Uncovering GSTK1 and PTGER3 as biomarkers for breast cancer prognosis through comprehensive analyses Authors: Lin Tan, Junlian Xiang, Yi Lu, Xiaoli Zhong Published: 2025 ABSTRACT AbstractBreast cancer threatens women’s health globally, and identifying prognostic genes is therefore critical for optimizing diagnosis and therapy. In this study, we retrieved datasets from the GEO database, and differential analysis subsequently identified 66 DEGs. GO and KEGG enrichment analyses revealed these DEGs were not only enriched in biological processes like angiogenesis and granulocyte differentiation but also in key pathways including Th17 cell differentiation and cancer transcriptional dysregulation. Subsequently, we evaluated 21 machine learning algorithms, among which RidgeCV and XGBoost significantly outperformed others. Specifically, prognostic models constructed based on these two algorithms exhibited excellent performance in decision curve analysis, confusion matrix evaluation, and feature importance assessment. Consequently, Cox regression analysis of the 29 overlapping genes identified by both models pinpointed GSTK1 and PTGER3 as key prognostic genes. Notably, GSTK1 and PTGER3 were underexpressed in breast cancer tissues, with high diagnostic potential and a strong correlation with pathological staging. Moreover, high expression of these two genes correlated with favorable patient survival, and a nomogram integrating them could effectively predict patient survival probabilities. Additionally, CIBERSORT and ssGSEA analyses linked GSTK1/PTGER3 to immune cell infiltration in the tumor microenvironment, further indicating their crucial roles in regulating the breast cancer immune landscape. In vitro experiments further confirmed that GSTK1 and PTGER3 are weakly expressed in breast cancer cells, and their high expression could inhibit the proliferation of breast cancer cells while predicting favorable patient survival outcomes. Collectively, this study identifies GSTK1 and PTGER3 as potential biomarkers for breast cancer precision medicine, thereby providing valuable insights into disease pathogenesis and novel therapy development.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12672-025-04329-7.Keywords: Breast cancer, Machine learning, GSTK1, PTGER3, Biomarker Introduction Breast cancer ranks among the leading causes of cancer incidence in women globally, significantly impacting patients’ quality of life and survival outcomes [1]. Despite ongoing advancements in diagnostic and therapeutic techniques, issues such as postoperative recurrence and drug resistance remain critical, and the overall prognosis for patients is still concerning. A deeper understanding of the prognostic factors of breast cancer is essential for early identification of high-risk individuals, enabling early diagnosis and treatment, which can ultimately improve survival rates and quality of life. Malignant tumor progression is intricately intertwined with the reprogramming of cellular energy metabolism [2, 3]. Among diverse tissue types, adipose tissue acts as a critical energy reservoir and endocrine organ that plays an indispensable role in the establishment of the tumor microenvironment and the provision of metabolic energy. This biological characteristic is particularly conspicuous in breast cancer [4, 5]. Specifically, within the fatty acid metabolic pathway, adipocytes release free fatty acids (FFAs) via lipolysis. These FFAs are subsequently internalized by tumor cells and exploited for mitochondrial β-oxidation, thereby furnishing the indispensable energy substrates to sustain tumor cell proliferation, invasion, and metastasis [6–8]. Furthermore, adipokines secreted by adipocytes, such as leptin and adiponectin, exert paracrine or autocrine regulatory effects on tumor cells to modulate their metabolic cascades [9]. Accumulating evidence corroborates that leptin, through the ER/JAK/STAT3, PI3K/AKT, and FAK/Src signaling network, acts as a core driver in breast cancer cell proliferation, invasion, epithelial-mesenchymal transition (EMT), and the maintenance of cancer stem cell (CSC) stemness [10–12]. In contrast, adiponectin exerts inhibitory effects on tumor cell growth and metabolic activity. Chung et al. demonstrated that adiponectin induces autophagy in breast cancer cells by activating the STK11/LKB1-AMPK-ULK1 signaling axis, which remarkably suppresses tumor growth in both in vitro cellular models and in vivo murine models [13]. Notably, with the continuous advancement of molecular biology technologies, a plethora of novel biomarkers have been progressively identified, providing robust support for the precise diagnosis of tumors, the assessment of therapeutic efficacy, and the formulation of individualized treatment strategies [14, 15]. Therefore, systematic exploration of key regulatory genes derived from the breast cancer adipose microenvironment harbors profound theoretical value and clinical significance. It facilitates an in-depth elucidation of breast cancer pathogenesis, enables the screening of potential therapeutic targets, and ameliorates patients’ clinical prognosis. This line of research is anticipated to pave an innovative avenue for the precise diagnosis and targeted therapy of breast cancer. With the rapid advancement of genomics, a wealth of tumor-related data has been accumulated, providing valuable resources for exploring cancer’s pathological mechanisms. Additionally, the widespread application of artificial intelligence and machine learning techniques in bioinformatics offers powerful tools for identifying key genes associated with disease prognosis. Different machine learning models can reveal potential relationships between genes and disease prognosis from various perspectives. By integrating multiple models, the risk of overfitting can be reduced, and the reliability of results can be improved, thus enhancing the possibility of discovering potential biomarkers. This study aims to leverage bioinformatics approaches to extract gene matrices from breast cancer-related databases and apply various machine learning algorithms to identify critical feature genes affecting breast cancer prognosis, and verify them in combination with in vitro experiments. The potential of these genes as biomarkers will be explored to provide new reference points for clinical diagnosis, prognosis, and treatment of breast cancer. Materials and methods Transcriptomic data acquisition and analysis The GSE153316 transcriptomic dataset (including 41 samples of adipose tissue adjacent to tumors and 41 samples of adipose tissue far from tumors) was retrieved from the Gene Expression Omnibus (GEO) for subsequent analysis. The baseline information of the samples is detailed in Table 1 of the supporting materials. First, Using the ID conversion tool available in SangerBox (http://sangerbox.com/), probe identifiers in the expression dataset were converted to official gene symbols, and probes with undetectable expression levels were removed. In cases where multiple probes corresponded to the same gene, the median expression value was retained for that gene. Differential expression analysis was performed using the limma package in R, and data preprocessing and statistical calculations were automated through Perl scripts. Differentially expressed genes (DEGs) were selected based on the following criteria: adjusted P-value < 0.05 (after Benjamini-Hochberg correction) and |log₂ fold change (FC)| > 1. Data visualization was conducted using the ggplot2, with a volcano plot employed to display the distribution and expression patterns of the differentially expressed genes. Table 1Cox regression analysis of key genesCharacteristicsTotal(N)Univariate analysisMultivariate analysisHazard ratio (95% CI)P valueHazard ratio (95% CI)P valueSULF11,086 Low542Reference High5441.206 (0.876–1.661)0.251IFI161,086 Low543Reference High5430.945 (0.685–1.304)0.732CPXM21,086 Low543Reference High5430.950 (0.689–1.308)0.752CLMP1,086 Low542Reference High5440.808 (0.586–1.114)0.193FAH1,086 Low543Reference High5431.027 (0.747–1.412)0.870CHST151,086 Low542Reference High5441.058 (0.769–1.454)0.729PTGER31,086 Low543ReferenceReference High5430.713 (0.518–0.982) 0.038 0.785 (0.565–1.092)0.150TM4SF11,086 Low543Reference High5431.136 (0.825–1.565)0.434ANGPT21,086 Low542Reference High5441.008 (0.733–1.386)0.961GSTK11,086 Low543ReferenceReference High5430.616 (0.447–0.851) 0.003 0.657 (0.473–0.913) 0.012 COPRS1,086 Low543Reference High5430.884 (0.641–1.218)0.449SLC43A31,086 Low543Reference High5430.968 (0.704–1.333)0.844TSHZ21,086 Low542ReferenceReference High5440.747 (0.540–1.035)0.0800.864 (0.594–1.256)0.444C1orf211,086 Low543Reference High5430.935 (0.680–1.286)0.680ASPN1,086 Low542Reference High5440.969 (0.704–1.333)0.846RGS161,086 Low542Reference High5441.033 (0.750–1.423)0.842CADM21,086 Low542Reference High5440.878 (0.638–1.208)0.425ADH51,086 Low542Reference High5440.913 (0.664–1.255)0.574CCL211,086 Low543Reference High5430.803 (0.583–1.105)0.178GPT21,086 Low543Reference High5431.009 (0.734–1.388)0.955TUBA81,086 Low543Reference High5431.064 (0.774–1.463)0.703PROX11,086 Low542Reference High5440.972 (0.707–1.337)0.862CFI1,086 Low543Reference High5430.903 (0.656–1.242)0.531SMOC21,086 Cox regression analysis of key genes Functional enrichment analysis Gene Ontology (GO) enrichment analysis for the DEGs was conducted using the Hipplot platform (https://hiplot.com.cn/), encompassing the three major domains: biological processes (BP), cellular components (CC), and molecular functions (MF). Terms with an adjusted p-value < 0.05, as determined by the Benjamini-Hochberg (BH) procedure, were considered statistically significant. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis was performed via the Metascape platform (http://metascape.org/), which integrates functional annotation with protein-protein interaction (PPI) network analysis. This approach aimed to identify key biological pathways and molecular mechanisms involved in the differential genes, particularly in the context of adipocyte biology in breast cancer. Machine learning algorithm integration and validation To identify robust prognostic biomarkers, a systematic evaluation of 21 machine learning algorithms was performed. The caret package was used to create radar plots to display the performance of each algorithm in terms of accuracy (Acc), F1 score, precision, and recall. The pROC package was used to construct Receiver Operating Characteristic (ROC) curves, and the area under the curve (AUC) was calculated to compare the performance of the algorithms. Ultimately, the algorithm demonstrating optimal performance and the largest AUC was selected for subsequent feature selection and model construction. Cox regression prognostic validation Using the survival package’s coxph function, single-factor Cox proportional hazards regression analysis was performed on the 29 intersecting genes identified by the RidgeCV and XGBoost models. The specific information of the 29 intersection genes is listed in Table 2 of the supplementary materials. A cohort of 1,086 breast cancer patients was curated from The Cancer Genome Atlas (TCGA) database (https://tcga-data.nci.nih.gov/). Genes were stratified into high- and low-expression groups based on the median expression level. Overall survival (OS) was defined as the duration from the date of initial breast cancer diagnosis to the date of death from any cause. Samples without a sample name suffix of “01A” or with missing survival time/status information were excluded from the analysis. Hazard ratios (HRs), 95% confidence intervals (95% CIs), and p-values were calculated to identify genes significantly associated with OS (p < 0.05). A multivariate Cox model was then constructed to adjust for clinical covariates, including the AJCC 8th edition TNM staging, molecular subtypes, and control for confounding variables, to assess whether candidate genes (such as GSTK1 and PTGER3) retained independent prognostic significance in the context of comprehensive clinical factors, Detailed clinical variables corresponding to the samples have been provided in Supplementary Table 3. (p < 0.05). Differential expression and pathological staging Differential expression analysis was performed using the Xiantao Academic Platform (https://www.xiantao.love/), with box plots and paired t-tests used to display expression differences of GSTK1 and PTGER3 between tumor and normal tissues. The Kruskal-Wallis test was applied to assess the correlation between gene expression and pathological stages (I–IV), identifying stage-specific expression patterns. Survival analysis and diagnostic performance ROC analysis was performed using the pROC package to calculate AUC values and evaluate the diagnostic performance of the genes in distinguishing tumor from normal tissues. Higher AUC values indicate stronger diagnostic accuracy. Kaplan-Meier survival curves were plotted, and log-rank tests were used to compare survival differences between high- and low-expression groups of the genes. Nomogram construction A nomogram was constructed using the rms package, integrating pathological T/N stages and the expression levels of GSTK1/PTGER3, to predict overall survival (OS), disease-specific survival (DSS), and progression-free interval (PFI). Weighted scores for each variable were calculated based on regression coefficients, and the total score was used to estimate 1- and 5-year survival probabilities, providing a visual tool for individualized prognostic assessment. Immune infiltration analysis Immune cell composition within the tumor microenvironment was assessed using the CIBERSORTx platform (https://cibersortx.stanford.edu/). This algorithm utilizes a validated gene marker matrix to quantify 22 immune cell subtypes (with default parameters: 1000 permutations, Q-value < 0.05). Pearson correlation analysis was performed to assess the relationship between gene expression and immune cell proportions (P < 0.05 indicating significant correlations). Additionally, single-sample Gene Set Enrichment Analysis (ssGSEA) was performed using the GSVA package to calculate enrichment scores for 56 immune cell types (from the Immune Gene Sets database). Spearman rank correlation analysis was used to explore the relationship between gene expression and ssGSEA scores, providing further insight into the regulatory role of GSTK1 and PTGER3 in the immune microenvironment. Cell culture and cell transfection The human breast cancer cell lines MCF7 and MDA-MB-231 were obtained from the cell bank of the Chinese Academy of Sciences (Shanghai, China). MCF7 and MDA-MB-231 were cultured in MEM medium (Gibco, Grand Island, NY, USA) and L-15 medium (Gibco). All mediums were supplemented with 10% fetal bovine serum, 1% l-glutamine, and 1% penicillin–streptomycin (Beyotime, Shanghai, China). The cells were maintained at 37 °C in a CO2 incubator (Thermo Fisher Scientific, Waltham, MA, USA). XMU-MP-1 was purchased from MedChemExpress (MCE, Monmouth Junction, NJ, USA). MCF7 and MDA-MB-231 cells were transfected with siRNA targeting GSTK1 (si-GSTK1 GGAUCAAGUUCAAGAACUUdTdT) and PTGER3 (si-PTGER3 GCAUCAGCUUCUUCAACUAdTdT) and the negative control siRNA (si-NC GGCAGAAUCCUGUAUGUAAdTdT). The transfection of siRNA was performed using the Lipo2000 transfection reagent (11668019, Invitrogen, Carlsbad, CA, USA) according to the supplier’s protocol. RNA isolation and qPCR Total RNA was extracted from cells according to the manufacturer’s protocol using TRIpure Total RNA Extraction Reagent (EP013, ELK Biotechnology). Subsequently, 1 µg of RNA was reverse transcribed into cDNA using a cDNA synthesis kit (Eq. 031, ELK Biotechnology). qPCR analysis was performed with SEnTurbo™ SYBR Green PCR SuperMix (Eq. 001, ELK Biotechnology) on the QuantStudio 6 Flex System (Life Technologies, Gaithersburg, MD, USA) instrument. The relative mRNA expression levels were normalized to ACTIN and calculated using the 2-ΔΔCT method. Cell proliferation assay Cells were seeded in 96-well plates and CCK-8 reagent was added at 48 h of culture. The absorbance at 450 nm was measured using a DR-200Bs microplate reader (Diatek, Wuxi, China), with solvent-treated cells as the control group and the blank well set to zero. Relative proliferation rate was calculated using the following formula: relative proliferation rate (%) = [(control group − blank) − (experimental group − blank)]/(control group − blank) × 100%. Statistical analyses All statistical analyses were performed using R (version 4.3.0). The ROC curves were visualized with the “pROC” package. Univariate Cox proportional hazards regression analysis was carried out via the “survival” package. A predictive nomogram was constructed utilizing the “rms” package. ssGSEA was implemented using the “GSVA” package. Unless otherwise stated, all volcano plots and general charts were generated with the “ggplot2” package or the base “plot” function, respectively. Data preprocessing and automated statistical computations were conducted using Perl (version 5.16). Correlation analyses were assessed by the Pearson method. Comparisons of continuous and categorical variables between two groups were performed using the Wilcoxon test and the chi-square test, respectively. A two-sided p-value of less than 0.05 was defined as statistically significant. Results Differential expression gene screening and enrichment analysis in breast cancer adipose tissue To explore the gene expression differences in adjacent and distant adipose tissues in breast cancer and to identify the biological functions and signaling pathways associated with differentially expressed genes, this study analyzed the GSE153316 dataset from the GEO database (Fig 1). Differential expression analysis identified 66 differentially expressed genes, of which 24 were upregulated in tumor-adjacent adipocytes, and 42 were downregulated (Fig. 2A-B). Enrichment analysis revealed that these genes were primarily involved in processes such as angiogenesis regulation, vasculature development regulation, granulocyte differentiation regulation, and epithelial cell fate commitment. At the cellular component level, the genes were significantly enriched in the extracellular matrix. At the molecular function level, they were closely associated with glutathione peroxidase activity and extracellular matrix structural constituent (Fig. 2C-E). KEGG pathway analysis further indicated that the main enriched pathways included Th17 cell differentiation and cancer transcriptional dysregulation (Fig. 2F). Flowchart of methodology performed in this research Differential expression gene screening and enrichment analysis in adipose tissue of breast cancer. A. Volcano plot showing differentially expressed genes (DEGs) between tumor-adjacent and distant adipose tissues. B. PCA plot showing differentially expressed genes (DEGs) between tumor-adjacent and distant adipose tissues. C–E. Gene Ontology (GO) enrichment analysis for differentially expressed genes: B. Biological Process (BP); C Cellular Component (CC); D Molecular Function (MF) F. KEGG pathway analysis illustrating the biological pathways involved in the differentially expressed genes Machine learning model construction To identify the most suitable machine learning algorithm, this study applied the differentially expressed genes to 21 machine learning algorithms and analyzed their performance using radar charts and ROC curves. The radar chart results showed significant differences in accuracy (Acc), F1 score, precision and recall across the algorithms. RidgeCV, XGBoost, KNN, SVM, and SGD performed particularly well on these metrics (Fig. 3A). ROC analysis indicated that RidgeCV and XGBoost outperformed other algorithms, with AUC values of 0.827 for RidgeCV and 0.791 for XGBoost (Fig. 3B). Based on these results, RidgeCV and XGBoost were selected for subsequent analysis. Performance Evaluation of Machine Learning Algorithms. A. Radar plot comparing the performance of 21 machine learning algorithms in terms of accuracy (Acc), F1 score, precision, and recall. B. ROC curves displaying the diagnostic performance of the 21 machine learning algorithms RidgeCV model validation To comprehensively evaluate the performance of the model based on differentially expressed genes, decision curve analysis was conducted to compare the clinical net benefit of the study model (Model) with the “treat all” and “treat none” strategies. The results showed that, within a threshold probability range of 20%-80%, the clinical net benefit of the RidgeCV model was significantly higher than both the “treat all” and “treat none” strategies (Fig. 4A). The standardized confusion matrix was then constructed to assess the model’s prediction accuracy, showing prediction accuracies of 0.80 and 0.82 for the “distant” and “adjacent” categories, respectively, with a low misclassification rate and strong discriminatory power (Fig. 4B). Feature importance was calculated based on mean decrease in impurity (MDI), revealing that genes such as GNAL and SMOC2 had high importance scores (Fig. 4C). Permutation importance analysis indicated that genes like COL14A1 and PROX1 had large absolute importance values, further confirming the key role of these genes in the model (Fig. 4D). Taken together, the RidgeCV model demonstrated strong performance and clearly identified key genes. RidgeCV Model Validation Results. A. Decision curve analysis illustrating the clinical net benefit of the RidgeCV model. B. Standardized confusion matrix displaying the model’s classification performance. C. Feature importance analysis based on mean decrease in impurity (MDI), highlighting key features influencing the model. D. Permutation importance analysis validating the significant impact of genes on the model’s prediction XGBoost model validation To assess the performance and gene importance of the XGBoost machine learning model in breast cancer research, decision curve analysis was performed, comparing its (red line) net benefit with the “treat all” (black line) and “treat none” (gray line) strategies. The results showed that within a certain threshold probability range, the XGBoost model provided higher net benefit, indicating its significant advantage in assisting clinical decision-making (Fig. 5A). The standardized confusion matrix showed prediction accuracies of 0.80 and 0.82 for the “distant” and “adjacent” categories, respectively, with a low misclassification rate and high discriminatory accuracy (Fig. 5B). Feature importance was evaluated based on MDI, identifying genes such as IFI16 and PROX1 with high importance scores, which played a crucial role in model classification (Fig. 5C). Permutation importance analysis further validated the key genes, with PROX1 and SMOC1 showing large absolute importance values, confirming their critical impact on model predictions (Fig. 5D). Overall, the XGBoost model demonstrated excellent performance and identified key genes. XGBoost Model Results. A. Decision curve analysis illustrating the clinical net benefit of the XGBoost model. B. Standardized confusion matrix showcasing the classification performance of the model. C. Mean Decrease in Impurity (MDI) analysis identifying key genes as the critical features for model classification. D. Permutation importance analysis assessing the significance of genes in the model’s predictions Cox regression analysis and clinical significance of key genes A Cox regression analysis was performed on the 29 common genes identified by both models to explore their relationship with prognosis and identify prognostically significant key genes. A total of 1,086 samples were included in the analysis, and both univariate and multivariate analyses were conducted (Table 1). In the univariate analysis, GSTK1 (HR = 0.616, 95% CI 0.447–0.851, p = 0.003) and PTGER3 (HR = 0.713, 95% CI 0.518–0.982, p = 0.038) exhibited significant protective effects (Fig. 6A). After adjusting for TNM stage and molecular subtype in the multivariate analysis, GSTK1 remained an independent prognostic factor (HR = 0.657, 95% CI 0.473–0.913, p = 0.012), while the prognostic association of PTGER3 did not reach statistical significance (hazard ratio = 0.785, 95% confidence interval = 0.565–1.092, p = 0.150) (Fig. 6B). Cox Regression Analysis and Clinical Significance of Key Genes. A. Univariate Cox regression analysis shows the hazard ratio (HR) and 95% confidence interval (CI) of the PTGER3 and GSTK1 genes. B. Multivariate Cox regression analysis shows the hazard ratio (HR) and 95% confidence interval (CI) of the PTGER3 and GSTK1 genes Differential expression of GSTK1 and PTGER3 in breast cancer To investigate the expression of GSTK1 and PTGER3 in breast cancer, boxplots and paired plots were used to analyze their expression levels in normal (blue) and tumor (red) tissues. Boxplots showed that the expression of both GSTK1 and PTGER3 was significantly lower in tumor tissues compared to normal tissues, with highly significant differences (Fig. 7A). Paired plots further confirmed that the expression of GSTK1 and PTGER3 was generally lower in most tumor tissues compared to normal tissues, reinforcing the significant downregulation of these genes (Fig. 7B). Differential Expression of GSTK1 and PTGER3 in Breast Cancer. A. Box plot showing that the expression of GSTK1 and PTGER3 is significantly lower in breast cancer tissues (red) compared to normal tissues (blue), with both differences being highly statistically significant (P < 0.001). B. Paired sample analysis comparing the differential expression of genes in tumor tissues and their matched normal tissues. In the figure, “ns” indicates no statistical significance (P ≥ 0.05), while P < 0.05, P < 0.01, and P < 0.001 are marked with *, **, and ***, respectively Clinical value of key genes in breast cancer To assess the clinical relevance of GSTK1 and PTGER3, ROC curves were constructed to evaluate their diagnostic efficacy. The AUC for GSTK1 was 0.641, while the AUC for PTGER3 was 0.788 (Fig. 8A). Boxplot analysis revealed that GSTK1 expression was lower in Stage I-III tumors compared to normal tissues, while PTGER3 expression was lower in Stage I-IV tumors compared to normal tissues (Fig. 8B and C). Survival analysis showed that high expression of GSTK1 was associated with a better prognosis (HR = 0.62, 95% CI 0.45–0.85, P = 0.003), and high expression of PTGER3 was also associated with better survival (HR = 0.71, 95% CI 0.52–0.98, P = 0.038) (Fig. 8D and E). To predict the survival probabilities of breast cancer patients more intuitively and conveniently, nomograms were constructed for overall survival (OS, Fig. 8F), disease-specific survival (DSS, Fig. 8G), and progression-free survival (PFI, Fig. 8H) based on pathological T stage, N stage, and the expression levels of GSTK1 and PTGER3. The scores of each factor were accumulated to obtain a total score, which was then used to predict the 1–5 year survival probability. When the pathological stage is advanced and the gene expression levels are unfavorable, a higher total score and larger linear predictor value may be associated with a lower survival probability. Clinical Value Assessment of Key Genes. A. ROC curves demonstrating the diagnostic efficacy of GSTK1 (AUC = 0.641) and PTGER3 (AUC = 0.788) for breast cancer. B, C. Box plots analyzing the relationship between gene expression and pathological stage. D, E. Kaplan-Meier survival curves: D. GSTK1 expression and its association with overall survival (OS); E. PTGER3 expression and its association with overall survival (OS). F–H. Nomogram integrating pathological T/N stage and gene expression levels to predict patient survival outcomes: F. 1–5 years overall survival (OS); G. Disease-specific survival (DSS); H. Progression-free survival (PFI) To comprehensively evaluate the performance of the GSTK1- and PTGER3-based prognostic models, a nomogram was constructed for analysis. Results demonstrated that PTGER3 scores exhibited a relatively concentrated distribution, whereas GSTK1 scores showed significant inter-individual variability. Meanwhile, tumor T stage and N stage were positively correlated with the total score, with higher pathological stages corresponding to elevated scores. Predictions of event occurrence probabilities at 1-year, 2-year, and 3-year follow-ups based on the total score revealed a gradual upward trend over time, which intuitively reflected the impact of time on prognostic risk. This indicated that the integrated model could achieve accurate quantitative prediction of risks at different time points (SI-Figure 1 A). Further validation via ROC curve analysis confirmed the excellent predictive efficacy of the model, with an AUC value of 0.742 for 2-year outcome prediction, indicating particularly prominent discriminative ability (SI-Figure 1B). Combined with decision curve analysis and survival probability calibration assessment, the predicted survival probabilities of the model showed high consistency with the actual survival rates, especially at the 3-year time point (SI-Figure 1 C-D). Immune infiltration correlation To investigate the relationship between GSTK1, PTGER3, and immune infiltration in breast cancer, CIBERSORT and ssGSEA algorithms were utilized. CIBERSORT analysis revealed that the GSTK1 gene is associated with the infiltration levels of various immune cells. Specifically, GSTK1 expression was positively correlated with resting mast cells (R = 0.128, P < 0.05), naive B cells (R = 0.105, P < 0.05) and activated NK cells (R = 0.104, P < 0.05). In contrast, GSTK1 showed a negative correlation with memory CD4 + T cells and macrophages. However, the correlation between GSTK1 and immune cells such as monocytes, memory B cells, and activated mast cells did not reach statistical significance (denoted as ns) (Fig. 9A). Immune Infiltration Correlation Analysis. A. CIBERSORT analysis showing the correlation between GSTK1 and immune cell infiltration. B. ssGSEA analysis revealing the correlation between GSTK1 and various immune cells. C. CIBERSORT analysis demonstrating the correlation between PTGER3 and immune cell infiltration. D. ssGSEA analysis indicating the correlation between PTGER3 and immune cells. In the figure, “ns” indicates no statistical significance (P ≥ 0.05), while P < 0.05, P < 0.01, and P < 0.001 are marked with *, **, and ***, respectively CIBERSORT analysis also indicated that the PTGER3 gene is associated with the infiltration levels of several immune cells. Specifically, PTGER3 exhibited a significant positive correlation with resting mast cells (R = 0.372, P < 0.05). Furthermore, PTGER3 was positively correlated with resting memory CD4 + T cells (R = 0.306, P < 0.05), naive B cells (R = 0.198, P < 0.05), and resting dendritic cells (R = 0.154, P < 0.05), among others. On the other hand, PTGER3 showed a negative correlation with M0 macrophages (R = -0.186, P < 0.05). Additionally, negative correlations were observed between PTGER3 and activated memory CD4 + T cells (R = -0.203, P < 0.05), follicular helper T cells (R = -0.167, P < 0.05), and regulatory T cells (R = -0.161, P < 0.05). It is noteworthy that the correlation between PTGER3 and certain immune cells did not reach statistical significance (denoted as ns), suggesting that these associations were not statistically significant in this analysis (Fig. 9C). ssGSEA analysis further demonstrated that the GSTK1 gene is correlated with the infiltration levels of various immune cells. Specifically, GSTK1 showed a significant positive correlation with NK cells (R = 0.221, P < 0.05). Additionally, GSTK1 was positively correlated with plasmacytoid dendritic cells (pDC, R = 0.213, P < 0.05), Th17 cells (R = 0.196, P < 0.05), and NK CD56bright cells (R = 0.169, P < 0.05). Conversely, GSTK1 exhibited a negative correlation with Th2 cells (R = -0.318, P < 0.05). Moreover, negative correlations were observed between GSTK1 and macrophages (R = -0.176, P < 0.05) as well as Th1 cells (R = -0.127, P < 0.05). It is important to note that the correlation between GSTK1 and certain immune cells did not reach statistical significance (denoted as ns), indicating that these associations were not statistically significant in this analysis (Fig. 9B). Similarly, ssGSEA analysis revealed that the PTGER3 gene is associated with the infiltration levels of various immune cells. Specifically, PTGER3 showed a significant positive correlation with mast cells (R = 0.446, P < 0.05). Additionally, PTGER3 was positively correlated with NK cells (R = 0.349, P < 0.05), eosinophils (R = 0.316, P < 0.05), and immature dendritic cells (iDC, R = 0.235, P < 0.05). Conversely, PTGER3 exhibited a negative correlation with Th2 cells (R = -0.238, P < 0.05), as well as with activated dendritic cells (aDC, R = -0.231, P < 0.05), Th1 cells (R = -0.231, P < 0.05), and regulatory T cells (TReg, R = -0.197, P < 0.05) (Fig. 9D). Effects of regulating GSTK1 and PTGER3 expression on cell proliferation in breast cancer models To investigate the roles of GSTK1 and PTGER3 in breast cancer cell proliferation, we manipulated their expression in MCF-7 and MDA-MB-231 cells via silencing (si-RNAs) and overexpression constructs, followed by qRT-PCR to verify gene expression and CCK-8 assays to assess cell growth. qRT-PCR results showed that si-GSTK1 significantly downregulated GSTK1 mRNA levels, while oe-GSTK1 markedly upregulated them in both cell lines (P < 0.001, Figs. 10A-B); similarly, si-PTGER3 dramatically reduced PTGER3 mRNA expression, and oe-PTGER3 substantially increased it in MCF-7 (P < 0.001, Figs. 10C) and MDA-MB-231 (P < 0.001, Figs. 10D) cells. Correspondingly, CCK-8 assays revealed that silencing GSTK1 or PTGER3 significantly promoted cell growth (P < 0.001), whereas their overexpression remarkably inhibited proliferation in both cell lines (Figs. 10E- H). Effects of Regulating GSTK1 and PTGER3 Expression on Cell Proliferation in Breast Cancer Models. A, B. Relative mRNA expression levels of GSTK1 in MCF-7 and MDA-MB-231 cells following its knockdown (si-GSTK) or overexpression (oe-GSTK1). C, D. Relative mRNA expression levels of PTGER3 in MCF-7 and MDA-MB-231 cells after knockdown (si-PTGER3) or overexpression (oe-PTGER3). E, F. Cell proliferation ability assessed by CCK-8 assay (OD450) in MCF-7 and MDA-MB-231 cells after modulation of GSTK1 expression. G, H. Cell proliferation ability measured by CCK-8 assay (OD450) in MCF-7 and MDA-MB-231 cells after modulation of PTGER3 expression. In the figure, “ns” indicates no statistical significance (P ≥ 0.05), while P < 0.05, P < 0.01, and P < 0.001 are marked with *, **, and ***, respectively Discussion The precise prognostic assessment and formulation of individualized treatment strategies for breast cancer heavily depend on the identification of specific biomarkers. In this study, we integrated transcriptomic data with machine learning algorithms to systematically analyze the prognostic value of GSTK1 and PTGER3 in the adipose microenvironment of breast cancer, as well as their regulatory roles in the immune microenvironment. This provides new scientific insights into the multifaceted roles of adipocytes in tumor progression. Adipose tissue, as a critical component of the breast cancer microenvironment, has been shown to be closely associated with tumor initiation, progression, invasion, metastasis, treatment response, and drug resistance [3]. In this study, we identified that differentially expressed genes in the tumor-adjacent and distant adipose tissue were significantly enriched in pathways related to angiogenesis regulation, granulocyte differentiation, and Th17 cell differentiation. This suggests that the adipose microenvironment may provide structural support for tumor growth by remodeling extracellular matrix components and, through the activation of the Th17 cell differentiation pathway, induce an immunosuppressive microenvironment. These findings indicate that the adipose microenvironment may shape a pro-cancer environment through multiple mechanisms. A key innovation of this study lies in the use of 21 machine learning algorithms for cross-validation, which allowed for the first identification of independent prognostic genes originating from the adipose microenvironment, specifically GSTK1 and PTGER3. Unlike single algorithm approaches, the intersection of genes identified through RidgeCV and XGBoost significantly improved the stability of key gene selection. Biologically, GSTK1, a core member of the glutathione S-transferase (GST) superfamily [16], exhibits a high binding affinity for glutathione (GSH). It exerts a critical role in scavenging oxidative stress by-products, such as lipid peroxides, thereby contributing to the attenuation of cancer susceptibility. Notably, adipose tissue serves as the predominant source of GSH within the tumor microenvironment (TME) [17]. Accumulating evidence has demonstrated that adipocyte-derived GSH promotes tumor progression through targeted regulation of the SCARB2-ARF1-mTORC1 signaling axis [18]. Furthermore, elevated GSTK1 expression has been shown to modulate lactate metabolism and oxidative stress homeostasis via the interaction between GPR81 and IGFBP6, which in turn mediates the suppression of breast cancer progression [8]. PTGER3, an important receptor for prostaglandin E2 (PGE2), plays a crucial role in regulating various physiological activities. PGE2, a bioactive molecule derived from polyunsaturated fatty acids, is involved in numerous physiological and pathological processes. When PTGER3 is normally expressed, PGE2 binds to it and activates downstream signaling pathways, such as G-protein coupling and modulation of second messenger levels, which influence various cellular functions, including digestion, nervous system activity, renal reabsorption, and uterine contraction. However, in cancers such as colorectal cancer, PTGER3 expression is downregulated [19]. This change may prevent PGE2 from binding to the receptor and activating downstream signals, disrupting the PGE2-cAMP signaling axis. Disruption of this pathway can lead to a series of downstream effects, including altered immune checkpoint molecule regulation, such as PD-L1 [20]. PD-L1, normally expressed on tumor cells and tumor-associated immune cells, inhibits T cell activation by binding to the PD-1 receptor on T cells, preventing immune overactivation. However, in the tumor microenvironment, aberrant PD-L1 expression enables tumor cells to evade immune surveillance [21, 22]. The downregulation of PTGER3 may disrupt the PGE2-cAMP axis, thereby increasing PD-L1 expression through mechanisms like altered transcription factor activity or epigenetic modifications. Excessive PD-L1 binding to PD-1 on T cells suppresses their activity, impairing their ability to recognize and kill tumor cells, thus promoting immune escape and facilitating tumor growth, proliferation, and metastasis. These findings not only expand our understanding of the molecular network governing adipose microenvironment regulation but also provide potential targets for therapies that modulate the metabolism-immune axis. From a clinical translation perspective, the nomogram developed in this study integrates TNM staging and gene expression data, significantly improving survival prediction accuracy for breast cancer patients and offering a quantitative tool for personalized treatment decisions. Notably, through siRNA-mediated gene knockdown experiments, the present study confirmed that overexpression GSTK1 or PTGER3 could significantly suppress the proliferative activity of breast cancer cells, thereby providing direct experimental evidence for the potential of GSTK1 and PTGER3 as predictive markers for therapeutic response in breast cancer. From the perspective of therapeutic targets, this study demonstrated that the expression of GSTK1 and PTGER3 was downregulated in breast cancer tissue. In recent years, it has been well established that microenvironment-regulating genes such as these represent promising therapeutic targets for breast cancer, as specific inhibition of their expression or blockade of their downstream signaling pathways may improve treatment efficacy. Several studies have indicated that high GSTK1 expression is associated with longer survival in patients with LumB breast cancer, likely due to its role in reducing reactive oxygen species and alleviating oxidative stress [23]. Moreover, PTGER3 has been proposed as a novel and promising biomarker and therapeutic target for triple-negative breast cancer (TNBC), since its low expression protects TNBC cells from ferroptosis and thereby promotes tumor progression [24]. Our results are consistent with previous studies and provide further evidence supporting the targeting of GSTK1 and PTGER3 for breast cancer treatment. Notably, the specific roles of GSTK1 and PTGER3 in immune infiltration regulation open new avenues for immunotherapy. The positive correlation between GSTK1 and NK cell infiltration suggests it may enhance innate immune surveillance to inhibit tumor metastasis, while the association between PTGER3 and mast cells and immature dendritic cells suggests its potential involvement in regulating antigen-presenting cell function, thereby influencing adaptive immune responses. These adipocyte-immune cell interactions indicate that targeting GSTK1 or PTGER3 could reshape immune cell recruitment and polarization in the tumor microenvironment, reversing the immunosuppressive state and enhancing the efficacy of immune checkpoint inhibitors. Nevertheless, this study has several limitations. The screening of differential genes was based on a single GEO dataset (GSE153316), and the generalizability of its results needs to be validated in multicenter cohorts such as TCGA-BRCA. In addition, although the biological functions of the key genes GSTK1 and PTGER3 were confirmed by in vitro experiments, this study did not fully explore the spatial heterogeneity of the adipose-tumor interface. Future studies should employ spatial transcriptomics to reveal the zonal distribution of gene expression and the spatiotemporal specificity of cell-cell interactions, thereby depicting the regulatory network of the adipose microenvironment in breast cancer more precisely. Conclusions This study conducted a comprehensive bioinformatics analysis combined with in vitro experiments, confirming the significant prognostic value of GSTK1 and PTGER3, key genes associated with the adipose microenvironment in breast cancer, as well as their involvement in immune regulation. These findings enhance our understanding of the multifaceted roles of adipocytes in tumor progression and provide a scientific basis for targeting the adipose–metabolism–immune axis in precision therapy. Future work should prioritize functional validation of these genes, multi-cohort clinical verification, and spatial heterogeneity investigations to facilitate the translation of these insights into clinical practice, thereby advancing personalized diagnosis and treatment strategies for breast cancer. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Acknowledgements Not applicable. Author contributions LT: Data curation, Validation, Writing–original draft, Writing–review & editing. LX: Data curation, Validation, Writing–original draft, Writing–review & editing. YL: Data curation, Validation, Writing–original draft, Writing–review & editing. XZ: Data curation, Funding acquisition, Resources, Supervision, Validation, Writing–original draft, Writing–review & editing. Funding Not applicable. Data availability The GEO public dataset used in this study is available in the data repository under the accession number GSE153316 (Microarray Data). Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Data Availability Statement The GEO public dataset used in this study is available in the data repository under the accession number GSE153316 (Microarray Data).