Ethics statement

The study involved secondary analyses of publicly available, de-identified datasets, including METABRIC, TCGA, SCAN-B, and publicly accessible single-cell RNA-seq data. Because all data were fully anonymized prior to access and no direct interaction with human participants occurred, ethical approval from an institutional review board or ethics committee was not required. The requirement for informed consent was waived as all data were fully anonymized prior to acquisition.

METABRIC dataset

The METABRIC dataset (n = 2,509) [30,31] was downloaded from cBioPortal (https://www.cbioportal.org/) [32,33] on September 2, 2024. Samples with available data for DSS, mRNA expression, and PAM50 subtype information (n = 1,479) were included in prognostic analyses. Based on PAM50 classification, tumors were categorized as LumA (n = 475), LumB (n = 348), Her2 (n = 175), Normal (n = 119), Basal (n = 184), and Claudin-low (n = 178). Additionally, for the recurrence risk analyses, samples with available relapse-free survival (RFS) data (n = 1,825) were used, including LumA (n = 699), LumB (n = 475), Her2 (n = 224), Basal (n = 209), and Claudin-low (n = 218) subtypes.

TCGA dataset

The TCGA Pan-Cancer Atlas breast cancer dataset (n = 1,084) [34] was downloaded from cBioPortal (https://www.cbioportal.org/) [32,33] on November 19, 2024. Samples with available DSS, mRNA expression, and PAM50 subtype information (n = 962) were included. Tumors were stratified based on PAM50 classification: LumA (n = 491), LumB (n = 192), Her2 (n = 77), Normal (n = 35), and Basal (n = 167). Additionally, for the recurrence risk analyses, samples with available DFS data (n = 823) were analyzed, including LumA (n = 432), LumB (n = 172), Her2 (n = 66), and Basal (n = 153) subtypes.

The SCAN-B dataset

The SCAN-B cohort (n = 3,678) [35] was downloaded from the Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo/) under accession number GSE96058. Samples with available overall survival (OS), mRNA expression, and PAM50 classification data (n = 3,052) were included. Tumors were stratified based on PAM50 classification: LumA (n = 1,657), LumB (n = 729), Her2-enriched (n = 327), and Basal (n = 339). For each gene, mRNA expression values were z-score normalized by subtracting the mean and dividing by the standard deviation across all samples for subsequent analyses.

Prognostic analysis

DSS was analyzed using Kaplan-Meier survival curves, log-rank (Cochran-Mantel-Haenszel) tests, and multivariate Cox proportional hazards regression analyses, as previously described [36,37]. Patients were stratified into high- and low-expression groups based on z-score-normalized expression of individual genes or CCJ signature scores. The optimal cutoff values were determined using receiver operating characteristic (ROC) curve analysis for DSS, maximizing the Youden index. Age at diagnosis was included as a confounding factor in all multivariate Cox models.

Identification of prognostic cell–cell adhesion genes

To identify cell–cell adhesion molecules with prognostic relevance in each subtype, a gene list was obtained from the Gene Ontology term “cell–cell junction organization” (GO:0045216) (S1 Table). In each intrinsic subtype within the discovery datasets (METABRIC and TCGA), patients were stratified into high- and low-expression groups for each gene using the same z-score-based approach described above. Because the TCGA dataset does not include a distinct Claudin-low subtype, genes identified as prognostic in either the Basal or Claudin-low subtypes of the METABRIC cohort were intersected with genes identified in the Basal subtype of the TCGA dataset. Genes showing a statistically significant association with DSS in both datasets (p < 0.05, with 95% CI for HRs not crossing 1) were defined as robust subtype-specific prognostic genes. These 179 genes were used only as an initial candidate pool for subtype-specific prognostic screening in the discovery cohorts and were not used directly as the input for the downstream GO enrichment analysis.

Calculation of the CCJ signature score

Subtype-specific CCJ signatures were constructed based on the selected prognostic genes, as previously described [38]. For each patient, the CCJ signature score was calculated using the following formula:

where represents the z-score-normalized expression value of the i-th gene associated with poor prognosis, represents the z-score-normalized expression value of the j-th gene associated with favorable prognosis, and n and m denote the total numbers of poor- and favorable-prognosis genes, respectively. Using these CCJ signature scores, patients in the discovery cohorts (METABRIC, TCGA) and the validation cohort (SCAN-B) were stratified into high- and low-score groups. Survival analyses, including Kaplan-Meier estimation and multivariate Cox proportional hazards modeling, were then performed. In the SCAN-B dataset analysis, the Her2 signature gene AJM1 and the LumA signature genes AFDN and INAVA were not available (NA) and were therefore excluded from the corresponding CCJ signature score calculations.

Histological subtype analysis

To examine whether the prognostic associations of the CCJ signatures were influenced by histological subtype, we performed a histology-focused analysis in the METABRIC and TCGA cohorts. Histological annotation was obtained from the dataset-specific pathology fields: the Tumor.Other.Histologic.Subtype column in METABRIC and the Tumor.Type column in TCGA. Because annotation schemes differed between datasets, histological categories were harmonized into simplified groups prior to analysis.

In METABRIC, cases annotated as “Lobular” were classified as invasive lobular carcinoma (ILC), and cases annotated as “Ductal/NST” were classified as non-ILC in the main comparative analyses. For descriptive analyses, “Mixed” was retained as a separate category, whereas all remaining annotations (including “Medullary,” “Metaplastic,” “Mucinous,” “Tubular/ cribriform,” “Other,” and “NA”) were grouped as “Other”. In TCGA, cases annotated as “Infiltrating Lobular Carcinoma” were classified as ILC, cases annotated as “Infiltrating Ductal Carcinoma” were classified as non-ILC, and “Mixed Histology (NOS)” was treated as “Mixed,” and all other categories (including “Breast Invasive Carcinoma,” “Infiltrating Carcinoma [NOS],” “Medullary Carcinoma,” “Metaplastic Carcinoma,” and “Mucinous Carcinoma,”) were grouped as “Other.”

Histological composition within each intrinsic subtype was summarized using 100% stacked bar plots. For the primary comparative analyses, only ILC and non-ILC cases were included, excluding “Mixed” and “Other” categories. CCJ signature scores were compared between ILC and non-ILC groups within each subtype using subtype-matched CCJ signature. Differences in CCJ signature scores were assessed using the Wilcoxon rank-sum test within each subtype, followed by Benjamini–Hochberg correction for multiple testing. To determine whether the association between CCJ signature classification and survival was independent of histological subtype, multivariable Cox proportional hazards models were fitted including CCJ signature class (high vs low), age at diagnosis, and simplified histology (ILC vs non-ILC) as covariates. Analyses were restricted to cases with complete data for survival, age, histology, and CCJ signature classification. Hazard ratios (HRs) with 95% confidence intervals (CIs) were estimated, and results were summarized using forest plots for each cohort.

Prognostic evaluation of EMT-related CCJ genes and CDH1

To further examine the relationship between the CCJ signatures and canonical epithelial–mesenchymal transition (EMT)-related genes, we assessed the overlap between the Gene Ontology (GO) term GO:0045216 (“cell–cell junction organization”) gene set and the HALLMARK_EPITHELIAL_MESENCHYMAL_TRANSITION gene set obtained from the Molecular Signatures Database (MSigDB) [39,40]. Because CDH1, a key epithelial marker with a well-established role in EMT, is not included in the HALLMARK EMT gene set, it was additionally incorporated into this analysis. Multivariable Cox proportional hazards analyses were then performed for the overlapping genes and CDH1 in the METABRIC and TCGA cohorts using the same subtype-specific framework described above. Hazard ratios (HRs) with 95% confidence intervals (CIs) were estimated, and statistical significance was defined as p < 0.05 with CIs not crossing 1.

Leave-one-gene-out sensitivity analysis of the CCJ signature

To assess the robustness of the CCJ signatures, we performed a leave-one-gene-out sensitivity analysis. For each subtype-specific signature, the original CCJ signature score was calculated as described above and then recalculated iteratively by omitting one gene at a time from the signature. For each recalculated signature, patients were reclassified into high- and low-score groups using receiver operating characteristic (ROC) analysis with the Youden index. Age-adjusted multivariable Cox proportional hazards models were then fitted using the same cohort-specific procedures as in the primary analyses. Hazard ratios (HRs) with 95% confidence intervals (CIs) derived from the original and leave-one-gene-out models were compared to evaluate the stability of the prognostic associations. To facilitate interpretation of potential redundancy among equally weighted genes, a subtype-specific integrated rank table was additionally generated from the discovery cohorts (S4 Table). This table was not used to modify the leave-one-gene-out models but was used descriptively to summarize the consistency of gene-level prognostic contributions across cohorts.

Gene-number downsampling analysis of the CCJ signature

To evaluate how many genes were required for each CCJ signature to retain prognostic information, we performed a gene-number downsampling analysis. To generate reduced signatures, genes were prioritized using a subtype-specific integrated rank table derived from the discovery cohorts (S4 Table). Gene-level prognostic ranks were defined based on the strength and consistency of associations with survival outcomes in each cohort. Briefly, within each subtype, gene-level prognostic ranks were calculated separately in the METABRIC and TCGA datasets and then averaged to obtain a mean rank. For the Basal context, METABRIC Basal and Claudin-low ranks were considered separately, and the better METABRIC rank was selected before averaging with the TCGA Basal rank. Reduced signatures were constructed by selecting top-ranked genes within each subtype. For each reduced gene set, CCJ signature scores were calculated using the formula described above, and patients were reclassified into high- and low-score groups using receiver operating characteristic (ROC) analysis with the Youden index. Age-adjusted multivariable Cox proportional hazards models were then fitted using the same cohort-specific procedures as in the primary analyses. In the SCAN-B dataset, only genes with available expression values were retained, and gene sets with fewer than two matched genes were excluded from analysis. Hazard ratios (HRs) with 95% confidence intervals (CIs) derived from the original and downsampled signatures were compared to evaluate the stability of prognostic associations as gene number decreased.

Recurrence risk analysis

Recurrence risk was evaluated using person-years approach, as previously described [36]. Patients were stratified into high- and low-score groups based on their CCJ signature scores, using optimal cutoff values determined by ROC analysis. Follow-up time was divided into predefined intervals to capture time-dependent recurrence patterns. For the METABRIC dataset, using RFS as the endpoint, intervals were defined as 0–5, 5–10, 10–15, and >15 years. For the TCGA dataset, using DFS as the endpoint, intervals were defined as 0–5, 5–10, and >10 years, reflecting the available follow-up distribution. Recurrence incidence rates were calculated as the number of events per person-year within each interval using the pyears function from the R survival package. To compare the risk between high- and low-score groups, the incidence rate ratios (IRRs) with corresponding 95% CIs were estimated. Statistical significance was assessed using the Poisson exact test. To account for multiple comparisons across time intervals, P-values were adjusted using the Holm method, and adjusted p < 0.05 was considered statistically significant.

Differential gene expression and functional enrichment analysis

To characterize biological differences associated with CCJ signature scores, differential gene expression (DEG) analysis was performed between high- and low-score groups in the TCGA and SCAN-B datasets. DEG analysis was restricted to the TCGA and SCAN-B cohorts because both datasets were generated using RNA-seq-based transcriptomic profiling, thereby reducing platform-related heterogeneity in fold-change-based comparisons; the METABRIC dataset was not included in this step. Because the available expression matrices differed in format, DEG analysis was conducted using different statistical frameworks: DESeq2 was applied to the TCGA count-like expression matrix, whereas limma was applied to the transformed SCAN-B expression matrix. As an internal positive control, we additionally evaluated whether each subtype-specific CCJ signature gene (i) was present in the expression matrix, (ii) passed expression filtering, and (iii) was identified as a significant DEG under the predefined criteria. Genes that were significantly and consistently altered in both cohorts (fold change [FC] > 2 or [FC] < 0.5, p < 0.05) were extracted. For commonly upregulated genes identified in the LumB and Basal CCJ signatures, GO enrichment analysis was performed using Metascape (http://metascape.org) [41] to identify associated biological processes and molecular functions.

Single-cell RNA-seq analysis of breast cancer samples

Publicly available single-cell RNA sequencing (scRNA-seq) data from a human breast cancer atlas [42] were obtained from the Gene Expression Omnibus (GEO; accession number GSE176078). Raw gene expression count matrices and associated metadata were used as input. All analyses were performed using the Seurat package (version 5.3.0). A Seurat object was constructed from the count matrices, followed by quality control (QC) filtering to remove low-quality cells. Cells with fewer than 200 detected genes (nFeature_RNA) or cells with mitochondrial transcript content exceeding 20% were excluded. To focus on untreated tumor microenvironments, cells from patients with a history of therapy (patient IDs: CID3963, CID4066, CID4398, CID4513, CID4523) were excluded from downstream analyses. The final cohort comprised 21 untreated patients, stratified by clinical subtype as follows: ER+ (n = 10), Her2+ (n = 3), Her2 + /ER+ (n = 1), and TNBC (n = 7).

Data normalization, integration, and clustering

Following QC, gene expression data were log-normalized using the NormalizeData function. The 2,000 most highly variable genes were identified using the variance-stabilizing transformation (vst) method implemented in FindVariableFeatures. Data were then scaled using ScaleData, and principal component analysis (PCA) was performed for dimensionality reduction. To correct for batch effects attributed to inter-patient variation, Harmony integration was applied to the first 50 principal components, using patient identity as the batch covariates. Harmony-corrected embeddings were used for all downstream analyses. For visualization, Uniform Manifold Approximation and Projection (UMAP) was conducted using the first 30 Harmony dimensions. Cell clustering was performed on the same dimensions by constructing a shared nearest neighbor (SNN) graph with FindNeighbors, followed by Louvain community detection using FindClusters at a resolution of 0.8.

Cell type identification and annotation

Cell types were identified using a combination of automated and manual annotation approaches based on established marker gene expression. For automated annotation, the SingleR R package was employed using the Human Primary Cell Atlas reference dataset, obtained from the celldex package. The Seurat object was first converted into a SingleCellExperiment object, after which the SingleR() function was applied to assign the most probable cell type label and an associated confidence score to each cell. To validate and supplement the automated annotations, the specific expression patterns of known marker genes were visualized using the DotPlot function. The marker genes used for cell type identification were as follows: epithelial cells (EPCAM, KRT18), T/NK cells (PTPRC, CD3D, NKG7, KLRD1), B-lineage cells (MS4A1, JCHAIN), myeloid cells (CD68, CD14, ITGAX), fibroblasts (COL1A1, DCN, PDGFRB), and endothelial cells (PECAM1, VWF). Final cell type annotations were determined by integrating both results. Specifically, the primary cell type assigned to each cluster by SingleR was first reviewed and then cross-referenced with the specific expression patterns of the marker genes, resulting in a final cell type label assigned to each cluster.

Cell type- and subtype-specific expression analysis of the CCJ signature

Using the cell type-annotated Seurat object, expression patterns of prognostic CCJ signature genes were evaluated in a subtype-specific manner. Cells corresponding to TNBC and Her2 + /ER+ subtypes were extracted from the untreated patient cohort, and separate Seurat subset objects were generated for each subtype. Expression of individual genes within the Basal CCJ signature (analyzed in TNBC samples) and the LumB CCJ signature (analyzed in Her2 + /ER+) was visualized using dot plots and UMAP-based feature plots generated with the FeaturePlot function. These visualizations enabled assessment of the specific cell populations contributing to CCJ gene expression within the tumor microenvironment of each subtype. Genes with undetected expression were displayed in gray across all cells to facilitate clear distinction.

Visualization of results

Overlap among prognostic genes across breast cancer subtypes was visualized using UpSet plots generated with the UpSetR package in R. Genes consistently associated with prognosis in both the METABRIC and TCGA datasets were additionally illustrated using Venn diagrams constructed with the VennDiagram package. Results from multivariate Cox proportional hazards analyses were visualized using forest plots generated with the forestplot package in R.

Statistical evaluation

All statistical analyses were performed using R software version 4.5.1 (R Core Team, R Foundation for Statistical Computing, Vienna, Austria).

DSS-based multivariate Cox analyses of “cell–cell junction organization”–related genes across PAM50 subtypes

We first investigated the prognostic impact of genes involved in cell–cell junction organization across intrinsic breast cancer subtypes. DSS-based multivariate Cox proportional hazards analyses were performed for 179 genes annotated to GO:0045216 (S1 Table) in METABRIC and TCGA cohorts, adjusting for patient age as a covariate. Analyses were stratified by PAM50 subtypes, including the Claudin-low group, to account for subtype-specific biological differences.

In both datasets, a substantial number of genes demonstrated significant associations with DSS, and these associations were largely subtype-specific, indicating that the prognostic relevance of CCJ-related genes differs markedly between molecular subtypes (Fig 2A–2D). Notably, no single gene was prognostic across all subtypes in either cohort, underscoring the heterogeneity of CCJ-related pathways in breast cancer progression. These findings provided the rationale for constructing subtype-specific CCJ gene expression signatures to capture coordinated transcriptional programs underlying subtype-dependent prognosis. To identify robust subtype-specific prognostic genes, we focused on genes consistently associated with DSS in both datasets. Genes associated with poor prognosis across both datasets included seven genes in LumA (KIRREL1, EXT1, AFDN, INAVA, MTSS1, STRN, GJC1), ten in LumB (CTNNA1, CLDN2, FERMT2, EFNB2, VCL, JAM3, CDH2, CDH3, CDH12, CDH13), five in Her2 (CTNNB1, ILDR1, CD2AP, DSG1, AJM1), and eight in Basal (FSCN1, RAMP2, CDH9, ZNF703, ACTB, MARVELD2, F2RL1, JAM3) (Table 1 and Fig 3A). Conversely, genes consistently associated with favorable prognosis included three genes in LumA (WNT11, GJA1, PARD6B), four in LumB (NF2, PARD6A, PARD6B, CLDN20), one in Her2 (NUMB), and one in Basal (PLEKHA7) (Table 1 and Fig 3A). No common prognostic genes were identified for the Normal subtype across both datasets. These findings suggest that CCJ-related genes may have a more limited or less consistent association with tumor progression and prognosis in the Normal subtype compared with other intrinsic subtypes. Although most prognostic genes were subtype-specific, JAM3 represented a notable exception, being associated with poor prognosis in both LumB and Basal subtypes (Fig 3B). Similarly, PARD6B emerged as a favorable prognostic gene shared between LumA and LumB subtypes (Fig 3C). Among the representative subtype-specific genes selected for subsequent biological interpretation, INAVA was reproducibly associated with poor prognosis in LumA tumors, CLDN2 and FERMT2 were reproducibly associated with poor prognosis in LumB tumors, and NUMB was the sole gene reproducibly associated with favorable prognosis in Her2 tumors (Table 1 and Fig 3A). These genes were therefore considered representative subtype-linked components of the CCJ signatures and are discussed further below in the context of established breast cancer biology.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Subtype-specific prognostic CCJ signature genes in breast cancer.

https://doi.org/10.1371/journal.pone.0352254.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Prognostic landscape of cell–cell junction organization genes across breast cancer subtypes.

Multivariate Cox proportional hazards analyses were performed for 179 genes annotated under the Gene Ontology term “cell–cell junction organization” (GO:0045216) to evaluate associations with DSS, adjusted for patient age. UpSet plots display the number of genes for which high expression was significantly associated with (A) poor prognosis and (B) favorable prognosis in the METABRIC dataset, and (C) poor prognosis and (D) favorable prognosis in the TCGA dataset. The vertical bars represent the number of significant genes shared among the indicated subtypes, highlighting the subtype-specific nature of prognostic associations.

https://doi.org/10.1371/journal.pone.0352254.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Identification of robust subtype-specific prognostic CCJ genes shared between METABRIC and TCGA datasets.

(A) Venn diagrams illustrating the overlap of significant prognostic genes between the METABRIC (blue) and TCGA (grey) datasets for each molecular subtype. Intersections represent robust prognostic genes consistently identified in both independent cohorts. No common prognostic genes were identified for the Normal subtype. (B) UpSet plot displaying the intersection of robust poor-prognosis genes across different subtypes (e.g., overlap between LumB and Basal). (C) UpSet plot displaying the intersection of robust favorable-prognosis genes across different subtypes. These plots visualize the degree of subtype-specificity among the validated prognostic genes.

https://doi.org/10.1371/journal.pone.0352254.g003

Prognostic significance of CCJ signatures in the discovery cohorts

Breast cancer is a highly heterogeneous disease in which diverse molecular pathways contribute to tumor development, progression, and treatment response. Reliance on individual gene expression is often insufficient to capture the complexity of tumor biology. Gene expression signatures that reflect coordinated activity of multiple genes provide a more robust framework for accurate prognostic and predictive assessment. Building on our previous findings [38], demonstrating the utility of information-theoretic gene signature analysis in breast cancer, we integrated subtype-specific prognostic CCJ genes to construct CCJ signatures for each intrinsic subtype by calculating CCJ signature scores, with the exception of the Normal subtype, which lacked common prognostic genes. The prognostic performance of these CCJ signatures was evaluated, using Kaplan-Meier survival analysis and multivariate Cox proportional hazards models for DSS, adjusting for age. In the METABRIC cohort, Kaplan-Meier analysis demonstrated that patients with high CCJ signature scores consistently exhibited significantly poorer DSS compared with those with low scores across all subtypes (LumA, p < 0.001; LumB, p < 0.001; Her2, p < 0.001; Basal, p = 0.013; Claudin-low, p < 0.001) (Fig 4A). Multivariate Cox analysis confirmed these findings, revealing that high CCJ signature scores were independently associated with worse DSS (LumA: hazard ratio [HR] = 2.29, 95% confidence interval [CI] = 1.63–3.22, p < 0.001; LumB: HR = 2.15, 95% CI = 1.57–2.96, p < 0.001; Her2: HR = 2.17, 95% CI = 1.41–3.35, p < 0.001; Basal: HR = 1.75, 95% CI = 1.12–2.74, p = 0.013; Claudin-low: HR = 2.69, 95% CI = 1.50–4.82, p < 0.001) (Fig 4B). Consistent results were observed in the TCGA cohort. Kaplan-Meier analysis showed significantly poorer DSS among patients with high CCJ signature scores across all subtypes (LumA, p = 0.002; LumB, p = 0.005; Her2, p < 0.001; Basal, p < 0.001) (Fig 4C). Multivariate Cox proportional hazards analysis also demonstrated that high CCJ signature scores independently predicted worse DSS in each subtype (LumA: HR = 3.86, 95% CI = 1.55–9.58, p = 0.004; LumB: HR = 5.19, 95% CI = 1.47–18.26, p = 0.010; Her2: HR = 8.10, 95% CI = 2.18–30.07, p = 0.002; Basal: HR = 8.58, 95% CI = 3.01–24.49, p < 0.001) (Fig 4D). The predictive accuracy of these signatures was further supported by ROC analysis (S2 Table). Collectively, these results indicate that high CCJ signature expression is associated with worse clinical outcomes across intrinsic subtypes.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Prognostic significance of subtype-specific CCJ signatures in METABRIC and TCGA datasets.

Kaplan-Meier survival analysis and multivariate Cox proportional hazards regression were performed to evaluate associations between CCJ signature scores and DSS. (A) Kaplan-Meier curves for the METABRIC dataset. Patients were stratified into high- (red) and low-score (blue) groups based on subtype-specific CCJ signatures across five subtypes (LumA, LumB, Her2, Basal, and Claudin-low). (B) Forest plot summarizing multivariate Cox regression results for the METABRIC dataset, adjusted for patient age. (C) Kaplan-Meier curves for the TCGA dataset across four available subtypes (LumA, LumB, Her2, and Basal). (D) Forest plot summarizing multivariate Cox regression results for the TCGA dataset, adjusted for patient age. Hazard ratios (HRs) and 95% confidence intervals (CIs) indicate the risk associated with high CCJ signature scores relative to low scores. P-values were calculated using the log-rank test (Kaplan-Meier) and the Wald test (Cox regression). In the Kaplan-Meier plots, red text indicates statistical significance (p < 0.05).

https://doi.org/10.1371/journal.pone.0352254.g004

Validation of CCJ signatures in the SCAN-B dataset

To independently validate the prognostic value of the CCJ signatures, we performed identical analyses using the SCAN-B dataset, which was not used during the initial gene extraction. Distinct subtype-specific patterns were observed. In the LumA subtype, no significant prognostic difference was detected between patients with high and low CCJ signature scores, as assessed by Kaplan-Meier analysis (p = 0.241) and multivariate Cox proportional hazards modeling (HR = 0.86, 95% CI = 0.57–1.29, p = 0.459) (Fig 5A and 5B). In contrast, the Her2 subtype showed an inverse pattern: patients with low CCJ signature scores exhibited significantly poorer survival in Kaplan-Meier analysis (p = 0.027). However, this association did not remain statistically significant after adjustment in the multivariate Cox model (HR = 0.53, 95% CI = 0.26–1.08, p = 0.082) (Fig 5A and 5B). Importantly, in the LumB subtype, patients with high CCJ signature scores were strongly associated with poor prognosis in the Kaplan-Meier analysis (p = 0.001) and multivariate Cox model (HR = 2.27, 95% CI = 1.45–3.57, p < 0.001) compared with those with low scores, which is consistent with the results from the discovery cohorts (Fig 5A and 5B). Similarly, in the Basal subtype, high CCJ signature scores were strongly associated with poor prognosis in the Kaplan-Meier analysis (p = 0.002) (Fig 5A). Because no disease-specific deaths occurred in the low-score group, multivariate Cox analysis could not be performed for this subtype. The predictive accuracy of these signatures was further supported by receiver operating characteristic (ROC) analysis (S2 Table). Collectively, these results indicate that high CCJ signature expression is associated with worse clinical outcomes in LumB and Basal tumors. The lack of concordant associations in LumA and Her2 subtypes likely reflects inter-cohort heterogeneity in clinical characteristics or transcriptomic profiling. Overall, these findings provide independent validation of the CCJ signature as a prognostic biomarker, particularly in the LumB and Basal subtypes of breast cancer.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Independent validation of the prognostic value of CCJ signatures in the SCAN-B dataset.

Kaplan-Meier survival analysis and multivariate Cox proportional hazards models were performed to validate the prognostic utility of CCJ signatures in the SCAN-B cohort. (A) Kaplan-Meier curves for OS across LumA, LumB, Her2, and Basal subtypes. Patients were stratified into high- (red) and low-score (blue) groups based on the subtype-specific CCJ signatures. (B) Forest plot summarizing the multivariate Cox regression results adjusted for patient age. Hazard ratios (HRs) and 95% confidence intervals (CIs) are shown. Multivariate analysis for the Basal subtype could not be performed (indicated as N/A) due to the absence of events in the low-score group. P-values were calculated using the log-rank test (Kaplan-Meier) and the Wald test (Cox regression). In the Kaplan-Meier plots, red text indicates statistical significance (p < 0.05).

https://doi.org/10.1371/journal.pone.0352254.g005

Histological subtype composition and histology-adjusted prognostic analysis of CCJ signatures

Because histological subtypes may influence the interpretation of the subtype-specific prognostic associations of the CCJ signatures, we performed additional histology-focused analyses in the METABRIC and TCGA cohorts. In particular, invasive lobular carcinoma (ILC), which is pathognomonically associated with loss of E-cadherin [43], was considered a potential confounding factor. First, we summarized the distribution of ILC and non-ILC cases across PAM50 intrinsic subtypes in both cohorts (S1A and S1B Fig). In both datasets, ILC cases were present across subtypes but consistently represented a minority compared with non-ILC cases. Next, we compared CCJ signature scores between ILC and non-ILC tumors within each intrinsic subtype (S1C and S1D Fig). In the METABRIC cohort, a significant difference was observed only in the LumA subtype, in which non-ILC tumors exhibited lower CCJ signature scores than ILC tumors. However, this finding was not reproduced in the TCGA cohort, and no consistent differences in CCJ signature scores between ILC and non-ILC cases were observed across subtypes. Finally, we performed multivariable Cox proportional hazards analyses including ILC status as a covariate (Fig 6A and 6B). Importantly, the prognostic associations of the subtype-matched CCJ signatures remained significant after adjustment for histological subtype, indicating that these associations were not explained solely by histology.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Histology-adjusted prognostic evaluation of CCJ signatures and EMT-related CCJ genes in the METABRIC and TCGA datasets.

(A, B) Forest plots summarizing multivariable Cox proportional hazards models for subtype-matched CCJ signatures adjusted for age and histological subtype in the METABRIC and TCGA datasets, respectively. (C) Venn diagram illustrating the overlap between genes annotated under the Gene Ontology term “cell–cell junction organization” (GO:0045216) and the HALLMARK_EPITHELIAL_MESENCHYMAL_TRANSITION gene set. (D, E) Forest plots summarizing multivariable Cox proportional hazards analyses of the overlapping genes and CDH1 across breast cancer subtypes in the METABRIC and TCGA datasets, respectively. Hazard ratios (HRs) and 95% confidence intervals (CIs) are shown. Only genes significantly associated with poor or favorable prognosis are displayed.

https://doi.org/10.1371/journal.pone.0352254.g006

EMT-related prognostic analysis of CCJ-associated genes

To further examine whether the prognostic relevance of the CCJ signatures could be explained by canonical EMT-related genes, we assessed the overlap between the GO:0045216 “cell–cell junction organization” gene set and the HALLMARK_EPITHELIAL_MESENCHYMAL_TRANSITION gene set from MSigDB [40], and additionally included CDH1 because of its central biological relevance to EMT (Fig 6C). We then performed subtype-specific multivariable Cox proportional hazards analyses for the overlapping genes and CDH1 in the METABRIC and TCGA cohorts (Fig 6D and 6E and S3 Table). Although several genes showed significant prognostic associations in individual subtypes, only a limited number were reproducibly associated with prognosis across both discovery cohorts. Specifically, GJA1 in LumA and CDH2 and FERMT2 in LumB showed concordant prognostic associations in both datasets. In contrast, CDH1 showed prognostic significance only in a subset of analyses and was not reproducibly associated across the two cohorts. These findings indicate that the prognostic relevance captured by the CCJ signatures is not recapitulated by E-cadherin alone and is not fully explained by canonical EMT-related genes.

Leave-one-gene-out sensitivity analysis of the CCJ signature

To evaluate whether the prognostic performance of each CCJ signature was driven by any single constituent gene, we performed a leave-one-gene-out sensitivity analysis in the METABRIC, TCGA, and SCAN-B cohorts (S2 Fig). In the discovery cohorts, omission of individual genes produced only modest changes in hazard ratios, and the prognostic associations of the subtype-matched CCJ signatures were largely preserved across subtypes (S2A and S2B Fig). These findings indicate that the prognostic relevance of the CCJ signatures in the discovery cohorts was not driven by any single gene, but rather reflected a collective signal. In the SCAN-B cohort, LumB showed the most stable pattern, whereas LumA and Her2 showed more limited or inconsistent effects. For the Basal subtype in the SCAN-B cohort, the original model is not shown because it was identical to the primary validation model and was not estimable owing to the absence of death events in the low-score group. In contrast, in leave-one-gene-out models that became estimable, hazard ratios varied substantially, suggesting that this subtype was strongly influenced by the limited number of events in the validation cohort rather than by the contribution of any specific gene (S2C Fig). Overall, these findings support that the principal prognostic associations of the CCJ signatures, particularly in the discovery cohorts and in the LumB validation analysis, were not driven by any single gene but instead reflected a robust collective signal.

Gene-number downsampling analysis of the CCJ signature

To determine whether a minimum number of genes was required for each CCJ signature to remain informative, we performed a gene-number downsampling analysis in the METABRIC, TCGA, and SCAN-B cohorts (S3 Fig). Reduced signatures were assembled according to the subtype-specific integrated rank table (S4 Table), such that genes with more consistently higher ranks across the discovery cohorts were preferentially retained in the top 7-, top 5-, and top 3-gene sets. In the discovery cohorts, the prognostic associations of most CCJ signatures were largely preserved as gene number was reduced. In METABRIC, the LumA, LumB, Her2, and Claudin-low signatures remained significantly associated with prognosis even in smaller gene sets. For the Basal signature, reduced gene sets yielded similar hazard ratios, although statistical significance became borderline in some versions. In TCGA, all evaluated signatures retained significant prognostic associations across the downsampled gene sets. In the SCAN-B cohort, the LumB signature remained significantly associated with prognosis across all evaluated gene sets, with hazard ratios of 2.27 (95% CI = 1.45–3.57, p < 0.001) for the original signature, 2.27 (95% CI = 1.40–3.69, p < 0.001) for the top 7-gene set, 1.67 (95% CI = 1.08–2.58, p = 0.021) for the top 5-gene set, and 1.85 (95% CI = 1.19–2.86, p = 0.006) for the top 3-gene set (S3C Fig). Although prognostic information was retained in the reduced gene sets, the original and top 7-gene versions showed the most stable performance, suggesting that retaining a larger number of genes may be advantageous for more robust prognostic stratification. For the Basal subtype, the original model could not be estimated because no death events occurred in the low-score group, whereas the downsampled gene sets showed significant associations with prognosis. These findings indicate that no single universal minimum gene number could be defined across all subtype-cohort combinations. Nevertheless, the overall preservation of prognostic associations across multiple reduced gene sets suggests that the CCJ signatures retain prognostic relevance, while the more stable performance of the original or moderately reduced signatures indicates that larger gene sets may be advantageous for more robust stratification.

Recurrence risk stratification using CCJ signatures

To evaluate recurrence risk with CCJ signature expression, recurrence incidence rates per person-year were calculated across defined time intervals (0–5, 5–10, 10–15, and >15 years for METABRIC; 0–5, 5–10, and >10 years for TCGA) (Figs 7 and S4 Fig and S5 Table). In the METABRIC dataset, patients with high CCJ signature scores consistently exhibited significantly higher recurrence rates during the first five years across LumA (p < 0.001), LumB (p < 0.001), Basal (p = 0.005), and Claudin-low (p = 0.001) subtypes. In the LumB subtype, recurrence rates in the high-score group remained elevated during the 5–10 and 10–15 year intervals, although these differences did not reach statistical significance (Fig 7). Consistent trends were observed in the TCGA cohort using disease-free survival (DFS) as the endpoint. In this dataset, the LumB subtype showed a trend towards higher event rates in the high CCJ signature group during the first 0–5 years (p = 0.051), and a similar pattern was observed in the Basal subtype (p = 0.091) (S4 Fig). Together, these findings highlight the ability of the CCJ signature to identify patients at increased risk of early recurrence (0–5 years), particularly within aggressive and luminal subtypes, suggesting its potential use in identifying patients vulnerable to early treatment failure.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Recurrence risk associated with the CCJ signature assessed by person-year analysis in METABRIC dataset.

Recurrence incidence rates (events per person-year) were calculated for patients stratified into high- (red) and low-score (blue) CCJ signature groups. (A) METABRIC dataset assessed over 0–5, 5–10, 10–15, and >15 year intervals. P-values derived from Poisson tests with Holm correction indicate statistical significance (p < 0.05) or trends. Error bars indicate 95% confidence intervals (CIs).

https://doi.org/10.1371/journal.pone.0352254.g007

Differential gene expression associated with CCJ signature scores

To further characterize the molecular properties associated with high CCJ signature expression, patients were stratified into high- and low-score groups, and differentially expressed gene (DEG) analyses were performed, using RNA-seq data from the TCGA and SCAN-B cohorts (Figs 8A–8D and S5). These cohorts were selected because both employed RNA-seq-based profiling and shared PAM50 subtype annotation, with exclusion of the Claudin-low subtype. DEG analysis revealed distinct subtype-specific transcriptional differences between high- and low-score tumors. For the LumB CCJ signature, 440 genes were upregulated and 384 were downregulated in the TCGA cohort (Fig 8A), whereas 249 genes were upregulated and 25 were downregulated in the SCAN-B cohort (Fig 8B). Among these, 93 genes were consistently upregulated and four genes (COX6C, DIO1, PHGR1, PARD6B) were consistently downregulated across both TCGA and SCAN-B datasets (S6 Table). For the Basal CCJ signature, the TCGA cohort exhibited 495 upregulated and 395 downregulated genes (Fig 8C), while the SCAN-B cohort exhibited 259 upregulated and 10 downregulated genes (Fig 8D). Of these, 60 genes were consistently upregulated and four genes (FABP7, KLHDC7B, SPIB, UBD) were consistently downregulated across both cohorts (S6 Table). Given the substantial number of commonly upregulated genes in the LumB (n = 93) and Basal (n = 60) signatures, Gene Ontology (GO) enrichment analysis was performed to elucidate the biological processes associated with high CCJ signature expression (Fig 8E and 8F). Both signatures showed strong enrichment for extracellular matrix (ECM)-related pathways. In the LumB CCJ signature, significantly enriched terms included “NABA CORE MATRISOME,” “extracellular matrix organization,” and “collagen degradation.” Similarly, the Basal CCJ signature was enriched for “NABA CORE MATRISOME,” “extracellular matrix organization,” and “regulation of collagen fibril organization.” In addition to ECM-related processes, the Basal signature was enriched for “cellular response to growth factor stimulus”.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Differential gene expression and functional enrichment analysis associated with high CCJ signature scores.

Volcano plots illustrate DEG comparing high- and low-CCJ signature score groups in the TCGA and SCAN-B datasets. Red dots represent significantly upregulated genes, blue dots represent significantly downregulated genes, and grey dots represent non-significant genes. Genes included in the CCJ signatures are annotated on the volcano plots with labels connected to the corresponding dots by leader lines. (A) LumB subtype in the TCGA dataset. (B) LumB subtype in the SCAN-B dataset. (C) Basal subtype in the TCGA dataset. (D) Basal subtype in the SCAN-B dataset. (E, F) Bar plots displaying the top enriched Gene Ontology (GO) terms and pathways for the genes that were commonly upregulated in both datasets for (E) the LumB CCJ signature and (F) the Basal CCJ signature.

https://doi.org/10.1371/journal.pone.0352254.g008

To further assess the representation and differential expression of CCJ signature genes, we evaluated whether each gene was (i) present in the expression matrix, (ii) retained after expression filtering, and (iii) identified as a significant DEG under the predefined criteria. In the TCGA cohort, 35 of 39 signature genes were detected in the DESeq2 results, whereas 29 of 39 genes were detected in the SCAN-B limma results (S7 Table). Among these, only a subset met the stringent DEG criteria (|log₂ fold change| > 1 and p < 0.05). Notably, all signature genes identified as significant DEGs exhibited a consistent direction of effect relative to the CCJ signature definition, supporting the biological coherence of the signatures.

Expression pattern of CCJ signature genes in cells of breast cancer tissues

Given the pronounced prognostic relevance of CCJ signatures in LumB and Basal subtypes, we subsequently examined the predominant cell populations expressing genes constituting these subtype-specific signatures. Using single-cell RNA-seq data from breast cancer patients (GSE176078), we analyzed expression patterns of genes in the LumB CCJ signature using data from an Her2 + /ER+ tumor (n = 1) as a representative LumB case. Likewise, because Basal tumors largely overlap with TNBC, we analyzed expression of genes in the Basal CCJ signature across TNBC patients (n = 7) (Figs 9 and S6). Within the LumB CCJ signature, PARD6B (favorable prognostic gene) and CDH3 (poor prognostic gene) were predominantly expressed in epithelial/cancer cell populations (Figs 9A and S6). In contrast, CTNNA1 (poor prognostic gene) displayed broader expression across multiple cell types, including epithelial/cancer cells, macrophages, fibroblasts, endothelial cells, and pericytes (Figs 9A and S6). For the Basal CCJ signature, MARVELD2 (poor prognostic gene) was mainly expressed in epithelial/cancer cells, whereas RAMP2 (poor prognostic gene) showed predominant expression in vascular endothelial cells (Figs 9B and S6). Additionally, ZNF703 (poor prognostic gene) was mainly expressed in stromal components, specifically in fibroblasts and pericytes (Figs 9B and S6). These observations indicate that CCJ signature genes exhibit heterogeneous cell type-specific expression patterns, reflecting contributions from both tumor-intrinsic epithelial cells and compartments of the tumor microenvironment.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Expression pattern of CCJ signature genes in cells of breast cancer tissues.

Dot plots visualize expression patterns of genes constituting the (A) LumB CCJ signature in an Her2 + /ER+ patient (n = 1) and (B) Basal CCJ signature in Triple-Negative Breast Cancer (TNBC) patients (n = 7), from the GSE176078 dataset. The x-axis lists the signature genes, and the y-axis represents identified cell types. Dot size indicates the percentage of cells expressing the gene (“Percent Expressed”), while color intensity represents the average expression level (“Average Expression”).

https://doi.org/10.1371/journal.pone.0352254.g009