Integrated single-cell analysis reveals distinct epigenetic-regulated cancer cell states and a heterogeneity-guided core signature in tamoxifen-resistant breast cancer

Abstract

Background

Inter- and intra-tumor heterogeneity is considered a significant factor contributing to the development of endocrine resistance in breast cancer. Recent advances in single-cell RNA sequencing (scRNA-seq) and single-cell ATAC sequencing (scATAC-seq) allow us to explore inter- and intra-tumor heterogeneity at single-cell resolution. However, such integrated single-cell analysis has not yet been demonstrated to characterize the transcriptome and chromatin accessibility in breast cancer endocrine resistance.

Methods

In this study, we conducted an integrated analysis combining scRNA-seq and scATAC-seq on more than 80,000 breast tissue cells from two normal tissues (NTs), three primary tumors (PTs), and three tamoxifen-treated recurrent tumors (RTs). A variety of cell types among breast tumor tissues were identified, PT- and RT-specific cancer cell states (CSs) were defined, and a heterogeneity-guided core signature (HCS) was derived through such integrated analysis. Functional experiments were performed to validate the oncogenic role of BMP7, a key gene within the core signature.

Results

We observed a striking level of cell-to-cell heterogeneity among six tumor tissues and delineated the primary to recurrent tumor progression, underscoring the significance of these single-cell level tumor cell clusters classified from scRNA-seq data. We defined nine CSs, including five PT-specific, three RT-specific, and one PT-RT-shared CSs, and identified distinct open chromatin regions of CSs, as well as a HCS of 137 genes. In addition, we predicted specific transcription factors (TFs) associated with the core signature and novel biological/metabolism pathways that mediate the communications between CSs and the tumor microenvironment (TME). We finally demonstrated that BMP7 plays an oncogenic role in tamoxifen-resistant breast cancer cells through modulating MAPK signaling pathways.

Conclusions

Our integrated single-cell analysis provides a comprehensive understanding of the tumor heterogeneity in tamoxifen resistance. We envision this integrated single-cell epigenomic and transcriptomic measure will become a powerful approach to unravel how epigenetic factors and the tumor microenvironment govern the development of tumor heterogeneity and to uncover potential therapeutic targets that circumvent heterogeneity-related failures.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13073-024-01407-3.

Background

Endocrine therapy, including tamoxifen, fulvestrant, and aromatase inhibitors (AIs), is currently a standard and effective treatment for estrogen receptor α (ER) breast cancer patients [1, 2]. Although endocrine treatment can significantly reduce the risk of recurrence, approximately a third of patients eventually develop endocrine resistance and relapse [3–5]. The mechanisms of endocrine resistance may include loss of ER expression [6], alterations in cell cycle and cell survival signaling molecules [7], altered metabolism [8], and crosstalk between ER and other signaling pathways [9, 10]. We and others have recently revealed that epigenetic regulation including three-dimensional (3D) chromatin organization might be linked to acquired tamoxifen resistance [11, 12]. Despite the great efforts have been made in understanding intrinsic underlying molecular mechanisms, the outcome of tamoxifen-associated endocrine treatment remains suboptimal [13, 14]. Inter-tumor (tumor by tumor) and intra-tumor (within tumor) heterogeneity of breast cancer has been shown to be associated with treatment resistance [15]. For example, ER expression variation among different tumors resulted in differential responses in endocrine treatment [16], and multiple driver mutations in HER2-amplified breast tumors had a direct influence on the treatment outcome [17].

Recent advances in single-cell multi-omics sequencing such as single-cell RNA sequencing (scRNA-seq) and single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) now allow us to explore inter- and intra-tumor heterogeneity at single-cell resolution [18–22]. Several studies have capitalized on the integration of scRNA-seq and scATAC-seq data to illustrate transcriptional and epigenetic changes during blood differentiation [18], cardiac progenitor cell fate decisions [19], and the enzalutamide treatment response and resistance [20]. However, such integrative single-cell analysis has not yet been demonstrated to characterize transcriptomes and chromatin accessibility in breast cancer endocrine resistance.

In this study, we conduct an integrated scRNA-seq and scATAC-seq analysis of ~ 82,400 breast tissue cells from two normal (NT), three primary (PT), and three tamoxifen-treated recurrent tumor tissues (RT). We uncover distinct tumor cell characteristics among breast tumor tissues at the single cell level, delineate inter- and intra-tumor cell heterogeneity, and further infer tumor progression from primary to recurrent tumor stages. We then define cancer cell states (CSs) by quantitatively measuring breast cancer gene expression modules and associate them with the epigenetic factors. We identify a heterogeneity-guided core signature by incorporating both expression and chromatin accessibility features. We further identify the signaling and metabolism pathways that mediate the communications between CSs and the tumor microenvironment (TME). We finally functionally examine one of core signature gene, BMP7, in tamoxifen-resistant breast cancer cells.

Methods

A cohort of human breast tissue samples

A cohort of eight human breast normal and tumor tissue samples were used for the study. Two human normal breast tissues were purchased from National Disease Research Interchange (NDRI, Philadelphia). Three human primary breast tumor tissues were procured from OriGene (OriGene Technologies Inc., Rockville, Maryland), and three tamoxifen-treated recurrent breast tumor tissues were obtained from Ontario Tumor Bank (Additional file 1: Table S1). All breast tumors are ER + , aged between 50 and 60 years old, grade between G1 and G3. The use of human materials has been reviewed by Medical College of Wisconsin (MCW)’s institutional review board. All tissues were subjective to scRNA-seq and scATAC-seq profiling and data analysis.

Single nucleus isolation

Nucleus suspension preparation was performed on ice throughout. Breast tissues (normal or tumor) were minced into 1–3 mm³ pieces in a sterile 35 mm culture dish, transferred to 1 ml of nuclear lysis buffer for 7 min and then centrifuged at 500 × g at 4 °C for 5 min. The supernatant was discarded, and the remaining cells were added to 5 ml medium. Subsequently, 250 μl of 20 × Collagenase II was added, and the mixture was incubated at 37 °C with 5% CO2 for 60 min. After confirming complete nuclear lysis, the cells were filtered through a 40 μm cell sieve. The collected strained cell suspension was centrifuged at 500 × g at 4 °C for 5 min, and the supernatant was discarded. The nuclei were resuspended in 5 ml of cold experimental medium and quantified for viable cell count using trypan blue. Single cells were subjective to live cell flow sorting using 7-AAD Viability Staining Solution [23].

Single-cell RNA-seq library preparation and sequencing

Samples were processed on a Sony SH800 Cell Sorter with a 100 mm sorting chip and nuclei were collected into 1.5 ml centrifuge tubes containing 10 µl of the Pre-FACS buffer. We collected ~ 10,000, ~ 15,000, and ~ 15,000 nuclei from the two breast normal, three breast primary tumor, and three breast tam-treated recurrent tumor tissue samples, respectively. Nuclei quality and quantity was evaluated using trypan blue on an Invitrogen Countess II device in duplicate, and a subset of nuclei was spun down in a fresh tube and resuspended in 10 × sample dilution buffer. Using a Chromium Single Cell 3’ Library and Gel Bead Kit v3 (10X Genomics), nuclei were immediately loaded onto a Chromium Single Cell Processor (10X Genomics) for barcoding of RNA from single nuclei. Sequencing libraries were constructed according to the manufacturer’s instructions and resulting cDNA samples were run on an Agilent Bioanalyzer using the High Sensitivity DNA Chip as quality control and to determine cDNA concentrations. The samples were combined and run on an Illumina NextSeq HO 150 run with 50 bp paired-end reads.

Single-cell ATAC-seq library preparation and sequencing

Samples were processed on a Sony SH800 Cell Sorter with a 100 mm sorting chip and nuclei were collected into 1.5 ml centrifuge tubes containing 10 µl of the Pre-FACS buffer. We collected ~ 10,000, ~ 15,000, and ~ 15,000 nuclei from the two breast normal, three breast primary tumor, and three breast tam-treated recurrent tumor tissue samples, respectively. Nuclei quality and quantity was evaluated using trypan blue on an Invitrogen Countess II device in duplicate, and a subset of nuclei was spun down in a fresh tube and resuspended in 10 × sample dilution buffer. Nuclei were then used for single-cell ATAC-seq library construction using the Chromium Single Cell ATAC Solution v1.0 kit (10 × Genomics) on a Chromium controller. Completed libraries were further quality checked for fragment size and distribution using an Agilent TapeStation prior to sequencing. Single-cell ATAC-seq samples were sequenced on an Illumina NextSeq HO 150 run with 50 bp paired-end reads.

Cell lines and reagents

Human breast cancer cell lines MCF7, T47D and their tamoxifen-resistant MCF7TR, T47DTR cells were derived from previous studies [24–26]. Cells were cultured in RPMI1640 medium (Thermo Fisher Scientific, Catalog #A1049101) supplemented with 10% fetal bovine serum (FBS) (Thermo Fisher Scientific, Catalog # SH30071.03) and 1% Penicillin–Streptomycin (Thermo Fisher Scientific, Catalog # 15,140,122). Tamoxifen-resistant cells were cultured in phenol-red free RPMI1640 medium (Thermo Fisher Scientific, Catalog # 11835030) with 10% charcoal-stripped FBS (Thermo Fisher Scientific, Catalog # 50–165-7328) and 1% Penicillin–Streptomycin and 100 nM tamoxifen (Sigma-Aldrich, Catalog #H7904-5MG). Tamoxifen was replaced every 48 h. All the cells were grown at 37 °C and 5% CO2 until they reach 90% confluence. Cell absorbance was recorded at different time points. For 4-hydroxytamoxifen (4-OHT, Millipore sigma) experiments, 3 µM of 4-OHT was added to MCF7TR and T47DTR cells.

Small interfering RNAs transfection

Small interfering RNAs (siRNA) targeting the BMP7 were purchased from Thermo Scientific. Two different BMP7 specific siRNAs: siRNA1 (Assay ID: s2035), siRNA2 (Assay ID: s2036) and Scramble siRNA (negative control-43–908-43) were transfected into cells using Lipofectamine® 2000 reagent (Invitrogen, Life Technologies). Briefly, MCF7, T47D, MCF7TR, and T47DTR cells (2 × 10⁵ cells/well) were seeded into six-well plates containing an antibiotic-free medium and were incubated overnight at 37 °C. For transfection, siRNA and Lipofectamine® 2000 mix (1:3) was prepared in serum free medium (Opti-MEM). The mixture was incubated at room temperature for 20 min to facilitate the formation of siRNA-Lipofectamine complex. The mixture was added to the cells in an appropriate volume of Opti-MEM to achieve a final concentration of 100 nm for each siRNA. After incubation for 6 h at 37 °C, RPMI supplemented with serum was added and cells were cultured for 24 h before analysis [27].

CCK-8 cell viability assay

Cell viability was measured by CCK-8 (CCK-8, Dojindo, USA) assay following the manufacturer’s instructions. In brief MCF7, T47D, MCF7TR, and T47DTR cells were harvested and plated at a density of 1 × 10³ cells per well in 96-well plates (Corning Inc.) and cultured in an incubator 5% CO2 incubator at 37 °C. After 24 h, the cells were transfected with different siRNAs. At the end of each time point, 10 μl of CCK-8 solution was added to each 96-well plates and the mixture was incubated for 1 h in the incubator at 37 °C. The optical density at 450 nm was measured at different time points using BioTek ELx800 Absorbance Microplate Reader [25, 26]. The experiments were repeated and analyzed three times separately.

RNA isolation and real-time PCR

Total RNA was isolated using Quick-RNATM Mini Prep (Zymo Research, USA) according to the manufacturer’s instructions. Five million cells from MCF7, T47D, MCF7TR, and T47DTR were lysed in RNA lysis buffer followed by eliminating the majority of gDNA with Spin-Away Filter. Then, the mixture of RNA and ethanol was loaded onto Zymo-Spin IIICG Column followed by DNase I treatment to remove the traces of DNA. The samples are washed twice with RNA wash buffer and the total RNA was eluted in 50 μl DNase/RNase-Free Water. Two microgram of RNA was used to convert to cDNA using a high-capacity cDNA reverse transcription kit (Applied Biosystem). qPCR was conducted with a Power SYBRTM Green PCR Master mix (Applied Biosystem). Real-time qPCR was performed on QuantStudio 3 Real-Time PCR system (Applied Biosystem, USA) according to the manufacturer’s instructions. The relative expression of RNAs was determined by the ΔΔCT method using ACTB as an internal control for quantification analyses of gene targets [25, 26]. Primers used are listed in Additional file 2: Table S2. Each PCR reaction was performed in triplicate, and the data presented were the average of three independent experiment results for all PCR reactions.

Western blotting

MCF7, T47D, MCF7TR, and T47DTR cells were seeded in six well plates of 2 × 10⁵ cells/well and grown for 24 h and then subjected to transfection. After transfection, the cells were lysed with cold RIPA lysis buffer (Pierce; Thermo Fisher Scientific, USA). The protein concentration was determined with the BCA Protein Assay kit (Pierce; Thermo Fisher Scientific, USA). On 12% SDS-PAGE, 25 µg of protein were separated and then transferred onto a polyvinyl difluoride (PVDF) membrane (Bio-Rad, USA). Membranes were then blocked with 5% bovine fetal serum at room temperature for 1 h. The membranes were subsequently incubated with primary antibodies at 4 °C overnight followed by secondary antibody for 1 h at room temperature. The blots were developed with detected by enhanced chemiluminescence detection kit (Bio-Rad Laboratories Inc.). The band intensity was quantified using NIH Image-J software (National Institutes of Health, Bethesda, MD, USA) [28]. Primary antibodies used were BMP7 (Proteintech, cat. no. 12221–1-AP, 1:500), MAPK (Cell signaling, cat.no. 9102, 1:500), p-MAPK (Cell signaling, cat.no. 9101, 1:500), and β-actin (Cell signaling, cat.no. 4967S, 1:1000). Horseradish peroxidase conjugated secondary antibodies against rabbit (cat. no. 656120; 1:5000) were purchased from Invitrogen.

Single-cell RNA-seq data preprocessing

Single-cell RNA-seq raw reads were aligned and assigned to Ensembl GRCh38 transcripts using the CellRanger v7.0.1 pipeline (10X Genomics Cloud analysis) with default parameters. Preprocessed data of each sample was then passed into the R package Seurat (v4.3.0) [29] with min.cells = 3 and min.features = 200. Each sample’s Seurat object (seu.obj) was then passed into our homemade quality control (QC) pipeline. The QC pipeline contains three steps: 1. R package scDblFinder (v.1.3.5) was used to identify doublets in each seu.obj; 2. 3 × median absolute deviation (MAD) was applied to automatically find the cutoffs for nCount, nFeature, and percent.mito; 3. Filters were applied to each tissue seu.obj to remove cells marked as doublets or outliers based on the results in QC steps 1 and 2. After the QC step, the seu.obj of NTs, PTs, and RTs were normalized and scaled with the SCTransform function, the linear dimensional reduction was performed on scaled data, and principal components (PCs) were identified by PCA analysis. For seu.obj integration, reference anchors were identified with the SelectIntegrationFeatures, PrepSCTIntegration, and FindIntegrationAnchors functions, and integrated data were then processed by the IntegrateData function. Clustering was performed using the Seurat functions FindNeighbors and FindClusters (resolution = 0.9). Clusters were then visualized with UMAP. All data was visualized with the SCT assay, and plots were generated using Seurat’s built-in visualization functions.

Cell type annotation

We first used the FindAllMarkers function in Seurat to identify the top 20 markers for each cluster and used Featureplot, Vlnplot, and Dotplot to visualize their expression levels. Meanwhile, we curated a set of breast normal and cancer cell marker genes from the PanglaoDB [30], CellMarker [31], ProteinAltas [32], and several literatures [33–35]. We then annotated each of clusters with the top 20 markers against the curated cell type markers, and finally derived the cell types for the clusters.

Tumor heterogeneity assessment

We assessed tumor tissue heterogeneity in two levels: cell type level and cell-to-cell level. We directly counted the number of cells in each tissue sample for assessing the cell type level of heterogeneity. For assessing the cell-to-cell level of heterogeneity, we followed the method described in the previous work [15]. In general, we built a cell-to-cell correlation matrix with each row and column representing a cell. The value for each element in the correlation matrix was the Pearson’s coefficient calculated from the scRNA-seq expression matrix with Pearson correlation analysis in R.

RNA velocity analysis

We performed RNA velocity analysis based on scVelo [36]. We first converted data from Seurat format into scVelo usable format. Then, we constructed spliced and unspliced count matrices for each tissue sample using velocyto [37]. We used the dynamic model in scVelo to compute RNA velocity based on the Seurat-converted clustering data and velocyto-generated matrices. We measured the speed and coherence of the velocities with velocity_confidence. The definition of the speed and the coherence of velocities are followed in a previous study [36]: The speed of velocities represents the rate of the differentiation equaling to the length of the velocity vector; the coherence of velocities measures how a velocity vector correlates with its neighboring velocities, which provides a measure of confidence. We computed transition probability from velocities with the velocity_graph function. Latent time was generated by recover_latent_time function, and the paga function orients the directionality of partition-based graph abstractions. All other downstream analyses followed instructions in scVelo standard pipeline.

Identification of cancer cell states

To identify cancer cell states, we first curated six gene expression modules associated with the essential biological processes in breast cancer: ER signaling, HER2 signaling, tumor invasion, proliferation, immune response, and angiogenesis from previous studies [38, 39]. The AddModuleScore function in Seurat was then used to add the module score of each module to all cells. For each cluster of cancer cells, we calculated the average module score and used the hierarchical clustering method to classify 13 cancer clusters into nine cancer cell states. PCA plot was used to visualize the cancer cell clusters in PC latent space. To obtain the expression features for RT-specific CSs, we identified differential markers between each RT_CS and all PT_CSs with the FindMarker function in Seurat and selected upregulated markers to represent the expression features for each of RT-specific CSs.

Single-cell ATAC-seq preprocessing

Single-cell ATAC-seq raw reads were demultiplexed using CellRanger-ATAC mkfastq (Cell Ranger ATAC, version 2.1.0, 10 × Genomics) and aligned to the GRCh38 reference genome and quantified using CellRanger-ATAC count function with default parameters. The preprocessed data of each sample was then passed into the R package Signac (v1.9.0). We first used Signac to create Seurat obj from scATAC-seq data. Each sample’s Seurat object (seu.obj) is then passed into our homemade quality control (QC) pipeline. The QC pipeline contained several QC matrices: 1. Nucleosome banding pattern (nucleosome_signal) was measured by counting the DNA fragment size. We quantified the approximate ratio of mono-nucleosome to nucleosome-free fragments for each cell. We removed the cells with nucleosome_signal > 4; 2. The transcriptional start site (TSS) enrichment score was measured by calculating the ratio of fragments centered at the TSS to fragments in the TSS-flanking region. We removed the cells with TSS enrichment score < 2; 3. The total number of fragments in peaks measured the cellular depth and complexity. We set a lower cutoff of 3000 and an upper cutoff of 25,000; 4. The fraction of fragments in peaks was calculated by the ratio of the fragments within peaks to all fragments. We dropped cells with less than 20% for this step; 5. Fraction of reads in genomic blacklist region. We dropped cells with more than 5% for this criterion. After the QC step, we first used reciprocal latent semantic indexing (LSI) projection by applying a series of functions in Signac (FindTopFeatures, RunTFIDF, RunSVD) to project cells into a share low-dimensional space. Then, we found integration anchors with the FindIntegrationAnchors function. After anchors identification, samples were integrated with corrected LSI embedding values calculated by IntegratEmbeddings [29].

Integration of scRNA-seq and scATAC-seq data

We used Signac [29, 40] to integrate scRNA-seq and scATAC-seq data by following steps: 1. Estimating the transcriptional activity of each gene by quantifying fragments in the 2000 bp upstream region and gene body and generating ACTIVITY assay by using GeneActivity function; 2. Normalizing and scaling the ACTIVITY assay and identifying anchors for integration with FindTransferAnchors; 3. Transferring annotations from scRNA-seq data onto scATAC-seq with TransferData and AddMetaData; 4. Imputing RNA expression into scATAC-seq based on anchors identified from step 2 and co-embedding scRNA-seq and scATAC-seq with merge function in Seurat.

Differential gene expression and peak analysis

To identify differentially expressed (marker) genes and peaks for clusters, cell types, and cell states, the functions FindAllMarkers (one.vs.all comparisons) and FindMarkers (one.vs.one) from the Seurat package were used with default parameters. Significant differentially expressed genes (markers) were selected as those with adjusted p values less than 0.05, average log2 fold change larger than 1, and percentage of cells with expression higher than 0.1. Significant differentially accessibility peaks were selected as those with adjusted p values less than 0.05, average log2 fold change larger than 0.25, and percentage of cells with expression higher than 0.1.

Identification of a heterogeneity-guided core signature

After integrating scRNA-seq and scATAC-seq data, cells from scATAC-seq were assigned to cancer cell states identified from scRNA-seq. We then detected differential peaks for each RT_CS as described above. We defined heterogeneity-guided accessibility features as the set of regions with increased accessibility in all RT_CSs. We further identified genes as part of the heterogeneity-guided core signature by meeting the following criteria: 1. the gene is listed in the heterogeneity-guided expression features; 2. the promoter of the gene is listed in the heterogeneity-guided accessibility features; and 3. the distal or proximal regions of the gene are listed in the heterogeneity-guided accessibility features.

Inferring co-accessibility for the heterogeneity-guided core signature

To calculate the co-accessibility of the heterogeneity-guided core signature, the Signac object was first converted into a Cicero object using the cell_data_et and make_cicero_cds functions with default parameters. The data frame of genome coordinates was generated with the BSgenome.Hsapiens.UCSC.hg38. Next, the run_cicero function was applied to estimate the co-accessibility of sites in the genome and calculate pairwise co-accessibility scores for each peak. We then performed the generate_ccans function to group pairwise connections into larger co-accessible networks. The ConnectionToLinks function was used to convert co-accessible networks into Signac-compatible links. Finally, the co-accessibility score of the heterogeneity-guided core signature was calculated by summing the co-accessibility scores of distal-promoter or proximal-promoter pairwise peaks with homemade python script.

Overrepresented TF motifs analyses

We performed TF motif analyses based on the genomic location and gene expression value. We first defined distal, proximal, promoter, and intergenic regions based on the distance to TSS. The definition of four genomic regions was followed in the previous study [41]. Specifically, the distal region was defined as 200 kb-upstream to 2 kb-upstream and 2 kb-downstream to 200 kb-downstream. The proximal region was defined as 2 kb-upstream to 200 bp-upstream and 200 bp-downstream to 2 kb-downstream. The promoter was defined as 200 bp-upstream to 200 bp-downstream. The intergenic region was defined as the region that does not belong to any distal, proximal, or promoter regions. We then used FindAllMarker with a logistic regression framework and custom Python script to identify the differential accessibility peaks (DAs) for all cell states in different genomic regions. To identify overrepresented TF motifs in DAs or HCS in distal and proximal regions, we acquired the DNA sequence motif information from JASPAR2022 [42] with getMatrixSet in TFBSTools [43] and added into scATAC-seq seu.obj with AddMotifs function in Signac. We then found background peaks with AccessiblePeaks, GetAssayData, and MatchRegionStats functions in Signac. Finally, we used FindMotifs to identify the overrepresented motifs.

Inference and analysis of cell–cell communication and metabolite-mediated cell communication

CellChat [44] was used to visualize and analyze the intercellular communications from scRNA-seq data. It infers cell-state specific signaling communications within a given scRNA-seq data using mass action models, along with differential expression analysis and statistical tests on cell groups, which can be both discrete states and continuous states along the pseudotime cell trajectory. To predict significant communications, CellChat identifies differentially over-expressed ligands and receptors for each cell group. To quantify communications between two cell groups mediated by these signaling genes, CellChat associates each interaction with a probability value. Finally, a set of biological signaling pathways was inferred to mediate the communication between cancer cell states and non-tumor cells.

MEBOSCOST [45] was applied to predict metabolite-mediated cell–cell communications from scRNA-seq data. It utilizes expression levels of enzyme and sensor genes to define sender and receiver cells based on metabolite efflux and influx rates from manually curated database. In general, we converted Seurat object into Scanpy object with Convert function in SeuratDisk. Then, a mebocost object is created by create_obj function with expression data and cell annotation information loaded. Next, the metabolite-mediated cell communication is inferred by infer_commu function with default parameters. Additionally, we performed COMPASS [46] to further constrain mCCC by efflux and influx rates as instructed, and update the commu_res in the mebocost object based on COMPASS output (secretion.tsv and uptake.tsv). Finally, the communications are visualized with several MEBOCOST visualization functions.

GO/pathway and survival analyses

We used DAVID functional tool [47] to perform the gene ontology (GO) and pathway analysis on the core signature. This tool has collected data libraries for transcriptional regulation, pathways and protein interactions, ontologies including GO and the human and mouse phenotype ontologies, signatures from cells treated with drugs, and expression of genes in different cells and tissues. To ensure the statistical significance of the analysis, we also included 58 downregulated genes common among three RT_CSs in the pathway analysis. We also used an online survival tool [48] to assess the effect of genes on breast cancer prognosis using microarray data of 1623 ER + breast cancer patients. In order to analyze the prognostic value of a particular gene, we first curated five cohorts [49–53] that contained both tamoxifen relapse and relapse-free patients. We then applied a faster version of BayesPrism [54], InstaPrism [55], to deconvolute the bulk cohort data and calculated the averaged expression matrix for RT/PT_CSs from cohort data. The cohorts were divided into two groups, relapse and relapse-free, according to the clinical information. And the comparison of core signature’s expression between relapse and relapse-free were visualized by Complexheatmap [56].

Results

Distinct tumor cell characteristics among breast tumor tissues

We conducted single-cell transcriptome profiling in 10 × Genomics multi-omics platform for two NTs, three PTs, and three RTs. A total of 39,525 single cells were sequenced for scRNA-seq data, with an average of 4940 cells per tissue (Fig. 1a). After quality control (QC) (see “Methods”), we removed low-quality cells and doublets (Additional file 3: Fig. S1) and obtained 26,184 high-quality singlets for further analysis.

Fig. 1.

Distinct tumor cell characteristics among breast tumor tissues. a An overview of scRNA-seq and scATAC-seq profiling and analysis on a total of eight breast tissues including two NTs, three PTs, and three RTs. b An UMAP (left panel) showing 12 clusters identified from two NTs by 17 markers and a violin plot (right panel) showing the different expression pattern of 17 markers in the clusters. c An UMAP (left panel) showing 15 clusters identified from three PTs by 20 markers and a dot plot (right panel) showing the expression pattern of 20 markers in the clusters. d An UMAP (left panel) showing 13 clusters identified from three RTs by 20 markers and a dot plot (right panel) showing the expression pattern of 20 markers in the clusters. e An UMAP showing six cell types annotated for 12 clusters in two NTs (left panel), seven cell types annotated for 15 clusters in three PTs (middle panel), and five cell types annotated for 13 clusters in three RTs (right panel), respectively

We utilized Seurat [29] to perform clustering on scRNA-seq data of NTs, PTs, and RTs separately, and identified 11, 15, and 13 clusters for NTs, PTs, and RTs, respectively (Fig. 1b–d—left panels). We then used the differential markers in each cluster (Fig. 1b–d—right panels, Additional file 4: Table S3) to annotate cell types for the clusters against several cell annotation databases [30–35]. We thus identified six cell types for NTs, luminal cells, basal cells, fibroblasts, endothelial cells, myeloid cells, and luminal progenitor (Fig. 1e—left panel); seven cell types for PTs (Fig. 1e—middle panel), breast cancer cells, myeloid cells, basal (non-cancerous) cells, adipocyte-like cells, luminal (non-cancerous) cells, neutrophils, and fibroblast-like cells; and five cell types for RTs (Fig. 1e—right panel), breast cancer cells, myeloid cells, luminal (non-cancerous) cells, stromal cells, and natural killer (NK) cells. Intriguingly, we found PTs and RTs have different breast cancer cell marker genes, including BAG1, BCL2, BRCA1, BRIP1, CCNB1, ESR1, KRT8, MKI67, and PGR for PTs and AURKA, CCNB1, ESR1, KRT19, MKI67, MUCL1, NEAT1, and XBP1 for RTs, indicating that distinct oncogenic programs exist between the primary and recurrent tumors (Fig. 1c, d—right panels). Additionally, we found that a proportion of luminal cells and luminal progenitors in NTs exhibited similar expression characteristics to breast cancer cells in PTs. The aberrantly expressed genes in these cells were enriched in signaling networks involved in cell cycle regulation, chromosome maintenance, spindle checkpoint, and ER-mediated signaling (Additional file 3: Fig. S2 and Additional file 5: Table S4). Moreover, we observed a higher number of immune-related cell types in the six tumor tissues (TTs) than NTs and a slight difference in the composition of cell types between PTs and RTs (Fig. 1e). Our data suggested that there existed not only distinct tumor cell characteristics, but also different TME components between primary and tamoxifen-treated recurrent tumors.

Tumor cell heterogeneity and progression between primary and recurrent stage

Next, we examined the composition of cell types in tumor tissues and found breast tumor cells were the major cell component in all of six TTs except RT3 (Fig. 2a). We then performed the cell-to-cell Pearson’s correlations for gene expression of six TTs and observed a higher similarity between cells from the same type of tissues, i.e., the correlations within PTs and RTs are more significant than the correlations between them (Fig. 2b). Further, for PTs, the correlations within tumor cells were higher than non-tumor cells and an overall correlation was enhanced by removing non-tumor cells, while for RTs, only the RT3 showed a higher correlation of tumor cells than non-tumor cells (Fig. 2b, c). To further understand the heterogeneity of the tumor cells, we extracted tumor cells of scRNA-seq data and re-mapped a total of 13 tumor cell-specific clusters, consisting of 7 PT-specific, 5 RT-specific, and 1 PT-RT-shared clusters (Fig. 2d), each with distinct breast cancer marker gene expression patterns (Fig. 2e, Additional file 3: Fig. S3a, and Additional file 6: Table S5). Our analysis thus revealed inter- and intra-tumor heterogeneity among tumors, where every tumor cell cluster included a mixture of cells originating from two or more distinct tumors (Fig. 2f).

Fig. 2.

Inter- and intra-tumor cell heterogeneity among breast tumor tissues. a A stacked bar plot showing the number of cells in each cell type in each of six TTs. b A cell-to-cell correlation matrix for single cells demonstrating low cell-to-cell correlations (Pearson’s r) between PTs and RTs (left). After separating tumor and non-tumor cells, cell-to-cell correlations were increased in PTs (right top) but unchanged in RTs (right bottom). Each row and column represent single cells. In the color panel on the far-right side, black represents tumor cells, and gray represents non-tumor cells. c Intra-sample correlations before (red boxes) and after (blue boxes) the removal of non-tumor cells. Each box shows the median and interquartile range (IQR 25th–75th percentiles), whiskers indicate the highest and the lowest value within 1.5 times the IQR, and outliers are marked as dots. d UMAPs showing 13 clusters (left) identified from tumor cells in six TTs and their annotation with tissue sample (right). e The stacked violin plots depicting the expression level of 15 known breast cancer-related genes in each cluster. f The proportional bar plot illustrating the ratio of tumor cells of six tumor tissues among 13 clusters. g The dimensional plots illustrating velocity confidence (coherence) of each cluster. h UMAPs of the latent time exhibiting the inferred transcriptional dynamics and tumor progression. i A PAGA graph showing the inferred directed abstracted representations of trajectories through a RNA velocity of tumor cells

We further performed RNA velocity analysis [36, 37] on the tumor cell clusters to infer the tumor progression from primary to recurrent tumors. We examined the proportions of spliced/unspliced counts for each cluster and found a higher percentage of unspliced molecules containing intronic sequences for PT-specific clusters at 31.1% than RT-specific clusters at 20.2% (Additional file 3: Fig. S3b). The speed and coherence measurement showed corresponding velocity speed patterns for PT- and RT-specific clusters and validation of the dynamic model (see “Methods,” Additional file 3: Fig. S3c, g). We further inferred the latent time for tumor cell clusters (Fig. 2h), where the latent time represented the cell’s internal transcriptional clock and approximated the cancer stage progression in real time. In consistent with our current knowledge, PT-specific clusters were at the early progression stage (low latent time value), the PT-RT-shared cluster has the middle level of latent time, while RT-specific clusters were at the late progression stage (high latent time value). We then inferred the trajectory of tumor progression from a partition-based graph abstraction (Additional file 3: Fig. 3i) and found that cluster 4 was the start point, clusters 5, 9, and 11 were the transition, and clusters 0 and 2 were the end point. We finally inferred driver genes for driving tumor progression via a high likelihood characterization in the dynamic model (Additional file 3: Fig. S3d, e). Together, our analysis demonstrated that single-cell level tumor cell clusters indeed reflected the underlying progression from primary tumor to recurrent tumor stage.

Cancer cell states of breast tumors

To further associate the 13 cancer cell clusters with breast cancer progression, we examined their relationship with six gene expression modules known to critical breast cancer biological processes, including ER signaling, HER2 signaling, proliferation, immune response, invasion, and angiogenesis (Additional file 7: Table S6). Interestingly, we found that each of six modules showed a preferred enrichment in one or two specific cell clusters, e.g., ER signaling and immune-response modules were enriched in PT-specific clusters while HER2 signaling and invasion modules were highly active in RT-specific clusters. The proliferation module was the most enriched in clusters 4 and 11 (Fig. 3a). Based on this enrichment combination of breast cancer modules, we were able to define nine breast cancer cell states (CSs) for breast tumors including five PT-specific cell states (PT_CSs), three RT-specific cell states (RT_CSs), and one PT-RT-shared cell state (PRT_CS) (Fig. 3b–d). When particularly examining the three RT_CSs, we identified 699, 640, and 394 differentially expressed genes (DEGs) for RT_CS1, RT_CS2, and RT_CS3 vs. all PT_CSs, respectively (Fig. 3e and Additional file 8: Table S7). We used upregulated genes in each of three RT_CSs to distinguish from each other (Fig. 3f), thus defined them as the RT_CS1-specific, RT_CS2-specific, and RT_CS3-specific gene expression features. We further identified an RT_CS-specific expression feature composed of 910 genes from all three RT_CSs’ gene expression features (Fig. 3g and Additional file 9: Table S8). In summary, we defined nine cancer cell states with distinct characteristics such that RT_CSs were highly enriched with HER2 signaling, invasion, angiogenesis, and higher latent time (Fig. 3h) as well as derived a heterogeneity-guided expression feature for RT_CSs.

Fig. 3.

Cancer cell states defined by breast cancer gene expression modules. a UMAPs visualizing the patterns of six breast cancer gene expression module scores. Six modules: ER signaling, HER2 signaling, proliferation, immune-response, invasion, and angiogenesis module. b A heatmap showing that K-mean hierarchical clustering identified nine cancer cell states based on six modules. c A PC plot illustrating the nine cancer cell states from b. Different colors representing different cell states. d A UMAP visualizing the nine cancer cell states. e Volcano plots depicting differentially expressed markers identified by comparing each RT_CS with all PT_CSs. f UMAPs visualizing the score of the expression features in each of three RT_CSs. g An upset plot presenting the intersection among expression features in three RT_CS. h A characterization of nine cancer cell states by distinct module scores and latent time

Epigenetic-regulated cancer cell states and a heterogeneity-guided core signature

We further conducted scATAC-seq profiling on the same eight tissue samples to study the epigenetic regulation. We sequenced a total of 42,290 cells with an average of 5286 cells per tissue and obtained 31,393 cells with highly confidently mapped reads after QC (Additional file 3: Fig. S4a–d). We acquired the open chromatin regions in the cancer cell states by integrating scATAC-seq data with scRNA-seq data (Fig. 4a and Additional file 3: Fig. S4e, f). We detected a total of 208,919 open chromatin regions located in four different genomic regions for all cell states (Fig. 4b and Additional file 10: Table S9) where the genomic regions, consisting of distal, proximal, and promoter as well as intergenic, were classified via the distance to annotated TSSs across the human genome (Additional file 3: Fig. S5a). By comparing with ENCODE candidate cis-regulatory elements (cCRE) registry [57], we found 61,064 novel open chromatin regions in cancer cell states with 21,611 PT_CSs specific and 27,074 RT_CSs specific as well as 12,380 in both cancer cell states (Additional file 3: Fig. S5b, Additional file 10: Table S9). To investigate the RT-specific chromatin accessibility, we first identified 18,457, 17,361, and 17,766 differential accessibility peaks (DAs) for RT_CS1, RT_CS2, and RT_CS3, respectively, in comparison to all peaks of PT_CSs (Additional file 3: Fig. S5c, Additional file 11: Table S10). Within all DAs, we identified 3292, 131, and 505 heterogeneity-guided accessibility features (HAFs) on distal, proximal, and promoter, respectively, by intersecting increased accessibility features across three RT_CSs (Fig. 4c). By integrating both heterogeneity-guided expression and accessibility features, we defined 137 genes (Additional file 12: Table S11) as a heterogeneity-guided core signature (HCS), which is highly expressed in RT_CSs and associated with increased accessibility at both promoter and proximal/distal, indicating synergistic promoter-enhancer epigenetic regulation of their expression (Fig. 4d). We further used a co-accessibility score defined by Cicero [58] to quantitatively measure the association between the expression and the accessibility for the HCS genes (Fig. 4e). Indeed, a higher co-accessibility score was associated with a higher expression and increased accessibility. We next conducted gene ontology (GO) and biological pathway analysis on this core signature. Interestingly, our analysis showed that morphogenesis of epithelium, epithelial to mesenchymal transition, response to estradiol, DNA repair, insulin signaling, and HER2 signaling were among the top enriched signaling pathways (Fig. 4f). To further study the potential trans-acting factors on driving tamoxifen resistance, we first searched the top overrepresented transcription factor (TF) motifs at the distal region DAs for each RT_CS (Additional file 3: Fig. S5d). We then narrowed down the regions to 524 distal/proximal increased accessibility regions of the core signature and detected the top TF motifs within those regions. Interestingly, we identified several TFs that are the known breast cancer oncogenic TFs such as ESR1, FOXA1, FOS, FOSL1::JUNB, and FOSL1::JUND, where FOS and FOSL1::JUNB were shown to be highly enriched in RT_CSs with a footprint analysis (Additional file 3: Fig. S5e). Finally, we confirmed that the chromatin accessibility of the HCS genes of RT_CSs were more accessible compared to those of PT_CSs (Additional file 3: Fig. S5f). Together, our integrative analysis identified a heterogeneity-guided core signature and key TFs associated with RT-specific cancer cell states, indicating their crucial role in driving tumor recurrence.

Fig. 4.

Epigenetic-regulated cancer cell states and a heterogeneity-guided core signature. a UMAPs illustrating an integration of the scRNA-seq and scATAC-seq data (left) and the geometric location of each of nine cancer cell states on co-embedded datasets (right). b The stacked bar plot displaying the number of peaks identified from scATAC-seq data across distal, proximal, promoter, and intergenic regions for each cancer cell state. c The upsetting plot presenting the number of UpAs in distal, proximal, and promoter regions that were common or specific among RT_CS1, RT_CS2, and RT_CS3. d A plot demonstrating the heterogeneity-guided core signature contains 137 genes resulted from incorporating both heterogeneity-guided expression and accessibility features. e A scatter plot showing the average log2 fold change of expression and accessibility as well as the co-accessibility score for 137 genes in HCS. f A bubble plot showing the top GO terms, KEGG/Wiki/REACTOME pathways for the core signature, respectively. g The top overrepresented TF motifs detected in 524 distal/proximal increased accessibility regions of 137 genes in HCS

Cell–cell communication between cancer cell states and non-tumor cells

To delineate functionally related signaling and metabolism pathways mediating intercellular communications between cancer cell states and non-tumor cells, we first applied CellChat [44], which mainly uses mass action models to infer cell–cell communications. We detected a total of 1226 significant interactions of ligand-receptor pairs between three RT-CSs and non-tumor cells, luminal, NK, myeloid, and stromal cells (Fig. 5a). We then ranked all identified signaling pathways with the probability either as source (ligand) (Fig. 5b) or as target (receptor) (Fig. 5c). Interestingly, as source, LAMININ, COLLAGEN, and FN1 were the top three common signaling pathways for all of three RT_CSs, but GRN, WNT, and EGF were respectively RT_CS1-specific, RT_CS2-specific, and RT_CS3-specific signaling pathway (Fig. 5d). However, as target, BMP, EGF, and LAMININ were among the top three common signaling pathways while OCLN, PSAP, and CDH1 were RT_CS1-specific, RT_CS2-specific, and RT_CS3-specific signaling pathway, respectively (Fig. 5e). When closely examining the enrichment of ligands and receptors within the signaling pathways, we found each signaling pathway utilized specific ligand-receptor pairs to communicate between RT_CSs and non-tumor cells (Fig. 5f, g and Additional file 3: Fig. S6). For example, in LAMININ signaling pathway, the top enriched ligands, LAMA5 and LAMC1, were mainly pairing with the enriched receptor, ITGB1 (Fig. 5f), while in BMP signaling pathway, BMP7 was the major ligand in non-tumor cells and BMPR1A, BMPR1B, ACVR1, and BMPR2 were the major receptors in RT_CSs (Fig. 5g). For five PT_CSs, we identified a total of 3496 significant interactions of ligand-receptor pairs and a different set of ranked signaling pathways as source or as target (Additional file 3: Fig. S7). Additionally, we also investigated metabolite-mediated cell communication by applying MEBOCOST [45]. Interestingly, we identified estrogen-ESR1 as a metabolite-mediated cell communication in both PTs and RTs (Additional file 3: Fig. S8, Additional file 3: Fig. S9), which has been reported to induce carcinogenesis [59]. Despite the receivers for estrogen-ESR1 are cancer cell states in both PTs and RTs, the senders are different, e.g., in PTs, fibroblast-like cells are the major senders, indicating the potential role of cancer-associated fibroblasts in the tumor microenvironment in supporting cancer through metabolite-mediated communication. In contrast, in RTs, the primary sender is PRT_CS1, suggesting a positive feedback loop between transitioning cancer cell states and recurrent cancer cell states. Additionally, we observed enriched palmitic acid/stearic acid-SLC27A5 communication between adipocyte-like cells and PT cancer states (Additional file 3: Fig. S8) and identified a high communication score for L-serine-SLC3A2 among RT_CSs (Additional file 3: Fig. S9). Together, our data suggested that distinct signaling and metabolite pathways mediated intercellular communications between cancer cell states and non-tumor cells.

Fig. 5.

Cell–cell communication between RT_CSs and non-tumor cells. a Number of ligand-receptor interactions among the three RT_CSs and four non-tumor cell types. b A bar plot showing the communication probability of all signaling pathways identified by CellChat for RT_CS1, RT_CS2, and RT_CS3 as source. c A bar plot showing the communication probability of all signaling pathways identified by CellChat for RT_CS1, RT_CS2, and RT_CS3 as target. d A circle network plot showing communication networks of three RT_CSs common signaling pathways, LAMININ, COLLEGAN, and FN1 and three RT_CSs specific signaling pathways, GRN, WNT, and EGF as source. The width of the edges indicates the communication probability. e A circle network plot showing communication networks of three RT_CSs common signaling pathways, BMP, EGF, and LAMININ and three RT_CSs specific signaling pathways, OCLN, PSAP, and CDH1 as target. The width of the edges indicates the communication probability. f A violin plot of the ligands and receptors involved in LAMININ signaling pathway. Purple color indicates the ligand and dark green color indicates the receptor. g A violin plot of the ligands and receptors involved in BMP signaling pathway. Purple color indicates the ligand and dark green color indicates the receptor

Functionally examination of a core signature gene

We next assessed the effect of the HCS genes on breast cancer prognosis and found many of those genes showed better survival on tamoxifen-treatment patients (Fig. 6a and Additional file 3: Fig. S10a). We further validated that the core signature showed differential expression between relapse-free and relapse clusters for Creighton (n = 60), Nagalla (n = 139), and Chin (n = 130) cohorts of breast cancer patients with tamoxifen treatment, respectively (Additional file 3: Fig. S10b).

Fig. 6.

Resensitization of TR cells to 4-OHT treatment upon the silence of BMP7. a K-M plots showing the relapse-free survival probability of BMP7 in systemic untreated patients vs. Tam-treated patients. b Knockdown of BMP7 by siRNAs in MCF7TR and T47DTR cells. Cells were transfected with siRNA. Relative mRNA expression levels were quantified by the qRT-PCR method, using β-actin as an internal control. The data were represented by the mean ± SD. ***p ≤ 0.0005, **p ≤ 0.001, *p ≤ 0.05 vs. control. c Effect of BMP7 siRNAs on cell proliferation in MCF7TR and T47DTR cell lines. Cell proliferation was measured by CCK-8 assay after transfection with BMP7 siRNAs. The data were represented as mean ± SD. ***p ≤ 0.005, **p ≤ 0.05 vs. control. d Effect of BMP7 siRNAs on cell proliferation in MCF7TR and T47DTR cell line treated with 4-OHT. Cell proliferation was measured by CCK-8 assay after transfection with BMP7 siRNAs. The data were represented as mean ± SD. ***p ≤ 0.005, **p ≤ 0.05 vs. control

We further performed in vitro functional validation on one of the core signature genes, BMP7, in tamoxifen-resistant (TR) breast cancer cell lines, MCF7TR and T47DTR. The selection of BMP7 is based on the following evidence: (1) BMP7 is the top log-likelihood gene for the majority of RT-specific clusters (Additional file 3: Fig. S3d); (2) BMP7 is one of the top RT_CS-specific core signature (Fig. 4e, Additional file 9: Table S8); (3) BMP7 is the main source and target in BMP signaling pathway identified as a major cell–cell communication between cancer cell state and non-tumor cells (Fig. 5c, e, g); and (4) BMP7 is among the top GO pathway, cell morphogenesis and its higher expression associated with poor survival in tamoxifen-treat patients (Figs. 4f and 6a). We used two BMP7 siRNA to direct against target regions of BMP7 mRNA. BMP7-siRNA1 was directed against a region on exon 2 whereas BMP7-siRNA2 targeted a region on exon 6. RT-PCR results demonstrated a significant reduction in BMP7 mRNA levels in MCF7TR cells following transfection with BMP7-siRNA1 and BMP7-siRNA2, resulting in decreases of 62% and 71%, respectively, when compared to the scramble siRNA (Scr-siRNA) and the non-transfected control group (p ≤ 0.0005, Fig. 6b). Similarly, in T47DTR cells, BMP7 mRNA expression was significantly reduced by 55% with BMP7-siRNA1 and by 30% with BMP7-siRNA2, compared to the Scr-siRNA and the control (p ≤ 0.0005, Fig. 6b). Both siRNA1 and siRNA2 exhibited higher efficacy in downregulating BMP7 expression in MCF7TR and T47DTR cells compared to MCF7 cells (Additional file 3: Fig. S11a) and T47D cells (Additional file 3: Fig. S12a). Subsequent cell proliferation assays on BMP7 siRNA-treated MCF7TR and T47DTR cells revealed a significant reduction in cell growth in BMP7 siRNA-treated MCF7TR cells compared to Scr-siRNA-treated MCF7TR and T47DTR cells in a time-dependent manner (Fig. 6c). Notably, both siRNA1 and siRNA2 demonstrated greater efficacies in downregulating BMP7 expression in MCF7TR and T47DTR cells compared to MCF7 and T47D (Additional file 3: Figs. S11c, S12c). Additionally, when BMP7 siRNA-treated MCF7TR and T47DTR cells were exposed to 4-OHT, an active metabolite of tamoxifen, a marked decrease in cell growth was observed compared to Scr-siRNA treatment (Fig. 6d).

Western blot analysis revealed a significant reduction in BMP7 protein levels following siRNA-mediated knockdown in MCF7TR cells, with siRNA1 and siRNA2 achieving reductions of 55% and 30%, respectively, compared to the control (p ≤ 0.005, Fig. 7a). In T47DTR cells, BMP7 protein expression was also significantly diminished by 28% with siRNA1 and 54% with siRNA2 relative to the control (p ≤ 0.005, Fig. 7b), with these reductions being more pronounced than those observed in MCF7 and T47D cells. Notably, both siRNA1 and siRNA2 were more effective in downregulating BMP7 expression in tamoxifen-resistant cells compared to their tamoxifen-sensitive cells (Additional file 3: Figs. S11b, S12b). To further elucidate the role of BMP7 in MAPK signaling within tamoxifen-resistant cells, we performed additional western blot analyses. Knockdown of BMP7 led to a significant reduction in MAPK protein levels, with decreases of 63% and 56% observed with siRNA1 and siRNA2, respectively. The phosphorylation of MAPK was also reduced, with decreases of 68% and 47% achieved by siRNA1 and siRNA2, respectively, in MCF7TR cells compared to MCF7 cells (p ≤ 0.005, Fig. 7c, Additional file 3: Fig. S11d, Additional file 13: Supplementary data 1). A similar pattern was observed in T47DTR cells, where MAPK protein levels were reduced by 38% with siRNA1 and 62% with siRNA2, and p-MAPK levels were decreased by 45% with siRNA1 and 41% with siRNA2, compared to T47D cells (p ≤ 0.005, Fig. 7d, Additional file 3: Fig. S12d). These data demonstrated that BMP7 plays an oncogenic role in tamoxifen-resistant breast cancer cells through modulating MAPK signaling pathways, suggesting that targeting BMP7 may offer a potential therapeutic strategy for overcoming tamoxifen-associated endocrine resistance.

Fig. 7.

Modulation of p-MARK/MAPK upon the silence of BMP7. a Representative western blot of BMP7 and β-actin proteins from cells transfected with siRNAs (top panel); the expression levels of BMP7 protein in MCF7TR cells transfected with siRNAs (bottom panel). The expression level of each band was measured by densitometry and normalized to corresponding β-actin. Results were expressed in relation to the control. The data were represented by the mean ± SD. ***p ≤ 0.005, **p ≤ 0.05 vs. control. b Representative western blot of BMP7 and β-actin proteins from cells transfected with siRNAs (top panel); the expression levels of BMP7 protein in T47DTR cells transfected with siRNAs (bottom panel). The expression level of each band was measured by densitometry and normalized to corresponding β-actin. Results were expressed in relation to the control. The data were represented by the mean ± SD. ***p ≤ 0.005, **p ≤ 0.05 vs. control. c Representative western blot of MAPK and p-MAPK proteins from cells transfected with siRNAs (top panel); the expression levels of MAPK and p-MAPK in MCF7TR cells transfected with siRNAs (bottom panel). The expression level of each band was measured by densitometry and normalized to corresponding β-actin. Results were expressed in relation to the control. The data were represented by the mean ± SD. ***p ≤ 0.005, **p ≤ 0.05 vs. control. d Representative western blot of MAPK and p-MAPK proteins from cells transfected with siRNAs (top panel); the expression levels of MAPK and p-MAPK in T47DTR cells transfected with siRNAs (bottom panel). The expression level of each band was measured by densitometry and normalized to corresponding β-actin. Results were expressed in relation to the control. The data were represented by the mean ± SD. ***p ≤ 0.005, **p ≤ 0.05 vs. control

Discussion

Recent advances in single-cell multi-omics sequencing allow us to explore tumor heterogeneity at single-cell resolution. In this study, we conducted an integrative single-cell analysis to characterize transcriptomes and chromatin accessibility in breast normal, primary, and Tam-treated recurrent tumor tissues. We identified a variety of cell types in breast tumor tissues and found a majority of cells were tumor cells for all six tumors except the third recurrent tumor (Fig. 2). We found that a subset of luminal cells and luminal progenitors in NTs exhibited breast cancer cell-like expression characteristics, which is consistent with previous studies [33]. Although surprisingly we failed to identify any T cells or B cells in our tissue samples, we were cautious about linking it to any biological functions since lymphoid cells are highly labile ex vivo and therefore are difficult to preserve with the single-cell isolation methods [60]. Nevertheless, our data may implicate low tumor-infiltrated lymphocytes in our tumor tissues. Further, we revealed nine distinct cancer cell states for breast tumors and identified epigenetic factors and novel biological signaling pathways that mediate their communications within the breast tumor microenvironment.

Interestingly, our RNA velocity analysis was able to delineate the breast tumor progression from primary to recurrent stages, underscoring the significance of these single-cell level tumor cell clusters classified from scRNA-seq data. Although such analysis could also infer the driver genes at each stage of tumor progression, we should point out that the driver genes are purely from a computational inference and should be cautiously used. Further analyses and experimental validations are needed to confirm their oncogenic roles in driving breast tumor progression, particularly in Tam-treatment recurrence.

We identified many interesting signaling and metabolite-mediated pathways, either as sources (senders) or as targets (receivers) that mediate the communications between cancer cells and non-tumor cells within TME (Fig. 5). Our data revealed the signaling pathways related to cancer cell adhesion and invasion, including LAMININ, COLLAGEN, and FN1, were the major signaling pathways to regulate the interactions between stromal cells and cancer cells, indicating they might play crucial roles in driving the resistance to endocrine therapy treatment. Indeed, stromal cells have been shown to promote breast cancer progression and endocrine resistance [61–63]. Further, we also observed NK cells had the fewest communications between cancer cell states and non-tumor cells, suggesting a very low immune response microenvironment. In the metabolite-mediated cell communication analysis, we not only revealed estrogen-ESR1 metabolite-mediated cell communication between in PTs and RTs, but also uncovered enriched palmitic acid/stearic acid-SLC27A5 communication between adipocyte-like cells and PT cancer states, recapitulating the process where fatty acids generated in adipocytes are taken up by breast cancer cells, supporting mitochondrial oxidation [64]. Interestingly, we observed a high communication score for L-serine-SLC3A2 between RT_CSs, emphasizing the essential role of this cysteine transporter in the development of tamoxifen-resistant breast cancer [65–67]. Nonetheless, all of these signaling and metabolite-mediated pathways will be a focus of future work to clarify their oncogenic functional roles in the context of tumor microenvironment and delineate the underlying mechanism that links endocrine resistance and tumor microenvironment.

Strikingly, we were able to identify a heterogeneity-guided core signature that could potentially have prognostic values in predicting tamoxifen-associated endocrine treatment. We specifically identified this core signature as tumor heterogeneity-guided because these commonly upregulated genes with increased accessibility at both promoter and distal/proximal regions were derived from the three distinct RT-specific cancer cell states, which are influenced by the heterogeneous tumor microenvironment (Fig. 3f, g). We then conducted an in vitro functional characterization on one of the core signature gene, BMP7. Interestingly, BMP7 has been extensively studied in various human cancers, where it is associated with metastasis and poor prognosis [68–71]. However, to the best of our knowledge, no existing studies have been reported on its role in tamoxifen-resistant breast cancer. Our functional examination demonstrated that BMP7 significantly inhibited the growth of MCF7TR and T47DTR cells and exhibited an additive effect in re-sensitizing these tamoxifen-resistant cells to 4-hydroxytamoxifen (4-OHT) treatment. To gain further insights into the mechanism by which BMP7 contributes to breast cancer tamoxifen resistance, we performed BMP7 knockdown using small interfering RNAs (siRNAs). This knockdown resulted in decreased BMP7 mRNA and protein expression in both tamoxifen-resistant cell lines. Further investigation into the underlying biological mechanisms revealed that BMP7 likely modulates the MAPK signaling cascade. BMP7 is known to activate MAPK pathways through Smad-independent mechanisms, potentially involving receptor-associated kinases or other intermediary signaling molecules [49, 50]. In tamoxifen-resistant cells, MAPK signaling is crucial for maintaining cell proliferation and survival despite the inhibitory effects of tamoxifen on estrogen receptor (ER) signaling. By reducing BMP7 levels, we observed diminished MAPK activation, leading to decreased MAPK phosphorylation and activity. This reduction may partially restore the sensitivity of these cells to tamoxifen or inhibit their growth and survival. Our findings underscore the pivotal role of BMP7 in maintaining MAPK signaling and suggest that targeting BMP7 could be a promising therapeutic strategy for overcoming MAPK-mediated resistance in breast cancer cells. Further studies are needed to further elucidate the molecular mechanistic role of BMP7 in driving breast cancer endocrine resistance.

As previously reported [72–74] and newly demonstrated in this study, breast cancer is highly intra-tumoral heterogeneous and intricately interacted with the tumor microenvironment, and the underlying mechanisms are diverse. We acknowledge that the sample sizes in this study might limit the generalizability of our conclusions to all tamoxifen-resistant breast cancers. A further longitudinal study examining pre- and post-treatment samples from the same patients would be necessary to draw comprehensive conclusions about the mechanisms of tamoxifen resistance. To the best of our knowledge, this is the first study to integrate scRNA-seq and scATAC-seq data to assess inter- and intra-tumor heterogeneity for primary and tamoxifen-treated recurrent breast cancer patients.

Conclusions

In summary, we observed a striking level of heterogeneity in all breast tumors with more than nine epigenetically characterized cancer cell states and identified novel biological/metabolism pathways that mediate the communications between cancer cell states and the tumor microenvironment. We further identified a heterogeneity-guided core gene signature that could potentially have prognostic values in predicting endocrine treatment outcomes. Overall, we envision that this integrated approach, combining single-cell epigenomic and transcriptomic analyses, will become a powerful tool for unraveling the interplay between epigenetic factors and the tumor microenvironment in driving tumor heterogeneity.

Abbreviations

scRNA-seq

Single-cell RNA sequencing

scATAC-seq

Single-cell ATAC sequencing

NTs

Normal tissues

PTs

Primary tumors

TTs

Tumor tissues

RTs

Tamoxifen-treated recurrent tumors

CSs

Cancer cell states

HCS

Heterogeneity-guided core signature

TFs

Transcription factors

TME

Tumor microenvironment

ER

Estrogen receptor α

AIs

Aromatase inhibitors

siRNA

Small interfering RNAs

PCs

Principal components

MAD

Median absolute deviation

DEGs

Differentially expressed genes

DAs

Differential accessibility peaks

QC

Quality control

NK

Natural killer cells

PT_CSs

PT-specific cell states

RT_CSs

RT-specific cell states

PRT_CS

PT-RT-shared cell state

cCRE

candidate cis-regulatory element

HAFs

Heterogeneity-guided accessibility features

GO

Gene ontology

Authors’ contributions

VXJ conceived the project. KF, AGO, and TL performed the data integration analyses. AGO and LC conducted the experiments with an assistance from Dr. Zhao Lai at Next Generation Sequencing Facilities at University of Texas Health San Antonio (UTHSA). VXJ, KF, AGO, TL, and LC wrote the manuscript, with all authors (BN, XY, QW, SK, GL) contributing to writing and providing the feedback. All authors read and approved the final manuscript.

Funding

This project was partially supported by grants from NIH R01GM114142 and Advancing A Healthier Wisconsin (AHW) Seed Grant.

Data availability

The datasets generated and analyzed during the current study are available in the GEO repository under accession number GSE240112 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE240112) [75]. Python/R scripts of this study are freely available at https://github.com/KunFang93/BRCA_TR_scRNAscATAC or at this DOI: 10.5281/zenodo.8247774 [76].

Declarations

Ethics approval and consent to participate

This study uses previously collected breast cancer patients’ pathological specimens, or diagnostic specimens from various biospecimen resources that have already been granted local Institutional Review Board (IRB) protocol approvals. All patients’ samples are de-identified but included with some basic clinical features. We will not have any keys that may link specific Health Insurance Portability and Accountability Act (HIPAA)-defined protected health information (PHI) to specific individuals. Medical College of Wisconsin has granted this study without an IRB; therefore, ethics approval was waived in this case. This study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.