droplet-based single-nucleus RNA-sequencing on 10X Genomics Chromium platform.

The single-nucleus RNA-sequencing data were processed and integrated using a standard Seurat workflow. Briefly, low-quality cells and genes were filtered out using the following criteria: cells expressing fewer than 200 or more than 2,500 genes, and cells with a mitochondrial gene percentage exceeding 15% were excluded. Genes detected in fewer than three cells were also removed. Subsequently, the gene expression matrices were normalized and scaled using the ‘ScaleData’ function, focusing on the top 3,000 highly variable genes and regressing out the influence of mitochondrial gene expression. Finally, to harmonize data across different batches or datasets, integration was performed using the ‘FindIntegrationAnchors’ and ‘IntegrateData’ functions.

The bulk tissue transcriptomic samples were selected based on the following criteria: (1) samples not preserved in formalin, (2) availability of complete clinical metadata, and (3) presence of definitive clinical outcome data with associated survival time.

Cell clustering analysis, visualization, and annotation

Cell-clustering and sub-clustering analyzes were performed with the FindClusters function of the Seurat package with proper resolutions. For the re-clustering of each type of cell cluster, cells with ribosome gene ratio higher than 35% were filtered. Uniform manifold approximation and projection (UMAP) was used to display identified cell clusters and sub-clusters. The cell clusters were annotated based on highly expressed genes, unique expressed genes, and reported markers.

Cell-type specificity validation using a public single-cell atlas

To confirm the cell-type specificity of the signature genes at single-cell resolution, a publicly available single-cell RNA sequencing dataset from the Human Breast Cell Atlas was utilized. The UMAP embeddings, encompassing all annotated major cell types in the breast (including adipocytes, fibroblasts, epithelial cells, and immune cells), were analyzed. The expression levels of core marker genes from the signature were projected onto this UMAP space to visually evaluate whether their expression was specifically enriched within adipocyte clusters, rather than in other stromal or parenchymal cell types.

Functional analysis

The Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways and the hallmark gene set enrichment analysis were conducted using the R package clusterProfiler and gene sets downloaded from the R package org.Hs.eg.db and the Molecular Signatures Database (MSigDB) with the aid of R package msigdbr. Differential expression rank order was used for subsequent Gene Set Enrichment Analysis (GSEA), performed using the clusterProfiler package in R. Pseudotime Trajectory Inference: The Monocle2 package was employed to reconstruct the differentiation trajectory and infer the pseudo temporal ordering of adipocyte subpopulations. Gene Set Variation Analysis (GSVA) was further used to assess the activity of specific biological pathways. Pathways with a p-value < 0.05 and false discovery rate (FDR) < 0.05 were considered significant.

CIBERSORTx and evaluation of immune cell infiltration

The subtype-specific expression profile from the snRNA-seq data was used to generate a signature matrix using CIBERSORTx. Default settings were used, with the exception of the minimum expression parameter, which was set to 0.25. This matrix was subsequently applied to deconvolute the TCGA-BRCA and METABRIC bulk RNA-seq data (in TPM format), estimating the relative abundance of each adipocyte subtype in every bulk sample. Two immune infiltration methods, including CIBERSORT and MCPCounter, were utilized to evaluate the immune cells infiltration in each sample.

Definition and validation of the subcluster gene signature

The subcluster gene signature was derived through a multi-step bioinformatic workflow applied to our integrated snRNA-seq atlas of human subcutaneous adipose tissue.

The subcluster gene signature was defined through a multi-step analytical workflow based on the integrated snRNA-seq atlas. First, a differential expression analysis was performed comparing one subcluster to all other adipocyte subclusters combined, using the FindMarkers function in Seurat (Wilcoxon rank-sum test; thresholds: |log2FC| > 0.5, adjusted p-value < 0.05). Subsequently, genes from this filtered candidate pool were ranked by both their average log2 fold-change (specificity) and their mean normalized expression within subcluster (activity level). The top 200 genes based on a composite score derived from these dual rankings were selected for the final signature, with classical cell cycle and common stress-response genes being excluded.

Calculation of the risk score

Prognostic variable selection was performed using univariable Cox regression analysis. Candidate variables identified through this screening were subsequently utilized to construct a series of machine-learning models for the prediction of breast cancer risk and prognosis. The evaluated algorithms included Random Survival Forest, Least Absolute Shrinkage and Selection Operator (LASSO) regression, Support Vector Machine, and eXtreme Gradient Boosting (XGBoost). Model performance was assessed and compared based on the Concordance Index (C-index) and the area under the receiver operating characteristic curve (AUC) to identify the optimal algorithm. Following its selection as the optimal model, the Lasso Cox regression model was employed to generate a risk score for each patient. A 10-fold cross-validation procedure (via the cv. glmnet function) was implemented to determine the optimal penalty parameter, lambda (λ.min). The RS of each patient was calculated by the mathematical formula:

Survival analysis

The optimal cutoff for stratifying samples into high- and low-expression groups was determined based on the TPM value of the target gene using the ‘surv_cutpoint’ function. Survival outcomes between the groups were compared using Kaplan-Meier survival curves, and the log-rank test was applied to assess statistical significance. All prognostic analyzes were performed using the KMsurv, survival, and survminer R packages.

Patients and adipose tissue samples

Human adipocyte samples were obtained from 54 BC patients and 12 healthy controls undergoing breast surgery at the Comprehensive Breast Health Center, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, between September 2022 and September 2025. CAAs and normal adipocytes (NAs) were isolated from each participant. Samples were processed for bulk RNA sequencing, immunohistochemistry (IHC), immunofluorescence (IF), adipocyte isolation, and cell culture as described below. The study protocol was approved by the Ethics Committee of Ruijin Hospital, Shanghai Jiao Tong University School of Medicine.

Bulk RNA sequencing and transcriptomic analysis

Bulk RNA sequencing of mammary adipose tissue samples and tumors of mice was performed by Novogene Co., Ltd. (Shanghai, China). The derived peritumoral tissue from breast cancer patients and normal tissue from healthy controls were processed through standard pipelines for quality control, alignment, and transcript quantification. Differentially expressed genes between comparative groups were identified using the DEGseq package with an MA-plot-based method and a random sampling model, applying thresholds of an absolute fold change ≥ 2 and an adjusted P-value < 0.05. These DEGs were used for functional and pathway enrichment analyzes in the GO and KEGG databases, with significance defined as P < 0.05 and the inclusion of at least two affiliated genes. To characterize the immune microenvironment, the relative infiltration levels of various immune cell types were estimated from the transcriptomic data using the CIBERSORT algorithm. Subsequently, a composite immune score was constructed based on these immune cell fractions and other relevant transcriptional features for inter-group comparisons. To specifically assess the conservation of the pre-defined adipocyte gene signature within the mammary tissue context, its expression was interrogated in these bulk transcriptomes. The overlap between this gene set and the expressed mammary adipose transcriptome (genes with TPM > 1) was analyzed and visualized, confirming the transcriptional activity of the signature’s core program in the target microenvironment.

Isolation, culture, and differentiation of adipocytes

Human adipose tissues were strictly isolated from peritumoral adipose tissue within 2 cm of the tumor margin following the protocol described in reference [33]. Tissues were washed with PBS, minced into fragments, and digested with type II collagenase (Sigma-Aldrich) at 37 °C for 30 min. The digested product was filtered through a 40 μm cell strainer (BD, USA) to obtain the stromal vascular fraction (SVF), which was then cultured in DMEM/F12 medium (Gibco, USA) medium supplemented with 10% FBS (Gibco, USA), 1% penicillin/streptomycin, and 1 mM L-glutamine.

For Mouse 3T3L1 cells which were obtained from the Type Culture Collection of the Chinese Academy of Sciences (Shanghai, China), were routinely maintained in DMEM medium (Gibco, USA) supplemented with 10% FBS (Gibco, USA), 1% penicillin/streptomycin, and 1 mM L-glutamine. Cells were incubated at 37 °C in a humidified atmosphere of 5% CO₂.

To obtain mature human adipocytes, the isolated primary preadipocytes were subjected to adipogenic differentiation. Specifically, cells were treated for 4 days with adipogenic induction medium containing 5 µg/mL insulin (Eli Lilly, USA), 1 µM dexamethasone (Sigma‒Aldrich, USA), and 0.5 mM IBMX (Sigma‒Aldrich, USA). Successful differentiation was morphologically confirmed by Oil Red O staining, which showed characteristic polyhedral morphology and abundant lipid droplet accumulation.

3T3L1 cells were grown to 100% confluence and maintained for an additional 48 h (contact inhibition). To initiate differentiation (designated Day 0), the medium was replaced with DMEM supplemented with 10% FBS, 5 µg/mL insulin (Eli Lilly, USA), 1 µM dexamethasone (Sigma‒Aldrich, USA), and 0.5 mM IBMX (Sigma‒Aldrich, USA). After 48 h (Day 2), medium was replaced with DMEM with 10% FBS and 5 µg/mL insulin. The medium was subsequently changed to fresh DMEM with 10% FBS and 5 µg/mL insulin every 48 h. Differentiation was carried out for up to 10 days, by which time over 90% of cells exhibited a mature adipocyte morphology with abundant lipid droplets.

Wild-type cells in logarithmic growth phase (about 50% confluence) were seeded in a 24-well plate. When approximately 50% confluence was reached, the cells were replaced with medium containing appropriate titers of lentiviral particles and 5–8 µg/mL Polybrene and incubated in an incubator at 37 °C. After 6–8 h, the culture was replaced with fresh complete medium and continued. Forty-eight hours post-transduction, the medium was changed to CM containing 2 µg/mL puromycin (Sigma-Aldrich, P8833) for selection. The selection medium was refreshed every 2–3 days. Puromycin-resistant, transduced cells were maintained under selection for 7 days to establish stable polyclonal populations.

Cell culture

The human breast cancer cell lines MDA-MB-231 and BT-549, along with the mouse breast cancer cell line 4T1, were obtained from the Type Culture Collection of the Chinese Academy of Sciences (Shanghai, China). MDA-MB-231 routinely cultured in high-glucose DMEM (Thermo Fisher Scientific, CA, USA), while 4T1 cells and BT-549 cells were maintained in RPMI-1640 medium (Thermo Fisher Scientific, CA, USA). All media were supplemented with 10% fetal bovine serum (Gibco, Life Technologies) and 1% penicillin/streptomycin. Additionally, the culture medium for BT-549 cells was supplemented with 10 µg/mL Insulin. All cell lines were kept in a humidified incubator at 37 °C with an atmosphere of 5% CO₂.

Co-culture assay

A 6-well Transwell system with 0.4-µm pore polycarbonate membranes (Corning) was applied in the study. CAAs or 3T3L1 cells at a concentration of 1 × 10⁵ cells/well were seeded into the upper chambers and then cocultured with MDA-MB-231, BT-549 or 4T1 cells (1 × 10⁵ cells/well) in the lower chambers.

Plasmids

The GFP vectors were obtained from the DNA library of Shanghai Jiao Tong University School of Medicine. Vector plasmids overexpressing GFP were co-transfected into HEK293T cells with packaging plasmids psPAX2 and pMD2. G to produce lentivirus according to the manufacturer’s instructions. The collected viral supernatants were used to infect 3T3L1, or CAA cells to establish stable GFP-expressing cell lines for subsequent in vitro and in vivo tracking experiments.

Wound healing assay

Cells were seeded in 6-well plates at a density of 4 × 10⁵ cells per well. Upon reaching confluence, a sterile micropipette tip was used to create a linear scratch in the cell monolayer. After treatment, cells were allowed to migrate for 48 h. Images of the wound area were captured at 0 and 48 h using a microscope, and migration distances were quantified with ImageJ software.

Cell migration and invasion assays

Cell migratory and invasive capacities were evaluated using Transwell chambers with 8 μm pore polycarbonate membranes. For invasion assays, membranes were pre-coated with Matrigel. Briefly, 2.5 × 10⁵ cells/mL in serum-free medium were seeded in the upper chamber, while the lower chamber contained medium with 10% FBS as a chemoattractant. After 36 h of incubation, cells that migrated or invaded through the membrane were fixed with 4% paraformaldehyde, stained with crystal violet, and quantified by microscopic counting.

Western blot analysis

Whole-cell lysates were extracted using RIPA buffer containing protease inhibitors. Proteins were separated by SDS-PAGE and transferred to PVDF membranes. The membranes were incubated with primary antibodies at 4 °C overnight, followed by incubation with HRP-conjugated secondary antibodies for 1 h at room temperature. After TBST washes, protein bands were visualized using an enhanced chemiluminescence detection system.

Pharmacological inhibitors and neutralizing antibodies

Dihydromunduletone (DHM) was used to inhibit ADGRG1 signaling. A working concentration of 15 µM was selected for in vitro assays based on its reported half-maximal inhibitory concentration (IC₅₀) of approximately 20.9 µM for ADGRG1. For AKT inhibition, cells were treated with MK-2206 at 10 µmol/L for 72 h.

For cytokine stimulation and neutralization experiments, recombinant human CXCL8 was used at 10 ng/mL for 72 h. For neutralization, CXCL8 (10 ng/mL) was combined with CXCL8 neutralizing antibody (10 µg/mL) for 72 h. In the murine system, recombinant MIP2 was used at 20 ng/mL for 72 h, with MIP2 neutralizing antibody (10 µg/mL) applied under the same conditions. For co-culture experiments, adipocytes were stimulated with Col3 (84nM) for 72 h, The stimulation medium containing Col3 was not removed; instead, cancer cells were directly added and co-cultured. To exclude potential direct effects of residual Col3 on cancer cells, a control experiment was performed in which tumor cells were treated with Col3 alone for 72 h in the absence of adipocytes. Cells were then harvested for Western blot analysis of EMT markers and AKT/mTOR phosphorylation, as well as qPCR analysis of SNAIL, TWIST, and ZEB1 expression.

All treatments were performed in serum-free medium unless otherwise specified. Control cells received equivalent volumes of vehicle (DMSO or PBS) as appropriate. Each experiment was repeated at least three times independently.

Immunohistochemistry and immunofluorescence staining

Formalin-fixed, paraffin-embedded tissue Sects.  (4–5 μm thickness) were deparaffinized and rehydrated. Antigen retrieval was performed followed by incubation with primary antibodies overnight at 4 °C. HRP-conjugated secondary antibodies were applied for 1 h at room temperature. Immunoreactivity was evaluated by two independent pathologists using a semiquantitative H-score system incorporating both staining intensity (0–3) and percentage of positive cells (0–4), with final scores ranging from 0 to 12.

For immunofluorescence, we stained the tissues with primary antibody. Then, the tissues on the coverslips were incubated with goat anti-rabbit or goat IgG secondary antibody for 1 h at room temperature in the dark. DAPI was used to stain the cell nuclei. Subsequently, the stained cells were photographed with a fluorescence microscope. For image analysis, Fiji ImageJ software was employed to quantify IHC staining intensity, particularly for nuclear signals. Statistical analyzes were performed using Prism GraphPad 10, in accordance with our earlier publication [34].

Image acquisition and colocalization analysis

Stained tissue sections were imaged using a confocal or fluorescence microscope. For the analysis of immunofluorescence images, they were processed and quantified using ImageJ/FIJI software (National Institutes of Health). Briefly, images were split into individual channels. The degree of co-localization between these two markers within the tumor stroma was quantitatively assessed by calculating the Pearson’s Correlation Coefficient (PCC). PCC values were determined using the Coloc 2 plugin within ImageJ/FIJI.

In vivo assays

Female wild-type BALB/c mice (5 weeks old) and female nude mice (5 weeks old) were purchased from Shanghai SLAC Co., Ltd. (Shanghai, China). On the day of in vivo experiments, in vitro‑differentiated CAAs (human or mouse) were mixed with corresponding tumor cells at ratios to ensure consistency across experimental groups. All quantitative data are presented as mean ± SEM, and each experimental group consisted of 5 mice.

For the nude mouse model, 1.0 × 10⁷ MDA-MB-231 cells were mixed with 2.5 × 10⁵ human CAAs or CAA-shADGRG1 in 100 µL of sterile PBS, resuspended thoroughly to form a homogeneous cell suspension, and then injected subcutaneously (n = 5 mice per group).

For the immunocompetent BALB/c mouse model, 1.0 × 10⁶ 4T1 cells were mixed with 2.5 × 10⁵ 3T3L1 cells (Ctrl, 3T3L1-shADGRG1, or 3T3L1-OEADGRG1, respectively) in 100 µL of sterile PBS, and the homogeneous suspension was injected into the mammary fat pad (n = 5 mice per group).

To maintain cell viability and injection accuracy, all cell suspensions were kept on ice and injected within one hour after mixing.

For combined therapy experiments, 4T1 cells (1.0 × 10⁶ cells/mouse) were injected into the mammary fat pad of BALB/c mice (n = 5 mice per group). When the tumor size reached 50 mm³, anti-PD-1 antibody (10 µg once) and Dihydromunduletone (40 µg once) were injected every 3 days [35, 36]. For CXCR1/2 inhibition experiments, Reparixin was intraperitoneally injected into nude mice at 5 mg/kg three times weekly (n = 5 mice per group). This experiment was conducted separately from the combined therapy studies.

The tumors of nude mice in the eighth week and BALB/c mice in the fourth week were harvested by surgical resection based on a margin of gross normal tissue. A schematic diagram illustrating the in vivo experiment workflow is provided in Additional file 1: Fig. S1.

The lung tissues and subcutaneous tumors were harvested for RNA sequencing, immunohistochemistry, and hematoxylin and eosin staining.

All animal procedures were approved by the Animal Ethics Committee of Ruijin Hospital and conducted in accordance with the NIH Guide for the Care and Use of Laboratory Animals.

Multiplex cytokine profiling

Concentrations of secreted cytokines in CAA supernatants were measured using the ABplex-100 multi-analyte flow assay system (cat# AI40001, ABclonal) according to the manufacturer’s instructions. The following cytokines were analyzed: IFN-γ, IL-2, IL-4, IL-6, IL-17 A, IL-12P70, IL-10, CXCL8, TNF-α, IL-1β, IL-5, and IFN-α. Briefly, 50 µL of each standard or diluted sample (25 µL sample diluent + 25 µL sample) was added to the reaction plate, followed by 5 µL of thoroughly mixed antibody-conjugated microsphere suspension. The plate was sealed and incubated at 37 °C for 60 min with shaking at 1200 rpm. After incubation, the plate was placed on a magnetic separator for 2 min, and the supernatant was removed by inverting and tapping on absorbent paper. Wells were washed with 100 µL wash buffer, shaken at 1200 rpm for 1 min, and magnetically separated again. Detection antibody solution (50 µL/well) was added, and the plate was incubated at 37 °C for 30 min with shaking at 1200 rpm, followed by magnetic separation and one wash step. Streptavidin-phycoerythrin solution (50 µL/well) was added and incubated at 37 °C for 15 min with shaking at 1200 rpm, protected from light. After magnetic separation and two washes, beads were resuspended in 55 µL wash buffer and analyzed on the ABplex-100 system. Data were acquired and analyzed using the manufacturer’s software, and cytokine concentrations were calculated from standard curves. All samples were measured in duplicate.

RT‒qPCR

Total RNA was isolated using TRIzol reagent (Invitrogen, Carlsbad, CA, USA). For the analysis of mRNAs, reverse transcription was performed using the PrimeScriptTM RT reagent kit (Cat# 4368813, Thermo Fisher Scientific) with random primers. The expression of RNAs was assessed by real-time quantitative polymerase chain reaction (RT-qPCR) in triplicate with SYBR Green Master Mix (Cat# 24759100, Roche) on an Applied Biosystems QuantStudioTM Real-Time PCR System. The primers (got from BioTNT) used in the study are listed in Additional file 1: Table S1. Comparative cycle threshold values (2 − ΔΔCt) were adopted to analyze the results.

GTP-Rho pull-down assay

The activation status of RhoA was assessed using a commercially available Active Rho Pull-down and Detection Kit (Beyotime, China, Cat# P2065), based on the specific affinity of the Rhotekin Rho-binding domain (RBD) for the active, GTP-bound form of RhoA. Differentiated CAAs and murine 3T3L1 were used.

Following starvation, cells were stimulated with 84 nM recombinant human (for CAAs) or murine (for 3T3L1) collagen type III (Col3) for 5 min at 37 °C. Control dishes received an equal volume of the vehicle. Stimulation was terminated by immediate placement on ice and rapid washing with ice-cold PBS.

Cells were then lysed in 1 mL of the provided Mg²⁺-containing lysis buffer supplemented with protease and phosphatase inhibitors. Clarified lysates were incubated with 30 µg of GST-Rhotekin-RBD fusion protein pre-bound to glutathione-sepharose beads for 1 h at 4 °C with gentle rotation. Beads were washed three times with cold lysis buffer to remove non-specifically bound proteins. Bound proteins and a portion of the total cell lysate were eluted in 2× Laemmli buffer, boiled for 5 min.

Statistical analysis

All quantitative data are presented as mean ± SEM. Before performing between-group comparisons, we first performed normality tests (Shapiro-Wilk test) and homogeneity of variance tests (Levene test) on the data from each experimental group. If the data met both normal distribution and homogeneity of variance (p > 0.05), parametric test was used: two-tailed unpaired Student t test was used for comparison between two groups, One-way ANOVA was used for comparison between multiple groups, supplemented by Tukey multiple comparison post hoc test. If the data did not meet the parametric test prerequisites, the corresponding nonparametric tests were used: Mann-Whitney U test was used for two-group comparison, and Kruskal-Wallis test was used for multi-group comparison supplemented by Dunn’s post hoc test. All analyzes were performed using R version 4.1.3. and GraphPad Prism 9.0 software, and the significance level was set at *p < 0.05, **p < 0.01, ***p < 0.005 and ****p < 0.001.

Results

Single-nucleus profiling reveals heterogeneous subpopulations in human adipose tissue

To delineate the cellular heterogeneity within human subcutaneous adipose tissue (hSAT), we performed an integrated analysis of a single-nucleus RNA sequencing (snRNA-seq) dataset. Following quality control and normalization, datasets were harmonized. An initial unannotated clustering of all nuclei confirmed successful batch correction and revealed the foundational transcriptional architecture prior to biological interpretation (Additional file 1: Fig. S2A, B). Following initial quality control and cell-type annotation, adipocytes were extracted for focused sub-clustering analysis. This approach allowed for an in-depth examination of the adipocyte compartment, independent of other stromal and immune cells (Fig. 1A).

Fig. 1.

Single-nucleus landscape and clinical relevance of adipocyte heterogeneity in breast cancer. A UMAP projection of major cell types identified in integrated snRNA-seq data from human subcutaneous and visceral adipose tissue. B Re-clustering of adipocytes only, visualized by UMAP (left) and t-SNE (right). C Heatmap displaying the expression of top marker genes defining each adipocyte subcluster. D GO functional enrichment analysis of differentially expressed genes in adipocyte subclusters. E KEGG functional enrichment analysis of differentially expressed genes in adipocyte subclusters. F Pseudotime trajectory analysis of adipocyte subpopulations. G Stacked bar plots showing the deconvoluted proportions of adipocyte subpopulations in TCGA-BRCA and METABRIC cohorts using CIBERSORTx. H Comparison of clinical characteristics (T stage, N stage, Clinical Stage, and Subtype) between patient groups stratified by the HAD7 signature score in the TCGA-BRCA cohort. Statistical significance was determined by the Kruskal-Wallis test. I Kaplan-Meier curves for OS of patients in the TCGA-BRCA cohort stratified by the HAD7 signature. P-value was determined by the log-rank test. J Kaplan-Meier curves for DSS of patients in the TCGA-BRCA cohort stratified by the HAD7 signature. P-value was determined by the log-rank test. K Immune cell infiltration profiles estimated by CIBERSORT. Statistical significance was determined by the Kruskal-Wallis test. L Kaplan-Meier curves for OS of patients in the TCGA-BRCA cohort stratified PCA score derived from HAD7 signature genes. P-value was determined by the log-rank test. M Kaplan-Meier curves for DSS of patients in the TCGA-BRCA cohort stratified PCA score derived from HAD7 signature genes. P-value was determined by the log-rank test. N Immune cell infiltration between PCA score-high and PCA score-low patient groups estimated by CIBERSORT. Statistical significance was determined by the Kruskal-Wallis test. O GSEA enrichment plots showing significant association of high PCA score with breast oncogenesis, bone metastasis, and hallmark pathways. DSS, Disease-Specific Survival; GO, Gene Ontology; GSEA, Gene Set Enrichment Analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; OS, Overall Survival; PCA, Principal Component Analysis; snRNA-seq, single-nucleus RNA sequencing; t-SNE, t-Distributed Stochastic Neighbor Embedding; UMAP, Uniform Manifold Approximation and Projection. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001; ns, not significant

The adipocyte subset was subjected to re-clustering, which resolved 10 (HAD1-HAD10) distinct subpopulations at an optimized resolution. These subpopulations were annotated as specific adipocyte subtypes based on the expression of established and novel marker genes. The resulting UMAP and t-SNE visualization clearly illustrated the segregation of these subtypes (Fig. 1B). This result demonstrates a previously underappreciated level of heterogeneity within the mature adipocyte compartment.

A heatmap of the top marker genes for each subpopulation not only confirmed their distinct transcriptional identities but also provided a molecular basis for their functional segregation, as these genes were linked to the pathway enrichments described below (Fig. 1C).

Functional enrichment analysis including GO and KEGG assigned specific biological identities to these subpopulations, explaining their transcriptional segregation. Based on pathway similarity and spatial proximity, cohesive functional groups were identified. HAD1 and HAD9 were associated with structural integrity and neurodegenerative pathways (Additional file 1: Fig. S2C, D). HAD2 and HAD3, though both metabolic, were treated as distinct entities due to their transcriptional and spatial separation, suggesting specialized metabolic roles (Additional file 1: Fig. S2E, F). HAD4 and HAD5 were unified by signatures of cellular repair and mitochondrial metabolism, indicative of a stress-responsive state (Fig. 1D, E, Additional file 1: Fig. S2G). HAD6 and HAD10 were linked to immune modulation and autophagy, positioning them as potential immunomodulatory adipocytes (Fig. 1D, E, Additional file 1: Fig. S2H). HAD8 was uniquely enriched for pathways related to neurodegenerative processes and viral response (Additional file 1: Fig. S2I). Critically, HAD7 was defined by enrichment in intracellular signaling and protein transport pathways, suggesting a role in active intercellular communication (Fig. 1D, E).

To model the developmental relationships among these states, pseudotime trajectory analysis was performed. The inferred trajectory presented a continuous path branching into three distinct fates, delineating a potential trifurcation in adipocyte differentiation (Fig. 1F, Additional file 1: Fig. S2J). The scatter plot shows cell distribution along this path, with the principal trajectory outlined (Additional file 1: Fig. S2K). The pseudotime ordering revealed a sequence of state transitions, positioning clusters such as HAD4 and HAD5 earlier in the trajectory, while other subpopulations like HAD7 and the HAD10 group resided later. This structure suggests that distinct metabolic and immunomodulatory phenotypes arise at different stages of a hierarchical differentiation process (Additional file 1: Fig. S2L). This model inferred the existence of progenitor-like states and more terminally differentiated states, providing a dynamic view of adipocyte differentiation and maturation within the tissue context.

Deconvolution of bulk tumors identifies a prognostic adipocyte subpopulation in breast cancer

To translate our single nucleus-based adipocyte taxonomy into clinically relevant insights, we investigated the abundance and impact of these subpopulations in large breast cancer cohorts. We applied the CIBERSORTx algorithm to bulk RNA-seq data from the TCGA-BRCA and METABRIC cohorts, using the expression profile of the defined hSAT adipocyte subtypes as a reference signature matrix. This deconvolution revealed a comprehensive landscape of adipocyte infiltration in breast cancer, demonstrating significant inter-patient heterogeneity in the composition of adipocyte subtypes.

Initial analysis of the deconvoluted proportions identified three adipocyte subpopulations—termed HAD4, HAD6, and HAD7—that were consistently prevalent across both the TCGA and METABRIC datasets (Fig. 1G). To determine which of these subpopulations held the strongest clinical influence, we utilized the differential gene expression signatures unique to HAD4, HAD6, and HAD7 to independently stratify patients within each cohort. For each signature, patients were dichotomized into high- and low-expression groups.

We then conducted a comprehensive comparative analysis of these groups within the TCGA-BRCA discovery cohort. This analysis encompassed the evaluation of adipocyte feature genes, the inferred proportions of other adipocyte subtypes, overall survival, and key clinical characteristics.

Stratification of the TCGA-BRCA cohort by the HAD7 signature score revealed associations with aggressive disease features. For example, patients in the HAD7-high group presented with more advanced pathological T stage and clinical stage, and the signature was enriched in TNBC subtypes (Fig. 1H). In contrast, the HAD4 signature showed no significant association with tumor stage, and the HAD6 signature showed weaker associations (Additional file 1: Fig. S3A). Consequently, only the HAD7-high group exhibited significantly worse Overall Survival (OS) and Disease-Specific Survival (DSS), establishing its independent prognostic value, while stratification by HAD4 or HAD6 did not yield significant survival differences (Fig. 1I, J, Additional file 1: Fig. S3B).

What’s more, the successful linkage of these signatures to adipocyte maturation markers confirmed the biological relevance of the transcriptomic stratifications (Additional file 1: Fig. S3C). Analysis of the tumor immune microenvironment via CIBERSORTx revealed that the HAD7-high group in the TCGA-BRCA cohort was characterized by a distinct immunosuppressive landscape, suggesting a potential mechanism for its poor prognosis (Fig. 1K). In contrast, the immune cell infiltration profiles associated with HAD4 and HAD6 signatures were less pronounced (Additional file 1: Fig. S3D). Notably, an alternative deconvolution method (MCPcounter) indicated comparable levels of general immune infiltration for both HAD4-high and HAD7-high groups (Additional file 1: Fig. S3E). This discrepancy suggests that the primary distinction of the HAD7 subgroup may not be the overall quantity of immune cells, but rather a specific shift towards an immunosuppressive functional phenotype, a nuance more readily captured by cell type-specific tools like CIBERSORTx. Together, these multi-faceted comparisons established HAD7 as the most potent and clinically relevant discriminator. Similar stratification in METABRIC recapitulated a significant association between high HAD7 signature expression and poorer overall survival (Additional file 1: Fig. S3F). Furthermore, the distinct immunosuppressive microenvironment linked to the HAD7-high phenotype was consistently observed in this validation cohort (Additional file 1: Fig. S3G). In contrast, parallel analyzes for the HAD4 and HAD6 signatures failed to yield significant, consistent associations with survival or immune contexture in either cohort, supporting the conclusion that their clinical influence is substantially weaker than that of HAD7.

While the dichotomization based on median signature expression (HAD7-signal) identified a clear prognostic trend, we sought to construct a more integrated and robust metric that could better capture the core biological program of this subpopulation. We therefore performed principal component analysis (PCA) on the HAD7 signature genes and defined the first principal component as the continuous HAD7-score (Additional file 1: Fig. S3H).

To validate that this PCA-derived score provided a superior metric to the simple HAD7- we performed a direct comparative assessment. The HAD7-score demonstrated enhanced prognostic discrimination in higher Harrell ‘s C-index and yielded a more specific association with an immunosuppressive microenvironment in our cohort, as detailed in the comparative analysis presented in Additional file 1: Table S2. Based on this methodological validation, we adopted the HAD7-score for subsequent analyzes.

Using an optimal cut-off, patients were stratified into HAD7-score-high and -low groups. This refined stratification revealed a better divergence in clinical outcome. Kaplan-Meier analysis showed that patients in the HAD7-score-high group had markedly poorer overall survival (Fig. 1L, M). We next investigated the tumor immune microenvironment underpinning this poor prognosis. Comparative analysis of immune cell infiltration using this optimized score revealed that the HAD7-score-high group was characterized by an immunosuppressive contexture, including reduced cytotoxic T cells and elevated immunosuppressive myeloid cells (Fig. 1N), suggesting a mechanistic link to immune evasion. Furthermore, GSEA enrichment analysis linked the high HAD7-score phenotype to signatures of “Breast Cancer Oncogenesis” and “Bone Metastasis”. Besides, in hallmark enrichment, gene set enrichment analysis revealed that tumors with a high HAD7 score were enriched for hallmark pathways including ‘HALLMARK_INFLAMMATORY_RESPONSE’ and ‘HALLMARK_MTORC1_SIGNALING’. (Fig. 1O).

Collectively, these data nominate the HAD7 adipocyte subpopulation, characterized by a specific signaling and immunomodulatory program, as a key component of the tumor microenvironment. Its abundance is consistently associated with aggressive disease features, immunosuppression, and inferior survival across independent cohorts, distinguishing it from other identified adipocyte subtypes.

A prognostic gene signature derived from the adipocyte subpopulation

Given the strong prognostic association of the HAD7 adipocyte subpopulation, we sought to develop a refined gene signature to quantify its clinical impact. We first performed univariable Cox regression analysis on the top 200 variable genes from the HAD7 signature within the TCGA-BRCA cohort. This initial screening identified a subset of genes significantly associated with overall survival (Fig. 2A). The hazard ratios (HRs) for these genes were predominantly greater than 1, indicating their expression is largely correlated with increased risk. This rigorous univariable screening was a critical first step to distill the extensive HAD7 gene signature into a core set with direct prognostic relevance, moving beyond mere identification to establish a tangible link with patient survival.

Fig. 2.

Construction and validation of a prognostic model based on adipocyte-derived signature genes. A Univariate Cox regression analysis identifying individual genes from the HAD7 signature pool significantly associated with overall survival in the TCGA-BRCA discovery cohort. B Construction of a multi-gene prognostic signature using LASSO Cox regression. Left panel: Coefficient profiles of HAD7 signature genes. Right panel: Ten-fold cross-validation for optimal penalty parameter (λ) selection, identifying the λ value (dotted vertical line) that minimizes the partial likelihood deviance (C) Kaplan-Meier survival analysis comparing overall survival between high-risk and low-risk groups stratified by the LASSO-derived risk score in the TCGA-BRCA. D Time-dependent ROC curves assessing the predictive accuracy of the risk score for overall survival in TCGA-BRCA. E Kaplan-Meier survival analysis comparing overall survival between high-risk and low-risk groups in the independent METABRIC validation cohort. F Time-dependent ROC curves assessing the predictive accuracy of the risk score for r overall survival in METABRIC datasets. G Forest plot displaying multivariate Cox regression analysis evaluating the independent prognostic value of the risk score alongside other clinical parameters. H Nomogram integrating the risk score and key clinical features for predicting overall survival probability. I GO functional enrichment analysis of differentially expressed genes between high- and low-risk groups. J KEGG pathway enrichment analysis of differentially expressed genes between risk groups. K GSEA revealing significant enrichment of immune suppression and therapy resistance-related pathways in the high-risk group. GO, Gene Ontology; GSEA, Gene Set Enrichment Analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; LASSO, Least Absolute Shrinkage and Selection Operator; ROC, Receiver Operating Characteristic. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001; ns, not significant

To construct a robust prognostic model, we employed a comprehensive machine-learning framework. The candidate genes from the univariable analysis were used to train multiple algorithms, including Random Survival Forest, CoxBoost, Support Vector Machine, LASSO regression, and XGBoost (Fig. 2B, Additional file 1: Fig. S3I). Model performance was rigorously evaluated and compared based on the Concordance Index and the area under the receiver operating characteristic curve. The LASSO Cox regression model demonstrated superior and stable predictive performance (Additional file 1: Table S3) and was therefore selected as the optimal model.

Using the coefficients from the LASSO model, a continuous risk score was calculated for each patient in the TCGA cohort. Patients were stratified into high-risk and low-risk groups based on an optimal cut-off value. This stratification revealed a highly significant divergence in survival probability between the two groups (Fig. 2C). The predictive accuracy of the risk score was further quantified by time-dependent ROC analysis, which showed high AUC values for predicting 1-, 3-, and 5-year overall survival in the training set (Fig. 2D). This validates the initial prognostic association, confirming that the HAD7-derived signature captures biologically meaningful risk.

The generalizability of the prognostic model was then validated in an independent cohort (METABRIC). The risk score consistently stratified patients into high- and low-risk groups with significant differences in survival outcomes (Fig. 2E) and shows predictive accuracy which is slightly inferior to the TCGA-BRCA one (Fig. 2F). This supports the model’s clinical applicability.

To facilitate potential clinical translation, we constructed a predictive nomogram by integrating the gene-based risk score with key clinical variables (Fig. 2G). The calibration curves demonstrated good agreement between the nomogram-predicted and observed survival probabilities (Fig. 2H).

To gain biological insights into the high-risk group, we performed functional enrichment analysis. GO and KEGG analyzes indicated that the high-risk phenotype was associated with the activation of multiple cancer-related pathways and biological processes linked to epithelial-mesenchymal transition (EMT) (Fig. 2I-J). Furthermore, GSEA revealed that the high-risk group was significantly enriched in gene signatures associated with resistance to endocrine therapy and anti-PD-1 immunotherapy (Fig. 2K, Additional file 1: Fig. S3J), suggesting that the HAD7-associated high-risk phenotype is driven by aggressive biological processes like EMT and, critically, is linked to therapy resistance pathways, which directly explains the poor survival and points to potential therapeutic vulnerabilities.

ADGRG1 is identified as a key upregulated oncogene in CAA

To establish the relevance of our subcutaneous adipose-derived HAD7 signature within the mammary tumor microenvironment, we first validated its conservation. Assessment of our bulk RNA-sequencing data from peritumoral and normal mammary adipose confirmed that all of the core HAD7 signature genes were expressed in this target tissue context, providing the fundamental transcriptional substrate (Additional file 2: Fig. S4A). Analysis of bulk RNA-sequencing data from peritumoral and normal mammary adipose tissue confirmed robust expression of the core HAD7 gene set. Interrogation of the Human Breast Cell Atlas (HBCA) at single-cell resolution further demonstrated that top HAD7 markers, exhibited specific expression within the adipocyte cluster of the healthy human breast (Fig. 3A), confirming the signature’s adipocyte-specific and cross-depot relevance [37]. This analysis confirms that the HAD7 transcriptional program is not an artifact of a specific depot but is physiologically relevant, as it is maintained in the target mammary adipose tissue, providing the biological context for its proposed role in the TME.

Fig. 3.

Identification and validation of ADGRG1 as a key upregulated gene in CAAs. A UMAP projection showed the expression of core HAD7 signature genes in a healthy human breast single-cell RNA-seq dataset (Human Breast Cell Atlas) (B) Venn diagram illustrating the intersection of differentially expressed genes (DEGs) from three sources: genes associated with the prognostic risk model, DEGs between 7 CAA and 4 normal adipocyte (NA) samples from our RNA sequencing, and DEGs between tumor and normal breast tissue from the TCGA-BRCA cohort. C Left: UMAP plot visualizing the expression of ADGRG1, confirming its specific enrichment in the HAD7 subcluster. Right: Violin plot showing the expression distribution of ADGRG1 across adipocyte subclusters from our snRNA-seq data. Each point represents a single nucleus. Statistical significance was determined by Wilcoxon rank-sum test with Bonferroni correction for multiple comparisons. D Representative multiplex immunofluorescence image of the tumor-adipose interface. The image shows co-staining for ADGRG1 (green), G0S2 (a HAD7 marker, red), FABP4 (pan-adipocyte marker, yellow), Pan-Cytokeratin (tumor cells, cyan), and αSMA (fibroblasts, magenta). Nuclei are counterstained with DAPI (blue). Scale bar, 25 μm. (n = 3) The bar graph (right) quantifies the PCC for pairwise colocalization analysis. Data are presented as mean ± SEM; individual data points (dots) are overlaid on the bars. E Scatter plot demonstrating the positive correlation between the immunofluorescence intensities of ADGRG1 and G0S2 from the IF analysis. The solid line represents the linear regression fit. F Volcano plot of DEGs from the CAA vs. NA RNA-seq analysis. (log2 fold change > 1, adjusted p-value < 0.05). The p-values were adjusted using the Benjamini-Hochberg method. G Representative IHC images of ADGRG1 protein expression in 12 paired NA and CAA samples. Scale bar, 50 μm. Data are presented as mean ± SEM (n = 12 biologically independent pairs). Statistical significance was determined by a paired two-tailed Student’s t-test. H Representative IF staining of ADGRG1 (red) in 12 paired NA and CAA samples. Nuclei are counterstained with DAPI (blue). Scale bar, 50 μm. Data are presented as mean ± SEM (n = 12 biologically independent pairs). Statistical significance was determined by a paired two-tailed Student’s t-test. I Western blot analysis of ADGRG1 protein levels in 12 paired NA and CAA lysates. The bar graph (right) shows the densitometric quantification of ADGRG1 protein levels normalized to GAPDH. Data are presented as mean ± SEM (n = 12 biologically independent patient pairs, each measured in 4 technical replicates across independent experimental runs, resulting in 48 total data points). Statistical significance was determined by a paired two-tailed Student’s t-test. J Representative IHC images of ADGRG1 staining in a larger cohort of 54 CAA samples, showing varying expression intensities. Scale bar, 50 μm. K Analysis of the association between ADGRG1 IHC score and key clinicopathological features in 54 breast cancer patients. For the tumor grade and molecular subtype panels, statistical significance was determined by one-way ANOVA with Tukey’s post-hoc test for multiple comparisons. The p-values for association with tumor stage and Ki-67 index were calculated using the Kruskal-Wallis test and Spearman’s correlation, respectively. CAA, Cancer-Associated Adipocyte; DAPI, 4’,6-Diamidino-2-Phenylindole; DEGs, Differentially Expressed Genes; IF, Immunofluorescence; IHC, Immunohistochemistry; NA, Normal Adipocyte; PCC, Pearson’s Correlation Coefficient; RNA-seq, RNA sequencing. *All quantitative data are presented as mean ± SEM from at least three independent experiments. Normality was assessed using the Shapiro-Wilk test before applying parametric tests. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001; ns, not significant

To bridge our computational risk model with actionable biology, a multi-dataset integration was performed to identify candidate effector genes. Differential expression analysis between RNA-seq profiles of freshly isolated cancer-associated adipocytes (CAAs, n = 7) and normal adipocytes (NAs, n = 4) generated a candidate list. Intersection of this list with prognostic genes from the machine-learning model and with differentially expressed genes from the TCGA-BRCA cohort revealed ADGRG1 as the sole gene common to all three datasets (Fig. 3B). This convergence nominated ADGRG1 as a high-priority candidate linking the HAD7-associated risk signature to CAA function.

Having established ADGRG1 as an HAD7 marker, its spatial distribution was investigated to infer functional relevance. Re-analysis of our single-nucleus data localized ADGRG1 expression predominantly to the adipocytes and HAD7 subcluster, as visualized by UMAP and quantified by violin plot (Fig. 3C, Additional file 2: Fig. S4B). Multiplex immunofluorescence was employed, co-staining for ADGRG1, the established HAD7-specific marker G0S2 (Additional file 2: Fig. S4C-D), the pan-adipocyte marker FABP4, tumor cells (Pan-Cytokeratin), and fibroblasts (αSMA). We found that ADGRG1 and G0S2 showed negligible co-localization with markers for tumor cells and cancer-associated fibroblasts. This absence of signal overlap confirms that the identified ADGRG1+/G0S2 + cells are distinct from these other major stromal and epithelial components within the TME. Both ADGRG1 and G0S2 demonstrated high correlation coefficients with FABP4, confirming their expression within the adipocyte lineage (Fig. 3D). ADGRG1 and G0S2 exhibited a positive correlation, with the corresponding scatter plot demonstrating a clear linear relationship between their fluorescence intensities (Fig. 3E). This high-degree co-localization provides evidence that ADGRG1 and G0S2 are co-expressed within the same adipocyte population, thereby validating their combined use as specific identifiers for the HAD7 subpopulation in situ.

To validate its dysregulation, ADGRG1 expression was first assessed in our primary samples. RNA-seq confirmed significant transcriptional upregulation in CAAs versus Nas (Fig. 3F). To confirm this specificity at the protein level, immunohistochemistry and immunofluorescence were done to define its subcellular localization, revealing predominant membrane-associated expression in CAAs, with minimal detection in NAs (Fig. 3G, H). Western blot analysis of an independent cohort of 12 paired samples demonstrated significantly higher ADGRG1 protein abundance in CAAs (Fig. 3I). These evidences robustly establish ADGRG1 not merely as a correlated gene but as a marker that is specifically enriched and upregulated in CAAs.

IHC analysis of ADGRG1 protein expression in CAAs from a cohort of 54 breast cancer patients revealed that high ADGRG1 levels were significantly associated with aggressive clinicopathological features, including advanced tumor stage and grade, high Ki-67 proliferation index, and the TNBC subtype (Fig. 3I; Table 1).Because patient age significantly differed between the ADGRG1‑high and ADGRG1‑low groups and age is a known prognostic factor in breast cancer, we performed multivariate logistic regression analyses adjusting for age to determine whether these associations were independent of age. After adjusting for age, high ADGRG1 expression remained significantly associated with high tumor grade (adjusted OR = 8.03, 95% CI: 1.88–34.3, p = 0.005), elevated Ki‑67 (adjusted OR = 10.68, 95% CI: 2.81–40.6, p < 0.001), TNBC subtype (adjusted OR = 7.94, 95% CI: 2.02–31.2, p = 0.003), and advanced tumor stage (adjusted OR = 5.72, 95% CI: 1.47–22.3, p = 0.012) (Additional file 2: Table S4). These results indicate that the aggressive clinicopathological features observed in the ADGRG1‑high group are not merely a reflection of younger patient age but are directly associated with ADGRG1 expression in CAAs. Finally, the clinical significance of this finding was assessed. To bolster the clinical relevance of the HAD7 signature, we also performed IHC for its core marker G0S2. Analysis of the cohort revealed that the averaged IHC score of ADGRG1 and G0S2 positively correlated with aggressive clinicopathological features. Furthermore, the individual scores for ADGRG1 and G0S2 were themselves significantly correlated, confirming their coordinated upregulation in aggressive disease contexts. This dual-marker validation strengthens the biological and clinical coherence of the HAD7 subpopulation (Additional file 2: Fig. S4E-F, Additional file 2: Table S5).

Table 1.

Clinicopathological features of breast cancer patients whose tissue was collected for IHC analysis (n = 54)

	Low	High	p
n	29	25
AGE (mean (SEM))	55.45 (10.52)	47.56 (10.49)	0.026
T (%)			0.055
1	12 (41.4)	3 (12.0)
2	12 (41.4)	16 (64.0)
3	5 (17.2)	6 (24.0)
N (%)			< 0.001
0	19 (65.5)	3 (12.0)
1	9 (31.0)	14 (56.0)
2	0 (0.0)	7 (28.0)
3	1 (3.4)	1 (4.0)
STAGE (%)			0.008
1	11 (37.9)	2 (8.0)
2	16 (55.2)	15 (60.0)
3	2 (6.9)	8 (32.0)
GRADE (%)			0.003
1	2 (6.9)	2 (8.0)
2	13 (44.8)	1 (4.0)
3	14 (48.3)	22 (88.0)
KI-67 (mean (SEM))	28.34 (15.16)	63.32 (19.82)	< 0.001
TYPE (%)			0.019
Basal	2 (6.9)	10 (40.0)
HER2	2 (6.9)	3 (12.0)
Luminal A	22 (75.9)	11 (44.0)
Luminal B	3 (10.3)	1 (4.0)

In summary, we establish ADGRG1 as an upregulated marker of the pro-tumorigenic HAD7 adipocyte subpopulation, localized at the critical tumor-adipose interface, and linked to aggressive pathology.

ADGRG1 in CAA drives tumor aggressiveness in vitro and in vivo

We first established functional adipocyte models by confirming the adipogenic differentiation capacity of both 3T3L1 preadipocytes and isolated primary human CAAs using Oil Red O staining (Fig. 4A). To interrogate ADGRG1 function, we established efficient knockdown in both primary CAAs and differentiated 3T3L1 cells using specific shRNAs (Additional file 2: Fig. S5A, B) and confirmed successful ADGRG1 overexpression in 3T3L1 cells (Additional file 2: Fig. S5C).

Fig. 4.

ADGRG1 in adipocytes promotes the aggressiveness of breast cancer cells in vitro. AOil Red O staining confirming successful adipogenic differentiation of 3T3L1 preadipocytes and primary human CAAs. Scale bars, 50 μm,10 μm. B Scratch wound healing assays of MDA-MB-231 and BT-549 cells co-cultured with control (shCtrl) or ADGRG1-knockdown (shADGRG1) CAAs. Representative images at 0 and 24 h are shown. Quantitative analysis of relative wound closure. Scale bars, 400 μm. Data are presented as mean ± SEM (n = 5 independent co-culture experiments); individual data points are overlaid. Statistical significance was determined by two-tailed Student’s t-test (C) Transwell migration (upper chamber uncoated) and invasion (upper chamber Matrigel-coated) assays of MDA-MB-231 and BT-549 cells co-cultured with shCtrl or shADGRG1 CAAs. Representative images are shown. Scale bars, 50 μm. Quantitative analysis of migrated or invaded cells per field. Data are presented as mean ± SEM (n = 5 independent co-culture experiments); individual data points are overlaid. Statistical significance was determined by two-tailed Student’s t-test (D) Scratch wound healing assays of 4T1 cells co-cultured with control 3T3L1 adipocytes (Vector), ADGRG1-overexpressing 3T3L1 adipocytes (OE), or OE adipocytes treated with the ADGRG1 inhibitor DHM (OE + DHM). Quantitative analysis of relative wound closure. Data are presented as mean ± SEM (n = 5 independent experiments); individual data points are overlaid. Scale bars, 400 μm.Statistical significance was determined by one-way ANOVA with Tukey’s post-hoc test for multiple comparisons (E) Transwell migration and invasion assays of 4T1 cells co-cultured with Vector, OE, or OE + DHM 3T3L1 adipocytes. Scale bars, 50 μm. Quantitative analysis of migrated or invaded cells per field. Data are presented as mean ± SEM (n = 5 independent experiments); individual data points are overlaid. Statistical significance was determined by one-way ANOVA with Tukey’s post-hoc test for multiple comparisons (F) Western blot analysis of epithelial (E-cadherin) and mesenchymal (N-cadherin, Vimentin) markers in MDA-MB-231 and BT-549 cells co-cultured with shCtrl or shADGRG1 CAAs. Quantification data are presented as mean ± SEM (n = 3 independent experiments); individual data points are overlaid. Protein levels were normalized to GAPDH. Statistical significance was determined by two-tailed Student’s t-test (G) Western blot analysis of EMT markers in 4T1 cells co-cultured with Vector, OE, or OE + DHM 3T3L1 adipocytes. Quantification data are presented as mean ± SEM (n = 3 independent experiments); individual data points are overlaid. Protein levels were normalized to Gapdh. Statistical significance was determined by one-way ANOVA with Tukey’s post-hoc test for multiple comparisons CAA, Cancer-Associated Adipocyte; EMT, Epithelial-Mesenchymal Transition; OE, Overexpression; shADGRG1, ADGRG1-knockdown; shCtrl, Short Hairpin RNA Control. * All quantitative data are presented as mean ± SEM from at least three independent experiments. Normality was assessed using the Shapiro-Wilk test before applying parametric tests. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001; ns, not significant

To establish a causal role for adipocyte ADGRG1, genetic and pharmacological loss-of-function experiments were conducted. Knockdown of ADGRG1 in primary human CAAs (shADGRG1) significantly impaired their capacity to stimulate the migration and invasion of co-cultured MDA-MB-231 and BT-549 triple-negative breast cancer cells, compared to control CAAs (shCtrl) (Fig. 4B, C). This establishes ADGRG1 activity as a necessary component of the pro-tumorigenic crosstalk emanating from CAAs.

Complementing the loss-of-function data, gain-of-function experiments demonstrated that ADGRG1 is also sufficient. 3T3L1-OEADGRG1 markedly enhanced the migratory and invasive capacities of co-cultured 4T1 murine mammary carcinoma cells (Fig. 4D, E). This confirms that elevated ADGRG1 signaling can actively drive a tumor-promoting adipocyte state.

We employed the small-molecule antagonist DHM, which inhibits ADGRG1 and its paralog ADGRG5. Pre-treatment of control CAAs with DHM significantly suppressed their pro-migratory effect on MDA-MB-231 and BT-549 cells. It induced significantly less migration of MDA-MB-231 and BT-549 cells in wound healing assays and reduced their invasion through Matrigel in Transwell assays (Additional file 2: Fig. S5D-F), effectively phenocopying the ADGRG1 knockdown phenotype. Crucially, DHM treatment also completely abrogated the enhanced pro-migratory activity conferred by ADGRG1 overexpression in 3T3L1 adipocytes (Fig. 4D, E). Western blot analysis confirmed that DHM treatment did not alter ADGRG1 protein levels (Additional file 2: Fig. S5G, H), indicating it acts as a functional antagonist of receptor activity, not an expression modulator. To definitively attribute the pharmacological effect of DHM to ADGRG1 inhibition in our system, we assessed the functional contribution of its known off-target, ADGRG5. We generated CAA-shADGRG5 and 3T3L1-shADRG5 (Additional file 2: Fig. S5I). Being co-cultured with MDA-MB-231, BT-549, and 4T1 cell lines, the loss of ADGRG5 did not attenuate the ability of CAAs and 3T3L1 to promote tumor cell migration and invasion. This stood in stark contrast to the potent inhibition consistently observed with ADGRG1 knockdown (Additional file 2: Fig. S5J, K). This differential genetic requirement provides compelling evidence that ADGRG5 is dispensable for this signaling axis. Therefore, the inhibitory efficacy of DHM in our co-culture assays can be confidently attributed to its on-target inhibition of ADGRG1.

To investigate the mechanism, the expression of EMT markers was analyzed. Knockdown of ADGRG1 in CAAs induced a more epithelial state in co-cultured cancer cells, characterized by upregulated E-cadherin and downregulated N-cadherin and Vimentin (Fig. 4F). Conversely, ADGRG1 overexpression in adipocytes promoted a mesenchymal shift, which was reversed by DHM inhibition (Fig. 4G).

These data demonstrate that ADGRG1 activity in adipocytes is both necessary and sufficient to enhance the aggressiveness of neighboring cancer cells. This effect is mediated, at least in part, through the regulation of the EMT program.

ADGRG1 in adipocytes promotes tumor growth, metastasis, and an immunosuppressive microenvironment in vivo

To substantiate the tumor-promoting role of adipocyte-specific ADGRG1 in a physiological context, we conducted in vivo assays.

First, in a syngeneic model, 4T1 mammary carcinoma cells were co-injected with control (Vector), ADGRG1-knockdown (KD), or ADGRG1-overexpressing (OE) 3T3L1 cells into Balb/c mice. Consistent with in vitro findings, tumors in the KD group exhibited significantly reduced growth kinetics, final tumor weight, and lung metastatic burden compared to the Vector control group (Fig. 5A-C). Conversely, the OE group demonstrated accelerated primary tumor growth and a marked increase in lung metastasis (Fig. 5D-F). This implied that ADGRG1 expression in adipocytes could modulate murine breast cancer cells. To validate these findings in a human context, we employed a xenograft model where MDA-MB-231 cells were co-injected with either shCtrl or shADGRG1 primary CAAs into nude mice. Mirroring the murine model, knockdown of ADGRG1 in human CAAs led to a substantial inhibition of both primary tumor growth and distant lung metastasis (Fig. 5G-I). These in vivo results definitively establish that ADGRG1 expression in adipocytes is a critical driver of both primary tumor progression and distant metastasis.

Fig. 5.

ADGRG1 in adipocytes drives tumor growth, metastasis, and immunosuppressive microenvironment suppression in vivo. A Representative images of tumors derived from 4T1 cells co-injected with control (shCtrl) or ADGRG1-knockdown (shADGRG1) 3T3L1 adipocytes into syngeneic mice. B Quantification of final tumor weight (left) and growth curves (right) for the experiment in (A). Data are presented as mean ± SEM (n = 5 mice per group); individual mouse data points are overlaid on the bar graph. Statistical significance for tumor weight was determined by a two-tailed Student’s t-test. For the growth curve, significance at the endpoint was assessed by t-test. C Left: Representative images of H&E-stained lung sections from mice in the shADGRG1 co-injection group (A), showing metastatic nodules Scale bar, 2 mm,200 μm,50 μm. Right: Quantitative analysis of the number of lung surface metastases. Data are presented as mean ± SEM (n = 5 mice per group); individual data points are overlaid. Statistical significance was determined by a two-tailed Mann-Whitney U test. D Representative images of tumors from 4T1 cells co-injected with control (Vector) or ADGRG1-overexpressing (OE) 3T3L1 adipocytes. E Quantification of final tumor weight (left) and growth curves (right) for the experiment in (D). Data are presented as mean ± SEM (n = 5 mice per group); individual mouse data points are overlaid. Statistical significance for tumor weight was determined by a two-tailed Student’s t-test. F Representative H&E-stained lung sections and quantification of lung metastases from the OE co-injection experiment (C). Scale bar, 2 mm,200 μm,50 μm Data are presented as mean ± SEM (n = 5 mice per group); individual data points are overlaid. Statistical significance was determined by a two-tailed Mann-Whitney U test (G) Representative images of xenograft tumors from MDA-MB-231 cells co-injected with control or ADGRG1-knockdown primary human CAAs in nude mice. H Quantification of final tumor weight (left) and growth curves (right) for the experiment in (G). Data are presented as mean ± SEM (n = 5 mice per group); individual mouse data points are overlaid. Statistical significance for tumor weight was determined by a two-tailed Student’s t-test (I) Representative H&E-stained lung sections and quantification of lung metastases from the experiment in (G). Scale bar, 2 mm,200 μm,50 μm Quantitative analysis of lung surface metastases. Data are presented as mean ± SEM (n = 5 mice per group); individual data points are overlaid. Statistical significance was determined by a two-tailed Mann-Whitney U test. J Representative IF staining of EMT markers in tumors from the 4T1/-3T3L1-shADGRG1 co-injection group (A). Nuclei are stained with DAPI. Scale bar, 50 μm. K Representative IF staining of EMT markers in tumors from the MDA-MB-231/CAA-shADGRG1 co-injection group (G). Scale bar, 50 μm. L Representative IF staining of EMT markers in tumors from the 4T1/3T3L1-OEADGRG1 co-injection group (A). Nuclei are stained with DAPI. Scale bar, 50 μm. M Representative IF staining of immune markers (CD4, CD8, PD-L1) in tumors from the 4T1/shADGRG1-3T3L1 co-injection group (A). Scale bar, 50 μm. N Representative IF staining of EMT and immune markers in tumors from the 4T1/3T3L1-OEADGRG1 co-injection group (C). Scale bar, 50 μm. CAA, Cancer-Associated Adipocyte; DAPI, 4’,6-Diamidino-2-Phenylindole; EMT, Epithelial-Mesenchymal Transition; H&E, Hematoxylin and Eosin; IF, Immunofluorescence; OE, Overexpression; *PD-L1*, Programmed Death-Ligand 1; shADGRG1, ADGRG1-knockdown; shCtrl, Short Hairpin RNA Control. *All quantitative data are presented as mean ± SEM from at least three independent experiments. For all panels, n represents biologically independent animals. Normality of data distribution was assessed using the Shapiro-Wilk test. The non-parametric Mann-Whitney U test was applied for metastasis count data, which was not normally distributed. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001; ns, not significant

To confirm durable engraftment and sustained ADGRG1 expression in adipocytes within the tumor stroma, a longitudinal immunofluorescence analysis was performed on tumors harvested at sequential time points. Adipocytes were identified using the specific marker FABP4. In the syngeneic model (4T1 + 3T3L1 cells), tumors analyzed at 7-, 14-, and 28-days post-injection similarly demonstrated persistent FABP4 + murine adipocytes. The ADGRG1 signal intensity was markedly elevated in adipocytes from the ADGRG1-overexpression group versus the control group (Additional file 2: Fig. S6A). In the xenograft model (MDA-MB-231 + human CAAs), tumors harvested at 14-, 28-, and 49-days post-injection confirmed the integration of FABP4 + adipocytes within the tumor stroma. ADGRG1 signal co-localized with these adipocytes and was nearly absent in the shADGRG1 group compared to controls (Additional file 2: Fig. S6B).

This spatiotemporal analysis provides direct evidence that the genetically modified adipocytes survived, integrated into the tumor microenvironment, and maintained their intended ADGRG1 expression states throughout tumor progression, thereby directly linking the manipulated adipocyte populations to the functional outcomes in each model.

Mechanistic analysis of the resected tumors revealed that the adipocyte ADGRG1 level shaped both tumor cell phenotype and the immune landscape. IF staining revealed that tumors grown with ADGRG1-knockdown adipocytes exhibited a molecular shift towards a less aggressive state, characterized by increased expression of the epithelial marker E-cadherin and decreased expression of the mesenchymal marker Vimentin (Fig. 5J). This shift was consistently observed in both the murine 4T1 and human MDA-MB-231 xenograft models (Fig. 5K, L). This in vivo correlation confirms that adipocyte ADGRG1 promotes a pro-metastatic phenotype in adjacent cancer cells Critically, analysis of the immune landscape revealed that ADGRG1 knockdown also led to an altered immune contexture, evidenced by reduced expression of the immune checkpoint protein PD-L1 and increased infiltration of both CD4 + and CD8 + T cells (Fig. 5M). In contrast, tumors associated with ADGRG1-overexpressing adipocytes displayed a pronounced pro-metastatic and immunosuppressive signature, with enhanced expression of mesenchymal markers, elevated cell proliferation, and upregulated PD-L1 expression (Fig. 5N). These findings indicate that adipocyte ADGRG1 not only drives tumor cell-intrinsic aggressiveness but also actively participates in remodeling the tumor immune microenvironment toward an immunosuppressive state.

Collectively, these in vivo data robustly demonstrate that ADGRG1 expression in adipocytes functions as a multi-faceted regulator of tumor progression. It drives disease aggressiveness by concurrently promoting a pro-metastatic EMT phenotype in cancer cells and fostering an immunosuppressive niche, thereby substantiating its role as a key stromal determinant of poor outcome.

ADGRG1 in CAAs activates the AKT pathway to drive pro-metastatic phenotypes in cancer cells

To define the global molecular changes induced by adipocyte ADGRG1, bulk RNA sequencing was performed on tumors derived from the 4T1/3T3L1 co-implantation model. Comparative analysis between the ADGRG1-OE and control groups identified a set of DEGs. Functional enrichment analysis of these DEGs revealed a profoundly altered tumor microenvironment. GO terms were significantly enriched for processes related to EMT (Fig. 6A). KEGG pathway analysis further highlighted activation of key oncogenic pathways, including the PI3K-AKT signaling pathway and pathways in cancer (Fig. 6B). Deconvolution of immune cell composition using CIBERSORTx indicated that the OE group was characterized by a significant increase in immunosuppressive myeloid populations and a concomitant decrease in cytotoxic lymphocytes (Fig. 6C). Gene Set Variation Analysis demonstrated enhanced activity of specific metabolic and pro-inflammatory pathways in the OE group (Fig. 6D). Finally, GSEA confirmed significant upregulation of gene signatures associated with inflammatory response and cytokine-cytokine receptor interaction (Fig. 6E). Collectively, these transcriptomic data demonstrate that ADGRG1 overexpression in adipocytes orchestrates a tumor microenvironment conducive to progression, marked by immune suppression, heightened inflammation, and activation of proliferative signaling.

Fig. 6.

ADGRG1 in adipocytes activates a CXCL8-AKT/mTOR signaling axis to remodel the tumor microenvironment. A, B Functional enrichment analysis of DEGs from bulk RNA-seq of tumors derived from 4T1 cells co-implanted with control versus ADGRG1-overexpressing 3T3L1 adipocytes. GO biological process terms (A) and KEGG pathways (B) significantly enriched in the OE group are shown. C Immune cell infiltration profiles in tumors from the experiment in (A), estimated using CIBERSORTx. Data are presented as mean proportion ± SEM for each cell type (n = 3 tumors per group). D GSVA illustrating the enhanced activity of specific metabolic and inflammatory pathways in tumors from the OE group compared to the Vector group. (E) GSEA demonstrating significant enrichment of cytokine and chemokine signaling pathways in the OE group. (F) Left: ELISA quantification of human CXCL8 secretion in conditioned media from shCtrl or shADGRG1 primary human CAAs, and shADGRG1 CAAs treated with the ADGRG1 inhibitor DHM. Right: ELISA quantification of the murine functional homolog MIP2 in conditioned media from control, ADGRG1-OE, and 3T3L1-OEADGRG1 treated with DHM. Data are presented as mean ± SEM (n = 3 individual experiments); individual data points are overlaid. Statistical significance was determined by one-way ANOVA with Tukey’s post-hoc test. G Western blot analysis of p-AKT, total AKT, p-mTOR, and total mTOR levels in cancer cells co-cultured with shCtrl or shADGRG1 CAAs, and shADGRG1 CAAs rescued with the AKT agonist SC79. Data are presented as mean ± SEM (n = 3); individual data points are overlaid. Protein levels were normalized to GAPDH; phosphorylation levels were normalized to total protein. Statistical significance was determined by one-way ANOVA with Tukey’s post-hoc test (H) Western blot analysis of p-AKT, total AKT, p-mTOR, and total mTOR levels in 4T1 cells treated with conditioned media from Vector, OE, or OE 3T3L1 adipocytes supplemented with an AKT inhibitor MK2206. Data presentation and analysis as in (G). I-K qPCR analysis of EMT transcription factor mRNA expression. MDA-MB-231 (I) and BT-549 (J) cells were treated with: control medium, recombinant human CXCL8, CXCL8 + a neutralizing anti-CXCL8 antibody, or CXCL8 + the AKT inhibitor MK-2206. 4T1 cells (K) were treated analogously with the murine homolog recombinant MIP2. Data are presented as mean ± SEM (n = 3); individual data points are overlaid. Statistical significance was determined by one-way ANOVA with Tukey’s post-hoc test. L, M Western blot analysis of EMT marker proteins under the same treatment conditions as in (I-K) for MDA-MB-231/BT-549 (L) and 4T1 ( M) cells. Data presentation and analysis as in (G). N, O Western blot analysis of AKT/mTOR pathway activation under the same treatment conditions as in (I-K) for MDA-MB-231/BT-549 (N) and 4T1 (O) cells, showing levels of p-AKT (Ser473), total AKT, p-mTOR (Ser2448), and total mTOR. Data presentation and analysis as in (G) AKT, Protein Kinase B; CAA, Cancer-Associated Adipocyte; DEGs, Differentially Expressed Genes; ELISA, Enzyme-Linked Immunosorbent Assay; GO, Gene Ontology; GSEA, Gene Set Enrichment Analysis; GSVA, Gene Set Variation Analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; mTOR, Mechanistic Target of Rapamycin; OE, Overexpression; p-AKT, Phosphorylated AKT; p-mTOR, Phosphorylated mTOR; shADGRG1, ADGRG1-knockdown; shCtrl, Short Hairpin RNA Control. *All quantitative data are presented as mean ± SEM from at least three independent experiments. Normality was assessed using the Shapiro-Wilk test before ANOVA. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001; ns, not significant

Based on the transcriptomic signature highlighting cytokine activity, we systematically profiled the secretome of CAAs to identify the key soluble mediator. Conditioned media from primary human CAAs, shADGRG1 CAAs, and shADGRG1 CAAs treated with the ADGRG1 inhibitor DHM were analyzed via a multiplex ELISA panel targeting 12 cytokines. This screening identified CXCL8 as the most prominently and consistently upregulated cytokine in CAA-conditioned media (Additional file 3: Fig. S7A). The elevation of CXCL8 was strictly dependent on ADGRG1 activity, as its secretion was significantly reduced upon ADGRG1 knockdown and abrogated by pharmacological inhibition with DHM (Fig. 6F). To assess this in the murine models, we measured the level of macrophage inflammatory protein-2 (MIP2), a functional homolog of CXCL8 in mice that signals through the same receptor (CXCR2). This approach was based on our previous studies which established MIP2 as a reliable and functionally analogous surrogate for CXCL8 in mouse tumor models [36]. Consistent with the human data, MIP2 was significantly elevated in the supernatant of 3T3L1-OEAdgrg1 cells (Fig. 6F). These results establish CXCL8 as a principal paracrine factor downstream of adipocyte ADGRG1 signaling. To functionally connect HA7 subpopulation to our mechanistic axis, we performed correlation analysis using primary human CAA isolates. We found a significant positive correlation between the baseline expression levels of ADGRG1 and G0S2 and the capacity of these CAAs to secrete CXCL8 upon stimulation (Additional file 3: Fig. S7B). This indicates that adipocytes bearing a stronger HAD7 molecular signature are functionally primed to execute the ADGRG1-CXCL8 paracrine pathway.

Given the well-established role of the CXCL8/CXCR2 axis in activating the AKT/mTOR signaling cascade, we hypothesized that this pathway serves as a key downstream effector of ADGRG1 [38, 39]. WB analysis of cancer cells co-cultured with adipocytes confirmed this: the presence of CAAs strongly activated the AKT/mTOR signaling axis, as evidenced by increased phosphorylation of both AKT and its key downstream target, mTOR. This activation was ADGRG1-dependent, as it was attenuated when cancer cells were co-cultured with shADGRG1-CAAs. Crucially, to confirm that AKT/mTOR activation is a critical downstream event, we demonstrated that the impaired phosphorylation in cancer cells co-cultured with shADGRG1-CAAs could be effectively rescued by the AKT agonist SC79 (Fig. 6G). These findings were corroborated in the murine system, where 4T1 cells treated with conditioned media from OEAdgrg1-3T3L1 cells showed heightened AKT/mTOR phosphorylation, which was reversed by the AKT inhibitor MK2206 (Fig. 6H). Collectively, these results demonstrate that the ADGRG1‑CXCL8 axis activates AKT/mTOR signaling in cancer cells, and that this pathway is functionally required for the enhanced migration, invasion, and EMT phenotypes.

The paracrine factor CXCL8 is sufficient and necessary to drive AKT/mTOR-dependent EMT

To test whether CXCL8 is the specific adipocyte-derived factor responsible for the observed pro-tumorigenic effects, we performed a series of complementary rescue experiments. Direct treatment with recombinant CXCL8 protein robustly enhanced the migratory and invasive capacities of MDA-MB-231 and BT-549 cells. Crucially, this effect was abolished by co-treatment with a CXCL8-neutralizing antibody, confirming that the phenotype was specifically attributable to CXCL8 signaling and not potential contaminants (Additional file 3: Fig. S8A-D). In the murine 4T1 model, the homologous chemokine MIP2 elicited a pro-migratory response, which was similarly blocked by its neutralizing antibody, underscoring the conservation of this chemokine axis across species (Additional file 3: Fig. S8E, F).

Having established specificity, we investigated the downstream pathway. CXCL8 treatment rapidly activated the AKT/mTOR pathway. To prove this activation was functionally necessary for the aggressive phenotype, we employed the AKT inhibitor MK-2206. Co-treatment with MK-2206 abrogated the enhanced migration and invasion induced by CXCL8 (Additional file 3: Fig. S8A-F). This pharmacological rescue experiment provides direct evidence that the AKT/mTOR axis is the primary signaling conduit for CXCL8’s effects.

We next investigated the mechanism by which this signaling leads to a sustained aggressive state. Transcriptional analysis by qPCR revealed that CXCL8 treatment orchestrated a pro-EMT transcriptional program, significantly upregulating key drivers SNAIL, TWIST, and ZEB1 in all cell lines (Fig. 6I-K). This shift was functionally cemented at the protein level, with CXCL8 inducing hallmark EMT changes, including downregulation of E-cadherin and upregulation of N-cadherin and Vimentin (Fig. 6L, M). Notably, both this transcriptional reprogramming and the resultant protein-level EMT signature were reversed by either CXCL8 neutralization or AKT/mTOR inhibition. This link between pathway activity and phenotypic reprogramming prompted us to validate the upstream signaling event directly. Western blot analysis confirmed that CXCL8/MIP2 treatment induced rapid and potent phosphorylation of both AKT and mTOR in all corresponding cell lines. This change could be effectively blocked by the respective neutralizing antibody, and the AKT inhibitor MK-2206 (Fig. 6N, O). This biochemical evidence provides the molecular connection that the CAA-derived CXCL8 could activate the core AKT/mTOR kinase cascade and then induce EMT.

Together, the CAA-secreted CXCL8 is sufficient to drive EMT-like molecular changes. These changes are both ligand-specific and dependent on AKT/mTOR signaling. CXCL8 receptor engagement directly initiates this required signaling pathway via AKT/mTOR phosphorylation. This cascade positions the ADGRG1-CXCL8 axis as a mechanistic bridge, converting a stromal receptor signal into a defined paracrine output that hijacks a canonical oncogenic pathway to transcriptionally enforce a pro-metastatic state in cancer cells.

In vivo blockade of the CXCL8/CXCR axis rescues ADGRG1-driven tumor promotion

To establish a direct causal link between adipocyte ADGRG1 and its downstream cytokine signaling in vivo, we performed a functional rescue experiment in an orthotopic mammary fat pad xenograft model. MDA-MB-231 tumor cells were co-injected with primary human CAAs under the following conditions: (1) tumor cells alone, (2) tumor cells alone with systemic Reparixin treatment, (3) tumor cells with control CAAs, (4) tumor cells with control CAAs plus Reparixin, and (5) tumor cells with ADGRG1-knockdown CAAs. This design allowed us to assess both the necessity of ADGRG1 and contribution of the CXCL8 signaling pathway.

Consistent with our previous findings, co-transplantation with control CAAs significantly promoted tumor growth compared to tumor cells alone (Additional file 3: Fig. S8G, H). Immunohistochemical analysis of these tumors revealed a pronounced EMT phenotype, characterized by downregulated E-cadherin and upregulated N-cadherin and Vimentin expression (Additional file 3: Fig. S8I). This pro-tumorigenic effect was attenuated upon co-injection with ADGRG1-knockdown CAAs, confirming that ADGRG1 in adipocytes is required to drive tumor progression in this setting (Additional file 3: Fig. S8G-I).

Critically, the pro-tumorigenic effects of control CAAs were also effectively rescued by interrupting the downstream cytokine axis. Systemic blockade of the CXCL8 receptors CXCR1/2 with Reparixin suppressed the enhanced tumor growth promoted by control CAAs (Additional file 3: Fig. S8G, H). The EMT marker expression in tumors from this group was also rescued (Additional file 3: Fig. S8I).

To figure out the possibility that this inhibition was merely due to a direct action of Reparixin on tumor cells, we included a control group receiving Reparixin alone. Two‑way ANOVA revealed a significant interaction between CAA co‑injection and Reparixin treatment for tumor weight, volume and EMT markers (Additional file 3: Table S6), indicating that Reparixin exerts an additional, CAA‑dependent inhibitory effect beyond its direct anti‑tumor activity.

These in vivo data demonstrate that the CXCL8 signaling pathway is functionally important for the tumor-promoting effects observed in our model, consistent with the in vitro evidence that adipocyte ADGRG1 drives CXCL8 secretion.

Collagen III activates ADGRG1 to trigger RhoA-dependent CXCL8 secretion

To identify the upstream signal activating ADGRG1 in CAAs, we focused on its known ligands. Prior literature identifies two principal ligands for ADGRG1: tissue transglutaminase 2 (TG2) and collagen type III, which can elicit distinct cellular responses [40, 41]. As Col3 binding is specifically linked to pro-inflammatory signaling via RhoA—a pathway aligned with our observed phenotype—we investigated its role [42].

To test the functional interaction, primary human CAAs were stimulated with purified recombinant collagen III (84 nM) in culture. Control cells received carrier solution alone. Following a 48-hour stimulation, conditioned media were collected for analysis. Col3 stimulation resulted in a significant increase in CXCL8 secretion from control CAAs, confirming it as a potent inducer. This induction was specific to ADGRG1 signaling, as it was abolished in ADGRG1-knockdown CAAs (Fig. 7A). In a parallel experiment using murine 3T3L1 cells, Col3 stimulation likewise triggered a robust increase in the secretion of MIP2 (Fig. 7A). This conserved response across species underscores the generalizability of this ligand-receptor axis.

Fig. 7.

Collagen III activates an ADGRG1-RhoA signaling axis in adipocytes to drive a paracrine CXCL8- AKT-mTOR-EMT circuit in cancer cells. A Left: ELISA quantification of human CXCL8 in conditioned media from primary human CAAs under three conditions: CAA, CAA stimulated with Col3, shADGRG1 stimulated with Col3. Right: ELISA quantification of the murine homolog MIP2 from 3T3L1 under analogous conditions: Control, Control + Col3, shADGRG1 + Col3. Data are presented as mean ± SEM (n = 6) from six independent experimental replicates. Statistical significance was determined by one-way ANOVA with Tukey’s post-hoc test for multiple comparisons. B Left: RhoA activation assay. Levels of active GTP-bound RhoA (RhoA-GTP) were measured by pull-down in human CAAs under the same three conditions as in (A, left). Right: Active RhoA levels in 3T3L1 cells under the same four conditions as in (A, right). Bar graphs show densitometric quantification normalized to total RhoA input. Data are presented as mean ± SEM (n = 3); individual data points are overlaid. Statistical analysis was by one-way ANOVA with Tukey’s post-hoc test. C Left: ELISA for CXCL8 in conditioned media from human CAAs under three conditions: untreated, treated with Col3, and treated with Col3 plus the RhoA inhibitor Rhosin. Right: ELISA for MIP2 from 3T3L1 cells under the analogous three conditions. Data are presented as mean ± SEM (n = 6 ); individual data points are overlaid. Statistical analysis was by one-way ANOVA with Tukey’s post-hoc test. D-F qPCR analysis of EMT transcription factor mRNA expression. MDA-MB-231 (D), BT-549 (E), and 4T1 (F) cells were subjected to co-culture under four conditions: with control CAAs, with Col3-stimulated adipocytes (+ Col3), with Col3-stimulated adipocytes in the presence of a CXCL8/MIP2 neutralizing antibody (+ Col3 + anti-CXCL8), and with Col3-stimulated adipocytes in the presence of the AKT inhibitor MK-2206 (+ Col3 + MK-2206). Expression of SNAIL, TWIST, and ZEB1 is shown. Data are presented as mean ± SEM (n = 3 independent co-culture experiments); individual data points are overlaid. Statistical significance was determined by one-way ANOVA with Tukey’s post-hoc test. (G-H) Western blot analysis of EMT marker proteins for MDA-MB-231, BT-549(G) and 4T1 cells (H) under the same four co-culture conditions described in (D-F). Protein levels of E-cadherin, N-cadherin, and Vimentin were normalized to GAPDH. Data are presented as mean ± SEM (n = 3); individual data points are overlaid. Statistical analysis was by one-way ANOVA with Tukey’s post-hoc test. I-J Western blot analysis of AKT/mTOR pathway activation for MDA-MB-231, BT-549 (I), and 4T1 (J) cells under the same four co-culture conditions. Levels of p-AKT, total AKT, p-mTOR, and total mTOR are shown. Phosphorylation levels were normalized to total protein. Data are presented as mean ± SEM (n = 3); individual data points are overlaid. Statistical analysis was by one-way ANOVA with Tukey’s post-hoc test. *For all panels, ‘n’ represents independent experiments. Normality of data distribution was confirmed using the Shapiro-Wilk test prior to ANOVA. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001; ns, not significant.

Having established Col3 as the activating ligand, we delineated the immediate downstream signaling. RhoA-GTP pull-down assays revealed that Col3 stimulation robustly increased active RhoA levels in control CAAs, but failed to do so in ADGRG1-knockdown cells (Fig. 7B). Pharmacological inhibition of RhoA signaling subsequently suppressed both basal and Col3-induced chemokine secretion, confirming RhoA’s role as a critical and necessary downstream node (Fig. 7C).

These data define a specific and conserved ligand-receptor interaction that initiates the pathogenic cascade. They demonstrate that Col3 binds to ADGRG1 on adjacent adipocytes. This engagement triggers RhoA activation, which in turn drives the secretion of CXCL8/MIP2, thereby positioning stromal collagen as a key upstream environmental cue licensing the pro-tumorigenic function of CAAs.

ADGRG1 in cancer-associated adipocytes activates a coherent signaling axis to drive tumor progression

Having established that Col3 specifically activates ADGRG1 to induce chemokine secretion, we next delineated the functional consequences of this paracrine signal on recipient cancer cells. To map the pathway from stromal cue to cellular reprogramming, we employed a systematic co-culture approach using primary human CAAs with MDA-MB-231 and BT-549 cells, and murine 3T3L1 adipocytes with 4T1 cells, under four conditions: control, Col3-stimulated, Col3-stimulated with CXCL8/MIP2 neutralization, and Col3-stimulated with AKT inhibition. qPCR analysis across all three cancer cell lines revealed that co-culture with Col3-stimulated adipocytes modulated a coherent pro-EMT transcriptional program, marked by significant upregulation of the master regulators SNAIL, TWIST, and ZEB1 (Fig. 7D-F). Crucially, this transcriptional reprogramming was abrogated by either neutralizing the secreted chemokine or inhibiting AKT, functionally linking the adipocyte-derived signal and the AKT pathway to this cell-fate decision. This transcriptional shift was faithfully executed at the protein level, as western blot analysis confirmed a hallmark EMT signature with downregulation of E-cadherin and upregulation of N-cadherin and Vimentin under Col3 stimulation, which was again reversed by both intervention strategies (Fig. 7G, H). Co-culture with Col3-stimulated adipocytes induced phosphorylation of both AKT and mTOR, and was prevented by CXCL8/MIP2 neutralization. (Fig. 7I, J). This implied that this activation was specific to the ADGRG1-CXCL8 axis. What’s more, AKT inhibition blocked downstream phenotypic and molecular changes, placing AKT/mTOR activation as the key signaling module transmitting the external chemokine cue. To formally exclude the possibility that residual Col3 in the co-culture medium could directly act on cancer cells, we treated tumor cells with Col3 alone in the absence of adipocytes (Additional file 3: Fig. S9A-F). Col3 alone did not alter the expression of EMT-related proteins nor the mRNA levels of SNAIL, TWIST, or ZEB1. Furthermore, no changes in AKT or mTOR phosphorylation were detected upon Col3 treatment alone. These results confirm that the EMT and AKT/mTOR activation observed in co-culture experiments are dependent on adipocyte-derived paracrine signals.

In a word, ADGRG1 on CAAs functions as the signal-transducing receptor; the secreted effector CXCL8/MIP2 is both necessary and sufficient for the subsequent effects; and in cancer cells, this signal converges on the AKT/mTOR pathway to drive a transcriptional EMT program.

Co-targeting ADGRG1 and PD-1 elicits superior anti-tumor efficacy

Given the dual role of ADGRG1 in activating the pro-angiogenic AKT/mTOR pathway and fostering an immunosuppressive microenvironment, we hypothesized that its inhibition would simultaneously impair tumor angiogenesis and enhance anti-tumor immunity, thereby potentiating the efficacy of immune checkpoint blockade. To test this, we treated mice bearing 4T1 tumors with: (1) vehicle control, (2) anti-PD-1 monotherapy, (3) the ADGRG1 inhibitor DHM monotherapy, or (4) a combination of DHM and anti-PD-1.

The combination therapy resulted in the most profound suppression of tumor growth, yielding the smallest final tumor volumes and weights, significantly outperforming either monotherapy (Fig. 8A-B). This superior effect was further evidenced by a reduction in lung metastatic burden, with the combination group showing the fewest and smallest metastatic nodules (Fig. 8C).

Fig. 8.

Combinatorial therapy targeting ADGRG1 and PD-1 demonstrates superior anti-tumor efficacy. A Representative images of tumors harvested from the 4T1 syngeneic mouse model treated with: vehicle control (Mock), anti-PD-1 monotherapy, ADGRG1 inhibitor (DHM) monotherapy, or the combination of DHM and anti-PD-1; B Quantification of in vivo therapeutic efficacy. Left: Final tumor weight from each treatment group. Right: Tumor growth curves over the treatment period. Data are presented as mean ± SEM (n = 5 biologically independent mice per group); individual mouse data points are overlaid on the bar graph (left). Statistical significance for tumor weight at endpoint was determined by one-way ANOVA with Tukey’s post-hoc test for multiple comparisons (C) Analysis of lung metastasis. Left: Representative images of H&E-stained lung sections from each treatment group; metastatic nodules are indicated by arrows. Scale bar, 2 mm,200 μm,50 μm Right: Quantitative analysis of the number of lung surface metastases. Data are presented as individual data points with median ± interquartile range (n = 5 mice per group). Statistical significance was determined by the Kruskal-Wallis test with Dunn’s post-hoc test; (D) Representative IHC staining of key molecular markers in tumor tissues. Scale bar, 50 μm. Data are presented as mean ± SEM (n = 5 biologically independent mice); individual animal data points are overlaid on the bars. Statistical significance was determined by one-way ANOVA with Tukey’s post-hoc test for multiple comparisons. E Representative IF staining of the same EMT and immune markers analyzed in (D). Scale bar, 50 μm. Data presentation and analysis as in (D). F Representative IF staining of the angiogenesis marker in tumor tissues from the four treatment groups. Scale bar, 50 μm. Data presentation and analysis as in (D). EMT, Epithelial-Mesenchymal Transition; H&E, Hematoxylin and Eosin; IF, Immunofluorescence; IHC, Immunohistochemistry; *PD-1*, Programmed Cell Death Protein 1. *All quantitative data are presented as mean ± SEM from at least three independent experiments. Normality of data distribution was confirmed using the Shapiro-Wilk test prior to ANOVA. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001; ns, not significant

Comprehensive IHC and IF analyzes revealed the multi-faceted basis for this efficacy. Tumors from DHM-treated groups exhibited significantly reduced density of the vascular marker CD31, indicating inhibition of angiogenesis, with the most pronounced effect seen in the combination group. Furthermore, tumors from the combination group displayed the most favorable anti-tumor profile across all measured parameters: an altered immune landscape characterized by enhanced T cell infiltration including increased densities of CD8 + and CD4 + T cells and reduced PD-L1 expression, coupled with a reversal of the pro-metastatic phenotype like upregulated E-cadherin and downregulated Vimentin (Fig. 8D-F).

In conclusion, pharmacological inhibition of the ADGRG1-CXCL8-AKT axis disrupts a pro-angiogenic and immunosuppressive signaling hub. When combined with anti-PD-1 therapy, it orchestrates an additive anti-tumor response by concurrently impairing blood supply, reversing EMT, and promoting T cell infiltration, thereby achieving superior control of tumor growth and metastasis.

Discussion

The breast cancer TME is a complex and dynamic ecosystem, within which adipocytes are not merely passive bystanders but active participants in tumor progression. As the most abundant stromal component in the breast, adipocytes engage in a continuous and reciprocal crosstalk with cancer cells [43, 44]. Under this influence, NAs undergo a phenotypic transformation into CAAs, which are known to promote tumor growth, invasion, and metastasis through the secretion of various cytokines and metabolites [12]. Compared to their normal counterparts, CAAs exhibit enhanced stemness properties and secretory capacity, producing a more complex repertoire of cytokines and chemokines that establish the foundation for therapy resistance [45]. This cellular reprogramming enables CAAs to serve as central coordinators of the immunosuppressive niche while simultaneously supporting cancer cell metabolic requirements and survival pathways. The convergence of increased differentiation potential and amplified secretory function in CAAs creates a feed-forward loop that sustains tumor progression and undermines therapeutic efficacy [46]. While our previous work and that of others have established the general tumor-promoting role of CAAs, a critical gap remains: the systematic classification of adipocyte heterogeneity in clinical samples and its direct impact on patient prognoses has been largely unexplored [36].

In this study, we first deconstructed the heterogeneity of human adipose tissue at single-nucleus resolution, identifying distinct adipocyte subpopulations. We discovered that one specific subpopulation, HAD7, characterized by inflammatory activity, was enriched in aggressive tumor contexts. By leveraging this subpopulation-specific gene signature to deconvolute bulk tumor samples, we constructed a novel prognostic model that effectively stratified patients based on the abundance of this adipocyte subset. This model successfully identified a subset of patients with significantly poorer outcomes, underscoring the clinical relevance of adipocyte heterogeneity. Subsequent machine-learning analysis of the genes underpinning this model pinpointed ADGRG1 as a key driver, whose expression in CAAs is instrumental in fostering an immunosuppressive microenvironment and driving tumor progression. Spatial validation using multiplex immunofluorescence co-staining for the HAD7 markers G0S2 and ADGRG1, alongside pan-adipocyte and tumor cell markers, confirmed that G0S2+/ADGRG1 + adipocytes are not randomly distributed. Instead, they are enriched at the tumor-adipose interface, a strategic positioning that facilitates efficient paracrine crosstalk. This spatial enrichment is particularly noteworthy given that these cells represent a minority subset within the total adipocyte population. Their functional contribution therefore likely arises from signaling potency and precise localization at the tumor border, rather than from numerical abundance. At this interface, paracrine signals such as CXCL8 need only diffuse very short distances to reach tumor cells, enabling efficient signaling despite the subset’s low frequency. The abundance of this spatially defined subpopulation further correlated with worse clinical outcomes, reinforcing its functional importance beyond mere proportional representation.

Immunohistochemical validation in an independent clinical cohort confirmed that high ADGRG1 protein expression in CAAs is significantly associated with advanced tumor stage, grade, and poor prognosis. Functional validation, both in vitro and in vivo, established ADGRG1’s essential role. Its knockdown in primary human CAAs or model adipocytes severely impaired their capacity to promote cancer cell migration, invasion, and EMT. Crucially, longitudinal immunofluorescence analysis of tumors from our in vivo co-injection models harvested at multiple time points confirmed the durable engraftment and persistent, manipulated ADGRG1 expression in adipocytes within the tumor stroma, ruling out transient effects and solidifying a direct causal link.

To connect these phenotypes to specific pathways, we leveraged transcriptomic profiling and cytokine analysis. This revealed that ADGRG1 activation in adipocytes initiates an ADGRG1-AKT-mTOR signaling axis, culminating in the robust secretion of the chemokine CXCL8. This CXCL8-driven axis provides a parsimonious explanation for the multi-faceted pro-tumorigenic niche: CXCL8 is a well-established recruiter of immunosuppressive myeloid cells, a potent angiogenic factor, and a direct promoter of cancer cell invasiveness [47]. We then delineated the downstream consequences for cancer cells. Recombinant CXCL8 treatment activated the AKT/mTOR pathway in tumor cells and recapitulated the pro-migratory and invasive phenotypes. The functional necessity of this axis was confirmed through rescue experiments: the pro-tumorigenic effects of conditioned media from ADGRG1-active CAAs, or of direct CXCL8 stimulation, were completely abolished by either a CXCL8-neutralizing antibody or the specific AKT inhibitor MK-2206.

We further elucidated the upstream trigger and immediate downstream signaling of this receptor. Guided by the literature on ADGRG1 ligands and our tumor microenvironment context, we identified Col3 as the relevant activating ligand. Stimulation of CAAs with Col3 triggered a marked, ADGRG1-dependent increase in CXCL8 secretion. This Col3-ADGRG1 interaction directly activated the small GTPase RhoA, and pharmacological inhibition of RhoA signaling potently suppressed CXCL8 production, positioning RhoA as the critical immediate downstream node. The specificity of this pathway to ADGRG1 was further underscored by the lack of a phenotypic impact following knockdown of its close paralog, ADGRG5, and by the use of the inhibitor DHM. While DHM can target related receptors, our functional data with ADGRG5 knockdown and the clear Col3-RhoA axis provide a coherent mechanism attributable to ADGRG1. This defines a coherent initiating axis, Col3 → ADGRG1 → RhoA activation → CXCL8 secretion. The secreted CXCL8 then acts on tumor cells to activate the AKT/mTOR pathway, a step verified as functionally necessary using the specific AKT inhibitor MK-2206, which abrogated CXCL8-induced migration, invasion, and EMT. This cascade drives a transcriptional reprogramming characterized by the upregulation of EMT regulators like Snail, Twist, and Zeb1, culminating in enhanced invasiveness. Nevertheless, it remains formally undefined whether CXCL8 activates AKT in tumor cells via PI3K-dependent, PI3K-independent, or convergent mechanisms in our model, as CXCR1/2 can engage multiple downstream adaptors. Future work is needed to delineate the precise upstream hierarchy. The therapeutic potential of disrupting this stromal axis is underscored by our finding that DHM, when combined with anti-PD-1 immune checkpoint blockade, additively inhibited tumor growth. This combination therapy not only reversed the ADGRG1-driven immunosuppressive microenvironment but also concurrently reduced EMT signatures and tumor angiogenesis, demonstrating a multi-pronged antitumor effect.

Thus, we define a complete, multi-step signaling cascade within the breast TME: Collagen III in the matrix activates ADGRG1 on CAAs, leading to RhoA-dependent transcription and secretion of CXCL8. This chemokine then acts on cancer cells to activate the AKT/mTOR pathway, driving a transcriptional EMT program and enhancing metastatic potential. This axis effectively positions ADGRG1 as a stromal master switch that coordinately regulates several hallmark capabilities of cancer, explaining the potent synergy observed when it is targeted alongside immune checkpoint blockade.

The functional role of ADGRG1 in cancer biology demonstrates remarkable context dependency, exhibiting both tumor-promoting and tumor-suppressive effects across different malignancies. In melanoma and glioblastoma, ADGRG1 acts primarily as a tumor suppressor, where its interaction with extracellular matrix components inhibits angiogenesis and metastatic progression through modulation of NF-κB signaling and other pathways [48]. Conversely, in breast, colorectal, and pancreatic cancers, ADGRG1 displays strong oncogenic properties, enhancing cancer cell proliferation, migration, and chemoresistance predominantly through activation of PI3K/AKT and RhoA signaling cascades [35, 49]. This functional duality underscores the importance of cellular context and tissue microenvironment in determining ADGRG1’s ultimate biological impact.

Our findings position ADGRG1 within CAAs as a novel mechanism contributing to its tumor-promoting functions in breast cancer. The identification of the ADGRG1-CXCL8-AKT/mTOR axis in CAAs reveals how this receptor can coordinate multiple aspects of tumor progression through stromal-epithelial crosstalk. This represents a significant expansion of ADGRG1’s known pro-tumorigenic mechanisms beyond its previously characterized cell-autonomous functions in cancer cells. The ability of ADGRG1 to orchestrate immunosuppression, angiogenesis, and enhanced invasiveness through modulation of the adipocyte secretome highlights its role as a master regulator of the tumor-supportive niche.

It is important to note that the bulk transcriptomic data from public cohorts (TCGA, METABRIC) used for clinical correlation and survival analysis do not consistently provide detailed metadata on patient body mass index (BMI) or metabolic health status. Furthermore, the adipose tissue samples utilized in the foundational snRNA-seq atlas were primarily derived from subcutaneous depots (abdomen/neck) of individuals undergoing bariatric or plastic surgery. While this provides a valuable map of adipocyte heterogeneity, these samples may not fully represent the spectrum of metabolic states or the specific microenvironment of tumor-adjacent adipose tissue in a breast cancer patient population with varying BMI. This is a key limitation of this study. Given the well-established and profound impact of obesity on systemic inflammation, adipokine secretion, and adipose tissue biology, it is plausible that the activity of the HAD7 program and its prognostic effect could be modulated by the host’s metabolic environment. For instance, an obesogenic milieu might amplify the ADGRG1-driven inflammatory axis we describe. Future studies are essential to directly interrogate the interaction between adipocyte subpopulations like HAD7 and patient BMI in well-annotated clinical cohorts. Such work could reveal whether the HAD7 signature holds universal prognostic value or is particularly potent in specific metabolic subgroups, paving the way for more personalized prognostic stratification and therapeutic strategies that consider both the tumor and its systemic host environment.

While our study delineates a novel stromal signaling pathway, several limitations warrant consideration. Firstly, the clinical validation cohort comparing CAAs and normal adipocytes was not matched for donor age, menopausal status, or BMI which known to profoundly influence adipocyte biology. Detailed information on these variables was not systematically collected for the healthy control group, so we cannot exclude potential confounding. Future studies should employ carefully matched controls or adjust for these factors. Secondly, we acknowledge a limitation regarding the interpretation of our in vivo rescue experiment. Because the CXCR1/2 inhibitor was administered systemically, it potentially acts on multiple cell types expressing these receptors, including neutrophils, macrophages, and other stromal cells within the TME. While our in vitro co-culture experiments demonstrate that adipocyte-derived CXCL8 (or murine MIP2) can directly activate AKT/mTOR signaling in cancer cells, the in vivo rescue effect cannot be definitively attributed to adipocyte-specific targeting. The systemic inhibition may also block CXCL8/CXCR2 signaling on immune cells or other stromal compartments, which could independently influence tumor growth. Therefore, we interpret these results as confirming that the CXCL8 signaling pathway is functionally important for the tumor-promoting effects observed in our model, while the precise cellular source and target populations in vivo warrant further investigation using cell-type-specific genetic approaches. Thirdly, our findings are primarily based on a limited set of models; future work should validate this axis across a broader spectrum of breast cancer subtypes and in more complex, immunocompetent murine models to fully assess its generalizability and therapeutic relevance. Finally, the ADGRG1-driven secretome is almost certainly more complex than a single chemokine. A more comprehensive, unbiased profiling of the adipocyte secretome upon ADGRG1 inhibition could reveal additional co-operating factors.

Conclusions

In summary, our work transcends the conventional, cancer-cell-centric view of oncogenic drivers by unveiling a pivotal role for the adhesion GPCR ADGRG1 within the stromal compartment of breast tumors. We establish that a specific adipocyte subpopulation utilizes ADGRG1 as a signaling hub to activate a CXCL8-AKT/mTOR axis, thereby orchestrating an immunosuppressive and pro-metastatic niche. The potent synergy achieved by co-targeting ADGRG1 and PD-1 not only validates the biological significance of this pathway but also illuminates a promising combinatorial strategy to reprogram the TME and overcome resistance to immunotherapy. By resolving ADGRG1’s context-dependent roles, our study positions stromal ADGRG1 as a compelling therapeutic target and underscores the critical importance of targeting the tumor ecosystem as a whole.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUC

area under the receiver operating characteristic curve

BC

breast cancer

BMI

body mass index

CAA

cancer-associated adipocyte

C-index

Concordance Index

Col3

collagen type III

DSS

Disease-Specific Survival

DHM

Dihydromunduletone

EBI

European Bioinformatics Institute

EMT

epithelial-mesenchymal transition

FDR

false discovery rate

GEO

Gene Expression Omnibus

GO

Gene Ontology

GSEA

Gene Set Enrichment Analysis

GSVA

Gene Set Variation Analysis

HBCA

Human Breast Cell Atlas

hSAT

human subcutaneous adipose tissue

HR

hazard ratio

IC₅₀

half-maximal inhibitory concentration

ICI

immune checkpoint inhibitor

IF

immunofluorescence

IHC

immunohistochemistry

KEGG

Kyoto Encyclopedia of Genes and Genomes

LASSO

Least Absolute Shrinkage and Selection Operator

MIP2

macrophage inflammatory protein-2

MSigDB

Molecular Signatures Database

NA

normal adipocyte

OS

Overall Survival

PCA

principal component analysis

PCC

Pearson’s Correlation Coefficient

RBD

Rho-binding domain

RT-qPCR

real-time quantitative polymerase chain reaction

snRNA-seq

single-nucleus RNA sequencing

TCGA

The Cancer Genome Atlas

TG2

transglutaminase 2

TME

tumor microenvironment

TNBC

Triple-Negative Breast Cancer

UMAP

Uniform manifold approximation and projection

XGBoost

eXtreme Gradient Boosting

Author contributions

KS and ZW mainly designed the article. ZZ, HW, RH and NZ contributed to the sample collection and collection of patient clinical information. ZZ and HZ did the experiments, analyzed the data, and wrote the initial manuscript draft. KS, ZW, RH, NZ and SW revised the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (82403376).

Data availability

The single-nucleus transcriptome profiles of human subcutaneous adipose tissue used in this study are available from the GEO under accession numbers GSE176171, GSE184687, and GSE225700, and from ArrayExpress under accession number E-MTAB-9199. The corresponding research papers for these datasets are cited as references [25–32]. Relevant omics data are available at the China National Center for Bioinformation (CNCB) under accession numbers OMIX012993 and OMIX012995. (https://www.cncb.ac.cn/). The raw mouse RNA-seq data are deposited at CNCB with accession number CRA035533 (https://www.cncb.ac.cn/). The raw human fat RNA-seq data are available at NCBI BioProject under accession number PRJNA1477840 (https://www.ncbi.nlm.nih.gov/bioproject/PRJNA1477840).The mRNA expression profiles and clinical information of multi-patient breast cancer samples were obtained from TCGA, http://portal.gdc.cancer.gov/) and cBioPortal (https://www.cbioportal.org/).

Declarations

Ethics approval and consent to participate

Tumor samples and clinical data were collected and used with the guidance of the Declaration of Helsinki and approved by the Ethical Committee of Ruijin Hospital, Shanghai Jiao Tong University School of Medicine. Ethics Approval number: (2022) Clinical Ethics Approval No. 2 A. Participants gave informed consent to participate in the study before taking part. All animal procedures were approved by the Animal Ethics Committee of Ruijin Hospital (Ethics Approval number: AUP-202505-04) and conducted in accordance with the NIH Guide for the Care and Use of Laboratory Animals.

Consent for publication

Not applicable.

Competing interests

Zheng Wang is a member of the BMC Medicine Editorial Board.