Single-cell RNA reveals a tumorigenic microenvironment in the interface zone of human breast tumors

This study was approved by the Medical Ethics Committee of Northwest University. Patients with primary, highly invasive, non-metastatic breast tumors diagnosed via core biopsy were recruited (Additional file 1: Table S1) and provided signed informed consent. Fresh tumor tissue samples were collected during surgical excision.

Breast specimen was rapidly transported to the research facility after resection in the operating room. An experienced anatomical pathologist grossly examined the tumor positioning and defined three zones in a cutting surface as described in our previous publications [5, 13], including a tumor zone within the tumor boundary, an interface zone within 5 mm of the outer tumor boundary and a distal normal zone at least 10 mm from the outer tumor boundary. For reproducible spatial identification, small invasive adenocarcinoma with an area less than 10 mm², clear tumor boundaries were included in this study. Given that tumors are randomly shaped objects, a breast sample was cut into a series of continuous and parallel sections, and each section had a 3 mm thickness. Only those tissues with three continuous sections with similar tumor boundary in the cutting surface were used (Additional file 1: Figure S1), which can ensure that three zones were precisely defined in a three-dimensional setting. Then, a fraction of tissues from all zones were fixed in formalin and embedded in paraffin for routine histopathological analysis to confirm that the tissue was correctly segmented and that tissues from the interface zone did not contain tumor cells. Finally, tissue pieces were extracted from the remainder of each zone for preparation of single-cell suspension, immunofluorescence examination and bulk RNA sequencing. In particular, tissue pieces derived from the tumor zone were extracted from the edge-most fraction of the tumor to avoid possible tissue necrosis that may be caused by hypoxia.

Single-cell suspension were prepared by mechanical dissociation and enzymatic digestion after removal of residual blood and surrounding adipose tissues. Cell suspension in which living cells accounted for more than 85% of the total cells was used for construction of single-cell RNA library and sequencing. Briefly, single-cell suspensions were parted into nanoliter-scale Gel Bead-In EMulsions (GEMs) to achieve single-cell resolution according to Chromium Next GEM Single Cell 3ʹ Reagent Kit v3.1 and Chromium Next GEM Chip G Single Cell Kit. Single cells in GEMs were lysed, and RNAs were reverse-transcripted and bar-coded by reverse transcriptase at 53 °C for 45 min and 85 °C for 5 min, and then held at 4 °C. Leftover biochemical reagents and primers in post-GEM reaction mixture were removed by Silane magnetic beads. cDNA was amplified to generate sufficient mass for library construction. Read 1 primer sequences were added during GEM incubation. Read 2 primer, a sample index, P5 and P7 sequences were added by end repair, A tailing, adaptor ligation and PCR during library construction. 16-bp bar code and 12-bp UMI were encoded in Read 1. Read 2 was used to sequence the cDNA fragment. Sample index sequences were incorporated as the i7 index read. Read 1 and Read 2 are standard Illumina® sequencing primer sites used in paired-end sequencing. Sequencing was performed by Illumina Novaseq 6000 in LC-bio Technology Co., Ltd (Hangzhou, Zhejiang, China). All samples were performed with the same process and kits.

Reads were aligned to genome by STAR software encapsulated in Cell Ranger and further aligned to exons, introns and intergenic regions. These reads aligned to the genome were used for UMI counting. Cells with a percentage of mitochondrial genes below 25% and a gene count exceeding 500 were used for further analysis. After quality control, a total of 88,548 single cells were obtained. Cell data were normalized by LogNormalize. The calculation formula is as follows:

Normalized data were used for clustering and visualizing based on their principal component analysis scores by Seurat (version 4.1.0), and a total of 87 clusters were identified. The following criteria were used to identify differential genes: (1) p value =  < 0.01; (2) log₂FC >  = 0.26. log₂FC means log₂ fold change; (3) the percentage of cells where genes were detected in a specific cluster more than 10%. A nonparametric manner called gene set variation analysis (GSVA, 1.34.0) was used to estimate hallmark pathway activation for clusters.

The analysis of transcription factor (TF) regulon activity was performed by the R package SCENIC (version 1.1.2.2), which infers the gene regulatory network based on co-expression. First, the log-normalized expression matrix was used as the input matrix. Second, co-expression networks of TFs and target genes were inferred by runGenie3 function. Then gene regulatory networks were constructed and scored by the functions of runSCENIC_1_coexNetwork2modules, runSCENIC_2_createRegulons and runSCENIC_3_scoreCells.

Cell lineage trajectories of macrophages, B cells and stem cells were inferred and characterized at the single-cell level. The UMI count matrix was used as the input. Dimensionality reduction was performed by using DDRTree algorithm based on differentially expressed genes; then, cell lineage trajectories between diverse cell subtypes based on cell cluster and pseudotime were inferred with the default parameters of Monocle (version, 2.22.0). Finally, “plot_genes_in_pseudotime” function was used to visualize dynamic changes of the cell.

CellPhoneDB (version, 3.1.0) was used to analyze communications among cell subtypes. First, the average expression level of each gene and the percentage of cells expressing the gene in every cell type were calculated. Then, based on the level of receptors and matched ligands in cell type, the possibility of cell interaction between cell types was inferred. Next, the cell type labels of all cells are randomly arranged 1000 times to calculate the significance of ligand–receptor pairs. Finally, the strength of cellular interactions was evaluated based on the expression of ligand–receptor pairs.

Gene set (top 10 marker genes of subcluster) signature was scored by mean value of log₂(TPM + 1) on TCGA and GTEx data using the GEPIA tool (http://​gepia2.​cancer-pku.​cn/​#index). Then, the correlation between prognostic impact and signature was calculated based on the score of gene set signature. The hazards ratio was calculated based on Cox PH Model. Samples were classified as high-signature or low-signature cohorts using percentile cutoffs of 50% or 50%. |log₂FC|> 1 and p-value < 0.01 were set as threshold.

For immunohistofluorescence staining of breast tissue sections, breast tissues were fixed with 4% paraformaldehyde for 1 day, dehydrated and embedded in paraffin for sections. The sections were dewaxed and rehydrated followed by antigen repair using a sodium citrate water bath at 98 °C for 40 min. After natural cooling, sections were rinsed with PBS 3 times for 5 min each time. Then, using 10% donkey serum at room temperature blocked sections for 1 h followed by incubation with primary antibodies: mouse mAb anti-KRT19 (1:400, Proteintech, #60187–1-lg, RRID:AB_2249705), rabbit pAb anti-IGFBP5 (1:100, Proteintech, #55205–1-AP, RRID:AB_2736835), rabbit pAb anti-DCN (1:400, Proteintech, #14667–1-AP, RRID:AB_2090265) or mouse mAb anti-MMP11 (1:50, Santa Cruz Biotechnology, sc-58381, RRID:AB_2144725) for overnight at 4 °C. After rewarming at 37 °C for 40 min, the sections were cleaned 3 times for 5 min each time. Sections were incubated with fluorochrome-conjugated secondary antibodies at 37 °C for 40 min and cleaned 3 times for 5 min each time. Nuclei were stained by DAPI. Finally, sections were photographed under OLYMPUS inverted fluorescence microscope.

MCF-7, MDA-MB-231 and SKBR3 human breast cancer cell lines and MCF10A human normal breast epithelial cell line were obtained from Procell Life Science&Technology Co., Ltd. (Procell, Wuhan, China). Short tandem repeat profiling was performed on each cell line. Mycoplasma was tested every two months. Cells were cryopreserved within the first 4 generations after acquisition and used within the 5th generation after each thaw. MCF-7, MDA-MB-231 and MCF10A cell lines were cultured in Procell-recommended media supplemented with 10% FBS (BI). SKBR3 cell line was cultured in RPMI-1640 media (Procell, PM150113) with 13% FBS (BI). Stable BMPR1B gene overexpression cell line was created with lentivirus-mediated gene transfection. Briefly, BMPR1B gene sequence was cloned into lentivirus expression vector (CD531B-1). Next, cell lines were infected with packaged lentivirus for 10 h followed by puromycin treatment. Finally, the level of BMPR1B transcript was detected by real-time quantitative polymerase chain reaction (RT-qPCR).

Total RNA was extracted from each sample using TransZol Up reagent (TransGen, Beijing, China), and reverse transcription was performed to generate cDNA using an EasyScript® One-Step gDNA Removal and cDNA Synthesis SuperMix (TransGen, AE311-02). Real-time PCR was performed using PerfectStart® Green qPCR SuperMix (TransGen, AQ101-01). Primer sequences targeting the bone morphogenetic protein receptor type 1B (BMPR1B) and housekeeping gene of GAPDH are listed in Additional file 1: Table S2. RT-qPCR data were calculated by the ΔΔCt of genes of interest relative to the housekeeping gene GAPDH.

Cell proliferation was performed using Cell Counting Kit-8 (TransGene, FC101-01). Cells were plated in 96-well plates and cultured in an incubator. At indicated time point after seeding, CCK-8 solution was added to the wells and incubated for 3 h; then, the absorbance was measured at 450 nm (BioTek, Epoch2 Instruments).

Cells were seeded in 6-well plates at a density of 1000 cells/3 ml per well. The medium was changed every 4 days. On the 6th day (MCF7 cells), 8th day (MDA-MB-231) and 20th day (SKBR3 cells), cell clones were photographed using OLYMPUS inverted fluorescence microscope.

A thin ruler was used to scrape a straight line in the confluent cell monolayer in 6-well plate to create a wound. Then detached cells were washed away with PBS and the cells were incubated in media with 1% FBS to allow cells to grow and close the wound. The scratches were photographed at 0 h,12 h, 24 h, 48 h, 72 h and 144 h using OLYMPUS inverted fluorescence microscope. The scratch area was calculated by the software OLYMPUS cellSens Entry, and the average value of the area was calculated by multiple scratches. The change rate of area represented the speed of cell migration.

Female BALB/cNj-Foxn1^nu/Gpt mice were purchased from corporation (Gempharmatech Co., Ltd, China). Mice were housed in specific pathogen-free facilities with ad libitum access to water and food under a 12-h dark and 12-h light cycle in animal housing with constant humidity (40–60%) and temperature (23 ± 1 °C). 5 × 10⁶ cells were collected and injected into subcutaneous location of both groin in nude mice. Cells expressing BMPR1B gene were injected on the left side, and control cells were injected on the right side. All programs have been approved by the Research Ethics Committee of Northwest University.

The log-rank test, Pearson’s correlation and two-sided paired Student’s t-test were used in the study. The method for score of gene signatures is one-way ANOVA. Unless otherwise specified, the quantitative data were presented as mean (± standard deviation) from three independent experiments. Significant differences were considered as follows: ns, no significance, p >  = 0.05; *p < 0.05; **p < 0.01.

We collected nine sets of fresh samples from the tumor zone, the interface zone and the normal zone of 3 patients with untreated and invasive breast cancer who underwent total mastectomy. After resection, samples were rapidly digested to single-cell suspension for sequence, analysis and functional study (Fig. 1a). A total of 88,548 cells were obtained from clean data. Among them, 28,718 cells originated from the tumor zone, 29,192 cells from the interface zone and 30,638 cells from the normal zone (Additional file 1: Table S3). Based on marker genes and artificial correction, all identified clusters could be assigned to known cell types: Fibroblasts, B cells, T cells, epithelial cells, stem cells, macrophages and endothelial cells, and different cell types had different transcriptional activities (Fig. 1b).

Meanwhile, the correlation analysis of single-cell RNA sequencing and matched bulk-seq showed that they had a good correlation (r = 0.86; Fig. 1c), indicating that gene expression was not significantly affected by cell dissociation. Overall, the data revealed the presence of a complex cellular heterogeneity containing a total of 87 different cell subclusters (Fig. 1d), and many subclusters had strong zone specificity, which implied that the interface zone has a unique microenvironment. Here, we mainly explored the zone-specific cell subtypes and functions of specific genes in epithelial cells.

T cells play crucial roles in the process of tumor inflammatory response. Here, we detected 12,941 T cells that were reclustered into 11 separate subsets (Fig. 2a), respectively, corresponding to natural killer cells (NK cells, cluster 9, KLRF1 +), CD8 + T cells (clusters 0, 4, 5 and 7, CD8A +), CD4 + T cells (clusters 1, 2 and 8, CD4 + /FOXP3-) and regulatory T cells (Tregs, clusters 3 and 6, CD4/IL2RA/FOXP3 +) (Fig. 2b). In addition, one cluster did not correspond to known cells (cluster 10, TOP2A +), and ontology analysis showed cell cycle pathway had strong enrichments in this cluster (Additional file 1: Figure S2a), suggesting cluster 10 had an active biological process.

T cell types of NK cells, CD8 + cells and CD4 + cells were detected in all three zones, with the exception of CD8 + cluster 7 and CD4 + cluster 8 mainly presenting in tumor zone. For Tregs, cluster 3 apparently enriched in tumor zone, and cluster 6 enriched in both tumor and interface zone (Fig. 2a and b). The enrichment of specific T cell clusters in tumor and interface zones suggested a transcriptome difference between T cell types that were influenced by microenvironment.

Subsequently, we compared levels of the pathway gene signatures among cells from the three zones (Fig. 2c). Analysis showed that overall pathway signals of CD4 + and CD8 + T cell were weak in the interface zone, which indicated that these T cells have weak bioprocess and functions. However, angiogenesis signal in the interface-derived Tregs was stronger than that in the normal zone, which might contribute to the induction and/or maintenance of Tregs in the interface zone [20]. Besides, some pathways differed between subclusters; for instance, CD8 + cluster 7 and CD4 + cluster 8 showed strong IFN-γ and IFN-α responses, enhanced complement activities, high allograft rejection activities (Fig. 2d) and granzyme expressions of GZMA, GZMM and GZMK (Fig. 2e), indicating these two subclusters of T cells showed some certain anti-tumor activities. For Tregs, compared with cluster 6, cluster 3 had stronger inflammatory response states, such as IFN response and complement reaction (Fig. 2d), although both of them expressed immunosuppressive molecules TIGIT and CTLA4 (Fig. 2e). However, survival analysis of marker genes showed that cluster 6 positively contributed to the prognosis of patients based on The Cancer Genome Atlas (TCGA) datasets, while cluster 3 had no relationship with patients’ prognosis (Fig. 2f), which indicated tregs had complex functions that need to be further explored.

In order to assess which transcription factors underlay differences in expression patterns of pathways among clusters, we performed single-cell regulatory network inference and clustering (SCENIC). SCENIC scanned transcription factor binding sites near differentially expressed gene positions and analyzed co-expression of these genes and transcription factors. There were significant differences in the expression of transcription factors (Fig. 2g). We noticed high expression levels of transcription factors STAT1, STAT2 and IRF7 in tumor-derived cluster 3, cluster 7 and cluster 8 (Fig. 2g). These factors are required for IFN production [21‐23], which suggested that they played important roles in inflammatory response. Upregulation of transcription factor BATF controlled the activation program of clusters 3 and 6 in tumor and interface zones [24].

Recently, immune-hot and immune-cold regions are considered additional elements of heterogeneity of TME [25, 26]. Here we found clusters 0, 1, 4, 5, 7 and 8 can express granzyme (Fig. 2e), which suggests that these clusters have immune activity, and thus, they belong to the immune-hot clusters. However, clusters 2, 3 and 6 had no obvious immune activity based on their low granzyme expression level, and they belong to the immune-cold clusters. Notably, we noticed that total cell numbers of immune-hot clusters were higher in the interface or the tumor zone than that in the normal zone in all three patients (Additional file 1: Table S4), indicating the interface zone share some characteristics with the tumor zone at the immune level. Altogether, the interface-derived T cells were different from those in normal and tumor zone.

Tumor-associated myeloid cells can differentiate into different cellular subsets based on their microenvironment, and every subset has unique characteristics and plays different roles [27]. A total of 2719 myeloid cells were obtained and reclustered into 11 subclusters (Fig. 3a). Clusters 3 and 5, mainly enriched in the tumor and interface zones, were characterized by genes S100A4, AP1S2 and S100A6 that defined monocytes [28, 29]. Cluster 7, across the three zones, was identified as dendritic cells (DCs) that selectively expressed DC markers IRF8, BIRC3, BASP1 and NAPSB [30, 31]. The remaining clusters corresponded to macrophages that expressed C1QA, C1QB, C1QC and APOE enabling the distinction from monocytes (Fig. 3b) [32, 33]. In addition, C1QA, C1QB and C1QC genes could well distinguish macrophages from monocytes and DCs, and commonly used markers MSR1, CD14 and CD163 could not achieve this discrimination (Fig. 3c).

According to the analysis of pathway gene signatures, macrophage subtypes had different pathway activities (Fig. 3d). Decreased signals of inflammatory response, IFN response, reactive oxygen species and allograft rejection corresponded to M2-like macrophage phenotypes in cluster 9 that expressed scavenger receptors (SCARA3) (Fig. 3d, Additional file 1: Figure S2b), which indicated cluster 9, across the three zones, had phagocytic function in the body [34]. For tumor-derived cluster 6, a high expression level of VEGFA and a strong hypoxia signal supported that cluster 6 played an important role in promoting angiogenesis (Fig. 3d, Additional file 1: Figure S2b).

Clusters 0, 2 and 10 were mainly enriched in the tumor and interface zones and displayed a specificity of zone distribution of macrophages with strong IFN response and reactive oxygen species (Fig. 3d). Despite all this, the poor separation of these clusters indicated they did not represent separate entities but diverse cell states, which was consistent with a spectrum of macrophage activation states [35]. In order to determine the differentiation state of macrophages, we performed a pseudotime analysis by Monocle and found that macrophage subclusters did not have a specific polarization state, except for M2-like cluster 9 (Fig. 3e and f). Therefore, clusters 0, 2 and 10 did not belong to polarized macrophages. SCENIC analysis showed PPARG transcription factor as classical M2 marker was upregulated in cluster 0 (Fig. 3g) [36], indicating cluster 0 might tend to polarize to M2 macrophages. Transcription factors IRF7, IRF1, STAT2 and STAT1, which boosted inflammatory response, were upregulated in cluster 2, indicating cluster 2 might tend to polarize to M1 macrophages. Moreover, cluster 10 with high expression level of EZH2 had a function in tumor promotion [37]. Together, results displayed macrophages with various states in interface zone, and cluster 0 and cluster 10 may provide an environment for tumor occurrence in the interface zone.

We detected 1,363 B cells and 122 mast cells that clustered near B cells. Given that TME could affect tumor infiltrating immune cells [16, 38], we explored the differences of B cells in three zones. Reclustering revealed 11 distinct subclusters (Fig. 3h), of which nine were enriched in tumor and interface zones and designated as plasma B cells (clusters 2 and 10, IGHG1 +), follicular B cells (clusters 0, 3, 6, 8 and 9, MS4A1 +), immature B cells (cluster 7, IL7R +) and mast cells (cluster 5, TPSAB1 +) (Fig. 3i). The remaining two clusters were derived from all three zones and designated as plasma B cells with high level of IGHA1 (clusters 1 and 4). Further, pseudotime analysis revealed the differentiation trajectory of B cells was divided into three branches corresponding to the main cell types follicular B cells, plasma B cells and mast cells (Fig. 3j). Wherein, immature B cells distributed along axes of the trajectory, indicating it was under a state of differentiation from follicular B cells to plasma B cells. Mast cells were independently located at one end of the differentiation trajectory because they did not belong to B cell lineage. Other clusters were well distributed at the corresponding location of cell types in the differentiation trajectory, indicating the differentiation process of follicular B cells to plasma B cells.

Pathway analysis showed no significant difference in plasma B cells among the three regions. However, interface-derived and tumor-derived follicular B cells expressed strong interferon response signal (Fig. 3k, Additional file 1: Figure S2c) and high levels of HLA class II molecules (Fig. 3l), which can enhance T cell-independent antibody response [39] and antigen presentation function. Additionally, interface-derived and tumor-derived mast cells had high expression levels of VEGFA and VEGFB genes (Additional file 1: Figure S2d), suggesting potential functions of promoting angiogenesis.

A total of 20,810 stem cells were obtained and classified into 13 subclusters, among which clusters 1 and 6 were enriched in the interface and tumor zones, while cluster 4 was mainly derived from the interface zone, and the remaining clusters distributed in all three zones (Fig. 4a).

Pseudotime analysis yielded a connected trajectory with three main branches (Fig. 4b). Clusters 1 and 6 were markedly enriched at the start branch of pseudotime, suggesting they represented mesenchymal stem cells (MSCs, Fig. 4b), which was further confirmed by expression of FN1, a marker gene of MSCs (Fig. 4c) [40]. Besides, expressions of maker genes that separated clusters 1 and 6 from other clusters were confirmed in TCGA data (Fig. 4d). Interestingly, clusters 1 and 6 displayed high expressions of collagen type IA1, A2 and type IIIA1 (Fig. 4c), which have been previously reported to promote tumor growth and invasion. Clusters 3 and 5 were enriched at three zones that were merged into the first branch point (Fig. 4b), revealing the multilineage potential of these two clusters; thus, they were identified as pluripotent stem cells (PSCs). The remaining clusters were enriched at one end branch of the trajectory (Fig. 4b) and expressed some epithelial marker genes (Fig. 4c), so they were defined as epithelial progenitor cells.

Analysis of pathway gene signatures showed that each subtype had specific pathway activities (Fig. 4e). Wherein, MSCs (cluster1 and 6) had a strong IFN response signal, indicating the immunosuppressive capacity [41]. Notch signaling and myc targets had different expression levels between clusters 3 and 5, both of which belong to PSCs, suggesting PSCs possessed different differentiation and proliferation capabilities [42, 43]. Notably, interface-derived cluster 4 showed a weak DNA repair ability, which might promote breast carcinogenesis indicating a special interface microenvironment [44].

Endothelial cells located between tissue and blood play a role in presenting tumor microenvironment signals. Here, we obtained 1258 endothelial cells and reclustered them into 10 clusters, and more interface-derived cells than normal-derived cells were observed among these clusters (Fig. 5a), indicating that endothelial cells in interface zone were strongly affected by tumor progression.

Based on marker genes for every cluster (Additional file 2: Table S5), we revealed one set of lymphatic endothelial cells (cluster 9, TFF3 +) and nine sets of vascular endothelial cells (FLT1 + , Fig. 5b): One was mostly interface-derived (clusters 2; HLA-DQA2 +) and three were mostly tumor-derived (clusters 4, 6 and 8; ROBO1 +) (Additional file 1: Figure S2e). The remaining clusters had no zone-specific enrichment. In tumor zone, we found that multiple pathway signals, especially related to angiogenesis, were enhanced compared with those in interface zone (Fig. 5c). Interestingly, angiogenesis-related molecules HSPG2, FLT1, TIE1, HIF1A and KDR gradually decreased from the tumor to the interface to the normal zone (Fig. 5d), suggesting that breast tissue in breast cancer patients should not be simply divided into tumor and normal zones, and that the interface zone should be considered. Additionally, interface-derived endothelial cells had increased expressions of MHC class II molecules HLA-DRB1, DPA1, DRA, DPB1 and DQB1 (Fig. 5d) and enhanced pathway activities of allograft rejection, hedgehog signaling and interferon response than those in normal zone (Additional file 1: Figure S2f), indicating that interface-derived endothelial cells were active in presenting antigen epitopes.

Given that clusters 2, 4, 6 and 8 were interface-specific or tumor-specific enrichment, we analyzed pathway gene signature and found a diversity of pathway activities among these clusters (Fig. 5e). Wherein, interface-derived cluster 2 had overall weak pathway signals. For tumor-derived clusters 4, 6 and 8, they also showed different pathway signal activities: Cluster 6 had weak interferon responses, but its G2M checkpoint and E2F target pathways were relatively strong compared with clusters 4 and 8, suggesting cells in cluster 6 were under a strong proliferative state.

Finally, we analyzed which transcription factors cause differences between tumor-derived clusters 4, 6, 8 and interface-derived cluster 2 using SCENIC (Fig. 5f). This identified KLF2 as a candidate transcription factor underlying the difference between cluster 6 and others, whereas genes regulated by ETS1 and ETS2 were responsible for specific cell phenotypes of clusters 4 and 8. Previous studies reported KLF2 promote cell proliferation [45] and ETS1 and ETS2 contribute to angiogenesis [46], indicating tumor-promoting functions of these tumor-derived clusters. Notably, interface-derived cluster2 expressed high levels of MYC transcription factor that was essential for vasculogenesis during tumor progression [47]. Therefore, we believed that some endothelial cells in both interface and tumor zones provided necessary conditions for the survival of tumor cells.

Although the function of fibroblasts in wound healing has been well understood, their role in cancer still needs to be explored [48, 49]. Here, a total of 7100 fibroblasts were detected, and reclustering revealed 11 subclusters (Fig. 6a). No interface-specific cluster was observed, but clusters 4, 7 and 10 significantly enriched in the tumor zone. Clusters 4, 7 and 10 preferentially expressed MMP11, COL11A1, NTM and KIF26B, whereas other clusters showed high expressions of CFD, VEGFD, CD34 and PDK4 (Fig. 6b). Average expression levels of the top 10 marker genes of clusters 4, 7 and 10 were confirmed in tumor and normal breast tissues based on TCGA database (Fig. 6c). We further verified this result by detecting the expression of a representative marker gene MMP11 in tissue sections through immunofluorescence staining (Fig. 6d). In additions, clusters 4, 7 and 10 expressed a variety of collagen proteins including COL1A1, COL1A2, COL3A1, COL6A1 and COL6A3 (Fig. 6e), which have been previously linked to abilities of proliferation, invasion, metastasis and drug resistance of tumor cells [50‐53].

According to the analysis of pathway gene signatures, we observed a significant phenotypic diversity among all clusters (Fig. 6f). Clusters 4, 7 and 10 had an increased glycolysis, which may be responsible for their collagen synthesis [54]. PI3K/Akt/mTOR signaling, which has been reported to increase the migration of fibroblasts [55], was highly activated in clusters 4, 7 and 10. Besides, clusters 7 and 10 displayed high allograft rejection signature (Fig. 6f) and high expressions of MHC-related proteins B2M, TAP1 and TAP2 (Fig. 6g), indicating that they had roles of antigen presentation in TME and might induce expansion of regulatory T cells [56]. Although we did not find fibroblasts that only enriched in the interface zone, we observed a very small portion of interface-derived fibroblasts shared some characteristics with tumor zone-derived fibroblasts in clusters 4, 7 and 10 (Fig. 6a).

Epithelial cells were the most abundant cells among all isolated cells from nine tissues. Twenty clusters were obtained by reclustering (Fig. 7a). The expression levels of epithelial markers revealed distinct cell phenotypes of these clusters (Fig. 7b). These clusters were further classified into eight luminal groups L1-L8 and three basal groups B1-B3 based on lineage marker gene expression patterns (Fig. 7c), and the reliability of these patterns was confirmed in breast epithelial cell lines by analyzing the data in a previous study (Fig. 7d) [57]. IGFBP5, a marker of tumor-derived clusters 6 and 16 (Fig. 7e), was further examined in tissue sections by immunofluorescence, and the result confirmed presence of these clusters as separate cellular entities (Fig. 7f). In particular, there are a total of 13,310 epithelial cells from the interface zone, of which 6,031 cells are from patient 1, 3,992 cells are from patient 2 and 3287 cells are from patient 3. Due to individual differences, the subtypes of these cells have patient preferences, for example, cluster 10 mainly derived from patient 3, clusters 4 and 12 mainly derived from patient 2, and cluster 2 derived from all three patients (Fig. 1d).

The heatmap showed the expression patterns of specific genes in epithelial subclusters (Fig. 7c). Group B1 was characterized by expressions of KRT5, ACTA2 and TAGLN and displayed high levels of MKI67 indicating a robust proliferative state. Group B2 was characterized by expressions of KRT5, KRT14 and KRT17 and showed a strong activity of immunogenicity and proliferation as indicated by receptor expressions of HLA-DRA, HLA-DRB, MET and EGFR. Wherein, about 54% of group B2 cells derived from interface zone.

Group B3 was characterized by expressions of ACTA2 and TAGLN, and it only consisted of few cells with high expressions of proliferative markers CDKN1B and BCL2. In particular, the number of interface-derived cells was higher than that of tumor or normal-derived cells in group B1, B2 and B3, suggesting a unique property of the interface zone of breast tissues in breast cancer patients. Groups L1 and L2 were characterized by the expression of CDH1, a marker of luminal subtype. Wherein, cluster 11 was mature HER2 + luminal B subtypes with a phenotype of ESR1 + PGR + HER2 + while cluster 7 was mature HER2 + subtypes with a phenotype of HER2 + PGR-ESR1-. Group L2 phenotypes showed high levels of ESR1 and HER2, indicating HER2 + luminal B subtypes. Meanwhile, group L2 expressed ACTA2 and TAGLN representing an epithelial-to-mesenchymal transition (EMT) state. Groups L3 and L4 cells derived from the tumor zone and had low expressions of ESR1/PGR/HER2. Besides, group L4 presented EMT phenotypes as indicated by co-expressions of KRT8/18, ACTA2 and TAGLN and proliferative phenotypes as indicated by KI67 expression. About 89% of group L5 cells derived from the interface zone, and these cells showed high levels of ESR1 and MKI67, representing a phenotype of luminal B cells. Group L6 showed a phenotype of luminal A cells with ESR1 + HER2-MKI67- and expressed a low level of immunogenicity-related gene HLA-DRB. In group L6, cluster 4 was composed of 85% interface-derived cells while cluster 9 was composed of 99% normal-derived cells. Group L7 and L8 across the three zones co-expressed EPCAM and ITGA6, characteristics of progenitor cells [58].

We further characterized the features of these groups with distinct phenotypes by analysis of pathway gene signatures (Fig. 7g). Results showed that most signal pathways in the basal groups were active, but G2M check point, E2F targets, notch signaling and oxidative phosphorylation pathway were not consistent indicating basal cells across the three zones were heterogeneous. For tumor-derived groups L3 and L4, their differences were obvious. Group L4 was enriched for hallmarks of cell cycle (e.g., G2M checkpoint and E2F targets) compared with group L3, which was consistent with the gene expression patterns analyzed above (Fig. 7c). For groups L5 and L6, interface-derived group L5 had higher G2M checkpoint and E2F target activities than group L6, suggesting that luminal cells were distinct under the influence of the microenvironment in the interface zone.

Considering the interaction between epithelial cells and stromal cells in the surrounding microenvironment, we further analyzed the cell–cell interaction network between all types in the normal zone, interface zone and tumor zone based on expressions of ligand–receptor pairs, respectively (Fig. 7h). Epithelial cells in the interface zone showed interactions with most cell types, and they showed especially strong interactions with macrophages and MSCs.

Meanwhile, macrophages and MSCs also showed strong interactions with epithelial cells in the tumor zone, while there is no such interaction in the normal zone. Based on these findings, we concluded that cell–cell interaction in the interface zone was under a more active state and some characteristics of the interface zone is shared with the tumor zone. Undoubtedly, cells in the tumor zone had their specific properties, for instance, fibroblasts that can be named CAFs, had stronger interactions with epithelial cells compared with other stromal cells.

Besides, the bubble plots displayed that the ligand–receptor pairs between the epithelial cells and the stromal cells are stronger in the interface and the tumor zones than that in the normal zone including TNF, HLA, collagen and their receptor families (Fig. 7i). Epithelial cells had a much greater effect on stromal cells than stromal cells have on epithelial cells in both the interface and tumor zones, which implied that epithelial cells play crucial roles in the formation of tumor or para-carcinoma microenvironment. For instance, interface- and tumor-derived epithelial cells expressed TNF, and endothelial cells expressed notch1, which was able to protect vascular endothelial cells from TNF-induced apoptosis [59], but endothelial cells almost had no effect on epithelial cells based on expressions of interaction pairs.

We performed an in-depth study on a specific gene of BMPR1B that was highly expressed in cluster 4 (luminal A) and 12 (luminal B), which were mainly enriched in the interface zone (Fig. 7a and b). By analyzing the differential genes in cluster 12, we further found that the average expression level of BMPR1B gene in tumor-derived epithelial cells was higher than that in interface-derived epithelial cells (Fig. 8a, left). In addition, the average expression level of BMPR1B was higher in both the interface- and tumor-derived epithelial cells in cluster 12 than in all luminal epithelial cells in the normal zone (Fig. 8a, middle and right). BMPR1B had the same expression trend in cluster 4 as in cluster 12 (Additional file 1: Figure S3a). As a whole, the average expression level of BMPR1B gradually increased from the normal to the interface to the tumor-derived luminal cells (Fig. 8b). The expression of BMPR1B gene between tumor and normal zones is consistent with the TCGA database (Fig. 8c).

To explore functions of BMPR1B gene, we first overexpressed BMPR1B gene in non-tumorigenic MCF10A breast cell line (Fig. 8d) and found that BMPR1B overexpression inhibited the proliferation (Fig. 8e, up) and migration (Fig. 8f, up) of MCF10A cells. However, tumor-bearing experiment showed BMPR1B did not change the non-tumorigenic property of MCF10A cells in vivo.

Next, we investigated functions of BMPR1B gene in breast cancer cells with different phenotypes including MCF-7 (luminal type), SKBR3 (HER2 + type) and MDA-MB-231 (triple-negative breast cancer) by stably overexpressing BMPR1B gene (Fig. 8d). For MCF-7 cells, BMPR1B overexpression could enhance proliferation (Fig. 8e, down), migration (Fig. 8f, down) and clone formation (Fig. 8g) of the MCF7 cells in vitro and could promote tumor formation and growth in vivo (Fig. 8h and i). For MDA-MB-231 cells, BMPR1B overexpression promoted cell proliferation (Additional file 1: Figure S3b) and clone formation (Additional file 1: Figure S3c), but it did not affect cell migration (Additional file 1: Figure S3d) and tumorigenicity in vivo (Additional file 1: Figure S3e). For SKBR3 cells, BMPR1B overexpression did not change proliferation (Additional file 1: Figure S3b), clone formation (Additional file 1: Figure S3c) or migration of the SKBR3 cells (Additional file 1: Figure S3d). These findings suggest that BMPR1B gene may play crucial roles in the development of luminal breast carcinoma.