Macrophage heterogeneity in breast cancer microenvironment and development of a prognostic model: exploring TFF1 as a novel therapeutic target

Abstract

Backgrounds

Breast cancer (BRCA) represents the most prevalent malignancy globally, with projections indicating 3.2 million new cases anticipated by 2050. Current treatment modalities, encompassing surgical intervention, chemotherapy, and immunotherapy, are markedly hindered by treatment resistance and the inherent complexity of BRCA, thereby impeding effective disease management. Consequently, this study is designed to elucidate cellular heterogeneity within the tumor microenvironment (TME) and to identify prospective therapeutic targets, thus enabling individualized treatment approaches for individuals with BRCA, which remains crucial.

Methods

GEO and TCGA databases were the sources of all analytical data employed in this investigation. Single-cell RNA sequencing data were employed to identify macrophage-associated genes in BRCA, followed by the development of a machine-learning-based prognostic model (PM) integrating data from TCGA and GEO databases. This model stratifies individuals into cohorts of either low or high risk, enabling a disparity evaluation in survival outcomes and tumor immune microenvironment characteristics. Gene co-expression patterns were examined through an analysis of gene co-expression network weighted by high dimension for screening pivotal central genes implicated in tumor immunity. Clinical PMs were subsequently constructed utilizing machine learning algorithms, with validation performed on training and test sets. Furthermore, XGBoost and LightGBM machine learning algorithms were implemented to pinpoint potential biomarkers. Ultimately, validation of these biomarkers was conducted through ELISA, CCK-8 assay, flow cytometry, Western blotting, and immunoprecipitation assays.

Results

Findings indicated a pronounced elevation of TAMs in BRCA patients, with their phenotypic attributes exhibiting a strong correlation with prognosis. The developed clinical PM demonstrated high accuracy and robust predictive capability concerning survival rates of 1, 3, or 5 years. The comprehensive evaluation determined that high-risk cohort’s immune microenvironment exhibited a greater inclination toward immune suppression, whereas the low-risk cohort was markedly predisposed to anti-tumor immune responses. Notably, the trifolium factor 1 (TFF1) gene was identified as a key determinant, with its overexpression being markedly associated with tumor invasiveness and immune evasion.

Conclusions

The study underscores cellular heterogeneity inherent in BRCA TME, while PM, predicated on macrophage subpopulation-related genes, offers a novel approach for risk stratification and personalized therapeutic interventions for BRCA patients. Moreover, TFF1 has been established as a prospective therapeutic target, yielding important directions for developing specific treatment interventions.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s00262-026-04354-7.

Introduction

As an aggressively invasive malignancy, breast cancer (BRCA) exceeded lung cancer globally in terms of detection [1]. Projections indicate that by 2050, approximately 3.2 million new cases of female BRCA will be recorded globally [2, 3]. Based on the stage of the disease and its pathological characteristics, a variety of BRCA treatment approaches have been developed, encompassing surgical procedures, chemotherapy, endocrine therapy, immunoantibody therapy, and radiotherapy [4, 5]. However, the intricate and heterogeneous nature of BRCA imposes considerable limitations on the extensive application of these therapeutic modalities [6, 7]. Alarmingly, resistance to chemotherapy, radiotherapy, and endocrine therapy continues to pose a severe threat to patient survival [8]. Even after undergoing standard diagnostic procedures and treatment regimens, certain BRCA patients still experience distant metastases [5, 9]. Consequently, ongoing investigations into the characteristics of high-risk BRCA populations, the development of a robust prognostic model (PM), and the identification of potential therapeutic targets remain imperative. These efforts can facilitate the selection of optimal treatment strategies and pharmacological interventions, ultimately advancing personalized therapeutic approaches and enhancing patient prognoses.

The tumor microenvironment (TME) has pivotal effect in terms of BRCA onset, progression, and metastatic potential. Within the TME, lymphocytes, dendritic cells (DCs), associated macrophages (TAMs), additional stromal components, and other infiltrating immune cells actively engage with cancer cells to establish a microenvironment conducive to tumor advancement, facilitating tumor proliferation and offering essential functional support [10, 11]. Macrophages constitute a critical component of the TME and participate in the oncogenic process from its initial stages to the progression of malignant tumors [12, 13]. Extensive prior research has indicated that macrophage infiltration within tumors correlates with unfavorable prognostic outcomes in cancer patients [14]. Mahmoud et al. observed that infiltration by the CD68 macrophage population is markedly linked to reduced disease-free survival [15]. TAMs contribute to tumor biology by fostering tumor growth and progression, as well as by inducing resistance to treatment [16]. These macrophages are split into two subtypes: M1 and M2. While M1-polarized macrophages exert anti-tumor effects by restricting tumor invasiveness, M2-polarized macrophages facilitate tumor proliferation and are associated with poor cancer prognosis [17]. When investigating potential drug targets for BRCA and developing corresponding therapeutic strategies, the equilibrium between the tumor-promoting and anti-tumor functions of TAMs must be carefully considered. An optimal therapeutic approach would involve suppressing the tumor-promoting properties of macrophages while augmenting their anti-tumor capabilities [18]. Although macrophage-targeted therapies for solid tumors present a promising avenue for treatment, substantial obstacles remain owing to undesirable effects and limited efficacy observed in clinical trials [19]. Given the substantial heterogeneity exhibited by macrophages, encompassing both tumor-promoting and anti-tumor functionalities, a deeper understanding of their heterogeneity and the context-dependent roles of TAMs may lead to more effective therapeutic interventions for BRCA.

With the rapid development of bioinformatics technology, advanced methods such as high-throughput sequencing and single-cell transcriptome analysis have been widely applied to explore potential biomarkers in tumor and non-tumor diseases, greatly promoting the study of disease molecular mechanisms and the discovery of clinical diagnostic biomarkers [20, 21]. Single-cell analysis examined macrophage-associated BRCA marker genes from this investigation. Additionally, clinical PM employing algorithms for machine learning was initiated based on datasets from GEO and TCGA. Through clinical PM, individuals were split as cohorts of either low or high risk, facilitating survival pattern variation assessment and the tumor immune microenvironment among BRCA patients. This PM effectively stratifies BRCA patients, enabling subgroup determination for potentially deriving immunotherapy benefits, thereby facilitating precise anti-tumor treatment and enhancing patient prognosis.

Methodologies and materials

Data download and processing

TCGA repository provided TCGA-BRCA information, containing mRNA gene expression patterns and associated clinical data derived from 113 healthy specimens and 1,118 BRCA cases. For external validation, the GEO dataset (GSE20685, GSE42568) was acquired from the GEO database. Specifically, GSE20685 consists of mRNA expression data along with clinical details from 327 BRCA samples, while GSE42568 includes mRNA expression profiles and associated clinical information derived from 17 normal samples and 104 BRCA samples. Additionally, from the GEO database, single-cell sequencing data from human BRCA tissues employed in this investigation were procured (GSE195861). GSE212966 provides information on sequencing among 13 BRCA tissues and one normal tissue.

Single-cell data processing

In this investigation, Seurat objects were constructed using the “Seurat” R package based on single-cell RNA sequencing data [22]. As the first step, data quality control excluded cells that expressed below 200 genes, displayed above 2500 genes, or showed above 10% mitochondrial gene levels. Subsequently, the top 2000 highly variable genes underwent principal component analysis, after which it selected ten principal components for clustering and dimensionality reduction. “Harmony” function combined sample information and reduced batch variations. Building upon Harmony outcomes, for dimensional reduction, Unified Manifold Approximation and Projection (UMAP) was performed via RunUMAP algorithm application, then FindClusters and FindNeighbors algorithms enabled proximity searches and cluster examination to reveal the underlying data organization. Ultimately, the clustering labels of the samples were determined using the identification function.

Analysis of high-dimensional WGCNA (hdWGCNA)

HdWGCNA functions as an extensive analytical framework engineered within high-dimensional transcriptome datasets to detect, examine, and elucidate networks of gene co-expression [23]. In this investigation, the hdWGCNA methodology was applied to construct a network (scale-free) at the resolution of individual cells. To achieve optimal network connectivity, we implemented the R package “hdWGCNA,” establishing a topology model (scale-free) with a > 0.85 threshold, while adopting 7 as the value for soft threshold. Pearson correlation coefficient was implemented to assess gene expression similarities, generating a topological overlap matrix, followed by clustering tree development utilizing the “PlotEndprogram” function. The genes were then classified into separate modules, and their associations with distinct cell clusters were evaluated through module genes for determining the module with the most significance. For within-module hub gene identification, we extracted the top 100 genes from each detected module. The final step involved conducting a functional enrichment examination, integrating Kyoto Encyclopedia of Genes and Genomes (KEGG) and Gene Ontology (GO), and employing “clusterProfiler” on these identified hub genes.

Differential expression analysis for screening differentially expressed genes (DEGs)

The study performed differential expression analysis among the TCGA-BRCA dataset BRCA and normal samples to identify DEGs. The “limma” R package was employed for this analysis, applying a threshold of logFC = 1 and P < 0.05 for the corrected p-value.

T cell-linked clinical PM development via a comprehensive machine learning algorithm

Subsequently, this study conducted intersection analysis among hdWGCNA-identified hub genes and DEGs to extract the intersecting genes. Based on these intersecting genes, for examining feature genes with prognostic significance, this study performed univariate Cox regression analysis. Ten computational learning techniques and 101 methodological combinations were integrated for developing a clinical PM with reliability and precision. The algorithms included stepwise Cox, random survival forest (RSF), elastic network (Enet), Lasso, Ridge, CoxBoost, supervised principal components (SuperPC), generalized boosted regression modeling (GBM), partial least squares regression for Cox (plsRcox), and survival support vector machine (survival-SVM).

The optimal model served as the one exhibiting the greatest mean C-index throughout all data collections. The specific process is as follows: Firstly, 101 algorithm combinations are applied to feature genes with prognostic significance, and a prognostic model is constructed based on the TCGA dataset through leave one cross validation (LOOCV). Secondly, based on the final PM genome filtered by each model, a multivariate Cox proportional hazards model is established, and the PM score of each sample is calculated using Cox regression coefficients and gene expression values. Subsequently, cross validation was performed on all models in the GSE20685 and GSE42568 queues. Finally, calculate the Harrell Consistency Index (C-index) on all validation datasets, and determine the model with the highest average C-index as the optimal model.

Assessment and verification of clinical PMs

TCGA-BRCA subjects underwent separation as cohorts of either high or low risk per established clinical PM. Kaplan–Meier curve was implemented to examine survival differences between these cohorts. For examining predictive capabilities of the model, receiver operating characteristic (ROC) curve’s area under the curve (AUC) was determined. Considering both PM clinical features and risk score (RS), Cox analyses (univariate and multivariate) ascertained prognostic elements influencing clinical outcomes in BRCA patients. Lastly, the same analyses were conducted on the validation sets (GSE20685, GSE42568) to confirm the prognostic capability of the clinical PM.

Construction of nomogram model

In this investigation, per RS procured from clinical PM as well as various clinical characteristics, a nomogram model was produced. Likelihood of survival at 1, 3, and 5 years in BRCA individuals was identified utilizing the “rms” and “survival” packages. Calibration curves were implemented to examine the discriminative power and precision of the nomogram.

Correlation between clinical PMs, immune microenvironment, and somatic mutations

CIBERSPORT algorithm calculated comparative immune cell levels infiltrating BRCA, after which the link between RSs and immune cell infiltration (ICI) in clinical PMs was assessed.

Molecular docking (MD) and recognition of key characteristics

For assessing model genes’ feature significance, LightGBM and XGBoost were utilized, enabling a recognition of crucial genes that most significantly influence the model. The SHAP value, an interpretative technique, was implemented to measure each feature’s influence on the model’s outcomes. For individual genes, the mean SHAP value across all samples was calculated to identify those exerting considerable effect on overall model efficiency. Additionally, for forecasting potential compounds that have small molecules and target genes possessing notable contributions, Connectivity Map (cMAP) database (https://portals.broadinstitute.org/cmap/) was applied. From PubChem database, small-molecule configurations were extracted, with obtained 3D configurations stored as SAF. From AlphaFold Protein Structure Database, protein configurations in 3D were accessed then stored as PDB. To enhance docking precision, CB-Dock2 (http://clab.labshare.co.uk/cb-dock/php/index.php) served as MD instrument in this investigation. A certain protein’s binding area is forecast by CB-Dock2 by implementing a curvature-based cavity detection approach to determine the center and size, combined with AutoDock Vina as an advanced software for docking. Moreover, binding patterns utilizing scores from Vina is classified by CB-Dock2, delivering 3D binding interaction representations. Absolute binding energy quantities underwent organization in diminishing sequence [24, 25]. The online CB-Dock2 platform was employed for MD, and a visualization examination was subsequently conducted.

Cell culture and transient transfection

Beijing Bena Biotechnology Co. (Beijing, China) was the source of MCF-10A, a human BRCA epithelial cell line, along with MCF7, MDA-MB-453, SK-BR-3, MDA-MB-231, and MDA-MB-361, the BRCA cell lines. DMEM/F-12 was the medium for maintaining these cell lines. Using Lipofectamine 2000 (Invitrogen, Thermo Fisher, USA), negative control and siRNA targeting trifolium factor 1 (TFF1) (Sagon, China) underwent transfection.

Gene	Target sequences (5’-3’)
si-TFF1#1	CTGACTTTGACTACTCAAAATTG
si-TFF1#2	CTCAAAATTGGCCTAAAAATTAA

Web database

From GEPIA2 database (http://gepia2.cancer-pku.cn/#index), the transcriptomic data of TFF1 utilized in this study were retrieved, while from the Human Protein Atlas (HPA) database (https://www.proteinatlas.org/), TFF1 immunohistochemical results were procured. Additionally, from STRING database (https://cn.string-db.org/), protein interaction network data were procured.

Enzyme-linked immunosorbent assay (ELISA)

JONLNBIO BIOTECHNOLOGY LIMITED (Shanghai, China) was the source of TFF1 ELISA kits (JL30890-96 T). The TFF1 assay kit was applied to ascertain TFF1 levels in the supernatant of the cell culture medium. Following a 6-h stimulation period in an FBS-free medium, the culture supernatant underwent 5,000 revolutions of centrifugation, then subsequently applied in ELISA analysis.

Cell viability

Per supplier protocols, the Cell Counting Kit-8 (CCK-8) assay (Beyotime, China) examined cell viability. With 1 × 10³ cells for each well, the cells exposed into several treatment kinds were put in 96-well plates. At designated time points, the CCK-8 solution was introduced. Following 37 °C cultivation for 2 h, a microplate reader (BioTek, USA) examined absorbance at 450 nm for individual wells.

Transwell assay

For assessing MCF7 and MDA-MB-231 cell lines’ invasive capacity, transwell assays were performed. To summarize the assessment of invasion, Matrigel-coated chambers (BD Biosciences, USA) served to store the cells (5 × 10⁴). The upper chamber contained a medium free from sera, while the lower compartment had a complete DMEM medium. Following cultivation for a full day, cells exhibiting migration or invasion were secured using 4% paraformaldehyde followed by staining using 0.1% crystalline violet.

5-Ethynyl-2′-deoxyuridine (EdU) assay

Employing the BeyoClick™ EdU Cell Proliferation Kit with Alexa Fluor 594 (Biotek, Shanghai, China), EdU assay was executed. After phosphate-buffered saline rinsing, cells were incubated for 2 h in EdU solution, and then, nuclei underwent DAPI solution staining. Subsequently, sample washing and examination were performed via an inverted microscope (ZEISS, Germany).

Apoptosis analysis by flow cytometry

Apoptosis analysis via Annexin-V Apoptosis Detection Kit (Beyotime, China) and flow cytometry (BD Biosciences, USA) was conducted. Cells were exposed to 20 μg/ml Annexin-V conjugated with FITC for 0.5 h followed by exposing for 5 min to 50 μg/ml PI at ambient conditions under dark conditions, following previously described protocols [30]. After the addition of 400 μl of mixing buffer, flow cytometric analysis was performed immediately according to standard procedures. Live cells remained unstained by both Annexin-V and PI; cells in early apoptosis displayed only Annexin-V labeling. Cells undergoing late apoptosis demonstrated positivity for both Annexin-V and PI, whereas physically disrupted cells exhibited PI staining exclusively.

Western blot (WB) analysis

Using RIPA lysis buffer (Biosharp, China), proteins were isolated. After determining concentration of proteins, aliquots were resolved through 10% SDS-PAGE gels, followed by transferring electrically into 0.45-μm PVDF membranes, which underwent blocking utilizing 5% skimmed milk solution over a 2-h period. Subsequently, the blocked membranes were exposed to designated primary antibodies throughout the night at 4 °C. These antibodies comprised anti-TFF1 (13,734–1-AP, Proteintech, 1/500), anti-ESR1 (21,244–1-AP, Proteintech, 1/500), anti-PI3K (20,584–1-AP, Proteintech, 1/1000), anti-AKT (60,203–2-Ig, Proteintech, 1/1000), anti-P-AKT (80,455–1-RR, Proteintech, 1/1000), anti-BCL2 (80,313–1-RR, Proteintech, 1/1000), anti-BAX (50,599–2-Ig, Proteintech, 1/1000), anti-GAPDH (ab8245, Abcam, 1/5000), anti-Histone H3 (17,168-1-AP, Proteintech, 1/5000). Following this, the membranes underwent incubation with secondary antibodies at ambient conditions for an hour. The immunoblots underwent processing utilizing ECL fluorescence detection kit (Beyotime, Shanghai) followed by examination via Tanon 4600 system (Tanon Science and Technology Co., Ltd.).

Co-immunoprecipitation (co-IP)

Co-IP is performed using the Co-IP kit according to the manufacturer’s instructions. Briefly, a portion of the lysate is used as input, and the rest of the lysate is immunoprecipitated by incubation with DYKDDDDK tag Monoclonal antibody (Binds to FLAG® tag epitope) (21,244-1-AP, Proteintech, 1/100), MYC tag Monoclonal antibody (60,003-2-Ig, Proteintech, 1/2000), or IgG antibodies overnight at 4 °C. Magnetic bead-conjugated antibody-protein complexes are then added. Samples are denatured by heating at 100 °C for 10 min and then eluted. The supernatant is collected for Western blot analysis.

Cellular immunofluorescence

Following paraformaldehyde fixation, the specimens underwent primary antibody treatment at 4 °C for 24 h using anti-ESR1 (21,244-1-AP, Proteintech, 1/100) and anti-TFF1 (13,734-1-AP, Proteintech, 1/100). The samples were then exposed to secondary antibodies consisting of Alexa Fluor® 594-labeled goat anti-rabbit antibody and Alexa Fluor® 488-labeled goat anti-mouse antibody (both from Abcam, UK). Nuclear staining was performed using DAPI (Biosharp, China) at ambient conditions for 30 min. A fluorescence microscope (ZEISS, Germany) was employed to observe positive cell signals at 400 × magnification.

Statistical analysis

R version 4.3.3 served as the software for computations of statistics and representations using graphs. For evaluating differences between paired sample groups, the Wilcoxon test was implemented, while Kruskal–Wallis test and one-way analysis of variance (ANOVA) probed variations across multiple cohorts. The Kaplan–Meier approach was adopted for survival investigations, and relationship evaluations were conducted via Pearson correlation analysis. Given the scope of statistical testing in this study, multiple comparison adjustments were made to p-values using the Benjamini–Hochberg false discovery rate (FDR) method. P < 0.05 was the threshold for statistical significance.

Results

Sequencing analysis of single-cell RNA demonstrates cellular heterogeneity among BRCA tissues

A systematic analysis of the single-cell dataset (GSE195861) was performed in this study. Following stringent quality control and dimensionality reduction clustering, 19 distinct cell clusters were identified. These clusters were subsequently annotated into five highly heterogeneous cell types. From Fig. 1B, as opposed to normal sample cohort, the proportion of macrophage subpopulations was markedly elevated in the BRCA sample cohort. Further examination of the macrophage subpopulation (Fig. 1C) identified 13 distinct cell clusters, among which clusters 0, 5, 6, 8, and 12 exhibited markedly higher proportions in BRCA samples relative to normal samples. These specific clusters serve as potential candidates for further investigation within the BRCA context (Fig. 1D–E).

Fig. 1.

Single-cell analysis identified the macrophage subpopulation as a key cell subpopulation in BRCA. A By employing UMAP analysis, an aggregate of 19 distinct cell clusters was discerned across all examined samples. B These 19 clusters were systematically categorized into five highly heterogeneous cell types. C A comparative analysis illustrates the proportion of these five types of cells between normal and BRCA sample cohorts. D To achieve further refinement, macrophage subpopulations underwent additional dimensionality reduction clustering analysis, resulting in the visualization of 13 distinct cell clusters via UMAP plots. E The distribution of macrophage subgroup-associated clusters was analyzed, delineating the differences between the BRCA sample cohort and the normal sample cohort

Identifying key modules related to macrophage subgroups in BRCA using the hdWGCNA algorithm

The key modules associated with the molecular characteristics of macrophage subgroups were identified using the hdWGCNA algorithm. A macrophage cell network (scale-free) was established using a soft threshold of 7, ensuring optimal connectivity, and ultimately, six gene modules were identified. Correlation analysis (Fig. 2A–D) demonstrated that clusters 0, 5, 6, 8, and 12 exhibited significant associations with specific gene modules, including turquoise, red, and blue. Subsequently, the top 100 genes from red, blue, and turquoise modules were the hub genes subject to following examination (Fig. 2E, F).

Fig. 2.

HdWGCNA illustrates five central modules that exhibit a strong correlation with the Macrophage subgroup in BRCA. A A soft power of 7 was selected to construct a scale-free network. B A hierarchical clustering tree diagram was employed for depicting six modules in the network. C Six gene modules were identified, with the most representative genes highlighted based on the hdWGCNA pipeline. D Analysis of correlation examined different gene modules’ relationships. E The distribution patterns of the six gene modules across distinct macrophage subgroups were analyzed. F Relevant bubble plots were generated to represent the five central modules that demonstrate a significant association with the Macrophage subgroup

Analysis and characterization of DEGs with functional annotations

Differential expression analysis was performed on the TCGA-BRCA dataset, identifying a total of 2,743 DEGs. Subsequently, analysis of intersection among these DEGs and 300 hub genes derived from hdWGCNA was performed, resulting in 71 intersecting genes. To investigate the biological pathways linked to these overlapping genes, this study applied evaluations of KEGG and GO pathway enrichment. GO assessment demonstrated such genes to partake in biological functions that comprise acute inflammatory response to antigen stimulation, response to oxidative stress, and negative regulation of external stimulus response. KEGG analysis (Fig. 3D) demonstrated those intersecting genes to be highly related to biological pathways, including Toll-like receptor signaling pathway, phenylalanine metabolism, and IL-17 signaling pathway (Fig. 3E).

Fig. 3.

Identification and functional annotation analysis of DEGs. A A heat map representing DEGs. B A volcano plot depicting the distribution of DEGs. C A Venn diagram illustrating 71 overlapping genes. D GO analysis was conducted for intersecting genes. E KEGG analysis was performed on the intersecting genes

Development and assessment of clinical PMs utilizing machine learning algorithms

Through examination of the overlapping genes previously identified, univariate Cox regression analysis revealed 33 essential genes linked to survival outcomes. Using the LOOCV methodology, 101 predictive frameworks were established in the TCGA-BRCA cohort, and every predictive system had their concordance indices determined. Notably, the model combination yielding the highest C-index was Lasso + GBM, achieving a C-index of 0.697. Ultimately, a clinical PM incorporating 11 key genes was established, as illustrated in Fig. 4A. Subsequently, the training set (TCGA-BRCA) was classified as cohorts of either low or high risk per clinical PM-derived median risk value. The low-risk cohort of PAAD patients exhibited a significant survival advantage (Fig. 4B). The predictive effectiveness of the PM was examined via ROC examination, yielding AUC measurements of 0.809, 0.822, and 0.823 corresponding to forecasts of survival of 1, 2, and 3 years (Fig. 4C). Additionally, Cox regression evaluations (univariate and multivariate) revealed the RS generated from this prognostic framework functions to be an autonomous factor affecting clinical outcomes in BRCA patients (Fig. 4D and E).

Fig. 4.

Development and assessment of a clinical prognostic model utilizing algorithms for machine learning. A C-index values corresponding to each prognostic model were generated from ten algorithms and 101 distinct combinations within training and testing datasets. C ROC curves for 1, 3, and 5 years were utilized to assess the clinical predictive performance of the model. D, E Univariate and multivariate Cox regression analyses probed risk scores related to the clinical prognostic models. I Univariate Cox regression analysis. J Multivariate Cox regression analysis

Validation of clinical PMs

The external clinical prediction model validation utilized GSE20685 and GSE42568 datasets. The validation outcomes aligned with those obtained from the training set. Utilizing the median risk values derived from this model, samples from GSE20685 and GSE42568 were split into cohorts of high and low risk, respectively. Remarkably, individuals with BRCA from low-risk cohort displayed a substantial survival benefit. In the GSE20685 cohort (Supplementary Fig. 1A and E), survival in 1, 2, and 3 years had respective AUC values of 0.541, 0.739, and 0.662. Similarly, in GSE42568 cohort (Supplementary Fig. 1B), the corresponding AUC values were 0.649, 0.647, and 0.747 (Supplementary Fig. 1F). Moreover, Cox regression analyses (univariate and multivariate) ascertained RS from this PM as an independent determinant influencing clinical results among BRCA patients (Supplementary Fig. 1C, D and G, H).

Development of BRCA-related nomogram

A novel nomogram has been constructed incorporating RSs and clinical variables, such as age, gender, tumor size, lymphatic metastasis, and disease stage. The calibration curve analysis presented in Fig. 5A illustrates that the column chart exhibits a strong predictive capacity for actual survival outcomes (Fig. 5B). Additionally, the findings derived from the risk accumulation chart indicate a notable increase in cumulative risk among high-risk patients as time progresses (Fig. 5C).

Fig. 5.

Establishment of a novel column chart associated with BRCA. A A column chart integrating clinical characteristics and risk scores was developed. B Calibration curves for the column chart were generated for 1, 2, and 3 years. C A chart on risk accumulation was constructed based on risk stratification

Link between RS of a clinical PM and TME, somatic mutations

CIBERSORT algorithm assessed RSs’ link with ICI. Significant variations in multiple immune cell populations have been detected between cohorts of low and high risk. Specifically, in the high-risk cohort, levels of macrophages M0 and M2 were significantly high, whereas T cells CD8, B cells naïve, macrophages M1, and T cells CD4 memory restoration depicted markedly elevated expression in the low-risk cohort (Supplementary Fig. 2A). Moreover, the RS demonstrated a significant positive correlation with B cell memory, neutrophils, macrophages M2, macrophages M0, and NK cells activated. Conversely, a substantial negative link was identified between the RS and diverse immune cells, including macrophages M1, DCs resting, B cells naïve, T cells CD4 memory resting, T cells CD4 memory activated, T cells CD8, and T cells follicular helper (Supplementary Fig. 2B–M).

Identification of important genes in models, drug prediction, and MD

The significance of 11 risk genes within the model was assessed through LightGBM and XGBoost, as well as respective model contributions were elucidated using SHAP values (Supplementary Fig. 3A–D). These risk genes were ranked in descending order based on their impact on model output, with TFF1 consistently appearing among the top three across both algorithms. This consistency suggests that TFF1 exerts a stable and significant influence on the model. Consequently, TFF1 was identified as a key gene for further analysis. To facilitate the prediction of small-molecule-targeted drugs associated with TFF1, the TCGA-BRCA cohort was stratified as TFF1 cohort of either low or high expression, followed by differential expression analysis. A sum of 150 upregulated DEGs was placed into the CMap database. Following connectivity score analysis, three targeted drugs (AG-957, CGS-15943, and pelitinib) were identified. To further evaluate the most effective small-molecule drugs interacting with the TFF1 protein, MD analysis was executed between TFF1 and the three highest-ranked compounds. Reduced binding energy scores reflect enhanced conformational binding strength. Specifically, binding energy values below −5 kcal/mol denote advantageous binding, while scores under −7 kcal/mol indicate robust activity. MD analysis outcomes suggested that all small-molecule drug compounds showed binding energies under −6 kcal/mol, indicating effective interaction alongside target protein TFF1. Among these compounds, AG-957 showed −6.5 kcal/mol binding energy, CGS-15943 exhibited −7.2 kcal/mol binding energy, and pelitinib displayed −7.7 kcal/mol as the lowest binding energy (Supplementary Fig. 3E–G).

Elevated TFF1 expression in BRCA

TFF1 levels across various tumor types were analyzed using the GEPIA2 database, revealing that TFF1 was upregulated in BRCA, COAD, ESCA, and LUAD, with the most pronounced elevation observed in BRCA (Fig. 6A, B). Immunohistochemical data obtained from the HPA database confirmed the presence of TFF1 expression in tissue microarrays of both Breast Duct carcinoma and Breast Lobular carcinoma (Fig. 6C). Additionally, tumor tissues and adjacent non-tumor ones were procured from four BRCA cases, and WB examination revealed an elevated level of TFF1 protein in the malignant samples (Fig. 6D, E). Collectively, these observations present strong support for the enhanced expression of TFF1 in BRCA.

Fig. 6.

Increased expression of TFF1 in BRCA. A Transcript levels of TFF1 across various tumor types were obtained from the GEPIA2 database. B TFF1 transcript levels, specifically in BRCA, were extracted from the GEPIA2 database. C Protein expression levels of TFF1 in BRCA, as documented in the HPA database. D, E A significant upregulation of TFF1 expression was observed in tumor tissues in comparison to adjacent para-tumor tissues. *** ≤ 0.001. Data are presented as mean ± S.E.M

Upregulated TFF1 promotes the malignancy of BRCA cell lines

The TFF1 levels in various BRCA cell lines were initially compared. WB analysis indicated that TFF1 expression underwent upregulation in human BRCA cell lines relative to human breast epithelial cell line MCF10A. Notably, the most pronounced upregulation of TFF1 was observed in MDA-MB-231 and MCF7 cell lines, leading to their selection for subsequent experiments (Fig. 7A).

Fig. 7.

Upregulated TFF1 promotes the malignancy of BRCA cell lines. A WB analysis of TFF1 expression in MCF-10A, MCF7, MDA-MB-231, SK-BR-3, MDA-MB-453, and MDA-MB-361. B, C Quantification of TFF1 concentration in the culture medium of MCF-10A, MCF7, and MDA-MB-231. D, E Levels of TFF1 in the culture medium of MCF7 and MDA-MB-231 following si-RNA treatment targeting different sequences. F, G Cell viability assessment of MCF7 and MDA-MB-231 post-si-RNA treatment against TFF1. H, I EdU fluorescence analysis of MCF7 and MDA-MB-231 after exposure to si-RNA targeting TFF1. J, K Evaluation of cell invasion capacity in MCF7 and MDA-MB-231 subsequent to si-RNA treatment against TFF1. ** ≤ 0.01, *** ≤ 0.001, **** ≤ 0.0001. The results are presented as mean ± S.E.M

Given that TFF1 is frequently secreted by cells as an exocrine protein, its concentration in the culture medium of MDA-MB-231 and MCF7 cell lines was analyzed. Compared to MCF10A, a substantial elevation in TFF1 levels was detected in the culture medium of those cell lines (Fig. 7B, C). Subsequently, this study assessed the inhibitory efficiency of two small interfering RNAs (si-RNAs) that targeted distinct sequences. ELISA results demonstrated that small interfering RNA#1 (si-RNA#1) effectively suppressed the upregulation of TFF1 concentration in the culture medium of MDA-MB-231 and MCF7. As a result, this si-RNA was selected for further study (Fig. 7D, E). Additionally, both CCK8 assays and cytofluorescence analysis from EdU indicated that si-RNA markedly inhibited the proliferative capacity of MDA-MB-231 and MCF7 (Fig. 7F–I). Similarly, transcriptional inhibition of TFF1 in those cell lines led to remarkable invasive ability suppression of BRCA cell lines (Fig. 7J, K).

TFF1 interacts with estrogen receptor 1 (ESR1)

To investigate the mechanism through which TFF1 exerts its function, potential interacting proteins were identified using the STRING database. The results from protein–protein interaction (PPI) networks suggested that ESR1 is the most likely interaction partner of TFF1 (Supplementary Fig. 4A). To validate this interaction, ESR1 levels were first examined in MCF10A, MDA-MB-231, and MCF7. WB analysis confirmed ESR1 protein expression to be markedly upregulated in MCF7 and MDA-MB-231 (Supplementary Fig. 4B–E). The results of co-IP as well as immunofluorescence indicated the interaction of ESR1 and TFF1 in MCF7 and MDA-MB-231 (Supplementary Fig. 4F–I).

TFF1 acts through the ESR1-PI3K-AKT signaling axis

The PI3K-AKT cascade functions as a downstream effector of ESR1 upon activation. To determine whether PI3K-AKT activation occurs through TFF1 via ESR1, recombinant TFF1 protein (10 μM) was introduced into the culture medium of MCF7 and MDA-MB-231 to enhance TFF1 stimulation. Following rTFF1 stimulation, there was substantial PI3K-AKT signaling cascade upregulation in MCF7 and MDA-MB-231, whereas this activation enabled NF-κB signaling pathway activity inhibition for both cell lines (Fig. 8A–D). Conversely, inhibition of TFF1 transcription through si-RNA in MCF7 and MDA-MB-231 resulted in the suppression of PI3K-AKT signaling cascade activity while promoting NF-κB signaling cascade activation (Fig. 8E–H). Restoration of NF-κB signaling pathway activity enhanced inflammatory apoptosis in MCF7 and MDA-MB-231, as evidenced by enhanced BAX expression, decreased BCL2 expression, and enhanced apoptotic ratio (Fig. 8I–L). Meanwhile, inhibition of TFF1 transcription restored CDH1 expression and suppressed CDH2, SNAI1, and Vimentin expression. EMT of BRCA was improved (Fig. 8M, N).

Fig. 8.

TFF1 acts through the ESR1-PI3K-AKT signaling axis. A Activity of the PI3K-AKT signaling pathway in MCF7 after rTFF1 protein stimulation. B Activity of the NF-κB signaling pathway in MCF7 after rTFF1 protein stimulation. C Activity of the PI3K-AKT signaling pathway in MDA-MB-231 after rTFF1 protein stimulation. D Activity of the NF-κB signaling pathway in MDA-MB-231 after rTFF1 protein stimulation. E Activity of the PI3K-AKT signaling pathway in MCF7 after si-TFF1 treatment. F Activity of the NF-κB signaling pathway in MCF7 after si-TFF1 treatment. G Activity of the PI3K-AKT signaling pathway in MDA-MB-231 after si-TFF1 treatment. H Activity of the NF-κB signaling pathway in MDA-MB-231 after si-TFF1 treatment. I, J Protein expression of BAX and BCL2 after si-TFF1 stimulation. K, L MCF7 and MDA-MB-231 apoptosis levels after si-TFF1 treatment. M, N Protein expression of CDH1, CDH2, SNAI1, and Vimentin after si-TFF1 treatment. ** ≤ 0.01. The results are presented as mean ± S.E.M

Discussion

This study systematically characterized cellular heterogeneity within BRCA tissues by employing single-cell RNA sequencing alongside various bioinformatics analytical approaches. A clinical PM was developed based on macrophage subgroup-associated genes, and potential therapeutic targets and drug candidates were explored. These findings not only enhance our comprehension of the BRCA TME but also offer novel insights for personalized therapeutic strategies.

Sequencing analysis of single-cell RNA revealed substantial cellular heterogeneity in BRCA tissues. Compared to normal samples, an increased proportion of macrophage subgroups was observed in BRCA samples, particularly in clusters 0, 5, 6, 8, and 12. This notable elevation suggests that these clusters may serve a pivotal function in BRCA initiation and progression. Further analysis using hdWGCNA identified key gene modules that exhibited strong correlations with these clusters, providing a foundation for screening potential functional genes. Macrophages serve a fundamental function in immune regulation within the TME, particularly M2 macrophages, which are frequently associated with immunosuppression and poor tumor prognosis. The findings of this investigation further reinforce the essential function of macrophages in BRCA development and progression. As innate immune cells, macrophages contribute to host defense, tissue homeostasis, and the regulation of inflammatory responses [26, 27]. TAMs constitute a significant component of inflammatory infiltration in breast tumors [28, 29]. By facilitating tumorigenesis, angiogenesis, and metastasis, macrophages promote tumor progression through multiple pathways [30]. Previous research has demonstrated that macrophages secrete factors that enhance breast tumor initiation. TAMs can directly stimulate tumor cell proliferation by releasing growth factors and inflammatory mediators [31]. Furthermore, TAMs have been shown to secrete TNF-α, which activates NF-κB in tumor cells, thereby inhibiting tumor cell apoptosis and enhancing invasiveness [32]. TAMs are critical in BRCA progression and metastasis, contributing to an immunosuppressive TME that may impede BRCA treatment. Therefore, targeting TAMs in BRCA through chemotherapy and/or immunotherapy is essential to improving therapeutic outcomes [27].

In this study, a clinical PM comprising 11 genes was constructed utilizing combined 101 machine learning algorithms. The model exhibited strong capability in predictivity for both training (TCGA-BRCA) and external validation cohorts (GSE20685, GSE42568). Kaplan–Meier survival analysis suggested high-risk subjects to have markedly diminished survival outcomes versus low-risk ones, while ROC curve further confirmed the model’s clinical utility in forecasting survival at 1, 3, and 5 years. Additionally, Cox regression examinations (univariate and multivariate) revealed the RS functions to be an independent prognostic indicator among BRCA individuals. The development of this model provides a novel tool for risk stratification and customized BRCA therapy approaches. Through the CIBERSORT algorithm, notable variations in levels of ICI were identified among cohorts of low and high risk. Specifically, macrophages M0 and M2 proportions were markedly high in the high-risk cohort, whereas T cells CD8, macrophages M1, and B cells naïve exhibited markedly higher proportions in the low-risk cohort. Such findings suggest the TME in high-risk patients to be possibly more inclined toward an immunosuppressive state, whereas the immune landscape of low-risk patients appears to be more conducive to an anti-tumor immune response. In BRCA, tumor cells are capable of recruiting and activating M2-like macrophages, which subsequently secrete M2-associated cytokines. This mechanism leads to the elongation of BRCA cells, the loss of contact inhibition, and an upregulation in vimentin expression [27]. Consequently, the increased presence of M0 and M2 macrophages in high-risk patients might support tumor advancement by promoting immune evasion of tumor cells. These insights emphasize the pivotal role of immune cells in BRCA prognosis and offer new perspectives for immunotherapeutic strategies targeting the TME.

Utilizing machine learning algorithms (XGBoost and LightGBM) combined with SHAP value analysis, TFF1 was identified as the key gene exerting the most significant influence within the PM. The elevated expression of TFF1 in BRCA may contribute to tumor invasiveness and immune evasion. TFF1 represents a member of three well-preserved proteins within the Trifolium Factor family and appears alongside mucins in the gastrointestinal system. Its primary expression occurs in the gastric mucosa [33]. Previous studies have shown that TFF1 manifests in multiple tumor types, and its upregulation is markedly correlated with the prognosis of patients with colorectal, pancreatic, gallbladder, and gastric cancers [34–37]. In BRCA, the presence of multiple estrogen-responsive elements in its upstream promoter has led to its classification as an estrogen-induced gene within the MCF7 cell line [38]. Prior research has also indicated that TFF1 facilitates migration and supports tumorigenesis in BRCA [39–41]. These observations correspond to our experimental outcomes, which highlight the pivotal role of TFF1 in promoting BRCA. Expression profiling from multiple databases, along with WB analysis, confirmed substantial TFF1 upregulation in BRCA tissues and cell lines, particularly in MCF7 and MDA-MB-231. Functional assays revealed that inhibiting TFF1 transcription substantially suppressed the proliferative and invasive capacities of these cell lines. Additionally, mechanistic investigations demonstrated that TFF1 interacts with ESR1. Through such interaction, TFF1 enables PI3K-AKT signaling pathway activation, which in turn suppresses the NF-κB signaling pathway, reduces inflammatory cell apoptosis, and enhances cell survival. Suppression of TFF1 transcription was found to reverse these effects, reactivating the NF-κB pathway, upregulating BAX expression, downregulating BCL2 expression, and promoting apoptosis. These findings underscore the significance of the TFF1/ESR1-PI3K-AKT-NF-κB signaling axis in BRCA progression and indicate that TFF1 could function as a promising therapeutic candidate.

To further explore the therapeutic potential of targeting TFF1, we employed molecular docking, a computational method widely used in modern drug discovery. As a core discipline of contemporary medicine, drug development reflects the deep integration of multidisciplinary technologies. The latest revolutionary advances in biotechnology and computational methods are reshaping the traditional paradigm of drug research. Molecular docking utilizes computational tools to predict the static binding conformation and affinity between ligands and their targets, enabling high-throughput screening of binding sites on biomolecules and significantly improving the efficiency of lead compound identification [42, 43]. In this study, we evaluated the potential of three targeted drugs (AG-957, CGS-15943, and pelitinib) to interact with TFF1. Among them, the binding of pelitinib to TFF1 is the most stable, indicating that it may be a candidate for further research in the context of TFF1-driven BRCA progression.

Although significant progress has been made in this study, there are still certain limitations. Firstly, the sample size of single-cell RNA sequencing data is relatively small, which may not fully capture the cellular heterogeneity of BRCA. Meanwhile, this study did not include tumor subtypes in the analysis scope, which to some extent limits the in-depth understanding of heterogeneity between different subtypes. Secondly, although the predictive performance of clinical PM has been tested on external datasets, their stability and universality need to be further validated in larger independent cohorts. Finally, although in vitro experiments were conducted to validate the results. However, functional validation has not yet been conducted through in vivo experiments, and there is a lack of relevant experimental data to support its therapeutic potential. Future research should aim to increase the sample size of single-cell sequencing data while improving the features of cell lineage and functional characteristics in BRCA TME. In addition, experimental validation and translational studies of TFF1 and its prospective targeted therapy are crucial for advancing precise treatment strategies for BRCA patients.

Conclusion

This study systematically characterized the cellular heterogeneity within the BRCA TME and constructed a clinical PM utilizing macrophage subgroup-associated genes, offering a novel tool for risk stratification and personalized treatment in BRCA. Furthermore, with the key gene being TFF1, along with prospective targeted therapeutics, this study provides valuable insights for future therapeutic strategies. These findings contribute significant theoretical support and practical guidance for both fundamental research and clinical applications in BRCA.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contribution

L.C. and G.Z. designed the study, analyzed the data, and wrote the manuscript. L.C., B.S. and Z.C. edited and revised the manuscript, generated figures. All authors read and approved the final manuscript.

Funding

None.

Data availability

All data can be access via correspondence authors.

Declarations

Conflict of interest

The authors declare no competing interests.

Consent for publication

This study has not been published before, and all authors approved this publication.

Ethical approval

This study has been approved by the Ethics Committee of Jiangsu Cancer Hospital.