CCL19 and CCR6 as diagnostic biomarkers for differentiating phyllodes tumors from breast fibroadenomas: a machine learning-driven approach integrating EMT biology

Abstract

Background

This study aimed to differentiate phyllodes tumors from fibroadenomas by identifying biomarkers through bioinformatics.

Methods

Gene expression datasets from the GEO database were analyzed, focusing on differential expression and epithelial-mesenchymal transition (EMT)-related genes between breast fibro-ma and phyllodes tumors.

Results

A total of 209 upregulated and 254 downregulated differentially expressed genes (DEGs) were identified. Intersection with 1142 EMT genes yielded 42 EMT-DEGs. Enrichment analysis linked these genes to epithelial cell proliferation, stem cell differentiation, and activities involving collagen-containing extracellular matrices, translation, repressor activity, and mRNA post-transcriptional regulation. Hub genes were identified using Weighted Gene Co-expression Network Analysis (WGCNA), Random Forest, and Lasso regression. The genes CCL19 and CCR6 emerged as potential biomarkers and demonstrated strong internal discriminative ability between fibroadenoma and phyllodes tumor samples, with Receiver Operating Characteristic (ROC) values of 0.962 and 0.915, respectively, in the GSE78071 dataset. The combined model showed favorable internal discrimination, with a mean absolute error of 0.006. Validation with six fresh tissue specimens confirmed higher expression of CCL19 and CCR6 in fibroadenomas compared to phyllodes tumors. Further analysis revealed their roles in chemokine signaling, cytokine interactions, and signaling pathways.

Conclusion

CCL19 and CCR6 were identified as diagnostic biomarkers for distinguishing breast fibroadenomas from phyllodes tumors.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-04451-0.

Introduction

Phyllodes tumors (PT) are a rare type of breast tumor, accounting for approximately 0.3%–1.0% of all breast neoplasms [1]. Clinically, PT occurs predominantly in women aged 35–55 years, with an earlier age of onset reported in Asian populations compared with other ethnic groups [2]. PTs are fibroepithelial lesions of the breast and share substantial cytological and histopathological similarities with fibroadenomas [3]. As biphasic tumors, PTs are characterized by a distinctive intraluminal growth pattern with leaf-like (phyllodes) architecture extending into dilated spaces, composed of epithelial elements containing glandular epithelium and myoepithelial cells embedded within a hypercellular stromal component [3]. Breast fibroadenomas (FB) consist of a benign proliferation of epithelial and stromal components, typically exhibiting well-circumscribed margins, absence of mesenchymal atypia, and no or only rare mitotic figures. In contrast, benign PTs may display leaf-like structures, mildly increased stromal cellularity, and mild nuclear atypia, although mitotic activity remains limited. According to the World Health Organization (WHO), PTs are classified into three categories based on histological criteria—including stromal cellularity and density, nuclear atypia, mitotic count, and tumor margins—as benign (approximately 60%), borderline (approximately 15%), and malignant (approximately 25%) [4]. The etiology of PT remains unclear, and definitive risk factors have not yet been established. Nevertheless, Previous genomic studies have reported recurrent chromosomal copy number gains in regions 1q, 5p, 7, and 8, as well as copy number losses in regions 9p, 10p, 6, and 13, in phyllodes tumors, reflecting underlying genomic instability rather than transcriptional expression changes [5]. In addition, an increased risk of PT has been observed in women with Li–Fraumeni syndrome, and MED12 mutations are frequently detected in PTs, occurring in 65.1% of benign, 65.6% of borderline, and 42.8% of malignant PTs, respectively [6]. Tan et al. further demonstrated that common genetic alterations in PT include mutations in MED12, TP53, RARA, and PIK3CA, while genes such as CDKN2A, HOXB13, PAX3, SIX1, HMGA3, and TGFB2 may contribute to malignant progression [7]. Epithelial–mesenchymal transition (EMT) is increasingly recognized as a transcriptional and phenotypic program regulating epithelial–stromal crosstalk, mesenchymal activation, and extracellular matrix remodeling, rather than being restricted to epithelial dedifferentiation alone [8, 9]. These EMT-associated biological processes are closely related to stromal plasticity and microenvironmental remodeling, which are key biological features distinguishing fibroadenomas from phyllodes tumors; Therefore, the present study proposes that this represents a key entry point for differentiating the two tumor types [10]. In current clinical practice, fine-needle aspiration cytology is commonly used for initial diagnosis [11]. Although core needle biopsy has been proposed as a tool for differentiating PT from FB, substantial histological overlap and sampling limitations continue to pose diagnostic challenges [12, 13]. Consequently, excisional biopsy or vacuum-assisted minimally invasive rotary biopsy is often required to obtain sufficient intact tissue for definitive pathological diagnosis, a scenario frequently encountered in routine clinical settings. To date, no studies have systematically explored EMT-related transcriptomic signatures, integrated with machine-learning approaches, as a genetic diagnostic strategy for distinguishing breast fibroadenomas from phyllodes tumors.

Machine learning techniques enable the identification of complex patterns within large, noisy, and high-dimensional datasets, making them particularly well suited for medical applications involving genomic and transcriptomic data, including cancer diagnosis and classification [14, 15]. To identify potential biomarkers and underlying molecular pathways differentiating breast fibroadenomas from phyllodes tumors, we analyzed transcriptomic data from the GEO database using established machine learning approaches, including Weighted Gene Co-expression Network Analysis (WGCNA), random forest, and LASSO regression. Using this integrative strategy, we identified robust diagnostic biomarkers for distinguishing breast fibroadenomas from phyllodes tumors.

Materials

Data collection and downloads

Gene expression data were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). The GSE78071 dataset was selected because it is currently the only publicly available dataset that includes both fibroadenoma and phyllodes tumor samples with clear pathological confirmation and balanced group sizes, enabling a direct transcriptomic comparison. The dataset GSE78071 contains transcriptome data for 24 fibroadenomas and 24 phyllodes tumors, with 24 fibroadenomas containing 4 complex fibroadenomas and 20 simple fibroadenomas, 4 of 24 phyllodes tumors as borderline phyllodes tumors, 4 malignant phyllodes tumors, and 16 benign phyllodes tumors.

Differential gene expression analysis

Prior to differential expression analysis, raw microarray data from the GSE78071 dataset were preprocessed using the limma package in R. Background correction and quantile normalization were applied to minimize technical variability and ensure comparable expression distributions across samples. The effectiveness of normalization and overall sample quality were evaluated by inspecting the distribution of normalized expression values across all samples, which showed highly consistent distributions without obvious outliers (Supplementary Fig. 1).

Probe identifiers were mapped to official gene symbols using the corresponding GEO platform annotation file. When multiple probes mapped to the same gene symbol, the probe with the highest mean expression value across samples was retained to represent that gene. Genes without valid annotation were excluded from downstream analyses.

Since all samples were derived from a single platform and dataset, potential batch effects were assessed based on normalized expression distributions, and no evident batch-driven bias was observed. Differentially expressed genes (DEGs) between fibroadenoma and phyllodes tumor groups were subsequently identified using the limma package in R (version 4.1.2), with a significance threshold of P < 0.05. Raw p-values were adjusted for multiple testing using the Benjamini–Hochberg false discovery rate (FDR) method. The results of the differential expression analysis were visualized using heatmaps and volcano plots.

EMT-DEGs

Functional enrichment analyses of the EMT-related differentially expressed genes (EMT-DEGs) were performed to investigate their associated biological functions and signaling pathways. Gene Ontology (GO) functional annotation and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were conducted using the “clusterProfiler” R package. Enrichment terms with a p-value < 0.05 were considered statistically significant.

Identification of diagnostic signature genes using three machine learning algorithms

To identify shared genes between the two conditions, three established machine learning algorithms were applied: Weighted Gene Co-expression Network Analysis (WGCNA), Least Absolute Shrinkage and Selection Operator (LASSO) regression, and Random Forest (RF).

For WGCNA, gene co-expression modules were constructed using the WGCNA package in R (version 4.0.2). Prior to network construction, outlier samples were removed to ensure data quality and network robustness. Genes with expression variances above the upper quartile were selected for subsequent analysis. A scale-free topology criterion was applied to determine the optimal soft-thresholding power (β = 4), and adjacency matrices were generated accordingly. These matrices were then transformed into topological overlap matrices (TOM), followed by calculation of dissimilarity measures (1 − TOM). Gene modules were identified using hierarchical clustering based on TOM dissimilarity, with modules assigned unique colors. Pearson correlation analysis was performed to evaluate the associations between module eigengenes and clinical traits, with modules exhibiting an absolute correlation coefficient (|R|) ≥ 0.4 and a p-value < 0.05 considered biologically relevant.

Random Forest analysis was performed using the randomForest package in R to classify genes expressed in the GSE78071 dataset and to identify disease-related signature genes. A Random Forest model consisting of 1000 trees was constructed in the discovery cohort. The optimal model parameters were determined based on cross-validation error rates. Gene importance was subsequently ranked, and the top 30 genes with the highest importance scores were selected for further analysis.

For LASSO regression, the 42 previously identified EMT-related differentially expressed genes (EMT-DEGs) were included as candidate variables. A penalized logistic regression model was built using the glmnet package in R with 10-fold cross-validation. The model was specified with the “binomial” family parameter. The optimal penalty parameter (λ) was selected based on the minimum cross-validated error criterion (λ.min). Coefficient profiles for the selected features were visualized using logarithmic λ plots, along with the corresponding partial likelihood deviance curves. In addition, the λ value within one standard error of the minimum (λ.1se) was calculated to further assess model stability and feature selection robustness.

Establishment and validation of diagnostic genes and diagnostic models

We intersected the Hub genes obtained by the above three machine algorithms and use the obtained genes as diagnostic genes. The internal discriminative ability of the identified hub genes was evaluated using receiver operating characteristic (ROC) analysis within the GSE78071 dataset. To further evaluate the combined diagnostic value of the identified hub genes, a diagnostic model was constructed using a logistic regression framework. The expression levels of CCL19 and CCR6 were included as independent variables, and tumor type (fibroadenoma vs. phyllodes tumor) was used as the binary outcome.Model construction and training were performed using the GSE78071 dataset as the discovery cohort. Prior to model fitting, no additional feature engineering or dimensionality reduction was applied, as hub gene selection had already been completed through the intersection of WGCNA, Random Forest, and LASSO regression analyses. The logistic regression model was fitted using standard maximum likelihood estimation. Model performance was internally evaluated by receiver operating characteristic (ROC) curve analysis and calibration assessment. Discriminative ability was quantified using the area under the ROC curve (AUC), while model calibration was examined using calibration curves comparing predicted and observed probabilities.

Due to the limited availability of independent public transcriptomic datasets for phyllodes tumors with comparable pathological annotation, external validation was not performed. Therefore, the diagnostic model was defined as an internally validated, exploratory model.

Human sample collection

On December 5, 2023, three patients with fibroadenoma and three patients with phyllodes tumor who had been pathologically diagnosed between April 1 and April 30, 2023, at the Department of Pathology, Ruikang Hospital Affiliated to Guangxi University of Traditional Chinese Medicine, were randomly selected for validation experiments. Reverse transcription–polymerase chain reaction (RT-PCR) was performed to validate the results of the transcriptomic analysis.

This study was approved by the Ethics Committee of Ruikang Hospital Affiliated to Guangxi University of Traditional Chinese Medicine. Total RNA was extracted from tissue samples using TRIzol reagent (Invitrogen™, Japan), and complementary DNA (cDNA) was synthesized using HiScript III RT SuperMix (Vazyme, Nanjing, China). Quantitative real-time PCR was subsequently conducted to measure the expression levels of CCR6 and CCL19 in fibroadenoma and phyllodes tumor samples.

Statistical comparisons between groups were performed using Student’s t-test. Data are presented as mean values, and a P value < 0.05 was considered statistically significant.

Results

Identification of differentially expressed genes between fibroadenomas and phyllodes tumors

This investigation utilized the GSE78071 dataset from the Gene Expression Omnibus (GEO) database, comprising samples from 24 patients with fibroadenoma and 24 with phyllodes tumors. After normalization and quality control, all samples exhibited highly comparable expression distributions without obvious outliers, supporting the reliability of subsequent differential expression analysis (Supplementary Fig. 1). Genes exhibiting a logarithmic fold change (log2 FC) greater than 1 and a p-value less than 0.05 were classified as differentially expressed genes (DEGs). A total of 463 DEGs were identified, encompassing 209 upregulated and 254 downregulated genes, as depicted in Fig. 1a. After Benjamini–Hochberg correction, application of a stringent threshold (FDR < 0.05) reduced the number of differentially expressed genes compared with the nominal p < 0.05 cutoff. However, the overall expression patterns and EMT-related biological signals remained consistent, and the key hub genes identified by integrative analyses were preserved (Supplementary Materials Table 2). A volcano plot was constructed using R programming to highlight the top 20 upregulated and downregulated genes. These genes were further visualized in a heatmap (Fig. 1b), where red indicated upregulated genes and blue represented downregulated genes. Subsequent analysis involved intersecting the epithelial–mesenchymal transition (EMT) gene set with the identified DEGs, resulting in the identification of 42 EMT-DEGs, as shown in Fig. 1c. The expression of EMT-DEGs is detailed in Supplementary Materials Table 1. Genome Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed on these 42 EMT-DEGs. The results indicated enrichment in various biological processes and pathways, including regulation of epithelial cell proliferation, stem cell differentiation, and activities related to collagen-containing extra-cellular matrices, translation, repressor activity, and mRNA base post-transcriptional activity regulation. KEGG pathway analysis revealed associations of these 42 genes with pathways including MicroRNAs in cancer, Axon guidance, and AGE-RAGE signaling pathway in diabetic complications, as illustrated in Fig. 1d. We further examined the correlations between the 42 EMT-DEGs, with the results presented in Fig. 1e. Additionally, a heatmap was created to elucidate the relationship between the different tumor types and the 42 EMT-DEGs (Fig. 1f). Finally, the expression levels of the 42 EMT-DEGs in the original dataset were depicted using a violin plot (Fig. 1g), providing a comprehensive visualization of the expression patterns.

Fig. 1.

a Volcano plot of differentially expressed genes (DEGs) in breast fibroadenoma and phyllodes tumor in GSE78071. Red nodes represent downregulated DEGs, blue nodes represent upregulated DEGs, and gray nodes represent genes with no significant differential expression. b Heatmap of expression levels of the top 20 differentially expressed genes (DEGs): blue represents fibroadenoma samples, gray represents phyllodes tumor samples, green indicates low expression, and red indicates high expression. c Venn diagram of differentially expressed genes in GSE78071 and genes related to Epithelial Mesenchymal Transition (EMT)(EMT-DEGs). d Bubble chart of enrichment analysis for 42 EMT-DEGs, categorized into CC, MF, and BP. Larger circles represent a greater number of genes enriched in a specific pathway. The color indicates the significance of the pathway, with darker colors representing smaller p-values and greater significance. BP = Biological Process, CC = Cellular Component, KEGG = Kyoto Encyclopedia of Genes and Genomes, MF = Molecular Function. e Heatmap of the correlation of 42 EMT-DEGs. f Clustered heatmap of 42 EMT-DEGs in different tumor samples. g Differential expression of 42 EMT-DEGs in breast fibroadenoma and phyllodes tumor samples in GSE78071. Cyan indicates fibroadenoma samples, and yellow indicates phyllodes tumor samples

WGCNA identifies a tumor-associated gene module strongly correlated with fibroadenoma–phyllodes tumor classification

In this study, we performed a comprehensive analysis of the GSE78071 dataset, which comprised expression profiles of 29,377 genes across 48 samples. To construct a gene co-expression network that satisfied the scale-free topology criterion, an optimal soft-thresholding power (β) of 4 was selected (Fig. 2a). At this threshold, the scale-free topology fit index (R²) reached 0.91 (Fig. 2b), indicating that the resulting network closely approximated a scale-free structure.

Fig. 2.

a Analysis of scale-free fit indices and mean connectivity for different soft-thresholding powers. b Histogram of connectivity distribution and scale-free topology check at β = 4. c An analysis of the interactive relationships among co-expressed genes. On the axes, different colors represent different modules, and the highlighted section in the center shows the extent of connectivity between these mod-ules. d The cluster dendrogram of genes in GSE78071. Each branch in the figure represents one gene, and every color below represents one co-expression module. e Heatmap of module characteristic genes and breast epithelial tumor grouping. f Scatter plot of module eigengenes in the grey60 module

Gene modules were subsequently identified using the Topological Overlap Measure (TOM) combined with hierarchical clustering and the Dynamic Tree Cut algorithm, with a minimum module size of 50 genes. Analysis of inter-module connectivity revealed no pronounced interactions among distinct modules, suggesting a high degree of independence between modules (Fig. 2c). After excluding modules containing fewer than 50 genes and merging highly similar modules using a cut height threshold of 0.25, Weighted Gene Co-expression Network Analysis (WGCNA) identified a total of 15 co-expression modules (Fig. 2d).

To determine the associations between these modules and epithelial tumor phenotypes, Pearson correlation analysis was performed between module eigengenes (MEs) and tumor group status. Among all identified modules, the grey60 module demonstrated the strongest correlation with tumor grouping (r = 0.52, p = 2 × 10⁻⁴), as shown in Fig. 2f, and exhibited the most prominent association in the module–trait relationship heatmap (Fig. 2e).

A total of 2,431 genes were extracted from the grey60 module for downstream analyses. Importantly, the module–trait relationship analysis demonstrated that the grey60 module exhibited the strongest and most statistically significant correlation with tumor grouping (fibroadenoma vs. phyllodes tumor) among all identified modules (Fig. 2e–f). This finding indicates that genes within the grey60 module collectively capture transcriptional programs most relevant to the biological differences between these two fibroepithelial lesions. Therefore, genes from the grey60 module were prioritized as the candidate gene set for subsequent feature selection using Random Forest and LASSO regression analyses.The strong association between this module and tumor classification highlights its potential relevance in elucidating the molecular characteristics underlying epithelial tumors. Overall, this large-scale gene co-expression network analysis provides a robust framework for identifying biologically meaningful gene modules associated with tumor phenotypes, thereby contributing to a deeper understanding of tumor-related molecular mechanisms.

Random forest analysis highlights candidate diagnostic genes associated with tumor classification

Based on the WGCNA results, genes from the grey60 module were considered highly relevant to tumor classification. To further identify key diagnostic genes at the individual gene level, these candidate genes were subsequently analyzed using a Random Forest classifier. We chose 1000 trees as the parameters of random forest model, set 67% of samples for prediction modeling, 33% of samples to validate the prediction model, and performed 150 samplings. Finally, we determined the final Hub gene by the relationship between error rate and the number of genes after multiple cross-validation. Eventually, we could see that the error rate tended to be stable when the number of genes was 3 after five cross-validations (Fig. 3b and c), so we identified the diagnostic genes as CCL19, CCR6 and ZXY by random forest method (Fig. 3a).

Fig. 3.

a Importance of genes in the prediction model constructed by random forest. b The relationship between the number of genes and the error rate in cross-validation. c The relationship between the number of genes and the error rate in 5-fold cross-validation

LASSO regression further refines EMT-related diagnostic gene signatures

In this study, the Least Absolute Shrinkage and Selection Operator (LASSO) regression algorithm was applied to the expression profiles of 42 epithelial–mesenchymal transition–related differentially expressed genes (EMT-DEGs). The coefficient trajectory plot and model performance are presented in Fig. 4a and b, respectively. The optimal regularization parameter was determined using the lambda.min (λ) criterion and subsequently adopted for model construction. Based on this optimal λ value, LASSO regression selected nine candidate genes that were incorporated into the diagnostic model (Fig. 4b).

Fig. 4.

a and b Screening of gene signatures from EMT-DEGs using the LASSO algorithm

The selected genes and their corresponding coefficients were as follows: TGFB2 (0.07242591), FN1 (Fibronectin 1; 0.23401272), TP63 (− 0.37358379), VASH2 (0.93907792), CCL19 (− 0.43531571), LHX2 (0.35342159), NKX6.1 (0.25733310), CCR6 (− 0.58267224), and CTHRC1 (0.01935721). These genes encompass diverse regulatory and functional roles relevant to the pathological processes under investigation. Negative coefficients indicate an inverse association with the disease condition, whereas positive coefficients reflect a direct association.

Overall, the LASSO-based feature selection strategy enabled efficient dimensionality reduction and robust identification of key diagnostic genes, providing a statistically rigorous foundation for subsequent model development.

CCL19 and CCR6 are identified as robust diagnostic hub genes and validated at transcriptomic and tissue levels

In our study, we employed a multifaceted approach, intersecting Weighted Gene Co-expression Network Analysis (WGCNA), Random Forest, and Lasso regression techniques. This methodology enabled us to identify two pivotal Hub genes, CCL19 and CCR6, as depicted in Fig. 5a. To assess the internal discriminative ability of the identified hub genes, we conducted a Receiver Operating Characteristic (ROC) analysis using the GSE78071 dataset. The analysis yielded ROC values of 0.962 for CCL19 and 0.915 for CCR6, as illustrated in Fig. 5b, indicating strong separation between fibroadenoma and phyllodes tumor samples within this cohort. To further refine this diagnostic capacity, we developed a diagnostic model based on logistic regression using the combined expression of CCL19 and CCR6, as shown in Fig. 5c. The model demonstrated favorable internal discriminative performance within the GSE78071 dataset, as assessed by ROC analysis and calibration curves (Fig. 5c–d). These results reflect internal, exploratory discrimination within a single discovery cohort and should not be interpreted as externally validated diagnostic accuracy.

Fig. 5.

Identification, diagnostic performance, and validation of hub genes CCL19 and CCR6: a Venn diagram showing the intersection of hub genes identified by WGCNA, Random Forest, and LASSO regression analyses. b Receiver operating characteristic (ROC) curves evaluating the individual diagnostic performance of CCL19 and CCR6 for discriminating fibroadenomas from phyllodes tumors in the GSE78071 dataset. c ROC curve of the combined diagnostic model constructed using multivariable logistic regression with CCL19 and CCR6 expression levels as independent variables and tumor type (fibroadenoma vs. phyllodes tumor) as the binary outcome. Model performance reflects internal discrimination within the GSE78071 discovery cohort. d Calibration curve of the CCL19–CCR6 logistic regression model, illustrating the agreement between predicted probabilities and observed outcomes. e Protein–protein interaction (PPI) network of genes associated with CCL19 and CCR6, constructed using the STRING database with a high-confidence interaction score (0.900). f Functional enrichment analysis of genes associated with CCL19 and CCR6, including Gene Ontology biological processes, cellular components, molecular functions, and KEGG pathways. g RT–qPCR validation of CCL19 and CCR6 expression in clinical breast fibroadenoma and phyllodes tumor tissues (n = 3 per group). Data are presented as mean ± SD; P < 0.05

We then integrated the diagnostic genes CCL19 and CCR6 into the STRING data-base, applying the highest confidence setting (0.900) and limiting the number of interactors to no more than 50. This process culminated in the formation of a Protein-Protein Interaction (PPI) network interaction map, as seen in Fig. 5e. Intriguingly, this map revealed a high degree of clustering within three gene families: the CCR/CCL and CXCL families. Subsequent enrichment analysis of these genes, depicted in Fig. 5f, highlighted several key areas: Biological Process (BP) encompassing cellular response to chemokine, response to chemokine, and cytokine-mediated signaling pathway; Cellular Component (CC) focusing on the external side of the plasma mem-brane; and Molecular Function (MF) involving chemokine receptor interaction, cytokine activity, and C-C chemokine receptor. Additionally, KEGG pathway analysis identified involvement in cytokine-cytokine receptor interaction and the chemokine signaling pathway, as well as viral protein interaction with cytokine and cytokine receptor. This enrichment information suggests a process of chemokine translocation across the cell membrane to bind with chemokine receptors, warranting further investigation to elucidate the underlying mechanisms.

To validate the diagnostic utility of CCL19 and CCR6, we conducted RT-PCR on pathological specimens from three patients with established fibroadenomas and three patients with phyllodes tumors. This analysis confirmed the gene expression levels of the two diagnostic biomarkers, CCR6 and CCL19. Our findings were consistent with the data analysis, showing that the expression of CCR6 and CCL19 was upregulated in fibroadenoma and downregulated in phyllodes tumors, as illustrated in Fig. 5g.

Discussion

Conventional histopathological criteria, including the degree of stromal cell proliferation, mitotic activity, and stromal heterogeneity, frequently result in overlapping features between fibroadenomas (FB) and phyllodes tumors (PT), particularly benign PT, thereby posing persistent diagnostic challenges. Although genomic studies have proposed MED12 mutations as potential discriminators [16–19], their nearly equivalent mutation frequencies in benign PT (65.1%) and FB (60–70%) [6, 20–22] underscore the limitations of relying solely on single-gene markers or histopathological assessment. To address this diagnostic gap, we leveraged transcriptomic data in combination with machine learning algorithms to identify CCL19 and CCR6 as novel diagnostic biomarkers. The strong diagnostic performance of these markers, reflected by high ROC AUC values (0.962 for CCL19 and 0.915 for CCR6), supports the robustness of our model and offers a quantitative and minimally invasive approach for preoperative differentiation—an important clinical advance given the potential consequences of misdiagnosis, including undertreatment of PT or overtreatment of FB [20–23]. The diagnostic potential of CCL19/CCR6 is supported by their mechanistic roles in epithelial-mesenchymal transition (EMT), a process critical for stromal remodeling. EMT is defined by the loss of epithelial markers (e.g., E-cadherin) and the acquisition of mesenchymal traits (e.g., vimentin) and is regulated through chemokine networks. Protein interaction analysis revealed that CCR6, a G protein-coupled receptor, activates the CXCR4-GNAQ-PLCβ signaling axis, triggering calcium-dependent cytoskeletal reorganization and increasing stromal cell motility.

Despite on-going efforts to differentiate these entities, an effective consensus has yet to be established. This study utilizes transcriptome data from the GEO database, analyzing fibroadenomas and phyllodes tumors through three machine learning algorithms. It focuses on EMT-related genes to identify those correlating with and effectively distinguishing between PT and FA. Fortuitously, we discovered significant differential ex-pression of CCR6 and CCL19 between PT and FA, which was corroborated in pathological tissue. Further, we ascertained that both CCR6 and CCL19 could independently discriminate between the two diseases. To enhance diagnostic efficacy, we developed a diagnostic model. To our knowledge, this is the first study to apply machine learning approaches to identify EMT-related diagnostic biomarkers distinguishing fibroadenomas from phyllodes tumors. However, this study is limited by the relatively small sample size and reliance on a single public dataset, which reflects the rarity of phyllodes tumors. External validation using independent cohorts will be required to confirm the generalizability of these findings. This internally validated model provides a molecular basis for improving the differential diagnosis of fibroadenomas and phyllodes tumors, although further external validation is required before clinical application [24, 25].This aligns with our findings in collagen-containing extracellular matrix (ECM) regulation (Fig. 1d), suggesting that CCR6-driven ECM deposition facilitates EMT. Conversely, CCL19, through CCR7 binding, recruits regulatory T cells (Tregs) and tumor-associated macrophages (TAMs) [26–29], which secrete TGF-β, further inducing EMT. Reduced CCL19 expression in PT (Fig. 5g) may weaken immune surveillance, promoting stromal hypercellularity, while elevated CCL19 in FB could suppress the PI3K/AKT/mTOR pathway via CXCR4, maintaining a quiescent tumor microenvironment (TME) (Fig. 5f). These dual mechanisms—stromal plasticity and immune evasion—align with the broader “EMT-metastasis” paradigm.

The role of chemokines in PT pathogenesis is further exemplified by CCL18 and CCL5. CCL18, secreted by TAMs, drives NF-κB/miR-21/PTEN/AKT-mediated fibroblast differentiation and PT progression, while CCL5-CCR5 interactions repolarize TAMs to release CCL18, exacerbating malignant phenotypes. While these cytokines play pivotal roles in PT biology, their utility in differentiating FB from PT remains unvalidated. In contrast, CCL19 and CCR6—traditionally studied in oncology for their roles in T cell homing, angiogenesis, and immune evasion [30, 31]—emerge as unique discriminators in this context. Our findings connect chemokine biology with EMT-driven stromal remodeling, shedding light on a previously underexplored axis in fibroepithelial tumors. While computational evidence (e.g., PPI clustering of GNAQ with CCR6/CCL19-associated genes; Fig. 5e) supports the CXCR4-GNAQ link, experimental validation is required. Previous studies in breast cancer indicate that CCR6/CCL20-driven GNAQ activation enhances VEGF production and angiogenesis, suggesting that similar pathways may drive PT stromal expansion. Likewise, CXCR4 inhibition has been shown to reverse EMT in triple-negative breast cancer by downregulating Snail and Twist [32]. emphasizing its therapeutic potential. However, this study is limited by its reliance on public datasets and restricted tissue validation. Future research should expand sequencing databases and experimentally validate mechanistic pathways to refine diagnostic models.

By integrating machine learning with EMT biology, this study positions CCL19/CCR6 as transformative biomarkers for FB/PT differentiation, offering both diagnostic precision and insights into stromal-immune crosstalk. This advancement enhances the molecular understanding of fibroepithelial tumors, addressing a long-standing clinical challenge. Several methodological limitations should be acknowledged. First, feature selection and ROC analysis were performed using the same transcriptomic dataset, which may introduce optimism bias and inflate AUC estimates due to circular validation. Accordingly, the reported ROC values should be interpreted as exploratory indicators of internal discrimination rather than definitive measures of diagnostic accuracy.

Furthermore, owing to the rarity of phyllodes tumors, independent external transcriptomic cohorts with comparable pathological annotations are currently unavailable. Although independent RT-qPCR validation in freshly collected clinical specimens demonstrates concordant expression patterns of CCL19 and CCR6 between computational analysis and clinical specimens, further validation in independent, multi-center cohorts will be necessary to establish generalizability and clinical utility.

Conclusions

This study delineates CCL19 and CCR6 as determinant diagnostic biomarkers for distinguishing phyllodes tumors (PT) from breast fibroadenomas (FB), addressing a critical gap in clinical differentiation. By integrating machine learning approaches (WGCNA, Random Forest, and LASSO regression) with bioinformatics focused on epithelial-to-mesenchymal transition (EMT), these chemokine-associated genes have been identified as highly effective discriminators, surpassing traditional markers like MED12 mutations, which show overlapping prevalence in FB and PT. Mechanistically, CCL19 and CCR6 function within chemokine signaling networks, potentially influencing EMT via the CXCR4-GNAQ-PLCβ pathway to affect stromal plasticity and immune microenvironment dynamics, offering fresh insights into tumor biology.

While this research advances fibroepithelial tumor diagnostics, its dependence on public databases and a limited validation cohort highlights the need for further refinement. Future studies should prioritize multi-center validation with larger sample sizes, the integration of proteomic data, and experimental investigations of the CXCR4-GNAQ pathway to enhance clinical applicability. These efforts will solidify CCL19 and CCR6 not only as diagnostic markers but also as potential therapeutic targets, bridging the gap between computational discovery and clinically actionable strategies.

Abbreviations

AKT

Protein Kinase B

BP

Biological Process

CC

Cellular Component

CCL

Chemokine (C–C motif) Ligand

CCR

Chemokine (C–C motif) Receptor

CXCL

Chemokine (C-X-C motif) Ligand

CXCR

Chemokine (C-X-C motif) Receptor

DEGs

Differentially Expressed Genes

ECM

Extracellular Matrix

EMT

Epithelial-Mesenchymal Transition

EMT-DEGs

Epithelial-Mesenchymal Transition-related Differentially Expressed Genes

FB

Fibroadenoma

GEO

Gene Expression Omnibus

GO

Gene Ontology

KEGG

Kyoto Encyclopedia of Genes and Genomes

LASSO

Least Absolute Shrinkage and Selection Operator

LIMMA

Linear Models for Microarray Data

MAE

Mean Absolute Error

MF

Molecular Function

mTOR

Mechanistic Target of Rapamycin

NF-κB

Nuclear Factor Kappa B

PAX3

Paired Box 3

PI3K

Phosphatidylinositol 3-Kinase

PIK3CA

Phosphatidylinositol-4,5-Bisphosphate 3-Kinase Catalytic Subunit Alpha

PLCβ

Phospholipase C Beta

PT

Phyllodes Tumor

PTEN

Phosphatase and Tensin Homolog

qPCR

Quantitative Polymerase Chain Reaction

RARA

Retinoic Acid Receptor Alpha

RF

Random Forest

ROC

Receiver Operating Characteristic

RT-PCR

Reverse Transcription Polymerase Chain Reaction

TAMs

Tumor-Associated Macrophages

TGF-β

Transforming Growth Factor Beta

TGFB2

Transforming Growth Factor Beta 2

TME

Tumor Microenvironment

TOM

Topological Overlap Measure

TP53

Tumor Protein p53

Tregs

Regulatory T Cells

Author contributions

Conceptualization, H. Liu; methodology, L. Q. Lv; software, L. Q. Lv; validation, H. Lv; formal analysis, H. Gao; data curation, L. Q. Li; writing—original draft preparation, Y. D. Lan; writing—review and editing, H. Gao; visualization, H. Gao; funding acquisition, H. Gao. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by Guangxi Health Commission. Grant number Z-A20230938.

Data availability

The datasets generated and analyzed during the current study are available in the GEO database (https://www.ncbi.nlm.nih.gov/geo/).

Declarations

Ethics approval and consent to participate

Ethical review and approval were waived for this study due to the data reproduced from Article utilized human tissue that was procured via our Pathology bank which provides de-identified samples. The Pathology Bank protocols are in accordance with the ethical standards of our institution and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.