Decoding adipocyte heterogeneity through single-nucleus transcriptomics unveils subtype-specific adipocytes orchestrate immunosuppressive niches in breast cancer.

Background
Breast cancer is the most common malignancy in women worldwide, which was responsible for over 2 million newly diagnosed cases and over 660,000 deaths in the year 2022.1 Current classification systems stratify breast tumors into distinct molecular subtypes—luminal-like (estrogen receptor (ER)-positive), human epidermal growth factor receptor-2 (HER2)-enriched, and triple-negative breast cancer (TNBC)—based on immunohistochemical profiling of ER, progesterone receptor, and HER2 expression.2 The molecular subtype is an important prognostic and predictive factor towards clinical outcomes and treatment decision-making, independent of tumor stage.2 Advances in molecular profiling and multiomics technologies have revolutionized breast oncology, driving a paradigm shift toward precision medicine. An increasing awareness has been raised of the need to consider the tumor microenvironment (TME) as a whole rather than focusing solely on neoplastic cells.
TME comprises an intricate network of tumor-infiltrating immune cells, such as lymphocytes, macrophages, dendritic cells, and supportive stromal cells including fibroblasts, endothelial cells, pericytes, and extracellular matrix (ECM). Cell–cell and cell-matrix interactions collectively orchestrate a permissive niche for tumorigenesis, invasion, and metastasis.3 As for the breast, adipocytes represent a defining feature of breast TME, comprising a large proportion of mammary stromal volume, a unique characteristic distinguishing breast malignancies from other solid tumors. Previous evidence has demonstrated mutual interactions between cancer-associated adipocytes (CAAs) and tumor cells. On the one hand, the morphology and function of CAAs are altered by tumor cells, characterized by dedifferentiation features and altered secretory functions.46 On the other hand, adipokines, cytokines, matrix metalloproteinases, and metabolites secreted by CAAs, in turn, regulate tumor cells and other TME cells, providing energy for tumor progression and inhibiting antitumor immune responses.710 However, given the great heterogeneity of CAAs, the exact mechanism underlying such mutual interactions remains elusive.
While single-cell RNA sequencing has revolutionized our understanding of cellular heterogeneity in both normal mammary tissue and breast TME,1113 its application to adipocyte biology remains methodologically constrained. The technical challenges posed by adipocytes’ cellular fragility, large lipid droplet content, and susceptibility to mechanical stress during dissociation limit the utility of conventional single-cell approaches.14 Single nucleus RNA sequencing (snRNA-seq), which processes only the nucleus, permits the interpretation of adipocyte complexity at the single-cell resolution.15 16 A previous study by Tang et al compared the cell proportion and function in paired cancer-adjacent and normal adipose tissues in patients with breast cancer using snRNA-seq and showed a decreased lipid uptake phenotype and an inflammatory state of CAA compared with normal adipose (NA).15 Nevertheless, evidence concerning adipocytes’ characteristics in breast-specific TME at the single-cell level remains scarce, and there is currently no literature available that simultaneously showcases the heterogeneity of both adipocytes and breast cancer cells.
In this study, we employed snRNA-seq to systematically profile the stromal compartment of normal breast adipose from healthy donors and peritumoral adipose from treatment-naïve patients with breast cancer stratified by molecular subtype (luminal, HER2+, TNBC). Our analytical framework integrates three key dimensions: First, comparative analysis of normal breast adipose against subcutaneous adipose depots to identify breast-specific adipocyte subsets. Second, high-resolution molecular mapping of cellular heterogeneity across normal and CAA microenvironments. Third, systematic characterization of cell–cell interactions between adipocyte subpopulations and tumor-infiltrating immune cells according to different molecular subtypes. By generating a spatially resolved transcriptomic atlas of human breast adipose ecosystems, this work delineates previously unrecognized mechanisms by which adipocyte subclusters orchestrate immune evasion and metabolic reprogramming in a subtype-dependent manner. These insights would advance our understanding of breast-specific TME dynamics and reveal actionable targets for stromal-directed therapies.

Methods

Study design
The primary aim of this study was to create a comprehensive single-cell atlas of adipose cells in breast cancer, encompassing various molecular subtypes, to better understand the role of adipose in breast-specific TME. We collected CAA tissue from nine treatment-naïve patients with primary breast cancer and NA tissue from three non-cancer donors in Ruijin Hospital, Shanghai Jiao Tong University School of Medicine (figure 1A). CAA was defined as adipose tissue within 1 cm from the tumor margin.17 The molecular subtype was classified according to the immunohistochemical status of ER, progesterone receptor, and HER2, determined by the experienced pathologists from the Department of Pathology, Ruijin Hospital following American Society of Clinical Oncology/College of American Pathologists guidelines.18 19 Detailed clinicopathological characteristics of each patient are listed in online supplemental table S1.

Single-nucleus RNA sequencing

Sample preparation
snRNA-seq was performed in the laboratory of NovelBio Bio-Pharm Technology. Freshly resected adipose tissue samples were snap-frozen in liquid nitrogen for intact nucleus isolation. The nucleus was isolated and purified as previously established: Briefly, the frozen tissue was homogenized in nuclei lysis buffer, which contained 250 mM sucrose, 10 mM Tris-HCl, 3 mM MgAc2, 0.1% Triton X-100 (Sigma-Aldrich, USA), 0.1 mM EDTA, 0.2 U/µL RNase Inhibitor (Takara, Japan). Diverse concentrations of sucrose were used to purify the nucleus before further sequencing.

Library generation and processing
The snRNA-seq libraries were generated using the 10x Genomics Chromium Controller Instrument and Chromium Single Cell 3’ V.3.1 Reagent Kits (10x Genomics, Pleasanton, California, USA). Concisely, nuclei were concentrated to approximately 1,000 nuclei/µL and then loaded into each channel to form single-cell Gel Bead-In-Emulsions (GEMs). After the reverse transcription step, GEMs were broken, and barcoded complementary DNA was purified, amplified, fragmented, A-tailed, ligated with adaptors, and index PCR amplified. The Qubit High Sensitivity DNA assay (Thermo Fisher Scientific) was employed to quantify the final libraries, while their size distribution was ascertained through the utilization of a High Sensitivity DNA chip on Bioanalyzer 2200 (Agilent). All libraries were sequenced on a 150 bp paired-end run by Illumina sequencer (Illumina, San Diego, California, USA).

Collection and re-analysis of publicly available datasets
Peer-reviewed publicly available snRNA-seq datasets were collected and included if meeting the following criteria: (1) known body mass index; (2) individual without obesity, diabetes, or cancer history; (3) female gender; and (4) cellranger data available for re-analysis (online supplemental table S1). Data were merged by taking the intersection of genes. Batch effect correction was accomplished using the mutual nearest neighbors (MNN) method.20

Single-nucleus RNA sequencing data analysis
snRNA-seq data analysis was conducted by NovelBio, using NovelBrain Cloud Analysis Platform (www.novelbrain.com). We employed fastp21 with default parameters to filter out adaptor sequences and eliminate low-quality reads, resulting in clean data. Subsequently, the feature-barcode matrices were derived by aligning the reads to the human genome (GRCh38 Ensembl: V.104) through the utilization of Cell Ranger (V.7.1.0),22 specifically with the intron-including mode tailored to the unique characteristics of snRNA-seq data. Through the cellranger aggr function, we processed the mapped reads to modify and obtain an aggregate expression matrix suitable for further analysis. For cell filtering, we used the Seurat package (V.4.1.1),23 24 applying criteria such as mitochondrial gene expression less than 20%, and gene counts ranging from 200 to 10,000. Doublets were identified using DoubletFinder (V.2.0.3) following parameters pN=0.25, resolution=0.8, and expected doublet ratio estimated according to the 10x Genomics guidelines. Dimension reduction was then performed to construct principal component (PC) analysis and Uniform Manifold Approximation and Projection (UMAP; variable feature=2,000; Max PC=50; PC use=top 10 significant PC). Graph clustering was conducted with the first 10 principal components and a resolution parameter of 0.8. To identify marker genes for each cluster, we applied the FindMarkers function in Seurat, employing the Wilcoxon rank-sum test with the following criteria: log2 fold change >0.25, adjusted p<0.05, and a minimum percentage of cells expressing the gene greater than 0.1. Following initial broad cell type annotation, major lineages such as adipocytes and myeloid cells were extracted for in-depth subclustering. To ensure high fidelity for this high-resolution analysis, we performed a second round of quality control and re-clustering to eliminate lineage-specific artifacts and residual low-quality cells prior to final cell state annotation. To assess transcription factor (TF) regulation strength, we applied the single-cell regulatory network inference and clustering (pySCENIC, V.0.9.5) workflow, using the cisTarget database (https://resources.aertslab.org/cistarget/) and GRNboost.25 Candidate TF-target interactions were selected if their importance score significantly exceeded permuted controls (permutation p<0.05). Those with significant motif enrichment in the target promoters (false discovery rate (FDR) <0.05) were defined as a regulon. A regulon was considered active in a cell if its area under the recovery curve score fell above the 95th percentile of the null distribution (p<0.05). To elucidate the key biological processes underlying each cluster, we performed Quantitative Set Analysis for Gene Expression (QuSAGE) (V.2.16.1) analysis,26 comparing each cluster against all others using hallmark gene sets from the Molecular Signatures Database in combination with Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. To gain insights into the biological significance of marker genes and differentially expressed genes, we conducted Gene Ontology (GO)27 and KEGG analysis.28 29 GO annotations were retrieved from various resources, including NCBI (http://www.ncbi.nlm.nih.gov/), UniProt (http://www.uniprot.org/), and the GO itself (http://www.geneontology.org/). Fisher’s exact test was used to identify statistically significant GO categories.30 For all gene-wise analyses and pathway-level analyses, the Benjamini-Hochberg method was used to correct for multiple comparisons. An FDR cut-off of <0.05 was uniformly applied to define statistically significant genes and pathways.

Pseudotime analysis
We conducted single-cell trajectories analysis employing Monocle3 (https://cole-trapnell-lab.github.io/monocle-release/monocle3/).31 First, within the UMAP-dimensionality-reduced space, a cell–cell similarity graph was constructed using the K-nearest neighbors algorithm. Subsequently, the SimplePPT algorithm was employed to optimize the graph structure, generating a principal graph encompassing backbone paths and branch nodes. The trajectory starting point was manually specified based on the biological context. Using this root node as the origin, Monocle3 calculated pseudotime by measuring the graph-based distance from each cell to the designated root node along the shortest path within the principal graph. These distances were used directly as pseudotime values. In multi-branched scenarios, pseudotime values were reallocated at branch points to maintain the developmental ordering of cells within each branch. Cells unreachable from the root node were labeled by Monocle3 as having “infinite pseudotime”. Monocle3 enables dynamic adjustments, automated smoothing and optimization steps during principal graph construction to ensure that pseudotime assignments align with biological expectations and enhance trajectory continuity. To visualize cellular composition dynamics along the trajectory, cells were ordered by pseudotime and subsequently grouped into 100 bins. Subsequently, based on the pseudotime analysis, we applied branch expression analysis modeling analysis to determine genes associated with specific branch fates.

Cell communication analysis
To facilitate a comprehensive analysis of cell–cell communication molecules, we used CellPhoneDB,32 a public repository encompassing ligands, receptors, and their interactions, for cell communication analysis. Proteins from various clusters across different time points, including membrane, secreted, and peripheral proteins, were annotated. Furthermore, we calculated significant mean and cell communication significance (p<0.05) by leveraging the interactions and the normalized cell matrix obtained through Seurat normalization.
Gene signatures were generated based on subtype-specific (pre-) adipocyte gene features and ligand-receptor pairs in ER+ and TNBC. Samples from the Cancer Genome Atlas (TCGA) database were then classified into High_risk and Low_risk groups, whose differences in immune microenvironment and biological behavior were compared by CIBERSORT and Gene Set Enrichment Analysis (GSEA).

Immunohistochemistry and immunofluorescence staining
Additional NA (N=20) and CAA tissue (N=20 for ER+disease, 10 recurrent and 10 recurrent-free; N=20 for TNBC, 10 recurrent and 10 recurrent-free) from patients with benign and malignant breast diseases were obtained from Ruijin Hospital, Shanghai Jiao Tong University School of Medicine. Immunohistochemistry (IHC) and immunofluorescence (IF) evaluation were accomplished in Wuhan Servicebio Technology.
For the IHC evaluation, tumor formalin-fixed paraffin-embedded slides were de-paraffinized, heated with citrate antigen retrieval solution for antigen retrieval, and incubated overnight at 4°C with primary antibodies, including anti-LEP (Servicebio, GB11309, 1:50) anti-CD34 (Servicebio, GB15013, 1:500), and anti-CD68 (CST, 76 437S, 1:800). Subsequently, the antibody solution was removed, and the slides were washed three times with wash buffer for 5 min each. Afterward, the slides were incubated with a secondary antibody for 60 min at room temperature. The detection of antigens was then performed using an IHC detection kit with 3,3'-diaminobenzidine (DAB), followed by counterstaining with hematoxylin and dehydration.
Primary antibodies applied for IF were anti-LDLR (Servicebio, GB11369, 1:2000), anti-LEP (Servicebio, GB11309, 1:2000), anti-LPL (Proteintech, 28602–1-AP, 1:2000), anti-CD34 (Servicebio, GB15013, 1:5000), anti-TGFBR1 (Proteintech, 30117–1-AP, 1:4000), anti-FABP4 (Servicebio GB115466, 1:3000), anti-KRT8 (Servicebio GB15231, 1:5000), anti-IL-6 (abclonal A21264, 1:2000), anti-p21CIP1 (Servicebio GB11153, 1:1000). After washing with phosphate-buffered saline, the secondary antibody was added before incubation at room temperature for 50 min. Between applying different primary antibodies, the slides were added with tyramide signal amplification dyes, incubated in the dark at room temperature for 10 min, placed in tris-buffered saline with Tween-20 (TBST) and washed three times 5 min each. Cell nuclei were restained with DAPI (Servicebio, G1012) at room temperature in the dark for 10 min. All slides were digitized by whole-slide scanning in a Pannoramic MIDI/250 digital pathology slide scanner (3DHISTECH) and images were taken with the aid of the software CaseViewer (V.2.4). For IF image analysis, fields of view (FOVs) were selected by experienced pathologists to ensure representative sampling. Specifically, for each tissue section from each donor, three non-overlapping FOVs were randomly chosen, covering the tumor center region, adipose region, and tumor-adipose junction region to avoid regional bias. Out-of-focus or artifact-containing FOVs were strictly excluded based on predefined criteria: FOVs were rejected if they showed blurred cellular boundaries (assessed by visual inspection of nuclear and cytoplasmic staining clarity), presence of tissue folding, bubbles, or non-specific staining aggregates (confirmed by comparing with negative control sections). This exclusion process was performed independently by three researchers (YT, ZW, and XF), with discrepancies resolved by consensus. Image analysis was then conducted using ImageJ. Regions of interest (ROIs) were defined as LEP/LPL-positive (for adipocytes) or CD34-positive (for pre-adipocytes). These ROIs were delineated via a segmentation process involving manual thresholding based on LEP/LPL or CD34 signal intensity to generate a binary mask. Prior to quantification, a maximum intensity projection was generated to represent the entire volume in a single two-dimensional image, ensuring consistent analysis across samples. For each FOV, 10 (pre-)adipocyte ROIs were quantified for mean fluorescence intensity (MFI) of LDLR or TGFBR1. To ensure accuracy, background subtraction was applied by subtracting the mean intensity of a cell-free region from the same image from all measurements. To verify reliability, inter-observer agreement was evaluated by having both researchers blindly score 20% of randomly selected FOVs for the presence and density of the adipocyte subpopulation, with an interclass correlation coefficient of 0.75 indicating excellent agreement. All observers were blinded to sample group labels during image acquisition, FOV selection, and quantification to minimize observer bias. Donor-level LDLR or TGFBR1 MFI was calculated as the mean of all ROIs per donor, resulting in a single representative value for each biological replicate.

Results

Single-nucleus profiling of adipose tissues unravels breast-enriched cell subsets
To systematically characterize the cellular composition of breast adipose tissue, we conducted snRNA-seq on fresh breast adipose specimens obtained from three non-cancer donors (designated as NA). These data were benchmarked against publicly available snRNA-seq datasets from subcutaneous adipose tissue (SAT) of three non-obese individuals (figure 1A; online supplemental table S1). Following rigorous quality control, our integrated analysis encompassed 46,852 high-quality nuclei (49.9% from in-house NA samples and 50.1% from public SAT repositories; online supplemental table S2). Inter-dataset batch effects were effectively mitigated using MNN alignment.
Through canonical marker-based annotation, we delineated seven major stromal cell lineages: adipocytes, pre-adipocytes, endothelial cells, myeloid cells, lymphocytes, pericytes, and smooth muscle cells (figure 1B, online supplemental figure S1A–G, online supplemental table S3). The upregulation of ADIPOQ, LEP, and PLIN1 jointly marked adipocyte lineage (online supplemental figure S1A,B). Comparative analysis revealed a statistically significant enrichment of adipocytes in breast NA compared with SAT (p=0.049, figure 1C).
Unsupervised clustering of the adipose compartment unveiled two principal differentiation states, adipocytes and pre-adipocytes, which were further resolved into nine transcriptionally distinct subpopulations (figure 1D,E, online supplemental table S4). The adipocyte lineage comprised mature adipocyte (Adi), and a transitional Adi_Transition subcluster. Pre-adipocyte heterogeneity was captured through seven subclusters, including two SAT-enriched populations (SAT_APOD and SAT_CD248) and five breast NA-predominant subtypes (Pre_Adi_ABCB5, Pre_Adi_EPHA3, Pre_Adi_FABP4, Pre_Adi_LGR4, and Pre_Adi_MMP16). Notably, breast NA exhibited significant enrichment of Adi, Pre_Adi_FABP4, and Pre_Adi_MMP16 subpopulations, while SAT-specific SAT_APOD and SAT_CD248 clusters were virtually absent in breast adipose (figure 1D,E).

Single-nucleus dissection of adipose tissue heterogeneity in breast
Next, we intend to compare the composition and function of adipose in non-cancer and breast cancer individuals. Nine freshly resected treatment-naïve breast tumor-adjacent adipose tissues were analyzed by snRNA-seq: three for TNBC, three for HER2+breast cancer, and three for luminal-like breast cancer (figure 1A; online supplemental table S1). After quality control, a total of 86,529 individual nuclei were included in further analysis (23.8% from luminal-like, 24.7% from HER2+, 25.3% from TNBC, and 26.3% from NA; onlinesupplemental tables S2 S7).
Initial clustering using canonical markers revealed comparable cellular composition between NA and CAA (figure 2A–C, online supplemental figure S2A,B; online supplemental table S3). When further compared among diverse molecular subtypes, Luminal-like breast cancers had a significantly higher proportion of adipocytes in tumor-adjacent tissue (50.8%, compared with 38.2% for HER2+, 33.5% for TNBC, and 47.3% for NA; figure 2D, online supplemental figure S2B, online supplemental table S7). Two major clusters of adipose cells were identified, namely LEP+ adipocytes and CD34+ pre-adipocytes (figure 2E), subsequently validated through IHC (figure 2F, online supplemental figure S2C). Furthermore, three subclusters of adipocytes, which were mature adipocyte (Adi), Adi_Transition, and Adi_LDLR (figure 3A,B, online supplemental table S4), as well as seven subclusters of pre-adipocytes, namely Pre_Adi_ABCB5, Pre_Adi_EPHA3, Pre_Adi_FABP4, Pre_Adi_LDLR, Pre_Adi_LGR4, Pre_Adi_LGR4_TGFBR1, and Pre_Adi_MMP16 (figure 4A,B) were annotated. Pseudotemporal reconstruction delineated a differentiation trajectory from pre-adipocyte progenitors to mature adipocyte states (figure 2G).

Comparison of normal and cancer-associated adipocytes identifies tumor-specific Adi_LDLR subtype
To unravel the heterogeneity of the adipose cell atlas in the breast, proportions of major adipose cell subpopulations among different molecular subtypes of breast cancer and normal adipose were then compared (online supplemental figure S3A). The canonical adipocyte population Adi (LPL+dominant) constituted 85.0% of NA but underwent a dramatic reduction in luminal-like tumors (21.8%, p<0.001; figure 3C,D, online supplemental figure S3B,C), indicating profound dedifferentiation of CAAs, particularly in ER+malignancies. Conversely, the Adi_LDLR subpopulation exhibited preferential accumulation in both ER+ (30.5%) and TNBC (29.7%) peri-tumoral adipose versus minimal presence in NA (2.8%, p<0.001; figure 3D). IF validation exhibited elevated LDLR MFI coupled with diminished LEP/LPL signals in ER+CAAs compared with NA (p=0.009; figure 3E, online supplemental figure S3D), whereas TNBC CAAs maintained LDLR levels comparable to NA (p=0.333). LDLR signal was generally co-localized with adipose markers but not with ductal marker KRT8 (online supplemental figure S3E). Notably, CAAs from relapsed patients displayed intensified LDLR signals in the peri-tumoral adipose region versus relapse-free counterparts in the ER+cohort (p=0.031; online supplemental figure S3E), suggesting Adi_LDLR expansion may correlate with disease progression. LDLR signals in the tumor center region were significantly less intense (p<0.001), and could not predict disease recurrence (p=0.778; online supplemental figure S3E).
Monocle pseudotime trajectory corroborated the transition from Adi_Transition to Adi and to Adi_LDLR (figures2G 3F). SCENIC-based transcriptional regulator analysis revealed bifurcated regulatory programs: Fate-1 branch genes (STAT3, IRF8, KLF4, MAFF) were enriched in “positive regulation of G1/S transition of mitotic cell cycle”, “cell migration”, and “epidermal growth factor receptor signaling pathway”, while Fate-2 branch effectors (LEP, LPL, LIPE, ADIPOQ) governed “lipid metabolic process”, “positive regulation of fatty acid beta-oxidation”, and “triglyceride biosynthetic process” (figure 3G, online supplemental figure S3F,G; online supplemental table S5). This transcriptional dichotomy underscores functional polarization within adipocyte subsets.
Secretome profiling identified Adi_LDLR as a predominant source of tumor-promoting factors, including MMP19, ANXA1/2, COL12A1, and THBS1—all established mediators of cancer progression, as compared with Adi (figure 3H).3339 QuSAGE was then performed to determine the biological functions of the adipocyte subsets. Cell senescence and senescence-associated secretory phenotype gene sets were significantly enriched in Adi_LDLR (figure 3I). IF further revealed that ER+patients with a higher proportion of Adi_LDLR had greater levels of IL-6 and p21 (online supplemental figure S3H), suggesting a disequilibrated senescence feature in the TME.

Pre_Adi_LGR4_TGFBR1 is overexpressed in TNBC and contributes to tumor progression
Among the seven pre-adipocyte subclusters, Pre_Adi_LGR4_TGFBR1, featuring TGFBR1, KIRREL3, and LGR4 upregulation, was significantly overexpressed in CAA (online supplemental figure S4A), especially in TNBC (18.1%, p<0.001; figure 4C,D, online supplemental figure S4B,C), but was rarely present in ER+tumors (1.7%) or NA (1.4%). IF demonstrated a significantly increased staining of TGFBR1 in TNBC patient-derived CAAs compared with ER+CAAs (p=0.034; figure 4E, online supplemental figure S4D) or NA (p=0.011). We also noticed a remarkable increase in MFI of TGFBR1 for patients with relapsed TNBC, as compared with those relapse-free (p=0.038; online supplemental figure S4D), linking Pre_Adi_LGR4_TGFBR1 to more aggressive disease states.
According to the pseudotime trajectory analysis, Pre_Adi_LGR4_TGFBR1 and Pre_Adi_LGR4 were identified as the pre-branch of all pre-adipocyte subsets (figures2G 4F). When compared with Pre_Adi_LGR4, Pre_Adi_LGR4_TGFBR1 exhibited a remarkable gene expression pattern associated with cancer cell growth, progression, and invasiveness (figure 4G, upregulated: NFATC2, LRRFIP1, ELL2, KIRREL3, SERPINE1; downregulated: DCN, ADAMTS5, IGFBP5).4047 As shown by KEGG pathway analysis, the molecular function of Pre_Adi_LGR4_TGFBR1 was marked by an activation of MAPK, PI3K-Akt, Wnt, TGF-β signaling, but a downregulation of fatty acid metabolism-related signaling pathways (figure 4H, online supplemental table S6). In addition, Pre_Adi_LGR4_TGFBR1 secreted higher levels of CSF1, PDGFC, VEGFA, and ADRGE5 (figure 4I),4852 as compared with other pre-adipocyte subclusters. Taken together, this subset of pre-adipocytes lacked normal adipose function and possibly contributed to fueling tumor progression in TNBC.
Besides, ER+tumor-adjacent adipose tissue exhibited a markedly elevated proportion of Pre_Adi_LDLR subpopulation (28.7%, p<0.001; figure 4C,D, online supplemental figure S4B,C), whose presence was validated via IF co-staining of CD34 and LDLR (online supplemental figure S4E). The function of Pre_Adi_LDLR was enriched in programmed death-ligand 1 expression and programmed cell death protein-1 checkpoint pathway, Th17 cell differentiation, T-cell receptor pathway, etc (figure 4J, online supplemental table S5), indicating that Pre_Adi_LDLR aids in forming a possible immunosuppressive niche in ER+breast cancer.

Specific adipocyte-macrophage communication leads to immunosuppressive microenvironment
Within the adipose stromal landscape, the myeloid lineage emerged as the predominant immune component (14,438 nuclei, 16.7%), second only to adipocytes (36,692 nuclei, 42.4%; online supplemental table S7). Six subclusters of myeloid cells were identified, which were Macro_TREM2, Macro_Cycling, Macro_LYVE1_FOLR2, Mono_CD16, Mono_FCN1, as well as Dc_CD1C (figure 5A,B, online supplemental figure S5A–C, online supplemental table S8). The myeloid subclusters were characterized by enriched CD163, MRC1, showing mutual exclusivity with adipocyte progenitor markers CD34, PDGFRA (online supplemental figure S5B,C).
When compared with NA and across molecular subtypes, luminal-like tumor-adjacent adipose tissues had a higher proportion of Macro_Cycling (marked by RRM2, p<0.001; figure 5C,D, online supplemental figure S5D). RRM2 was a reported marker towards M2 macrophage polarization.53 QuSAGE analysis linked the molecular function of Macro_Cycling towards immune checkpoint activation-related pathway, and hypoxia-related pathway (figure 5F, online supplemental table S9).
Furthermore, we intended to explore the differences in characteristics and function of Macro_TREM2, the functional macrophage subset in the TME, between tumor-adjacent adipose and normal adipose tissue. In Macro_TREM2 from ER+patients, we observed an upregulation of genes majorly associated with ECM (COL12A1, COL4A1, COL4A2, LAMA4, SPON1, ADAMTS12, VCAN; online supplemental figure S5E), cell adhesion and migration (COL4A1, COL4A2, SPP1, GLDN, VCAN).5462 KEGG pathway analysis revealed dual regulatory reprogramming in ER+Macro_TREM2 versus normal counterparts, including activated ECM engagement (focal adhesion, ECM-receptor interaction) and suppressed immunoregulatory capacity (B-cell receptor signaling, antigen processing and presentation, Th differentiation, and chemokine signaling pathways; figure 5G, online supplemental table S10). Meanwhile, in Macro_TREM2 from patients with TNBC, we observed an upregulation of genes majorly associated with epithelial-mesenchymal transition (ANKRD28, NHSL1), cell migration and metastasis (NHSL1, FN1, DAB1; online supplemental figure S5F). KEGG pathway analysis also established an increased ECM engagement (upregulated ECM-receptor interaction, focal adhesion pathways) and immunosuppression (downregulated antigen processing and presentation, B-cell receptor signaling pathway, Fc gamma R-mediated phagocytosis, natural killer cell-mediated cytotoxicity pathways) in TN Macro_TREM2 versus normal counterparts (figure 5H, online supplemental table S11). This transcriptional shift suggested that Macro_TREM2 in CAA adopted a protumorigenic phenotype through concurrent ECM restructuring and immune-evasion programming.
We hypothesized that (pre-) adipocyte-macrophage interactions underpin the establishment of immune-evasive niches. We identified 130 and 130 functionally related signaling pathways between adipocytes and macrophages and pre-adipocytes and macrophages, respectively, comprising 982 validated molecular interactions (online supplemental table S12). In ER+tumors, Adi_LDLR was the dominant signaling source, while Macro_TREM2 was among those with the greatest incoming interaction strength (figure 6A,B, online supplemental figure S6A). Adi_LDLR, which, as mentioned above, was significantly increased in luminal-like tumor-adjacent adipose, showed strong potential interactions with Macro_TREM2 via SPP1-Integrin, VEGFA-NRP1, and VEGFA-FLT1 (figure 6C). These ligand-receptor pairs have been previously reported for their functions in immunosuppression, stromal remodeling, angiogenesis, and tumor migration.6164 Stratifying ER+cases in TCGA by Adi_LDLR-associated signatures revealed distinct microenvironmental states, where the low-risk group showed an enhanced cytotoxic immunity (increased CD8+T cells, T follicular helpers, activated natural killer (NK) cells; figure 6D, online supplemental figure S6C), and the high-risk group exhibited ECM remodeling dominance (enrichment in epithelial-mesenchymal transition, and ECM receptor interaction pathways; figure 6E; online supplemental table S13). In TNBC, Pre_Adi_LGR4_TGFBR1 showed the strongest outgoing signal (online supplemental figure S6B) and regulated Macro_TREM2 mainly through LGALS3-MERTK, NGF-SORT1, FN1-Integrin, and LAMC1-Integrin (figure 6F). These ligand-receptor pairs all have established pro-proliferative, pro-metastatic, and immunosuppressive functions.6570 Using a gene signature composed of Pre_Adi_LGR4_TGFBR1 features and corresponding ligand-receptor pairs with TCGA data in TNBC, we found that the Low_risk group had significantly greater infiltration with CD8+T cell, T follicular helper, and activated NK cells, while the High_risk group had more M2 macrophages (figure 6G, online supplemental figure S6D). GSEA analysis showed an enrichment of epithelial-mesenchymal transition, ECM receptor interaction, as well as a downregulation of immune response-related pathways in the High_risk group (figure 6H, online supplemental table S14). To summarize, the (pre-) adipocyte-macrophage communications formed an immunosuppressive niche in breast cancer that would promote tumor progression.

Discussion
In the current study, we intended to explore the heterogeneity of adipose cells in the breast TME by using snRNA-seq techniques. We identified three cancer-specific (pre-) adipocyte clusters Adi_LDLR, Pre_Adi_LDLR, and Pre_Adi_LGR4_TGFBR1, which were virtually absent in normal breast adipose. Adi_LDLR and Pre_Adi_LDLR were significantly enriched in ER+CAA and were functionally related to cell senescence and immunosuppression as shown by secretory protein and pathway analysis. Pre_Adi_LGR4_TGFBR1 was mainly found in TNBC CAA, which was featured by a downregulation of fatty acid metabolism-related signaling pathways and an upregulation of cell proliferation and progression pathways. Both Adi_LDLR and Pre_Adi_LGR4_TGFBR1 were significantly associated with disease development and recurrence, as confirmed by IF staining, suggesting the formation of immunosuppressive niches by these cancer-specific adipose subsets in both subtypes. Cell communication analyses further revealed a specific (pre-) adipocyte-macrophage interaction via ligand-receptor pairs involved in stromal remodeling and tumor migration for ER+tumors, and tumor proliferation, tumor metastasis for TNBC. To our knowledge, this is the first evidence concerning the heterogeneity in cell composition, function, and cell–cell interactions of adipocytes in breast cancer TME of diverse molecular subtypes, which may provide a theoretical basis for future studies.
Adipose cells serve as pivotal regulators within the distinctive TME of breast cancer. With the impact of tumor and stromal cells in TME, NA undergoes malignant transformation into CAA, manifested by dedifferentiation, altered secretory function, increased release of free fatty acids, etc, which provides external energy for tumor cells and promotes cell proliferation and migration.46 Meanwhile, by secreting adipokines, cytokines, matrix metalloproteinases, etc, CAA can alter the biological functions of breast cancer cells, promote stemness, and mediate therapeutic resistance.9 10 Recent advancements in snRNA-seq and single-cell transcriptomic analysis have revolutionized our understanding of TME dynamics at cellular resolution. For instance, Tang et al found that breast cancer induced a fundamental metabolic reprogramming in adipocytes, shifting their phenotype from lipolytic to lipogenic states. This transformation was marked by expansion of lipid biosynthesis-enriched subclusters (ELOVL5, ACSL4, ACSL1, ACSL3, FASN), and contraction of subpopulations associated with oxidative phosphorylation (MT-ND1, MT-ND2, MT-ND3, MT-CO1, MT-CO2) and lipid transport functions (GULP1, NEGR1, ABCA9, ABCA6, ABCA10).15 Another study identified four clusters of (pre-) adipocytes in subcutaneous adipose tissue, namely DEPP1+Pre_Adi, KCND2+Pre_Adi, DPP4+Adipose stem cell, and ADIPOQ+Adi, with the last cluster significantly associated with a poorer prognosis in breast cancer.71 Here in the current study, we established a novel classification of three tumor-enriched (pre-) adipocyte subpopulations—designated as Adi_LDLR, Pre_Adi_LDLR, and Pre_Adi_LGR4_TGFBR1 based on their surface marker profiles—and provided the first evidence of subtype-specific distribution patterns across molecularly heterogeneous breast carcinomas. Our pioneering identification revealed significant variations in the prevalence of these adipocyte clusters among luminal, HER2-enriched, and triple-negative subtypes, suggesting their potential roles in shaping subtype-specific TMEs.
Pre_Adi_LDLR and Adi_LDLR both remarkably augmented in ER+CAA but were scarce in NA. snRNA-seq data suggested that Pre_Adi_LDLR and Adi_LDLR were featured by disequilibrated cell senescence, immunosuppression, and tumor promotion. Pseudotime trajectory analysis corroborated the transition from Pre_Adi_LDLR to Adi_LDLR. Further IF confirmed the association between LDLR staining intensity with ER+cancer development and recurrence. Similar to our findings, previous evidence revealed a relationship between LDLR expression on adipose tissue with adipose dysfunction, systemic inflammation, and insulin resistance.72
Pre_Adi_LGR4_TGFBR1, mainly found in TNBC, was characterized by deficient normal adipose functions and tumor-promoting potency. TGFBR1 staining intensity was significantly associated with TNBC development and recurrence, as established by IF. Several studies have also reported a relationship between TGFBR1 with decreased adipogenesis and impaired metabolism.7375 Pre_Adi_LGR4_TGFBR1 secreted higher levels of CSF1, PDGFC, VEGFA, and ADRGE5, with all these proteins involved in the regulation of TME macrophages, inhibition of apoptosis, stimulation of stromal activation, and epithelial to mesenchymal transition.4852 These findings provide compelling evidence for the TME-remodeling capacity of CAA subpopulations and identify promising therapeutic targets for TNBC treatment.
The dynamic crosstalk between adipocytes and immune cells constitutes a critical regulatory axis within the adipose microenvironment. Our findings align with previous studies indicating that macrophages are the primary immune cell types in adipose tissues.15 71 Specifically, we demonstrated that Macro_TREM2 from ER+patients was functionally associated with increased focal adhesion, ECM-receptor interaction, and cell migration, as well as downregulated immune reactions, suggesting a protumor immunosuppressive TME.5462 Consistent with our findings, Tang et al demonstrated a significant augmentation of MARCO+PLAUR+ macrophages, which was marked by metastasis-driving and ECM degradation gene signatures, and a decrease of HPGDS+SLC40 A1+ macrophages, featured by inflammatory gene signatures in CAA compared with NA.15 Mechanistically, we identified tumor subtype-specific paracrine networks: Adi_LDLR-driven communication via SPP1-Integrin, VEGFA-NRP1, and VEGFA-FLT1 in ER+tumors, and Pre_Adi_LGR4_TGFBR1-mediated regulation through LGALS3-MERTK, NGF-SORT1, FN1-Integrin, and LAMC1-Integrin in TNBC. All these ligand-receptor pairs have been previously reported for immunosuppression, stromal remodeling, and migration-promoting functions.6170 A previous study from Liu et al has identified a special subpopulation of macrophages, namely lipid-associated macrophages, mostly distributed in tumor-adipose junctional regions. These lipid-associated macrophages, which highly expressed macrophage markers (CD163, SPP1), lipid metabolism-related genes (FABP4, LPL), and lipid receptors (LGALS3, TREM2), demonstrated M2-like phenotype and inhibited the antitumorigenic effect of breast cancer treatment.65 Our findings helped us understand the complex paracrine communication between (pre-) adipocytes and macrophages in breast-specific TME, to pave the way for future therapeutic approaches by targeting (pre-) adipocyte-macrophage crosstalk.
While our study provides novel insights into adipocyte heterogeneity within breast cancer, several considerations merit discussion. First, the rather limited cohort size of snRNA-seq may limit statistical power, particularly when stratifying samples by molecular subtypes. Second, the cohort enrolled in this study did not include obese patients (defined as body mass index ≥30). Given that prior studies have well established that the biological function, inflammatory profile, and metabolic state of adipose tissue in obese individuals differ substantially from those in lean or overweight individuals, the findings of this study may not be fully generalizable to patients with breast cancer with obesity. Third, technical constraints inherent to adipocyte biology—specifically their large cytoplasmic volume and membrane fragility—precluded conventional flow cytometry validation, necessitating complementary spatial validation through IF and IHC quantification of CAA versus NA compartmentalization. Next, the absence of parallel single-cell profiling of malignant epithelial cells prevents direct interrogation of adipocyte-tumor cell crosstalk. This knowledge gap could be addressed through emerging spatial transcriptomics platforms that preserve topological relationships between adipocytes, immune cells and adjacent tumor nests. Finally, our findings reveal but do not mechanistically resolve the tripartite interplay between (pre-) adipocytes, macrophages, and breast cancer cells—a complexity demanding future multiomic integrations.

Conclusions
In conclusion, this study illuminated a distinct landscape in the breast cancer adipose microenvironment at single-nucleus resolution, unveiling previously uncharacterized cancer-specific (pre-) adipocyte subpopulations (Adi_LDLR, Pre_Adi_LDLR, Pre_Adi_LGR4_TGFBR1) with subtype-dependent prevalence. The crosstalk between (pre-) adipocytes and adipose tissue macrophages fostered an immunosuppressive and tumor-promoting microenvironment in certain subtypes. These mechanistic insights position adipocyte subpopulation dynamics as actionable biomarkers and potential targets. Future translation should prioritize spatial multiomics validation and adipocyte-centered combination regimens to disrupt protumorigenic niche signaling.

Supplementary material
10.1136/jitc-2025-012711online supplemental figure 110.1136/jitc-2025-012711online supplemental figure 210.1136/jitc-2025-012711online supplemental figure 310.1136/jitc-2025-012711online supplemental figure 410.1136/jitc-2025-012711online supplemental figure 510.1136/jitc-2025-012711online supplemental figure 610.1136/jitc-2025-012711online supplemental table 110.1136/jitc-2025-012711online supplemental table 210.1136/jitc-2025-012711online supplemental table 310.1136/jitc-2025-012711online supplemental table 410.1136/jitc-2025-012711online supplemental table 510.1136/jitc-2025-012711online supplemental table 610.1136/jitc-2025-012711online supplemental table 710.1136/jitc-2025-012711online supplemental table 810.1136/jitc-2025-012711online supplemental table 910.1136/jitc-2025-012711online supplemental table 1010.1136/jitc-2025-012711online supplemental table 1110.1136/jitc-2025-012711online supplemental table 1210.1136/jitc-2025-012711online supplemental table 1310.1136/jitc-2025-012711online supplemental table 1410.1136/jitc-2025-012711online supplemental file 1