Location matters: Spatial-dependent effects of myofibroblasts modulate survival in breast cancer

Summary

While cancer-associated fibroblasts (CAFs) have been historically considered pro-tumorigenic, several studies found that myofibroblast-like CAFs (myCAFs) may have tumor-restraining properties. Here, we show that myofibroblasts exhibit location-dependent effects in breast cancer progression. Deconvolution of bulk RNA sequencing data from The Cancer Genome Atlas showed that high intratumoral myCAF abundance in ER⁺/PR⁺/HER2⁻ breast tumors is associated with poor patient overall survival, whereas an inverse association was observed for peritumoral myCAFs in tumor-adjacent normal tissues. Single-nucleus RNA sequencing of 2 ER⁺/PR⁺/HER2⁻ invasive breast carcinomas and their matched peritumoral tissues identified high abundance of myCAFs in the larger tumor, whereas a distinct myofibroblast-like sub-population was found in the peritumoral tissues of the smaller tumor. Three-dimensional co-culture experiments demonstrated that myofibroblasts positioned outside the tumor spheroid reduced breast cancer cell invasion. These data highlight that the spatial location of myofibroblasts determines their effects and provides an explanation for their dual role in cancer biology.

Graphical abstract

Highlights

• •SnRNA-seq identifies distinct myofibroblast subtypes in peritumoral tissue
• •Intratumoral myCAFs associate with poor survival in ER⁺/PR⁺/HER2⁻ breast cancer
• •Peri-tumoral myCAFs associate with improved overall survival in breast cancer
• •Peri-tumoral myofibroblasts restrain invasion via increased collagen deposition

Immunology; Biological sciences; Cancer

Introduction

Cancer-associated fibroblasts (CAFs) are a major component of the tumor microenvironment (TME) and have attracted great research interest for their roles in tumor initiation and progression.¹ Recent single-cell RNA sequencing (scRNA-seq) studies have identified prognostic signatures and models that predict outcomes in breast cancer patients.^2,3,4 These studies have also provided important insights into CAF heterogeneity, revealing myofibroblast-like CAFs (myCAFs) and inflammatory CAFs (iCAFs) sub-populations across multiple cancer types.^1,5,6,7 The alpha-smooth muscle actin (αSMA)-expressing myCAFs are characterized by an upregulation of extracellular matrix (ECM) remodeling and collagen biosynthesis genes. In contrast, iCAFs are characterized by high levels of cytokines, chemokines, and stem cell markers.⁶ In vitro studies in pancreatic ductal adenocarcinoma (PDAC) found that myCAFs are driven by TGF-β signaling, while iCAFs are regulated by IL1-induced JAK/STAT signaling. Importantly, these subtypes are not fixed and can be reprogrammed and interconverted using TGF-β receptor or JAK inhibitors,^8,9 highlighting the influence of the TME on their functional states.

While CAFs are historically considered pro-tumorigenic through their secretion of ECM proteins, growth factors, cytokines, and exosomes,¹⁰ some studies found that myCAFs inhibit tumor growth in PDAC co-culture organoid models, suggesting tumor-restraining properties.^8,9 Further, in transgenic mouse models of PDAC, depletion of αSMA+ myofibroblasts accelerated tumor progression and decreased survival.¹¹ This study also found that fewer myofibroblasts in tumors of PDAC patients was correlated with reduced survival.¹¹

This tumor-restraining role of myofibroblasts may stem from collagen deposition, which acts as a physical barrier to tumor cell invasion and dissemination.¹ Deleting type 1 collagen (Col1) specifically in myofibroblasts accelerated the emergence of PDAC and decreased overall survival (OS) in a dual recombinase genetic mouse model of spontaneous PDAC.¹² Similarly, Col1 deletion in CAFs in mice injected intrasplenically with different colorectal and pancreatic cancer cell lines led to increased liver metastasis.¹³ However, in breast cancer, which similarly feature a dense reactive stroma,¹⁴ the functional duality of myofibroblasts has yet to be determined.

In our previous study, we performed multi-regional transcriptomic profiling by microarray of 8 estrogen receptor (ER)⁺/progesterone receptor (PR)⁺/human epidermal growth factor receptor 2 (HER2)⁻ breast tumors and peritumoral stroma (up to 7 cm from tumor margins), identifying 2 major peritumoral stromal subtypes: an adipose-enriched subtype, and a myofibroblast-enriched subtype.¹⁵ By deconvoluting tumor-adjacent normal tissue (NAT) RNA sequencing (RNA-seq) samples from The Cancer Genome Atlas breast invasive carcinoma (TCGA-BRCA), we estimated the proportions of these identified subtypes and found that the myofibroblast-enriched subtype was associated with better OS. This led us to hypothesize a tumor-restraining role of peritumoral myofibroblasts in breast cancer.

In this study, we first used deconvolution of TCGA breast tumor and NAT RNA-seq samples to demonstrate that the enrichment of myCAFs within the tumors compared to the surrounding normal regions is differently associated with OS of breast cancer patients in a location-dependent manner. Then, we performed single-nucleus RNA sequencing (snRNA-seq) on 2 ER⁺/PR⁺/HER2^- invasive breast carcinomas, their matched peritumoral tissues, and 1 reduction mammoplasty (RM) sample. From the snRNA-seq analysis, we identified a high abundance of activated myCAFs within the larger tumor and a distinct myofibroblast-like sub-population in the peritumoral tissues of the smaller tumor. Our analyses also revealed distinct transcriptomic profiles between peritumoral and RM adipocytes, indicating that tumors can alter the gene expression of surrounding tissues even at considerable distances. Additionally, we developed a three-dimensional (3D) co-culture model of MCF7 breast cancer cells and BJ myofibroblasts, showing that the presence of peritumoral myofibroblasts restrained cancer cell invasion, concurrent with an increase in extracellular collagen deposition. These findings suggest that myofibroblasts may have tumor-promoting or -restraining roles depending on their location within or surrounding the tumor.

Results

myCAF abundance is associated with OS in TCGA breast cancer patients

Due to conflicting reports in literature on whether myofibroblasts have tumor-promoting or tumor-restraining functions, we hypothesized that their effects may be dependent on their spatial context. To investigate this, we first examined how myCAF abundance within breast tumors or in the normal-adjacent regions is associated with patient survival. CIBERSORTx deconvolution was performed to infer myCAF and iCAF proportions in TCGA-BRCA ER+/PR+/HER2-tumors (n = 419) and NAT (n = 50) RNA-seq samples (Data S1A, S1B, S5, and S6). Patients with high inferred myCAF proportions within the tumor had poorer OS compared to those with lower myCAF abundance (HR = 1.93, 95% confidence interval, CI: 1.09 to 3.40, p = 0.02) (Figure 1A), adjusting for age, tumor size (T stage), lymph node involvement (N stage), and distant metastasis (M stage).

Figure 1.

myCAF abundance is associated with overall survival in TCGA breast cancer patients

(A) Overall survival curves of TCGA BRCA patients with ER⁺/PR⁺/HER2^- tumors (n = 419) divided into high and low myCAF groups based on CIBERSORTx deconvolution proportions of RNA-seq data from breast tumors.

(B) Overall survival curves of TCGA BRCA patients with ER⁺/PR⁺/HER2⁻ tumors and NAT sample available (n = 50) divided into high and low myCAF groups based on CIBERSORTx deconvolution of RNA-seq data from NAT tissues.

On the other hand, higher myCAF abundance in NAT was associated with better OS (HR = 0.124, 95% CI: 0.037 to 0.504, p = 0.003) compared to patients with low NAT myCAF levels (Figure 1B), suggesting that myCAFs surrounding the tumor may restrain cancer progression. Only 1 patient had distant metastasis; this patient had a high myCAF-tumor sample and low myCAF-NAT sample.

Similar results were observed when the analysis was performed on breast tumors across all subtypes (n = 733) and adjusting for ER, PR, and HER2 status (HR = 1.51, 95% CI: 1.01–2.27, p = 0.046) (Data S1C) and NAT (n = 89) obtained from patients across all breast tumor subtypes (HR = 0.332, 95% CI: 0.142–0.776, p = 0.01) (Data S1D).

Therefore, our results indicate that the spatial context of myCAFs in breast tumors influences their impact on cancer progression, with intratumoral (i.t.) myCAFs associated with poorer survival and peritumoral myCAFs linked to better survival, highlighting the importance of distinguishing and understanding the intratumoral and peritumoral environment.

Cellular composition of two ER + PR + HER2-breast tumors and their peritumoral environment

As the previous analyses were based on bulk transcriptomics, we next sought to examine the breast intratumoral and peritumoral microenvironments at a single-cell resolution to understand how they may impact tumor progression. snRNA-seq was performed on frozen tissues from 2 ER+/PR+/HER2-primary breast tumors and their matched peritumoral tissues from patients after mastectomy, and from 1 patient after a cosmetic RM for macromastia (Table 1).

Table 1.

Summary of patient characteristics

	Age	BMI	Tumor Size (mm)	Grade	ER	PR	HER2	Positive lymph node	Local recurrence
Patient 1	62	29.4	75	1	pos (>90%)	pos (50%)	neg (10%)	2 of 41	nil
Patient 2	82	27.7	13	2	pos (99%)	pos (80%)	neg (1%)	0 of 1	nil
RM	52	32.3	–	–	–	–	–	–	–

Tumor samples had been obtained from multiple regions of each tumor mass as previously described,¹² and were combined into a single tumor sample per patient (Tum1 and Tum2 in Figure 2A). Similarly, peritumoral samples surrounding the tumor (1–7 cm from tumor margins) had been obtained and were combined into a single peritumoral sample for each patient (Peritum1 and Peritum2 in Figure 2A). Five non-directional samples were also obtained from the RM patient and combined into a single RM sample. Patient 1 tumor had a diameter of 75 mm while patient 2’s tumor had a diameter of 13 mm.

Figure 2.

Cellular composition of two ER⁺PR⁺HER2^- breast tumors and the peritumoral environment

(A) Schematic representation of the breast tumor and peritumoral samples obtained from each patient for snRNA-seq. Cross indicates sample was collected for snRNA-sequencing. Peritumoral samples had been collected 1–7 cm from the tumor margins.

(B) UMAP visualization of all 25,150 cells obtained from all five samples and colored by cluster number.

(C) UMAP visualization of all cells colored by cell type annotation.

(D) Bar plots showing the number of cells (left) and proportion (right) of each cell type within each sample.

(E) Bar plots showing the percentage of different NK/T cell subsets within each sample.

(F) Bar plots showing the percentage of different myeloid cell subsets within each sample.

(G) Inferred copy number alteration regions in luminal and basal cells estimated by inferCNV and bar plots showing the number of inferred normal and malignant cells in tumor 1.

(H–J) Peritumor 1, (I) Tumor 2, and (J) Peritumor 2. Red indicates amplified regions and blue indicates deleted regions.

(K) Pathway enrichment results for genes significantly upregulated in Peritum1 adipocytes compared to RM adipocytes.

(L) Pathway enrichment results for genes significantly upregulated in Peritum2 adipocytes compared to RM adipocytes.

In total, 25,150 high-quality nuclei were obtained across all 5 samples and 17 clusters were found (Figure 2B), which were annotated by cell type (Figure 2C) through the expression of manually curated canonical cell type marker genes (Data S2A and S2B). As the expression of canonical immune cell type markers were generally low across all clusters (Data S2A), SingleR automatic cell type annotation²¹ was performed using the Human Cell Primary Atlas dataset²² and a TNBC scRNA-seq dataset² as reference (Data S2C and S2D), identifying, cluster 9 as myeloid cells, cluster 10 as natural killer (NK)/T cells, and cluster 16 as B cells. Notably, many of the top upregulated genes for each cell type were not classical cell type marker genes (Data S2E), suggesting biological differences between cytoplasmic and nuclear RNA pools. All annotated cell types were detected in each sample with varying proportions (Figure 2D and Data S3).

Further sub-clustering of the NK/T cell cluster identified 3 clusters (Data S2F). Expression of established T cell markers were very low across all clusters (Data S2G) because their low abundance in the nucleus; hence, we were unable to differentiate them into CD4⁺, CD8⁺ or regulatory T cells. High expression levels of cytolytic genes (GNLY, KLRF1, NCAM1, KLRD1) were identified in one of the clusters (Data S2H), while the remaining 2 clusters showed high expression of T cell-specific genes (THEMIS, IL7R) (Data S2I). One T cell cluster expressed significantly higher levels of CELF2 and PTPRC involved in T cell activation and signaling^30,31 and was annotated as activated T helper cell (Data S2J). The Peritum2 sample showed the highest percentage of activated T helper cells and cytolytic NK/T cells (Figure 2E). Sub-clustering of the myeloid cluster was similarly performed (Data S2K), and cells were annotated as M1 or M2 macrophages, or dendritic cells based on the expression of established markers (Data S2L and S2M). Peritum2 also showed the highest percentage of dendritic cells and M2 macrophages (Figure 2F).

Using the inferCNV (copy number variation) method,²³ we estimated the presence of large-scale CNVs within all luminal hormone-responsive (LumHR), luminal secretory (LumSec), and basal cells in the tumor and peritumor samples, which are indicative of malignant cells (Figures 2G–2J). A high number of malignant LumHR cells were identified in both Tum1 and Peritum1 samples while no CNVs were identified in the LumSec and basal cells of these samples (Figures 2G and 2H). On the other hand, malignant LumHR, LumSec, and basal cells were detected in Tum2, with few malignant cells detected in Peritum2 (Figures 2I and 2J).

Lastly, we compared the transcriptomic profiles of peritumoral adipocytes against those from the RM sample to investigate the influence of tumor on the surrounding tissue. Differentially expressed genes (DEGs) were identified (log2FC > 0.1, adj. p < 0.05) (Data S8), followed by pathway enrichment analysis on the top 200 DEGs and gene set enrichment analysis (GSEA) using the Hallmark and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases.

The ECM organization Reactome pathway, which included genes involved in ECM synthesis (e.g., COL1A1, COL3A1, ELN, MFAP5), ECM degradation (MMP7), and integrin-mediated assembly (e.g., LAMA3, LAMB3, LAMC1), was upregulated in Peritum1 adipocytes compared to RM adipocytes (Figure 2K). On the other hand, genes involved in cell-cell adhesion (e.g., CDH1, CDH13), and cell-substrate adhesions (e.g., FBLN2, ITGB8, TNC) were downregulated in Peritum1 adipocytes compared to RM adipocytes (Data S2N). Apelin signaling was upregulated in Peritum1 adipocytes (Figure 2K), and previously reported to inhibit adipogenesis of pre-adipocytes and lipolysis in mature adipocytes.³² GSEA using Hallmark pathways identified an enrichment of estrogen response (Data S2O), which has been shown to reduce adipogenesis in adipocytes.³³ These pathways may have contributed to a downregulation in fatty acid metabolism and steroid hormone biosynthesis as identified by GSEA using KEGG pathways (Data S2P).

Similarly, Peritum2 adipocytes, compared to RM adipocytes, upregulated genes involved in ECM synthesis (e.g., COL1A1, COL1A2, COL3A1) and ECM degradation (e.g., MMP2, ADAMTS5) (Figure 2L), whereas genes involved in focal adhesion and cell junction assembly were downregulated (Data S2Q). Interestingly, immune activation and immune response-regulating signaling pathways, such as IL2-STAT5 signaling, interferon response, and NK cell-mediated cytotoxicity, were enriched in Peritum2 adipocytes compared to RM adipocytes (Figure 2L, Data S2R, and S2S).

Altogether, these findings highlight not only that the two tumors exhibit substantial genomic diversity but that these differences could influence the impact that the tumor has on peritumoral adipocytes, suggesting that their tumor-mediated altered secretome modulates cancer progression. Moreover, these two patients displayed notable variations in the cellular environments surrounding and within the tumors, including the stromal and immune components.

Re-clustering of fibroblasts reveals distinct fibroblast subtypes

We next sought to examine fibroblast heterogeneity in more detail to investigate whether there is intratumoral or peritumoral-specific subpopulations, which may add insights in their association with different patient prognosis. Subclustering of fibroblasts identified 5 clusters of at least 100 cells each (Figure 3A). We first investigated whether these clusters reflect myCAFs and iCAFs classification, as previously identified in scRNA-seq studies on breast tumors. Using the AUCell method, we performed scoring across the clusters using 2 published myCAF and iCAF gene signatures (Data S9),^2,3 but did not identify a clear dichotomy in CAF subtypes (Figures S3A and S3B). DEGs for each cluster compared to all the other clusters were identified (log2FC > 0.5, adj. p < 0.05) and pathway enrichment analysis was performed on top 200 genes (Figure 3B, Data S3C–S3G, and S10).

Figure 3.

Re-clustering of fibroblasts reveals distinct fibroblast subtypes

(A) UMAP visualization of 2,003 fibroblasts obtained from all five samples and colored by subtype.

(B) Heatmap showing the top 10 DEGs for each fibroblast cluster compared to all other fibroblasts.

(C) Bar plots showing the proportion of each fibroblast subtype within each sample.

(D) Heatmap showing the communication probabilities of key interactions between fibroblasts and epithelial cells within the tumor or immune cells within the same sample.

(E) Graphical summary of patient 1 and 2 tumors and peritumoral tissues.

Twelve genes were significantly upregulated in cluster 0 but 117 genes were downregulated, many of which were involved in ECM synthesis (e.g., COL1A1, COL3A1), ECM degradation (e.g., MMP2, MMP7), and integrin signaling. This cluster was named “quiescent fibroblasts” and made up approximately 50% of fibroblasts in all samples except for Peritum2 (34.2%) (Figures 3B, 3C, Data S3C, and S6). Cluster 1 showed an upregulation of Rho GTPase signaling and enrichment of genes involved in ECM synthesis (e.g., CCN1, EXT1) and ECM remodeling (e.g., ADAMTS3, ADAMTS5). As fibroblasts from this cluster were present in all samples, including RM, they were termed “maintenance fibroblasts” (Data S3D).

Approximately 30% of Tum1 fibroblasts were found in cluster 2 (Figure 3C), which was enriched for genes involved in ECM synthesis (e.g., COL8A1, COL11A1) and published CAF markers (POSTN, FN1, FAP) and was named “activated (Tum1) fibroblasts” (Data S3E). Cluster 3 fibroblasts, which were mainly present in the Peritum2 sample, showed adipocyte-like features and upregulation of triglyceride metabolism genes (FABP4, PLIN1, ADIPOQ) (Data S3F). The top upregulated gene in this cluster is PDE3B, which is known to play a key role in regulating lipolysis, energy homeostasis, and insulin secretion.¹⁶ As PDE3B is typically expressed in cells such as adipocytes and hepatocytes, this suggests that the origin of these fibroblasts may be adipocyte-derived.

Lastly, cluster 4 was enriched for genes involved in ECM synthesis (e.g., COL1A1, COL3A1, COL6A2) and insulin-like growth factor binding proteins (IGFBPs) (e.g., IGFBP3/5/6), and was labeled “activated (Peritum2) fibroblasts” (Data S3G). Interestingly, CD34 was also found to be upregulated in these fibroblasts. While CD34 is commonly used as a marker to identify hematopoietic stem cells, CD34/vimentin dual-positive reactive fibroblasts have been observed in the cancer microenvironment of various human cancers, including breast, colon, lung, and pancreatic cancers.¹⁷ In one study, CD34⁺ and SMA+ CAFs were found to be an independent prognostic factor for progression-free survival and OS in stage 1–3 non-small cell lung cancer.¹⁸ Moreover, stromal loss of CD34⁺ fibroblasts was significantly associated with lower overall and disease-free survival rates in invasive lobular carcinomas.¹⁹ Further research would be interesting to fully elucidate their functional role in tumor progression.

We next performed DEG analysis across samples, comparing how the transcriptomic landscape of intratumoral and peritumoral fibroblasts differed from each other and from RM fibroblasts (Data S10). Pathways enriched in Tum1 fibroblasts compared to Peritum1 and RM fibroblasts were mainly related to ECM synthesis and assembly, collagen biosynthesis, and TGF-β signaling (Data S3H and S3I), reflective of the presence of more activated myCAF-like fibroblasts in this sample. Fibroblasts from the Peritum1 sample also showed upregulation of Rho GTPase signaling, estrogen signaling, and tight junctions relative to RM fibroblasts (Data S3J).

Interestingly for patient 2, the Tum2 sample had a more similar distribution of fibroblast subtypes as the RM sample compared to Peritum2, which had a significant proportion of cluster 4-activated fibroblasts. Pathway enrichment analysis identified an upregulation of ECM organization genes in both intratumoral and peritumoral fibroblasts compared to RM (Data S3K and S3L); however, there were key differences in the types of collagens upregulated between them. These include an upregulation of COL7A, COL15A1, and COL27A1 in Tum2 fibroblasts, whereas COL1A1, COL1A2, and COL3A1 were more highly expressed in Peritum2. While collagens I and III are classical fibrillar collagens seen in fibroblasts-rich interstitial tissues, type 7 and 15 collagens are non-fibrillar collagens with minimal association to CAF activity. Collagen VII localizes to the epidermal-dermal junction in skin, where it plays a key role in skin integrity.²⁰ Collagen XV is primarily found in the basement membrane and helps maintain tissue integrity, particularly in vascular walls and skeletal muscle.²¹ Although type 27 collagen is a novel member of fibrillar collagen subfamily, its role and function have been primarily characterized in cartilage development, rather than associated with CAFs.²² Furthermore, upregulation of IGFBPs and immune-related pathways such as inflammatory response, TNFα signaling, IL6/JAK/STAT3 signaling, chemokine signaling, and antigen processing and presentation, were specifically identified in Peritum2 fibroblasts compared to Tum2 or RM (Data S3L and S3M).

To further understand how these fibroblast sub-populations interact with other cells within the TME, the CellChat tool²⁷ was used to infer cell-state specific signaling communications based on known ligand-receptor interactions (Data S11). We focused specifically on the interactions of fibroblasts with LumHR, LumSec, and basal cells within the respective tumors. The highest communication probabilities were identified between collagens and fibronectin expressed by Tum1-activated and Peritum2-activated fibroblasts with CD44 expressed by the epithelial cells within the respective tumors (Figure 3D). However, strong interactions between these ECM proteins produced by Peritum2-activated fibroblasts with SDC4 in Tum2 epithelial cells were specifically identified as well (Figure 3D). Although integrins remain the principal receptors for cell attachment, SDC4 has been shown to contribute to focal adhesion formation and to reduce breast cancer cell invasion in a 3D collagen matrix.³⁴ This finding suggests that the ECM components from peritumoral fibroblasts might interact with cancer cells via SDC4, potentially altering their migratory behavior. Additionally, several interactions, which have been reported to promote cell proliferation and migration were specifically identified between Tum1-activated fibroblasts and LumHR cells within the tumor (MPZL1-MPZL1, JAG1-NOTCH2, THBS1-CD47) (Figure 3D). Several unique interactions with immune cells were also identified between Peritum2-activated fibroblasts and immune cells of the same sample. The LPAR1-ADGRE5 and CD99⁻CD99 interactions may be involved in promoting immune cell infiltration and contributed to the higher proportions of immune cells observed in this sample (Figures 2E and 2F), whereas the COL3A1-ADGRG1 interaction between Peritum2 fibroblasts and cytolytic NK/T cells may negatively regulate NK cell effector functions.^35,36

Therefore, our snRNA-seq analysis revealed 5 unique fibroblast sub-populations within breast tumors, peritumoral, and RM tissues with distinct transcriptomic features and interactions with surrounding cells. Our findings demonstrate the heterogeneity in both intratumoral and peritumoral cell type composition between these two patients (Figure 3E), that may have contributed to differences in their outcomes. Patient 1 who had a larger tumor (75 mm) and more invading malignant cells in the peritumoral regions, showed high levels of an activated myCAF-like sub-population within the tumor and low immune cell infiltration. In contrast, patient 2 who had a smaller tumor (13 mm) and few peritumoral malignant cells detected had higher levels of myofibroblasts and immune cells in the peritumoral region.

Peritumoral myofibroblasts reduce breast cancer cell invasion in 3D in vitro co-culture

We examined in vitro whether the location of myofibroblasts influences cancer cells proliferation and migration. Using the hanging drop technique and OrganiX 3D tissue culture platform, we developed a 3D co-culture model of breast cancer cells and myofibroblasts to recapitulate the intratumoral and peritumoral microenvironments (Figure 4A).

Figure 4.

Peritumoral myofibroblasts reduce breast cancer cell invasion in 3D in vitro co-culture

(A) Schematic representation of experimental set up.

(B) Representative confocal images (10× magnification) of spheroids composed of MCF7 cells alone or with BJ cells after 48 h of hanging drop formation and 72 h in OrganiX followed by staining with anti-Ki67-FITC antibody. Scale bars, 300 μm.

(C) Percentage of Ki67-positive cancer cells and (D) total number of cancer cells as analyzed using Imaris software in 3 independent repeats of the experiments.

(E) Representative confocal images (10× magnification) of spheroids composed of MCF7 cells alone or with BJ cells after 48 h of hanging drop formation and 72 h in OrganiX followed by staining with collagen pan anti-body. Scale bars, 300 μm.

(F and G) Percentage of invading cancer cells and (G) dispersion of cancer cells as analyzed using Imaris software in 3 independent repeats of the experiments. Mann-Whitney U test p values are shown.

(∗p ≤ 0.05, ∗∗p ≤ 0.01, ∗∗∗p ≤ 0.001). Means with standard deviations are shown.

We formed spheroids composed of MCF7 ER⁺/PR⁺/HER2^- breast cancer cells alone or in co-culture with BJ fibroblasts from normal human foreskin. BJ fibroblasts were selected as they possess a myofibroblast-like phenotype and express higher levels of myCAF markers compared to iCAF markers based on publicly available RNA-seq data (Data S4A) and as shown in our reverse transcription quantitative PCR (RT-qPCR) validation with selected marker genes (Data S4B). A 1:2 ratio of MCF7 to BJ was used in the hanging drops to consistently form spheroids with high viability over 2 days, whereas a high number of dead cells was observed on day 4 (Data S4C–S4E), probably because of the lack of sufficient nutrients to support their fast growth in the hanging drop for more than 2 days. Therefore, the spheroids considered for the analysis were formed in 2 days.

Once formed, the spheroids were seeded into the central hydrogel port of the OrganiX inserts, embedded either in collagen alone (3 mg/mL rat tail collagen I), or collagen mixed with BJ fibroblasts (Figure 4A). After 72 h, cells were stained with DRAQ7 dye to ensure that the selected culture conditions were optimal for cell survival and we observed very low levels of DRAQ7-positive cells (<5% of cells) (Data S4F and S4G). As a positive control for cell death, two spheroids from each condition were treated with 20% ethanol into the media throughout the 72 h. Interestingly, we observed that co-culture spheroids were more resistant to ethanol-induced cell death compared to MCF7 only spheroids, irrespective of whether peritumoral BJ were present (Data S4F and S4G), suggesting that intratumoral fibroblasts may play a role in protecting cancer cells from apoptosis.

We immunostained the cells for Ki67 to examine their proliferation after 72 h in culture within the OrganiX inserts and found that spheroids with intratumoral BJ had significantly lower percentage of Ki67-positive cancer cells compared to MCF7-only spheroids (Figures 4B and 4C), similar to what we observed after spheroid formation in hanging drops (Data S4E). During hanging drops, the presence of fibroblasts in co-culture may have affected the proliferative rates of cancer cells due to competition for nutrients. Hence, subsequent analyses focused on comparing the effects of having or not peritumoral fibroblasts around the spheroids in either monoculture or coculture condition, rather than the effect of intratumoral fibroblasts on the cancer cells.

The presence of peritumoral BJ cells did not show a significant effect on cell proliferation as measured by the percentage of Ki67-positive cancer cells after 72 h in 3D co-culture, irrespective of whether there were intratumoral BJ cells (Figure 4C). However, the total number of cancer cells were reduced when the spheroids were co-cultured in 3D with BJ peritumorally (Figure 4D), indicating that cell survival might be influenced.

Moreover, we found that peritumoral BJ significantly reduced the percentage of cancer cells, which invaded into the surrounding collagen matrix after 72 h (Figures 4E and 4F), and this effect was more pronounced when intratumoral BJ cells were present. Dispersion of cancer cells, which reflects their spread relative to the centroid of the spheroid, confirmed a significantly reduced invasion of cancer cells when co-cultured with BJ peritumorally (Figure 4G).

To understand how peritumoral fibroblasts might lead to reduced cancer cell invasion, immunofluorescence using a human collagen pan antibody was performed after 72 h of incubation in the OrganiX plate without cell permeabilization to visualize extracellular collagen. No staining was observed in wells with hydrogel only (Data S4H). The MCF7-only spheroid conditions with and without peritumoral BJ showed relatively sparse and weak collagen staining (Figure 4E). While the MCF7+BJ co-culture spheroid without peritumoral BJ showed stronger collagen staining, this was similarly sparse and dispersed. On the other hand, the condition with both intratumoral and peritumoral BJ showed strong, dense, and continuous collagen staining, which enveloped the spheroid, suggesting that an increased collagen deposition acting as a physical barrier to impede cancer cell invasion may be one of the tumor-restraining mechanisms of peritumoral myofibroblasts.

To validate our findings, all cells from the wells—including spheroids and BJ fibroblasts embedded in the gel—were dissociated at the end of the experiment, and RNA was extracted for quantification of several collagen genes by RT-qPCR. We observed significantly increased expression of COL1A1, COL3A1, COL4A2, and COL5A2 in conditions, which contained peritumoral BJ fibroblasts compared to those without (Data S4I). Notably, COL1A1 and COL3A1, which encode fibrillar collagens involved in structural matrix deposition, showed the most pronounced upregulation when both intratumoral and peritumoral BJ fibroblasts were present. In contrast, the expression of COL4A2 and COL5A2, which correspond to non-fibrillar (type 4) and regulatory fibrillar (type 5) collagens, was comparatively less pronounced. These results suggest that the increased collagen staining observed is primarily driven by fibrillar collagens, which may contribute to a more restrictive ECM.

To identify the source of collagen at the tumor-stroma interface, we also performed a fibroblast isolation step using anti-fibroblast microbeads and assessed collagen gene expression by qPCR in BJ fibroblasts isolated from both the spheroids and surrounding gel (Data S4J). The results showed that COL1A1 and COL3A1 were highly expressed by BJ fibroblasts, whereas COL4A2 and COL5A2 were expressed at lower levels, consistent with the qPCR data from unsorted whole-well cell populations. Notably, collagen expression was further elevated in BJ fibroblasts isolated from 3D co-culture conditions compared to those grown in 2D monolayer culture. These findings indicate that the fibrillar collagens observed at the tumor-stroma interface are predominantly produced by fibroblasts rather than MCF7 tumor cells.

Discussion

In this study, we have investigated whether the location of myofibroblasts intratumoral or peritumorally in breast cancers determine if they have tumor-promoting or -restraining effects. Taking a multi-model approach by performing deconvolution and survival analyses in the TCGA-BRCA cohort, snRNA-seq of 5 samples, and 3D co-culture experiments with MCF7 and BJ cells, we demonstrate that peritumoral myofibroblasts may be able to restrain cancer cell invasion through increased ECM deposition, leading to better prognosis among patients with higher abundance of peritumoral myofibroblasts.

While several studies involving PDAC mouse models suggested that collagen deposition by myofibroblasts may act as a mechanical barrier for tumor cell invasion and dissemination,¹ a dense ECM can also activate integrin signaling in cancer cells and promote their survival while creating a barrier to anti-cancer drugs and immune cell infiltration, thus playing a dual role in cancer.^23,24 Moreover, fibroblasts in the TME also exert tumor-promoting effects by secreting soluble factors such as growth factors, cytokines, and chemokines.^1,23 Our findings from this study suggest that peritumoral myofibroblasts, which surround the tumor externally, may have more pronounced tumor-restraining effects which overcome their pro-tumor effects, leading to improved OS observed among patients with myCAF-enriched NAT. On the other hand, the pro-tumor effects may be dominant for intraumoral myofibroblasts, which are in close or direct contact with cancer cells, leading to poorer survival among breast cancer patients with myCAF-enriched tumors.

In our snRNA-seq analyses, we identified 5 fibroblast sub-clusters which differed in their enriched pathways and distribution across the samples. While the majority of fibroblasts across all samples were mainly quiescent and maintenance fibroblasts, two activated myofibroblast-like sub-populations were identified, one of which was mainly found in the larger-sized Tum1 while the other was primarily found in Peritum2. While both showed an upregulation of ECM genes, which were strongly predicted to interact with CD44 on tumor epithelial cells, there were distinct differences between them as well. For example, several commonly used CAF markers such as FAP, POSTN, and ACTA2^25,26 were more highly expressed in Tum1-activated fibroblasts than those from Peritum2. Several cell proliferation- and migration-promoting interactions were also specifically identified between Tum1-activated fibroblasts and LumHR cells within the tumor,^27,28,29 supporting their tumor-promoting role.

On the other hand, one of the unique characteristics identified in Peritum2-activated fibroblasts, largely found in the peritumoral tissues of the smaller tumor, was an upregulation of insulin-like growth factor (IGF)-binding proteins (IGFBP-3/4/5/6/7), which may be induced by elevated local production of TGF-β1³⁰. In particular, IGFBP3 has been shown to play a crucial role in fibroblast-to-myofibroblast differentiation of prostatic stromal cells and synergistically potentiate TGFβ1-mediated stromal remodeling.³⁰ Other IGFBPs were also shown to induce ECM production, and interact with integrins to decrease FAK signaling and inhibit IGF1R signaling in lung cancer cells due to their ability to sequester IGFs, thus functioning as tumor suppressors.^30,31,32,33 Another possible mechanism by which these fibroblasts might influence tumor behavior is through interactions between secreted ECM molecules and SDC4 on cancer cells. Although integrins are the major mediators of cell-ECM adhesion, SDC4 has been implicated in focal adhesion dynamics and may contribute to the modulation of cell adhesion and migration. Some studies have suggested that SDC4 can reduce breast cancer cell invasion in 3D collagen matrices^34,35,36 and its higher expression has been associated with better relapse-free survival in breast cancer patients across all molecular subtypes.³⁷ These observations raise the possibility that SDC4-mediated interactions could play a role in the tumor-suppressive effects of peritumoral fibroblasts, although further investigation is needed to clarify this mechanism.

Limitations of the study

While we have highlighted some of the key differences between these 2 myofibroblast-like sub-populations and provided a list of DEGs between them (Data S10), further studies with larger sample sizes will be needed to validate which genes can be used as biomarkers to precisely distinguish between these two subtypes. Moreover, though our in vitro results support our in-silico findings pointing to tumor-restraining functions of peritumoral myofibroblasts, a caveat that needs to be highlighted is that BJ fibroblasts used are not breast-derived. These cells were chosen for their availability and stability during model development. Future studies using primary mammary fibroblasts or patient-derived breast CAFs will be valuable for validating our findings in an organ-specific context. While there may be some tissue-specific interactions between breast cancer cells and mammary fibroblasts, the tumor-suppressive effects observed with BJ fibroblasts may still apply to fibroblasts from other origins that exhibit a myofibroblast phenotype and high ECM production.

Nonetheless, one noteworthy feature of the study is that the snRNA-seq analysis has provided an additional advantage of capturing nuclei from adipocytes, which is currently lacking in scRNA-seq breast studies, due to adipocytes incompatibility with conventional droplet-based single-cell technologies.^38,39 We found that peritumoral adipocytes downregulated lipid metabolism-related pathways, but upregulated ECM-related genes and immune activation pathways compared to RM adipocytes, further highlighting how the tumor can impose its influences on the surrounding tissue over a substantial distance. Previous studies have indeed reported histological changes to tumor-adjacent adipocytes including adipocyte de-differentiation, loss of lipids, and acquisition of fibroblast-like features,⁴⁰ while in vitro co-culture of breast cancer cells with adipocytes identified reduced lipid droplets, ECM modifications, and increased secretion of cytokines and growth factors.⁴¹ Moreover, these tumor-modified adipocytes may acquire a fibroblast-like morphology and represent an intermediate form of emerging CAFs.^40,42 We indeed identified adipocyte-like fibroblasts in our study, which may reflect their dedifferentiation toward fibroblast-like cells.

In summary, our results suggest that myofibroblasts have tumor-promoting or -restraining roles depending on their location within or surrounding the tumor, which may be one reason why clinical trials involving widespread depletion of CAFs based on traditional CAF markers have been unsuccessful.^43,44 Further studies with much larger sample sizes, spatial transcriptomics, or by profiling myofibroblasts isolated from tumor and peritumoral tissues using fluorescence-activated cell sorting, are still required to identify specific cell surface markers for selective targeting of tumor promoting-CAFs only. Nevertheless, given this new paradigm that the location of myofibroblasts affects their functions, novel drug delivery systems capable of penetrating through the dense peritumoral ECM and specifically targeting i.t. myofibroblasts might be therapeutically beneficial and can be explored.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Kanaga Sabapathy (kanaga.sabapathy@ntu.edu.sg).

Materials availability

This study did not generate new unique reagents.

Data and code availability

The raw and processed snRNA-seq data generated in this study have been deposited in the Gene Expression Omnibus database (GEO) under GSE271813. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This research was supported by the: National Medical Research Council Singapore; NCC Cancer Fund (all to K.S.); and NCC Research Fund (NCCSPG-YR2015-JAN-7) (to K.S. and V.K.M.T.). H.S.H.L. is supported by the NUS Research Scholarship and NUS Medicine Graduate Studies Award. G.A. is supported by the Singapore Immunology Network (SIgN), Agency for Science, Technology and Research (A∗STAR) and by the A∗STAR Skin Research Labs (A∗STAR SRL). We thank Dr. Giulio Giustarini and Bi Jiaming from G.A. lab for their assistance in using the confocal microscope and helpful suggestions.

Author contributions

K.S. and H.S.H.L. conceived the idea of evaluating the spatial-dependent roles of myofibroblasts in breast cancer; H.S.H.L. and G.A. designed the 3D co-culture experiments; H.S.H.L. performed the analyses with the help of A.G. H.S.H.L. and I.P. performed the experiments and prepared the figures; V.K.M.T. coordinated the collection and provided the tissues used in the study; H.S.H.L., G.A., and K.S. wrote the manuscript with input from all others. All authors reviewed and commented on the manuscript.

Declaration of interests

G.A. is co-inventor of the OrganiX^TM plate (International patent application WO 2022/197254 A1), licensed to AIM Biotech Pte Ltd.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Mouse Alexa Fluor 647 anti-human Ki67 antibody	Invitrogen	Cat# MA5-44140; RRID:AB_2913072
Rat anti-Ki-67-FITC (clone SolA15)	Invitrogen (eBioscience)	Cat# 11-5698-82; RRID:AB_11151330
Rabbit anti-Collagen-Pan	Invitrogen	Cat# PA1-85324; RRID:AB_930067
DyLight 488 Donkey anti-rabbit IgG (Poly4606)	Biolegend	Cat# 406404; RRID:AB_1575130
Biological samples
Primary breast cancer tumor tissue and peri-tumoral breast tissue (snRNA-seq; 2 patients, female, ER + PR + HER2−)	National Cancer Center Singapore (NCCS)	CIRB2015/2722
Reduction mammoplasty tissue (snRNA-seq; 1 subject)	National Cancer Center Singapore (NCCS)	CIRB2015/2722
TCGA breast tumor bulk RNA-seq data (n = 733 tumors; n = 89 NAT)	UCSC Xena	GDC TCGA Breast Cancer (BRCA)
Chemicals, peptides, and recombinant proteins
Dulbecco’s Modified Eagle Medium (DMEM)	Sigma-Aldrich	N/A
Fetal Bovine Serum (FBS)	ThermoScientific	Cat#10082147
Penicillin/Streptomycin 100 U/mL/100 μg/mL	Gibco	Cat#15140122
Sodium pyruvate	Gibco	Cat#11360070
GlutaMAX supplement	Gibco	Cat#35050061
0.25% Trypsin-EDTA	Gibco	Cat# 25200056
CellTracker Green CMFDA Dye	ThermoFisher	Cat# C2925
CellTracker Orange CMRA Dye	ThermoFisher	Cat# C34551
CellTracker Deep Red Dye	ThermoFisher	Cat# C34565
Rat tail collagen type I gel	Corning	Cat# 354249; 3 mg/mL
Image-IT DEAD Green Viability Stain	ThermoFisher	Cat# I10291
DRAQ7 dye	Biolegend	Cat# 424001
4% Paraformaldehyde	Standard supplier	N/A
Triton X-100	Standard supplier	N/A
Bovine Serum Albumin (BSA)	Standard supplier	N/A
Hoechst 33342	Invitrogen	Cat# H3570
Trizol reagent	ThermoFisher	Cat#15596026
Collagenase	ThermoFisher	Cat#17100017
TryplE (1×)	ThermoFisher	Cat#12604013
Critical commercial assays
GEXSCOPE Single Nucleus RNAseq Kit	Singleron Biotechnologies	N/A
Qiagen RNeasy Kit	Qiagen	N/A
SuperScript II Reverse Transcriptase	Invitrogen	N/A
QuantiFast SYBR Green PCR Kit	Qiagen	N/A
iScript cDNA Synthesis Kit	Bio-Rad	N/A
SsoAdvanced PreAmp Supermix	Bio-Rad	N/A
Anti-Fibroblast Microbeads	Miltenyi Biotec	N/A
PicoPure RNA Isolation Kit	ThermoFisher	N/A
Deposited data
Raw and processed snRNA-seq data	This study	GEO: GSE271813
Experimental models: Cell lines
MCF-7 (human breast adenocarcinoma; female)	ATCC	CVCL_0031
BJ (normal human foreskin fibroblasts; male)	ATCC	CVCL_3653
Oligonucleotides
qRT-PCR primers for ACTA2, FAP, COL1A1, MMP1, IL1A1, IL6, CXCL1, CXCL12, GAPDH, HPRT-1, COL3A1, COL4A2, COL5A2 (see Table 2 in Method Details)	This paper	N/A
Software and algorithms
CeleScope tools v1.14.0	Singleron Biotechnologies	https://github.com/singleron-RD/CeleScope
STARsolo	Github	https://github.com/alexdobin/STAR
cutadapt v3.7	Github	https://github.com/marcelm/cutadapt/releases
Seurat	CRAN	https://satijalab.org/seurat/
Harmony	CRAN	https://CRAN.R-project.org/package=harmony
SingleR	Bioconductor	https://bioconductor.org/packages/release/bioc/html/SingleR.html
inferCNV	Bioconductor	https://bioconductor.org/packages/release/bioc/html/infercnv.html
ReactomePA	Bioconductor	https://bioconductor.org/packages/release/bioc/html/ReactomePA.html
clusterProfiler	Bioconductor	https://bioconductor.org/packages/release/bioc/html/clusterProfiler.html
singleseqgset	Github	https://github.com/arc85/singleseqgset
AUCell	Bioconductor	https://bioconductor.org/packages/release/bioc/html/AUCell.html
CellChat	Github	https://github.com/jinworks/CellChat
CIBERSORTx	Newman et al.45	https://cibersortx.stanford.edu
R v4.2.1	R Core Team	https://cran.r-project.org/
GraphPad Prism v8.4.3	GraphPad Software	https://www.graphpad.com/
IMARIS v9.7.1 (3D image analysis)	Oxford Instruments	https://imaris.oxinst.com/versions/9-7
ImageJ v2.9.0	Schneider et al.	https://downloads.imagej.net/fiji/releases/2.9.0/
Olympus FluoView software (image stitching)	Olympus	N/A
BioRender (figure illustration)	BioRender	https://biorender.com
Other
Olympus FV3000 inverted confocal microscope	Olympus, Tokyo, Japan	N/A
Illumina NovaSeq6000 (paired-end 150 bp sequencing)	Illumina	N/A
ViiA7 Real-Time PCR System	ThermoFisher Scientific	N/A
C1000Touch Thermal Cycler	Bio-Rad	N/A
Applied Biosystems ProFlex PCR System	ThermoFisher	N/A
NanoPhotometer NP80	Implen, Germany	N/A

Experimental model and study participant details

Primary breast tissue

Surgically resected tissues used for snRNA-seq were obtained from mastectomy specimens from two patients with primary breast cancer and one subject who had undergone breast reduction surgery for macromastia at the National Cancer Center Singapore (NCCS) between 2016 and 2017. Both breast cancer patients were female, of Chinese ethnicity and ER+/PR+/HER2-status was confirmed by standard IHC. Patient 1 was 62 years old while patient 2 was 82 years old.

All patients had been identified prior to surgery and provided written informed consent for tissue harvest. This study complied with all relevant ethical regulations including the Declaration of Helsinki, and ethical approval was obtained from the National Cancer Center Singapore Institutional Ethics Board (approval number: CIRB2015/2722).

Free-hand dissection using sterile scalpel blades was performed to obtain tumor and peri-tumoral samples from different zones (Zone 1: 1–3 cm; Zone 2: 3–5 cm; Zone 3: 5–7 cm from the tumor margin). Samples were kept frozen at −80°C until further processing.

Cell lines

Human MCF-7 breast adenocarcinoma cells (female, CVCL_0031) and BJ normal foreskin fibroblasts (male, CVCL_3653) were obtained from ATCC. Both cell lines were cultured in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% fetal bovine serum (FBS, ThermoScientific), penicillin/streptomycin (Gibco; 100 U/mL/100 μg/mL), sodium pyruvate (Gibco; 1 mM), and GlutaMAX supplement (Gibco; 2 mM). All cells were maintained at 37°C with 5% CO₂. Cell lines were not independently authenticated in this study; identity was assumed based on supplier certification. All cell lines were routinely screened for mycoplasma contamination and tested negative throughout the study.

Method details

TCGA analysis

Bulk RNA-seq expression data from TCGA breast tumors (n = 733) and normal adjacent to tumor (NAT) tissues and their clinical information (n = 89) were obtained from UCSC Xena⁴⁶ (Data S5). Updated overall survival data of TCGA patients generated by Liu et al.⁴⁷ was used for survival analyses. Deconvolution of CAF subtypes was performed with CIBERSORTx webtool⁴⁵ using a signature matrix adapted from Li et al.,⁴⁸ which was originally built from PDAC scRNA-seq data from Elyada et al.⁶ (Data S6). The number of permutations for significance analysis was set to 100. Patients were divided into myCAF-high/low groups based on whether their inferred myCAF proportions were above or below the median of the cohort. The survival and survminer R packages were then used to perform univariate and multivariate Cox regression analysis, with the multivariate Cox regression model including age, TNM staging, and ER, PR and HER2 status.

snRNA-sequencing

Single nuclei were isolated from approximately 50 mg of frozen tissue for each tumor sample and 100 mg for each peri-tumoral and RM sample by Singleron Biotechnologies using the GEXSCOPE Single Nucleus RNAseq kit according to the manufacturer's instructions. Briefly, for each library, the nuclei suspension of specified concentration was loaded onto a microfluidic chip for capture. The single-nuclei capture, lysis and mRNA capture steps were automated using Singleron Matrix NEO system. The final single-nuclei RNA sequencing libraries were sequenced on Illumina NovaSeq6000 flow cell with paired-end 150 bp.

Data processing

Fastq files were preprocessed using the CeleScope tools (version 1.14.0); www.github.com/singleron-RD/CeleScope, assay RNA, Singleron Biotechnologies), to generate gene expression matrix files using default parameters. Briefly, cellular barcodes in Read 1 were used to demultiplex and identify reads of the same cell origin. Low quality and adapter sequences were removed using cutadapt (version 3.7). The mapping was done using STARSOLO (https://github.com/alexdobin/STAR/blob/master/docs/STARsolo.md) against the human genome build (GRCh38) with ENSEMBL Gene Annotation (Version 99). The reads were assigned to genes using the featureCount tool, and the cell calling was performed by fitting a negative bimodal distribution and determining the threshold between empty wells and cell-associated wells. The gene count matrix was then generated, providing the number of unique molecular identifiers (UMIs) for each gene and cell. Nuclei with <300 genes detected (potential empty droplet) or have number of genes detected >3 standard deviations above the median (potential doublet) were filtered out. Samples were integrated using the Harmony package⁴⁹ followed by UMAP visualization and clustering using the Louvain clustering algorithm in Seurat⁵⁰ (dimensions = 15, resolution = 0.7). NK/T cells, myeloid cells, and fibroblasts were independently re-clustered (dimensions = 10, resolution = 0.3).

Cluster annotation

Clusters were annotated into major cell types using manually curated canonical cell type markers.^5,51,52 The SingleR package⁵³ was used to perform automatic annotation method using the Human Cell Primary Atlas dataset⁵⁴ and a triple negative breast cancer (TNBC) scRNA-seq Data S5 as reference. For annotation of re-clustered NK/T cells, myeloid cells, and fibroblasts, the FindAllMarkers function in Seurat was used to find all differentially expressed genes (DEGs) from each cluster (log2FC > 0.5, adj. p < 0.05).

InferCNV

Malignant epithelial cells were distinguished from normal epithelial cells using the inferCNV package.⁵⁵ The analysis was run for each sample separately using immune cells within the sample as normal reference and CNVs were predicted within all LumHR, LumSec and basal cells by InferCNV’s Hidden Markov model. Only CNVs with posterior probability less than 0.5 are included.

Differential gene expression, pathway enrichment, and GSEA

The FindMarkers function in Seurat was used to find all DEGs between samples of interest (log2FC > 0.1 or 0.5, adj. p < 0.05). If more than 200 DEGs were identified, only the top 200 were used for pathway enrichment performed using the ReactomePA⁵⁶ and clusterProfiler packages⁵⁷ based on the Reactome, Kyoto Encyclopedia of Genes and Genomes (KEGG) and Gene Ontology (GO) databases. The p-value cut-off was set to 0.05 and q-value cutoff was set to 0.2 in all analyses and top 5 enriched pathways were plotted. Gene set enrichment analysis (GSEA) using Hallmark and KEGG databases was performed with singleseqgset package. Scores for known gene sets were calculated using the AUCell package⁵⁸ and scaled from 0 to 1. myCAF and iCAF gene signatures used were obtained from Wu et al.⁵ and Elyada et al.⁶ (Data S9).

Cell-cell communication analysis

The CellChat package⁵⁹ was used to quantitatively infer and analyze intercellular communication networks and ligand-receptor interactions. The manually curated CellChatDB database containing approximately 3,300 validated molecular interactions was used and all significant interactions can be found in Data S11.

Cell culture

Human cell lines MCF-7 (breast adenocarcinoma) and BJ (normal foreskin fibroblasts) were cultured in Dulbecco’s modified Eagle’s medium (DMEM) with 10% fetal bovine serum (FBS, ThermoScientific), penicillin/streptomycin (Gibco, 100 U/mL/100 μg/mL), sodium pyruvate (Gibco, 1 mM), and GlutaMAX supplement (Gibco, 2 mM). All cells were cultured at 37°C with 5% CO₂.

Quantitative real-time PCR (qRT-PCR)

For the quantification of myCAF marker expression in BJ cells, total RNA was prepared from BJ fibroblasts using Qiagen RNeasy kit according to the manufacturer’s instructions. 1 μg of total RNA was reverse transcribed into cDNA using Superscript II (Invitrogen). qRT-PCR was performed with gene-specific primers for myCAF and iCAF markers and QuantiFast SYBR Green PCR Kit using the ViiA7 real-time PCR system (ThermoFisher Scientific, USA) according to manufacturer’s instructions. Relative gene expression was normalized with GAPDH expression.

For the quantification of collagen expression, spheroids were prepared for RNA extraction by dissociation with Tryple (1×, Thremofisher, USA) and Collagenase (10 μg/mL, Thermofisher, USA) in 1:1 ratio. Following dissociation, 12 spheroids per condition were further processed with Anti-Fibroblast Microbeads (Miltenyi Biotec, Germany) for fibroblast isolation. For gene expression analysis, total RNA was harvested using Trizol reagent (Thermofisher, USA) and the PicoPure RNA isolation kit (Thermofisher, USA), according to the manufacturer’s instructions. NanoPhotometer NP80 (Implen, Germany) was used to measure the concentration of the extracted RNA and to assess RNA quality and purity. Measurements were subjected to quality control minimum standards of A260/230 > 2 for quality and A260/280 > 1.8 for purity. First strand complementary DNA (cDNA) was synthesized from 150 ng of the total RNA, using the iScript cDNA Synthesis Kit (Bio-Rad, UK) according to manufacturer’s protocol sing Applied Biosystems ProFlex PCR System (Thermofisher, USA). Pre-amplification step was performed using SsoAdvanced PreAmp Supermix (Bio-Rad, UK) according to manufacturer’s instruction for 12 cycles. qRT-PCR was performed using C1000Touch Thermal Cycler (Bio-Rad, UK) and SYBR Green master mix (Thermofisher, USA). The comparative threshold cycle (CT) method was used to measure the expression level of each target gene and the mean of the reference gene HPRT1 using 1/(2ˆdeltaCT) in which deltaCT is the difference between the CT value of the target gene and the CT value of the reference genes. The sequences of primers used are shown in the following table.

qRT-PCR primers used.

Gene symbol	Primer sequence 5′-3′
ACTA2	Forward	AAGACGGGAATCCTGTGAAGC
	Reverse	GTCACCCACGTAGCTGTCTT
FAP	Forward	CTGGGAATATTACGCGTCTGTC
	Reverse	CTGGTCAGAGTACCACATTGC
COL1A1	Forward	CTGGAAGAGTGGAGAGTACTGG
	Reverse	TACTCGAACTGGAATCCATCGG
MMP1	Forward	GGACTTAGTCCAGAAATACCTGG
	Reverse	CAATCCTGTAGGTCAGATGTGT
IL1A1	Forward	GCTGAAGGAGATGCCTGAGA
	Reverse	GACCTACGCCTGGTTTTCCA
IL6	Forward	TCAGCCCTGAGAAAGGAGAC
	Reverse	GGCAAGTCTCCTCATTGAATCC
CXCL1	Forward	CCGAAGTCATAGCCACACTCAA
	Reverse	GTTGGATTTGTCACTGTTCAGC
CXCL12	Forward	CTCAACACTCCAAACTGTGC
	Reverse	TTTCCTCGAGTGGGTCTAGC
GAPDH	Forward	ACCCCTTCATTGACCTCAAC
	Reverse	CAGCGCCAGTAGAGGCAG
HPRT-1	Forward	TGACACTGGCAAAACAATGCA
	Reverse	GGTCCTTTTCACCAGCAAGCT
COL1A1	Forward	GAGGGCCAAGACGAAGACATC
	Reverse	CAGATCACGTCATCGCACAAC
COL3A1	Forward	GCCAAATATGTGTCTGTGACTCA
	Reverse	GGGCGAGTAGGAGCAGTTG
COL4A2	Forward	TTATGCACTGCCTAAAGAGGAGC
	Reverse	CCCTTAACTCCGTAGAAACCAAG
COL5A2	Forward	GACTGTGCCGACCCTGTAAC
	Reverse	CCTGGACGACCACGTATGC

Cell labeling

After trypsinizing (0.25% Trypsin-EDTA, Gibco, cat. n. 25200056) and pelleting BJ cells, 1 mL of PBS containing CellTracker Green CMFDA (ThermoFisher, cat. n. C2925, 1:500 dilution), CellTracker Orange CMRA (ThermoFisher, cat. n. C34551, 1:1000 dilution), or CellTracker Deep Red Dye (ThermoFisher, cat n. C34565, 1:1000 dilution) was used to stain the cells according to experimental requirements. The cell pellet was resuspended in PBS containing the dyes and incubated at 37°C for 30 min, before washing twice with 5 mL of medium and resuspending them in DMEM for spheroid formation or mixing with collagen gel.

Hanging drop spheroid formation

35 μL of cell suspension media containing 6000 MCF7 cells alone or with an additional 12000 BJ cells were deposited in each well of a custom-made polydimethylsiloxane (PDMS) support, supported by the lid of a 96-well plate. The lid was then inverted over a 96-well plate and incubated at 37°C with 5% CO₂ for a consistent formation of multicellular tumor spheroids.⁶⁰

OrganiX seeding

Pre-formed spheroids were collected by spinning the 96-well plate at 300 g × 1 min. Each spheroid was resuspended in 50 μL of rat tail collagen type I gel mixture (Corning, cat. n. 354249, diluted with distilled H₂O to 3 mg/mL, pH adjusted to 7.5) without or with pre-stained BJ fibroblasts (0.2 × 10⁶ cells/mL), before seeding in the OrganiX insert.⁶¹ The plate was incubated at 37°C for 30 min to allow the gel to polymerize. 35 μL of DMEM was added into each media channel to hydrate the channels before adding 400 μL of media into one side of the reservoir and 200 μL in the other reservoir. The plate was then incubated at 37°C with 5% CO₂ for 72 h.

Cell viability assessment and immunofluorescence

For assessment of cell viability, spheroids collected after hanging drop formation were stained with Image-IT DEAD Green Viability Stain (ThermoFisher, cat. n. I10291) at a final concentration of 100 nM in DMEM for 30 min at 37°C, while spheroids in OrganiX were stained with DRAQ7 dye (Biolegend, cat. n. 424001, 1:1000 dilution in DMEM) for 1 h at 37°C. Spheroids were then washed thrice with PBS before fixation with 4% paraformaldehyde for 30 min at room temperature.

For Ki67-staining, fixed spheroids collected after hanging drop were permeabilized with 0.5% Triton X-100 in PBS for 30 min and blocked with 0.5% Triton X-100 + 3% BSA in PBS for 2 h with shaking at room temperature. Spheroids were then incubated overnight at 4°C with 1% BSA in PBS containing mouse Alexa Fluor 647 anti-human Ki67 antibody (Invitrogen, cat. n. MA5-44140, 1:40 dilution) + 1 μg/mL Hoechst 33342 (Invitrogen, cat. n. H3570, 1:1000 dilution).

For spheroids in OrganiX, fixed spheroids were similarly permeabilized and blocked, before incubating overnight at 4°C with rat anti-Ki-67-FITC (Invitrogen, SolA15, cat. n. 11-5698-82, 1:300 dilution) 1% BSA + 1 μg/mL Hoechst 33342 in PBS. For extracellular pan collagen staining, the permeabilization step was not performed. Overnight staining at 4°C with shaking was performed with rabbit anti-Collagen-Pan (Invitrogen, PA1-85324, 1:70 dilution) + 1% BSA + 1 μg/mL Hoechst 33342 in PBS, followed by washing thrice with PBS and blocking as before for 2 h. Overnight incubation at 4°C was then performed with DyLight 488 Donkey anti-rabbit IgG (Biolegend, Poly4606, cat. n. 406404, 1:1000 dilution) + 1% BSA in PBS. Spheroids were washed thrice with PBS before image acquisition.

Microscopy and image analysis

Images were acquired using the inverted confocal microscope Olympus FV3000 (Olympus, Tokyo, Japan). To capture the entire well, 4 images were taken per well and stitched together within Olympus FluoView software. Images were processed using IMARIS software (v. 9.7.1) and ImageJ (v. 2.9.0). Individual nuclei were identified using the function “spots” based on Hoechst 3342 signal by setting the diameter to 8.5 μm. Intra- and peri-tumoral fibroblasts were identified by selecting spots positive for the appropriate fluorescence signal emitted by the CellTracker dyes within the identified nuclei. MCF7 cells were identified as those positive only for Hoechst 3342. The “surfaces” function was used to render a 3D image of the spheroid and detect cells which had invaded into the surrounding gel. To calculate cancer cell dispersion, the center of the spheroid was first calculated by finding the mean x, y, and z coordinates of all nuclei within the spheroid. The standard deviations for all cancer cell nuclei in the x, y, and z directions were calculated next. Dispersion was then calculated by taking the square root of the sum of the x, y, and z standard deviations.

Quantification and statistical analysis

All graphs and statistical analyses were performed using R (v4.2.1) or GraphPad Prism (v8.4.3). Statistical significance for experimental comparisons was assessed using the Mann–Whitney test, as applicable. Statistical parameters are reported in the figure legends.

Published: July 1, 2026