Development and validation of an immune-related molecular subtyping model based on a four-gene prognostic biomarker signature and multi-omics integrative analysis in Luminal B breast cancer

Abstract

Background

Luminal B breast cancer (LBBC) is characterized by marked molecular heterogeneity and unfavorable therapeutic outcomes, underscoring the need for concise and biologically interpretable biomarkers that can support reliable prognostic stratification. However, existing biomarker signatures are often redundant, weakly generalizable, and insufficiently validated in multi-omics regulatory consistency.

Methods

Using a minimum prognostic-redundancy feature identification strategy, we identified an immune-related four-gene prognostic biomarker signature and determined its optimal classification scheme. A molecular subtyping model was subsequently developed and externally validated to distinguish two prognostic immune subtypes: the immune-barren type (IBT) and the immune-enriched type (IET). Functional enrichment analysis, tumor immune microenvironment profiling, and multi-omics integrative consistency assessments were performed to elucidate the biological characteristics and regulatory underpinnings of the predicted subtypes. Moreover, the clinical independence, prognostic robustness, and cross-cohort reproducibility of the model were comprehensively evaluated, and independent clinical variables were incorporated to construct an individualized decision-support tool.

Results

The four-gene biomarker signature demonstrated strong prognostic discrimination and subtype classification capability (max AUC = 0.86, P < 0.05). Across extensive multi-source external cohorts (n = 1,841), the molecular subtyping model robustly stratified patients into immune-related subtypes with significantly different survival outcomes and consistently high predictive performance (ACC ≥ 0.93, AUC ≥ 0.97, all P < 0.001), while remaining independent of conventional clinical variables (P < 0.001). Furthermore, the predicted subtypes exhibited concordant immune functional characteristics and multi-omics regulatory consistency (all P < 0.05). In addition, the individualized decision-support tool showed significant prognostic relevance (P = 0.004) and favorable comparative performance relative to previously reported prognostic models (AUC = 0.86).

Conclusion

This immune-related molecular subtyping model, derived from a four-gene prognostic biomarker signature and supported by multi-omics integrative evidence, provides a robust and clinically independent model for prognostic stratification in LBBC. These findings offer a biologically interpretable basis for individualized risk assessment and may facilitate precision-oriented clinical decision support.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13058-026-02342-4.

Introduction

Breast cancer is one of the most common malignant tumors in women globally [1], and Luminal B breast cancer (LBBC) that accounts for approximately 20–40% [2–4] is a subtype of hormone receptor-positive (HR+) breast cancer with a relatively poor prognosis [5]. Characterized by high tumor heterogeneity and complex biological features, classic immunohistochemical (IHC) diagnostic methods struggle to accurately reflect the molecular characteristics and prognostic differences of the tumor [6], leading to suboptimal treatment outcomes for LBBC. This subtype of breast cancer responds relatively poorly to endocrine therapy and is more prone to recurrence and metastasis [7], indicating a greater complexity in the diagnosis and treatment of LBBC.

Clinical trials have shown that even after comprehensive treatment, some patients with LBBC still have a poor prognosis [8], but they may benefit from immunotherapy [9]. In recent years, immunotherapy has been proven to be a novel and effective treatment for malignant tumors [9]. However, only a small proportion of breast cancer patients respond to immunotherapy [9]. Studies have indicated that HR+ breast cancer patients may exhibit a trend of reduced immunosuppressive cells after treatment with endocrine therapy combined with immune checkpoint inhibitors (ICIs) [10], suggesting that HR+ breast cancer patients could benefit from immunotherapy. Compared to Luminal A breast cancer, LBBC has immunogenic characteristics (higher tumor-infiltrating lymphocytes, higher tumor mutation burden, and immune checkpoint molecule expression), which could increase sensitivity to immunotherapy [11–14]. Moreover, anthracyclines, as effective inducers of immunogenic cell death, have garnered clinical attention because they can transform the tumor immune microenvironment into a tumor state, thereby attenuating the activity of ICIs [15, 16]. Multiple studies have shown that estrogen promotes tumor progression through various effects on immune cells, while aromatase inhibitors or ovarian function inhibitors promote the formation of an anti-tumor immune microenvironment, which is the biological rationale for combining immunotherapy with endocrine therapy [17–19]. Maria et al. [17] found that 16.3% patients achieved pathological complete response (pCR) in a study of 43 premenopausal LBBC patients treated with neoadjuvant chemotherapy combined with immunotherapy, with a tumor residual burden rate of 25.6%, significantly higher pCR rates compared to Luminal A subtype. As an immune-related gene expression indicator, specific tumor-infiltrating lymphocytes (TILs) subsets are closely related to pCR. Therefore, LBBC with basal cell molecular characteristics and immune activity shows significant effects with anthracyclines and anti-PD-1 therapy [17]. Increasing evidence suggests that the immune system plays an important role in controlling the occurrence and progression of breast cancer. However, there is currently a lack of sufficient basic research and clinical immune classification diagnostic models to evaluate the indications and benefits of immunotherapy for LBBC.

Currently, machine learning demonstrates immense potential in the fields of genomic analysis and disease diagnosis [20]. In the process of breast cancer diagnosis and prognosis assessment, machine learning algorithms are capable of extracting complex features from vast amounts of medical genomic and clinical data, aiding physicians in making more accurate diagnoses [20]. Guo et al. [21] integrated genomic and imaging data of invasive breast cancer from the TCGA-ESCA database and selected the radio-genomic features with the best predictive performance through Lasso regression and Logistic regression analysis, thereby improving diagnostic efficacy in pathological staging, ER and PR status prediction. Jiang et al. [22] discovered that HR+/HER-2- subtype breast cancer patients could be subdivided into four subtypes: canonical luminal, immunogenic, proliferative, and receptor tyrosine kinase (RTK)-driven. The proliferative subtype, mainly LBBC patients, showed significant gene expression heterogeneity. However, this study was not specifically detailed to reveal the mechanisms associated with the strong heterogeneity of treatment and prognostic outcomes in LBBC patients. Zhao et al. [23] reported that the prognosis of patients with LBBC was closely associated with tumor immune activity, as well as alterations in cell-cycle regulation and metabolic signaling pathways, based on multi-platform integrative analyses of TCGA data. These findings highlight the critical role of immune-related biological processes in shaping disease progression and clinical outcomes in LBBC. However, due to the lack of a concise and interpretable classification diagnostic model, this classification research with significant prognostic implications has not been effectively applied in clinical practice.

To address these challenges, we integrated large-scale multi-omics analyses and external validation to construct and validate an independent immune-related molecular model for prognostic subtyping based on an identified set of four optimal cancer biomarkers, together with a clinically oriented decision-support tool. We further explored and validated the immune functional features and multi-omics regulatory consistency of the predicted LBBC subtypes, thereby improving prognostic assessment and providing auxiliary support for precision and individualized treatment.

Materials and methods

Patient data collection of LBBC

All the datasets enrolled into this study was from the Cancer Genome Atlas (TCGA), cBioPortal for Cancer Genomics, Array Express, Gene Expression Omnibus (GEO), and real-world data (RWD) from China. Specifically, a total of 1,841 patients were included in this study, and TCGA (n = 187) was used as the training and testing set, while METABRIC/Nature 2012 (n = 263), E-MTAB-6703 (n = 71), GSE96058 (n = 656), GSE54275 (n = 86), GSE20685 (n = 75), GSE2109 (n = 105), and China (n = 398) as external independent validation sets. The TCGA dataset for LBBC patients included multiple omics data types: whole-exome DNA sequences (n = 187), DNA methylation array (n = 129), mRNA sequences (n = 187), miRNA sequences (n = 181), lncRNA sequences (n = 187), and reverse phase protein array (RPPA, n = 162). The single-cell data were obtained from GSE158399, GSE188807, GSE173634, and Zenodo.13,743,374.

Statistical analysis

The statistical methods in this study were performed in the R statistical environment (v4.3.1). Shapiro-Wilk test was used to evaluate the normal distribution before DEGs analysis. For comparison of continuous variables, clinical factor frequency, and survival status between two LBBC subtypes, Wilcoxon test (non-normal distribution data), Fisher’s Exact test, log-rank tests, or Renyi test (cross-survival curve) were utilized to determine the significance of the statistical difference. DESeq2 was used for differential analysis of genomic expression profiles (mRNA, miRNA, and lncRNA), with adjustment of False Discovery Rate (FDR) using the Benjamini-Hochberg algorithm (|log2FoldChange| > 1, padj < 0.05) [24]. Univariate and multivariate Cox proportional hazard regression models and logistic regression models were used to assess the clinical independence of prognostic and classification factors, respectively. Spearman correlation coefficient was calculated to assess the significance of the correlation between two continuous variables. All p-values or adjusted p-values (if applicable) < 0.05 were considered statistically significant.

Identification of the optimal prognostic diagnostic gene set

To solve the issue of insufficient generalization ability of prognostic factors in high-dimensional GEP data, we proposed a minimum prognostic redundancy feature (MPRF) identification method to accurately identify the best diagnostic signature of LBBC, which mainly includes the following three steps:

Step 1: calculate the mutual information matrix (MIM) of features from cancer marker, and compute the mutual information (MI) based on the similarity between feature variables, where I and P respectively represent the MI and correlation coefficient between feature variables, and and respectively represent the two feature variables to calculate mutual information.

1

Step 2: evaluate the redundant information of features from cancer marker, and quantify the redundant information of each marker by calculating the average mutual information value of each feature variable in the feature set S.

2

Step 3: evaluate the prognostic effectiveness of features from cancer marker, calculate the prognostic diagnostic power of each feature through survival analysis, where represents the significance of survival analysis of feature variables, and  sign is the  signum function.

3

Step 4: calculate the prognostic redundancy score of the biomarker features. Based on the redundancy information and prognostic effectiveness, calculate the prognostic importance of each feature variable and rank it.

4

Diagnostic performance evaluation of the optimal four-gene set

First, Principal Component Analysis (PCA) was utilized to analyze the overall distribution of gene expression profiles to verify the ability of the optimal four-gene set to classify and diagnose LBBC subtypes. Then, the area under ROC curve (AUC) and accuracy rate were used to quantify the classification diagnosis ability of each gene for LBBC subtype.

Construction and validation of LBBC subtypes diagnostic classification model and prognostic prediction tool

In this study, we evaluate the inter-group survival difference and diagnostic error rate of the LBBC subtypes classification diagnostic model constructed by 11 machine learning algorithms (Classification and Regression Tree, Conditional Inference Tree, C5.0 Algorithm, Decision Tree, Iterative Dichotomiser 3, K-Nearest Neighbor, Naïve Bayes, Linear Discriminant Analysis, Random Forest, Logistic Regression, and Support Vector Machine) through 5-fold cross-validation comparison [25–33]. Then, we compared the minimum error rate and the maximum AUC of each gene, and then developed a diagnostic model based on the optimal gene set. Finally, a total of 7 external independent datasets were used to validate biological features consistent with the optimal model, including AUC, accuracy, and molecular expression differences of the diagnostic model. In addition, the R package rms [34] was used to construct a new clinical diagnostic tool for prognostic prediction based on this optimal model and was compared with 6 recently published nomograms. The cutoff value was determined using the maximum sensitivity and specificity points on the predicted ROC curve of the four-gene model, as calculated by the Youden index.

DNA copy number and mutation feature analysis

Firstly, DNA copy number alterations were identified using the GISTIC_2.0 (Genomic Identification of Significant Targets in Cancer) analysis method [35]. Then, copy number gains (log2 copy ratio > 0.3) or losses (log2 copy ratio < − 0.3) were visualized using the ComplexHeatmap R package [36]. Finally, the Wilcoxon test was used to analyze the difference in copy number changes of each key gene between the two molecular subtypes.

DNA methylation analysis

First, the ChAMP (The Chip Analysis Methylation Pipeline) R package was applied to identify 45,894 significant cancer methylation sites among 192,023 DNA methylation sites (P-value < 0.05) [37]. Then, a total of 8,727 significantly differentially expressed methylation sites were identified between IBT (n = 65) and IET (n = 64) subtypes (P-value < 0.05). Finally, the significant differential methylation sites corresponding to the optimal feature set were identified and their regulatory expression mechanisms were studied.

Analysis of reverse phase protein array (RPPA)

A total of 144 cancer-related differential proteins (P-value < 0.05) were identified between tumor (n = 162) and normal (n = 22) patients using the limma R package [38]. Then, 29 significantly different proteins were identified between the predicted immune-barren type (PIBT, n = 84) and predicted immune-enriched type (PIET, n = 78). Similarly, significantly different proteins associated with the optimal feature set identified in previous analyses were selected as potential therapeutic targets for further investigation. In addition, the expression levels and domain mutation status of key proteins were investigated using data from cBioPortal [39].

Functional enrichment analysis

Firstly, a gene set enrichment analysis (GSEA) was performed for 443 genes with significant differences in high expression associated with four-gene model between IBT and IET molecular subtypes using the Molecular Signatures Database (MSigDB, FDR q-value < 0.05 and minimum gene set size ≥ 3) [40]. Then, the significant pathway enrichment and key gene coverage status were visualized using the R packages ggplot2 and fgsea [41, 42].

Construction of competitive endogenous RNA (ceRNAs) network

To construct the ceRNA network (lncRNAs: 5, miRNA: 1, mRNAs: 48, pathway: 11), we first used difference analysis and Spearman correlation analysis algorithm to identify key candidate genes that were significantly associated with the four-gene model between IBT and IET subtypes in LBBC patients. Then, some key databases (starBase, LncRNA2Traget v2.0, LncTarD, and LncBase_Predicted_v2) were used to predict the interactions between lncRNA-mRNA, lncRNA-miRNA and miRNA-mRNA [43–46]. Finally, the ceRNA network was visualized using Cytoscape software 3.10.1 [47], which contains key lncRNAs, miRNA, mRNAs, and enrichment pathways.

Results

Identification of the optimal prognostic biomarker set and classification method for LBBC subtypes

To identify the optimal biomarkers set capable of accurately distinguishing the prognostic molecular subtypes of LBBC, we evaluated the genes that were significantly differentially expressed between the IBT and IET subtypes (P-value < 0.001, Figure S1A) based on the proposed MPRF method (Figure S1B). An optimal four-gene biomarker set (CACNA2D2, CCDC24, CORIN, and MMP1) was ultimately determined, which exhibited low collinearity redundancy and high prognostic representativeness in 187 LBBC patients from TCGA (0 < Score ≤ 0.4, Fig. 1A, B).

Fig. 1.

Identification of the optimal prognostic biomarker set and classification method. A Feature selection based on minimal prognostic redundancy feature algorithm. Node size represents the minimum prognostic redundancy score of each gene biomarker indicates candidate cancer biomarkers. B The Venn diagram showed that the four overlap gene biomarkers were identified between IBT and IET subtypes by the minimal prognostic redundancy feature screening algorithm based on the minimal redundancy (MR) and survival analysis algorithm. C Classification ability of multiple machine learning methods for the four-gene signature and 20-gene signature evaluated by 5-fold cross-validation. The points represent the significance of the survival difference in the predicted outcomes. D Compare the classification diagnostic power of all gene combinations in a training set, a testing set, and a validation dataset. The horizontal coordinates represent the number of genes in different feature biomarker sets. The left vertical axis referred to classification minimum error rate (histogram), and the right vertical axis referred to prediction outcome maximum AUC (line). Colored cells indicate the optimal gene set for each gene number combination

Meanwhile, to determine the optimal classification method, we compared the discriminant error rate (5-fold cross-validation) and prognostic significance across 11 machine-learning algorithms using both the four-gene and twenty-gene signatures. Support vector machine (SVM) achieved the lowest classification error and exhibited significant diagnostic power in distinguishing the two LBBC subtypes (P-value = 0.032, HR: 2.247, Fig. 1C).

Furthermore, to determine whether the four-gene set represented the minimal effective combination, we evaluated its predictive performance relative to other gene combinations. Across all datasets, the four-gene signature consistently achieved the lowest classification error rate (ER = 0.07) and stronger predictive power (AUC = 0.97) compared with alternative combinations (Fig. 1D, Table S1), thereby supporting its identification as the minimal effective biomarker set for LBBC subtype stratification.

Validation of the discriminative capability of the optimal biomarker set

To validate the discriminative capability of the four-gene set, we first performed principal component analysis (PCA) in 187 LBBC patients using the integrative representative ability of the biomarker set. The results showed that this set effectively separated the samples into two distinct molecular subtypes (Fig. 2A). Subsequently, median expression profiling and receiver operating characteristic (ROC) analysis were conducted to evaluate the independent classification ability of each biomarker. All four biomarkers exhibited high mRNA expression abundance (Fig. 2B) and strong predictive ability (AUC ≥ 0.80, Fig. 2C), indicating that each biomarker could independently predict the molecular subtypes with high sensitivity and specificity.

Fig. 2.

Validation of the discriminative capability of the optimal biomarker set. A PCA showing that the optimal four-gene biomarker separates patients into IBT and IET subtypes. B Distribution of genes expression abundance across samples. C Discriminative ability of individual biomarkers for subtype identification. D Classification contribution of each biomarker

Meanwhile, to more precisely evaluate the classification capability of the biomarkers, the performance of each gene was quantified using accuracy rate and mean expression value. The results showed that CORIN and MMP1 were predominantly distributed in patients with the IBT subtype, with high accuracy rates (HAR ≥ 0.83) and high mean expression values (MEV > 2.5), whereas CCDC24 and CACNA2D2 were mainly distributed in the IET subgroup (HAR = 0.89, MEV > 6.0, Fig. 2D).

In addition, survival analysis and multi-omics integrative analysis were performed to verify the prognostic and multi-omics classification capability of each biomarker. Each gene significantly stratified patients into two subgroups with distinct clinical outcomes (P-value ≤ 0.042, Figure S2A-D) and revealed consistent subtype-associated multi-omics differences, including somatic mutations, copy-number variation (CNV), DNA methylation, lncRNAs, miRNAs, mRNAs, proteins, and pathways (P-value < 0.05, Figure S2E). Taken together, these findings indicate that the four-gene biomarker set confers strong diagnostic capability in assigning cases to the IBT or IET subtype.

Construction and validation of the optimal classifier for two molecular subtypes of LBBC

To develop a generalized classifier capable of accurately predicting LBBC molecular subtypes in complex multi-source heterogeneous data, we first applied the interpretable four-gene signature as prior knowledge to guide the optimization of the classification SVM kernel for constructing a four-gene classifier. The results indicated that the polynomial kernel achieved the highest diagnostic accuracy by effectively capturing and integrating subtype-associated features through mapping the four-gene signature into a unified high-dimensional space (Table S2).

Meanwhile, all results consistently showed that CCDC24 and CACNA2D2 were highly expressed in the PIBT subtype, whereas CORIN and MMP1 were highly expressed in the PIET subtype (P-value < 0.001, Fig. 3A–I, top), across the testing set and seven independent external validation cohorts (n = 1710, Table S3). Moreover, the classifier demonstrated strong and stable diagnostic performance across all datasets (AUC ≥ 0.97, ACC ≥ 0.93, P-value < 0.001, Fig. 3A–I, middle and bottom), indicating excellent generalization capability in complex multi-source heterogeneous data. Taken together, these findings suggest that the four-gene classifier represents a reliable predictive model with high sensitivity and specificity for identifying LBBC molecular subtypes with significant prognostic relevance in clinical practice.

Fig. 3.

Construction and validation of the optimal classifier for two molecular subtypes of LBBC. A−I Develop the new classified model with the optimal four-gene signature and validate it in the testing set and another 7 independent datasets. In the top panel, the optimal four gene biomarkers showed differential expression profiles between PIBT and PIET subtype. In the middle panel, the visualized predictions of the optimal four-gene signature model, where the decision boundary thresholds were defined using the support vector machine algorithm. In the bottom panel, the prediction performance of the optimal four-gene signature model was demonstrated through the area under of ROC curve (AUC) and the confusion matrix (accuracy)

Validation of functional relevance and multi-omics consistency of the optimal four-gene model

To validate the biological functional and expression consistency of the subtypes predicted by the optimal four-gene model, we first performed integrative analyses including functional enrichment, immune infiltration, and single-cell data analysis based on genes significantly correlated with the model (|r| > 0.3, P-value < 0.001). These results demonstrated a high degree of consistency across functional pathways, the immune microenvironment, and single-cell expression patterns, thereby supporting the robustness and biological interpretability of the four-gene model–defined subtypes.

Specifically, the PIBT subtype was predominantly enriched in endocrine regulation pathways, particularly early and late estrogen response (Figs. 4A and S3B). Consistently, immune infiltration analysis revealed a relatively higher abundance of plasma cells (P-value < 0.05, Figure S4A). Furthermore, single-cell data analysis showed that CCDC24 and CACNA2D2 were significantly enriched in B cells (Fig. 4B, C), which was concordant with increased plasma cell infiltration and estrogen response–related functional enrichment observed in the PIBT subtype.

Fig. 4.

Validation of functional relevance and multi-omics consistency of the optimal four-gene model. A Pathway enrichment of biomarker-associated differential genes across subtypes. B t-SNE plot of 55,677 filtered single cells from 11 Luminal B breast samples. C Density of biomarkers. D Multi-omics specific differences in methylation, transcriptomic expression, and copy number alterations of the biomarker between subtypes. E Distribution of subtype-specific somatic mutations along the biomarker protein sequence

In contrast, the PIET subtype was predominantly enriched in oncogenic signaling and cancer immune–related pathways, including epithelial–mesenchymal transition, cytokine–cytokine receptor interaction, and IL6–JAK–STAT3 signaling (Figs. 4A and S3B). Consistently, immune infiltration analysis revealed increased infiltration of resting dendritic cells, M1 macrophages, fibroblasts, T cells, and monocyte lineage cells (P-value < 0.05, Figure S4A). Furthermore, single-cell data analysis showed that MMP1 and CORIN were more broadly expressed in macrophages and tissue stem cell populations (Fig. 4B, C), which was consistent with increased infiltration of fibroblasts, monocytic lineage cells, and M1 macrophages, as well as activation of stromal oncogenic signaling and dysregulated immune-related pathways in the PIET subtype.

Meanwhile, we investigated the multi-omics expression consistency of the four-gene signature at the molecular level. Significant and concordant multi-omics differences were observed between the two molecular subtypes (P-value < 0.05, Figs. 4D and S3A). Specifically, in RNA-seq data, CACNA2D2 and CCDC24 were highly expressed in the PIBT subtype, whereas CORIN and MMP1 were highly expressed in the PIET subtype. In CNA data, CACNA2D2 and MMP1 exhibited copy-number loss patterns that were directionally consistent with their RNA-seq expression differences between the two subtypes. By contrast, in DNA methylation data, the differential methylation patterns at the CCDC24 and CORIN loci also showed significant differences, but in an inverse regulatory direction relative to their RNA-seq expression levels, consistent with the canonical regulatory relationship. In addition, effective mutation sites of CACNA2D2 and CORIN were identified in the available proteomic data (Fig. 4E). Taken together, these results provide strong cross-omics regulatory concordance evidence, demonstrating that the subtype-associated RNA-seq expression differences are supported by CNV alterations, DNA methylation regulation, and proteomic variation.

Comprehensive analysis of clinical independence and prognostic value for the optimal four-gene model

To assess the clinical prognostic independence of the four-gene model, the univariate and multivariate Cox regression analyses were performed in 187 LBBC patients. The model remained significantly associated with overall survival after adjustment for 15 clinical variables (HR = 3.71, 95%CI 1.53–9.03, P-value = 0.004, Fig. 5A), indicating that it serves as an independent prognostic factor in LBBC. It is also worth noting that the model can stratify prognostic outcomes based on other independent clinical factors (P-value < 0.058, Figure S3C-D). Meanwhile, 5-year prognostic prediction analysis showed that the four-gene model exhibited higher predictive ability than other independent clinical factors and achieved the highest prognostic ability when combined with age and drug therapy (AUC = 0.82; Fig. S3E). In addition, multiple analysis confirmed that the four-gene model was independently associated with LBBC molecular subtype status after adjustment for 14 clinical variables (P-value < 0.01, Fig. 5B, Table S4), further supporting its clinical independence in subtype classification.

Fig. 5.

Comprehensive analysis of clinical independence and prognostic value for the optimal four-gene model. A Cox regression analyses assessing the prognostic independence from conventional clinical factors. B Logistic regression analyses assessing the classification independence for LBBC subtype assignment. C Validation of the prognostic performance in three independent external datasets. D Evaluation of subtype-specific prognostic value in LBBC

To further evaluate its prognostic value, we conducted survival analysis in three independent LBBC cohorts (n = 1,106). The model robustly stratified patients into two subtypes (PIBT and PIET) with significantly different survival outcomes (P-value < 0.05; Fig. 5C), which consistently showed that the PIET subtype had higher risk and lower 5-year OS compared with the PIBT subtype across all datasets (HR: 1.58–2.25). Meanwhile, subtype-specific analysis showed that the model did not significantly stratify survival in Luminal A, HER2-positive, or triple-negative breast cancer (P-value > 0.05; Fig. 5D), indicating its significant specificity for LBBC.

Construction and evaluation of a novel auxiliary model for survival prediction

To develop a quantitative approach for estimating individual survival probabilities in LBBC patients, we constructed a novel nomogram (PDT-LBDA, Fig. 6A) based on the four-gene model and two significant clinical risk factors from multivariate Cox analysis (Fig. 5A). The calibration curves demonstrated good agreement between predicted and observed survival probabilities, with comparable C-index values at 3 and 5 years (mean C-index = 0.834 ± 0.04, P-value < 0.001; Fig. 6B), indicating favorable calibration performance. Further validation was performed using bootstrap in the TCGA cohort (Fig. 6C). The integrated Brier score of the PDT-LBDA (0.120) was lower than its out-of-bag (OOB) estimate (0.137) and the four-gene model alone (0.128) across the integration interval [0, 99.8) months.

Fig. 6.

Construction and evaluation of a novel auxiliary model for survival prediction. A Integrating the four-gene model and independence clinical risk factors for predicting 3 and 5-year survival probabilities. B Calibration plots showing agreement between predicted and observed 3 and 5-year survival. C Bootstrap cross-validation. D Comparison of 5-year prognostic performance with six risk models and nomograms. E Decision curve analysis evaluating 5-year net benefit

To further assess its prognostic superiority, we compared the predictive performance of PDT-LBDA with six previously reported risk-based prognostic models for breast cancer [48–53]. PDT-LBDA achieved significantly higher 5-year prognostic accuracy (Risk score = 0.79, C-index = 0.834, AUC = 0.86) than all comparator models (max Risk score = 0.61, max C-index = 0.828, max AUC = 0.85, P-value < 0.001, Fig. 6D; Table S5). Decision curve analysis further demonstrated that PDT-LBDA provided consistently greater net clinical benefit and higher stability across a wide range of threshold probabilities. Specifically, PDT-LBDA yielded the greatest net benefit when the decision threshold exceeded 39%, supporting its utility in high-risk clinical decision-making scenarios (Fig. 6E).

Moreover, we explored the relationship between the risk score and immune checkpoint-related genes across different cellular components of the tumor microenvironment. The results showed that the risk score was positively correlated with fibroblast-associated markers PDPN (r = 0.42) and FAP (r = 0.32), macrophage-related immune regulators PDCD1LG2 (r = 0.38), CD80 (r = 0.37), and CD86 (r = 0.31), as well as the T cell exhaustion marker HAVCR2 (r = 0.36). These findings suggest that an elevated risk score is associated with coordinated activation of immune checkpoints in stromal, myeloid, and T cell compartments, reflecting an immune-activated tumor microenvironment, consistent with the findings described above (Table S4B).

Collectively, these findings indicate that the PDT-LBDA nomogram exhibits strong generalizability and robust prognostic predictive ability based on the four-gene model and may serve as a practical auxiliary tool to support individualized adjuvant treatment decision-making in patients with LBBC.

Discussion

LBBC exhibits marked tumor heterogeneity and distinct immunogenic features compared with other breast cancer subtypes, including substantial variability in therapeutic prognosis, tumor-infiltrating lymphocyte abundance, tumor mutation burden, and immune checkpoint expression [12–14, 23]. However, current clinical diagnostic systems lack immune-related molecular classification methods or prognostic tools capable of accurately distinguishing survival differences among LBBC patients. Therefore, developing a biologically meaningful immune-related molecular model for prognostic subtyping is essential to advance individualized treatment decision-making in LBBC.

To accurately characterize prognostic molecular heterogeneity in LBBC, we first integrated a minimal prognostic redundancy feature strategy with multiple subtype discrimination evaluation approaches to identify and validate an optimal four-gene biomarker set (CACNA2D2, CCDC24, CORIN, and MMP1) as informative prior knowledge for subtype delineation. Subsequently, we constructed an optimal four-gene classification model that was independent of conventional clinical diagnostic and prognostic factors, which was selected through systematic comparison of alternative diagnostic methods and kernel functions. The results demonstrated that the model adhered to a “minimum-gene principle” and consistently stratified LBBC patients into two molecular subtypes with high accuracy across TCGA and seven independent external validation cohorts (ACC ≥ 0.93, AUC ≥ 0.97, P-value < 0.001, n = 1,841), while concomitantly exhibiting marked immune-functional phenotypic differences and high consistency in multi-omics regulatory expression between the subtypes. Overall, these findings indicate that the model not only captures robust subtype-specific molecular patterns but also demonstrates strong generalizability across heterogeneous data sources, thereby conferring significant prognostic value.

Meanwhile, our findings further support the consistent subtype-specific regulatory mechanisms of the four-gene model. Multi-omics regulatory analysis revealed significant cross-omics expression consistency between the two subtypes. At the functional level, functional enrichment and immune infiltration analyses revealed distinct biological regulatory patterns: PIBT was mainly enriched in endocrine regulation pathways, whereas PIET was primarily enriched in stromal oncogenic-signaling and dysregulated immune-related pathways, and these functional patterns were also concordant with the single-cell data analysis. At the molecular level, CACNA2D2 and CCDC24 were highly expressed in PIBT and were associated with more favorable prognosis, whereas CORIN and MMP1 were highly expressed in PIET and were associated with poorer outcomes, accompanied by matching CNV alterations, methylation-expression and proteomic evidence (P-value < 0.05), consistent with previous reports [54–57]. Notably, hsa-miR-144 appeared to play a key role in multi-omics regulation, exhibiting a significant negative association with survival, consistent with previous findings [58].

In addition, to support individualized treatment and precise prognostic evaluation in LBBC patients, we first validated the significant prognostic stratification and predictive capability of the four-gene model (P-value < 0.05, AUC = 0.79). Subsequently, we developed a novel visual auxiliary prognostic nomogram (PDT-LBDA) that integrates the four-gene model with two additional independent clinical prognostic factors. Notably, the nomogram represents the first auxiliary decision-support tool for the 3 or 5-year prognosis across different age groups based on a minimal feature set, and it exhibits superior prognostic performance and greater clinical net benefit compared with previously reported risk models (P-value < 0.001, C-index = 0.834, AUC = 0.86). These findings suggest that PDT-LBDA may assist in individualized risk assessment for patients with LBBC to help identify those who are more likely to benefit from chemotherapy-predominant.

In summary, our study demonstrates that the four-gene biomarker model provides refined and robust discriminatory capability for prognostic molecular subtyping in LBBC from both algorithmic and biological perspectives. Furthermore, we developed an interpretable and clinically applicable PDT-LBDA tool that facilitates individualized prognostic assessment and supports precision-guided treatment decision-making. Through integrated multi-omics analyses, we also elucidated potential regulatory mechanisms and proposed candidate molecular targets for future translational exploration. Future work will focus on validating these findings in larger multi-center cohorts and further exploring these proposed pathological mechanisms and targets through molecular biology experiments.

Conclusions

In this study, we first identified and validated an optimal four-gene biomarker set as prior knowledge for LBBC molecular diagnosis using a minimum prognostic redundancy feature filtering strategy. Subsequently, we constructed a high-precision molecular subtype diagnostic model and confirmed its strong generalization ability across seven independent external validation cohorts. Moreover, we demonstrated that the four-gene model exhibits both diagnostic and prognostic independence from conventional clinical factors at algorithmic and biological levels. On this basis, we developed a prognostic auxiliary decision-support tool (PDT-LBDA) applicable to LBBC patients across age groups, which may assist individualized risk assessment.

In addition, comprehensive multi-omics analyses revealed subtype-specific regulatory mechanisms and functional characteristics associated with the immune-related four-gene model, elucidating the consistent regulatory mechanisms of prognostic heterogeneity in LBBC and providing candidate multi-omics targets for future therapeutic investigation. Although these findings offer new perspectives for prognostic diagnosis and treatment stratification in LBBC, further molecular and experimental validation is still required to substantiate the proposed regulatory mechanisms and therapeutic targets, which will be a focus of our future research. In particular, validating the four-gene signature in clinical cohorts of patients receiving immune checkpoint inhibitors (ICIs) will be essential to directly evaluate its predictive value for therapy response.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

JLo performed the biostatistical analysis, developed the predictive model, and drafted the manuscript. BL and JLi provided guidance on research methodology and paper organization. SZ supervised this study. The others assisted with samples collection.

Funding

This study was partially supported by the National Natural Science Foundation of China (Grant No: 62076015), the Beijing Natural Science Foundation – Haidian Original Innovation Joint Fund (Project No. L252088), and the REaDI Fund of Massey University.

Data availability

All publicly available datasets used in this study were obtained from GEO, TCGA, cBioPortal, and ArrayExpress. Other datasets are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This research complied with the Declaration of Helsinki. Most datasets used in this study are publicly available, with anonymization performed by the data providers. Similarly, private datasets were also anonymized, and their collection and use were reviewed and approved by the Institutional Review Committee.

Consent for publication

Not required.

Competing interests

The authors declare no competing interests.