Introduction

Breast cancer remains a clinical challenge1. Despite improvements in care, a significant proportion of breast cancer patients relapse2,3. During the past decade, molecular profiling of tumors has been implemented in the clinical setting to improve prognostication and treatment selection3,4,5,6. One example is the PAM50 Risk of Recurrence (ROR) score classification is now clinically used worldwide for a subgroup of breast cancer. The PAM50 ROR score is validated for postmenopausal breast cancer patients with ER+ /HER2− tumors receiving five years of endocrine therapy. The ROR score provides prognostic information and can be used to select patients for adjuvant chemotherapy4,7,8. However, further prognostication and treatment prediction refinement is still needed3. Additional information may be gained by looking beyond the malignant cells of the tumor, which most molecular risk scores are based on, and incorporating information from the tumor microenvironment (TME).

The TME has gained increasing attention for its role in breast cancer development and treatment response9,10. It has become increasingly clear that the tumor depends on its surroundings to be able to grow, survive, and metastasize9,10. Currently, there are few prognostic markers derived from the TME and these markers are mostly related to immune cells, such as tumor infiltrating lymphocytes9,10,11. Stromal cells, in particular cancer-associated fibroblasts (CAFs), constitute a significant part of the TME and play a key role in modulating various processes such as epithelial-mesenchymal transition (EMT), hypoxia, and angiogenesis, all important for the development of metastasis12,13.

Challenges remain in translating findings related to the TME into relevant biomarkers useful in clinical practice, since the TME is highly heterogenous and comprises several distinct cell types9,10,14. Depending on the composition of the TME and its interplay with malignant cells, the TME can either be tumor promoting or suppressing9,10,12. An interesting biomarker of an active TME is Caveolin-1 (CAV1). CAV1 is a master regulator of cell signaling and vesicular transport and is located in cholesterol-rich plasma membrane raft domains, known as caveolae15,16. CAV1 modulates key pathogenic processes involving the TME, including drug internalization, tumor-stroma interactions, hypoxia response, cellular metabolism, inflammation, and EMT15,16,17.

Furthermore, studies have reported that CAV1 protein expression in stromal cells may serve as a prognostic biomarker in breast cancer16,18,19,20,21. However, the prognostic impact of CAV1 is both context and localization dependent as previously reported18,19. It is unclear how the interplay between CAV1 and clinically used molecular risk scores, such as PAM50 ROR, relates to prognosis. Herein, we investigate the role of CAV1 gene expression in relation to PAM50 subtypes, ROR scores, and their joint impact on clinical outcome in two large breast cancer cohorts.

Results

Relationship with clinicopathological variables

For SCAN-B, 5326 of 7743 patients were assessed in the analysis, Fig. 1. For METABRIC the entire dataset of 1980 patients was used. All samples in both SCAN-B and METABRIC expressed CAV1. The distribution of CAV1 across PAM50 subtypes was similar for both cohorts, with the CAV1 expression being highest in Normal-like and followed by Luminal A subtype (both Ps < 0.001), Fig. 2A,B. Likewise, the correlations between CAV1 expression and the eight gene modules were similar in both cohorts, showing strong correlations between CAV1 expression and the Lipid and Stroma modules and ROR category dependent correlation with the Steroid and Immune response modules, Supplementary file 1: Supplementary Fig. 1. Notably, there was an inverse correlation with ROR category in both cohorts (both r <  − 0.34 and P < 0.001), Fig. 2C,D, Supplementary file 1: Supplementary Fig. 1. Even after adjusting for other predictors of PAM50 ROR category, the highest expression of CAV1 (T3) was strongly negatively associated with ROR high in both SCAN-B (adjusted OR 0.26 95% CI 0.19–0.35, P < 0.001) and METABRIC (adjusted OR 0.32 95% CI 0.21–0.47, P < 0.001), Supplementary file 1: Supplementary Fig. 2. CAV1 gene expression was also negatively correlated with most genes that are a part of the PAM50 ROR, except for KRT14, KRT15, KRT5, and SFRP1, Supplementary file 1: Supplementary Fig. 2. CAV1 tertiles in relation to clinicopathological factors are presented in Table 1 for SCAN-B and METABRIC.

Figure 1
figure 1

Flowchart of included and excluded patients in SCAN-B.

Full size image
Figure 2
figure 2

CAV1 expression by PAM50 and ROR category. CAV1 expression (continuous) by PAM50 molecular subtype in SCAN-B (A) and METABRIC (B). CAV1 expression (continuous) by PAM50 ROR category in SCAN-B (C) and METABRIC (D). Kaplan–Meier estimates of PAM50 ROR category among all patients in relation to distant metastasis-free interval in SCAN-B (E) and breast cancer-specific survival in METABRIC (F). Kaplan–Meier estimates of combined ROR category and CAV1 expression (in tertiles) among all patients in relation to distant metastasis-free interval in SCAN-B (G) and breast cancer-specific survival in METABRIC (H). The number of patients is indicated at each time-point.

Full size image
Table 1 Descriptive statistics of CAV1 tertiles in relation to clinicopathological factors in SCAN-B and METABRIC.
Full size table

Survival analysis

In SCAN-B, the median follow-up for the 4158 patients still at risk was 5.45 (IQR 5.07–8.15) years. Follow-up was restricted to ten years in METABRIC and all events after ten years were censored. This was done for two reasons; to make METABRIC more comparable to SCAN-B and because the PAM50 ROR score was developed to predict the risk of distant metastasis within 10 years22. The median follow-up for the 1089 patients still at risk in METABRIC was 10.0 years (IQR 10.0–10.0). The hazards were proportional for the tertiles for all endpoints.

In the univariable survival analyses of the complete cohorts, patients with ROR high had an increased risk of distant metastasis and breast cancer-specific survival compared to ROR low as expected, Fig. 2E,F. Moreover, the addition of CAV1 expression further stratified the distant metastasis-risk and breast cancer-specific survival in the univariable models Fig. 2G,H. The highest expression of CAV1 (T3) was associated with lower risk of recurrence, HR 0.74 (95% CI 0.60–0.92) in SCAN-B and HR 0.66 (95% CI 0.54–0.80) in METABRIC; lower risk of distant metastasis, HR 0.65 (95% CI 0.51–0.84) in SCAN-B and HR 0.61 (95% CI 0.50–0.76) in METABRIC; and lower risk of death, HR 0.68 (95% CI 0.58–0.80) in SCAN-B and HR 0.70 (95% CI 0.59–0.84) in METABRIC, Fig. 3. The highest expression of CAV1 (T3) also conferred a lower risk of breast cancer-related death HR 0.70 (95% CI 0.57–0.87) in METABRIC, Fig. 3.

Figure 3
figure 3

Univariable survival analyses of CAV1 expression. Kaplan–Meier estimates of CAV1 expression (in tertiles) among all patients in relation to recurrence-free interval in SCAN-B (A) and METABRIC (B), distant metastasis-free interval in SCAN-B (C) and METABRIC (D), overall survival in SCAN-B (E) and METABRIC (F), and breast cancer-specific survival (G). The number of patients is indicated at each time-point.

Full size image

In the multivariable analyses, the highest expression of CAV1 (T3) in SCAN-B instead conferred an increased risk of recurrence and distant metastasis but not death, Fig. 2. and Supplementary file 1: Supplementary Fig. 3. CAV1 tertiles were therefore adjusted for each variable used in the multivariable model, one at a time to see which variable affected the hazard ratio the most, which was the ROR category. Subsequently, interaction analyses were performed between ROR category and CAV1 tertiles on RFI and DMFI, revealing significant additive interactions and effect modifications of both ROR category on CAV1 and vice versa, Table 2. In SCAN-B, when stratifying by ROR category, the highest expression of CAV1 (T3) conferred increased risk of recurrence adjusted HR 1.57 (95% CI 1.10–2.24) and distant metastasis adjusted HR 1.60 (95% CI 1.08–2.37) only in patients with tumors classified as ROR High but not in ROR Low/Intermediate tumors, Supplementary file 1: Supplementary Fig. 4 and 5.

Table 2 Full report of interactions between ROR High and CAV1 T3 on DMFI.
Full size table

The distribution of ROR categories in the two cohorts differed with a larger proportion of tumors classified as ROR High in METABRIC than in SCAN-B. Considering the interactions between ROR categories and CAV1 expression in SCAN-B, interaction analyses between ROR categories and CAV1 tertiles for all four endpoints were performed to investigate if potential effect modifications previously seen in SCAN-B was the underlying reason for the discrepant findings on prognosis in METABRIC. (In METABRIC, associations between CAV1 tertiles and either one of the endpoints were not statistically significant in the multivariable analysis, Supplementary file 1: Supplementary Fig. 3.) Subsequently, it became clear that there were significant multiplicate and additive interactions between CAV1 T3 and ROR category on DMFI, OS, and BCSS in METABRIC, Table 2 and Supplementary file 2: Supplementary Table 1, 2, and 3. Similarly, there were effect modifications of both ROR category on CAV1 and vice versa concerning prognosis, Table 2 and Supplementary file 1: Supplementary Table 1, 2, and 3. Interestingly, in both SCAN-B and METABRIC, CAV1 tertiles could identify tumors where the predictive potential of ROR was the highest, Table 2 and Supplementary file 1: Supplementary Table 1, 2, and 3. Similar to DMFI in SCAN-B, when stratifying by ROR category in METABRIC, the highest expression of CAV1 (T3) conferred borderline increased risk of breast cancer-specific death adjusted HR 1.24 (95% CI 0.95–1.62) only in patients with tumors classified as ROR High but not in ROR Low/Intermediate tumors, Table 2, Fig. 2, Supplementary file 1: Supplementary Fig. 6 and 7. The additive interaction was present in both cohorts but was stronger in SCAN-B RERI 1.18 (95% CI 0.10–2.26, P = 0.016) than in METABRIC RERI 0.54 (95% CI 0.20–0.89, P = 0.001) Table 2. For METABRIC, the additive interaction was even stronger when breast cancer-specific survival was used as endpoint RERI 0.77 (95% CI 0.43–1.12, P < 0.001), Fig. 2H. Furthermore, CAV1 tertiles could also delineate in which group the ROR category was prognostic. In the CAV1 T1 tumors, the ROR category did not predict risk of distant metastasis, Supplementary file 1: Supplementary Fig. 8. In CAV1 T2 tumors, the ROR category predicted distant metastasis risk in SCAN-B but not METABRIC Supplementary file 1: Supplementary Fig. 9. For the CAV1 T3 tumors, ROR category was strongly associated with distant metastasis risk in both SCAN-B and METABRIC Supplementary file 1: Supplementary Fig. 10.

DGE and GSEA analysis for CAV1 Tertile 3 vs Tertile 1

To elucidate potential biological explanations behind the differential impact of CAV1 according to ROR category, DGE analyses were performed separately in ROR categories for tumors with the highest (T3) versus the lowest (T1) CAV1 expression.

In ROR high tumors, a total of 223 genes were found to be upregulated in high expressing (T3) vs low expressing (T1) CAV1 tumors, and no genes were downregulated, Supplementary file 2: Supplementary Table 4. Notably, several other genes coding for proteins involved in caveolae formation, e.g. CAV2 and CAVIN2 were higher expressed in ROR high/CAV1 high tumors, supporting a potential association with caveolae abundance. In ROR Low/Intermediate tumors, 450 genes were upregulated, and 18 genes were downregulated in high expressing (T3) vs low expressing (T1) CAV1 tumors, Supplementary file 2: Supplementary Table 5. In both ROR categories, genes related to stromal activation, EMT, CAFs and adipogenesis (SOX10, STAC2, FGF2, PTGFR, IGF1, IGF2, GDF10, ADAM33, CD36, PLIN4, PLIN1, MME, PENK, among others) were upregulated, Supplementary file 2: Supplementary Table 4 and 5. Only in ROR low/intermediate tumors were some potential tumor suppressor genes down-regulated (EEF1A2, GRM4, ROBO2 CHGB, CEACAM5, among others), Supplementary file 2: Supplementary Table 5.

Significantly enriched gene sets in high expressing (T3) CAV1 tumors in both ROR categories included EMT, TGF-β signaling, fatty acid metabolism, hypoxia, myogenesis, angiogenesis, xenobiotic metabolism among others, Fig. 4 and Supplementary file 2: Supplementary Tables 6 and 7. In low expressing (T1) CAV1 tumors regardless of ROR category, the MYC targets gene set was enriched, Fig. 4 and Supplementary file 2: Supplementary Tables 6 and 7. The main differences in gene set enrichment between high expressing (T3) CAV1 tumors and low expressing (T1) CAV1 tumors were related to immune response. Among high expressing (T3) CAV1 tumors, interferon-α response and complement hallmarks were enriched only in ROR high, while interferon-γ response hallmark were enriched only in ROR low/intermediate, Fig. 4 and Supplementary file 2: Supplementary Tables 6 and 7. Similar patterns were seen regarding GO terms, Supplementary file 2: Supplementary Tables 8 and 9.

Figure 4
figure 4

Molecular analyses of CAV1 expression. Volcano plot showing significant up- and downregulated genes (red) in high expressing (T3) in relation to low expressing (T1) CAV1 separately by ROR High category (A) and ROR Low/Intermediate category (B). Gene Set Enrichment Analysis (GO categories) of genes ranked by fold change (log2FC) and p-value < 0.05, up- and downregulated genes (red) in high expressing (T3) in relation to low expressing (T1) CAV1 tumors separately by ROR High category (C) and ROR Low/Intermediate category (D). Categories found enriched in both subgroup analyses are indicated by red text. Dot plot showing activated and suppressed Hallmark Signatures in high expressing (T3) in relation to low expressing (T1) CAV1 tumors separately by ROR High category (E) and ROR Low/Intermediate category (F).

Full size image

Discussion

Herein, we report that high CAV1 gene expression conferred an especially poor prognosis in patients whose tumors were classified as ROR high. In addition, ROR gave markedly different prognostic information depending on the underlying CAV1 expression, even after taking PAM50 subtype, other clinical predictors, and treatments into account. To our knowledge, this is the first study to examine the CAV1 mRNA gene expression in relation to molecular subtypes and prognosis in large breast cancer cohorts.

Moreover, CAV1 expression was associated with extracellular matrix remodeling, EMT myogenesis, hypoxia, angiogenesis, and stromal activation in both ROR High and Low/Intermediate classified tumors, as corroborated by both GSEA results and their correlations with the stromal gene module. It is known that CAV1 can remodel the extra cellular matrix through activation of stromal cells, elongating and facilitating invasion and metastasis23. Functionally, alterations in caveolae in stromal cells of the TME promote paracrine tumor growth via TGFβ, which activates EMT and myofibroblast differentiation, favoring tumor growth and metastasis24,25. EMT and myogenesis are markers of increased cell motility and loss of adhesion, both required for metastasis26.

Furthermore, CAV1 is linked to angiogenesis, endothelial permeability, and vascular endothelial growth factor (VEGF) response, which is required for tumor survival and the ability to enter the circulation. However, the exact role of CAV1 is unclear27. Hypoxia and angiogenesis are interlinked, and CAV1 is a direct transcriptional target of hypoxia-inducible factors 1α and 2α (HIF1α and 2α) that lead to increased dimerization and phosphorylation of the epidermal growth factor receptor (EGFR) conferring enhancing proliferative, migratory, and invasive capacities of malignant cells28. Further, hypoxia induces metabolic reprogramming in the tumor and CAV1 alterations confer a shift from mitochondrial respiration to tumor promoting aerobic glycolysis through attenuation of MYC expression29, corroborated by our data as seen in the downregulation of MYC response.

So far, the role of CAV1 in the immunomodulatory properties of the TME remains unexplored and further studies are warranted. The types of immune signals enriched in high expressing CAV1 tumors were dependent on ROR category and activation of immune response appeared higher in ROR low/intermediate tumors. The stromal microenvironment has immunomodulatory functions and can inhibit immune cells and decrease their efficacy in targeting and killing malignant cells, through regulation of extravasation and local immune cell replication9,10,11.

Our findings and existing literature suggest that CAV1 promotes metastasis and relapse through several critical pathways regardless of genomic risk classification. CAV1 can be considered as an essential protein that regulates paracrine signaling and the interplay between the malignant cells and TME. However, CAV1 expression only yielded additional prognostic information in tumors considered ROR high. A potential explanation for this finding might be that malignant cells that already acquired the intrinsic potential to metastasize still need an active and tumor promoting environment to do so. It might explain why high CAV1 expression in tumors identified patients for whom the ROR score provided most prognostic information. In contrast, the ROR score only gave little prognostic information in patients whose tumors had low CAV1 expression. It has been hypothesized that both an active tumor promoting TME, and oncogenic intrinsic features of malignant cells are needed for the tumor to be able to metastasize9,10, which is in line with our findings.

Our study examined mRNA rather than protein levels. In addition to mRNA expression, protein levels are also affected by translation, post-translational modifications, and regulation of the rate of protein decay30. The global correlation between mRNA and protein is expected to be high30. We have previously reported a correlation of Rs = 0.47 for CAV118. Consequently, the present study results must be interpreted in the context of the biological phenotype related to high CAV1 mRNA expression.

The standard treatment regimens differ between SCAN-B and METABRIC31,32,33,34,35 mainly due to samples being collected during different time periods31,34,35. Therefore, there is a large discrepancy in the type of treatments between the older METABRIC and the contemporary SCAN-B cohort. Differences in treatments could explain why the results regarding prognosis were not fully replicated. CAV1 has been shown to modulate treatment efficacy of chemotherapy (including epirubicin and taxanes) and trastuzumab in breast cancer and other cancers36,37,38,39. These treatments were rarely or not at all used in METABRIC. Our findings in SCAN-B may partly be explained by how CAV1 modulates these treatments since patients with ROR high tumors are more likely to receive chemotherapy and trastuzumab, where treatment efficacy partly depends on CAV1 expression. Unfortunately, a lack of more detailed information on treatments makes it hard to evaluate the role of CAV1 expression in response to specific treatments in our study. Assessment of CAV1 expression in tumor samples from previous randomized clinical trials is warranted to confirm whether CAV1 expression may further refine ROR score prediction.

It should be mentioned that the PAM50 ROR score is used clinically for risk prediction in postmenopausal patients with ER+ /HER2− tumors to identify patients where the recurrence risk is low enough to omit chemotherapy4,7,8. However, studies have shown that in ER negative disease, (ER− /HER2+ and TNBC), both PAM50 subtype and ROR score can predict neoadjuvant treatment response40,41,42,43. Further, in TNBC disease, PAM50 subtype categorization could predict sensitivity to taxanes and capecitabine treatment44. Similarly, PAM50 subtype and ROR score were shown to be prognostic in HER2+ disease45. Therefore, PAM50 ROR could play a role in more than one clinical subgroup of breast cancer, and we believe it is of interest to study the ROR score in a broader context. Our data indicate that CAV1 expression could identify tumors where ROR score was prognostic, potentially broadening the applicability of the PAM50 ROR score, beyond the subgroup of ER+ /HER2− tumors.

Beyond the type of treatment, there could be several other reasons why the results regarding prognosis were not precisely replicated in the METABRIC. First, the METABRIC is a smaller cohort, hence, sample sizes in the tests are smaller and potential survival associations may not be as readily detectable. METABRIC also consists of more advanced tumors and is not population-based46, which is reflected by differences between the cohorts in the distribution of clinicopathological variables, including ROR category. Since nodal status is the key factor for determining ROR category and nodal status was substantially higher in METABRIC than in SCAN-B, this fact may in part explain why the prognostic impact of CAV1 gene expression differed somewhat between the cohorts. The underlying risk of recurrence and death in METABRIC is considerably higher than in SCAN-B, making direct comparisons regarding prognosis difficult46. The derived CAV1 tertile classifications are relative to a population and not based on absolute cut-offs for each tumor. The tertile cut-offs were applied separately for each cohort, meaning that some tumors would be reclassified if a unform cut-off had been applied. One might expect that relatively more tumors in METABRIC would have be classified as low CAV1 expressing due to the inverse association between CAV1 expression and tumor aggressiveness46.

It should be noted that the gene expression data is derived from bulk tumors, which reflects the averaged gene expression across thousands of cells and different cell types47. Therefore, it was not possible to definitively infer which cell types CAV1 was located in47 and, subsequently, the role CAV1 plays in these cell types in the context of breast cancer. Unfortunately, to date, no available assays can apply single-cell resolution for large-scale cohorts such as SCAN-B and METABRIC to evaluate prognostic biomarkers.

Nonetheless, the population-based contemporary SCAN-B cohort offers unique advantages, which lies in the large-scale RNAseq analysis of consecutively enrolled breast cancers32,33,48. To our knowledge, this is the largest cohort of its kind to date. Due to a rigorous population-based approach with consistently high inclusion rates because of seamless integration of patient enrollment and tissue sampling incorporated into routine clinical practice, the study cohort can be considered representative of the general patient demographics in the catchment area32,33,48. Therefore, SCAN-B allows for the evaluation of biomarkers in a contemporary real-world setting. Further, most findings were confirmed, showing stable associations of CAV1 expression with clinicopathological factors and tumor biology, consistent with the literature. In both cohorts, similar additive interactions with ROR regarding clinical outcome were shown as well as the underlying CAV1 expression being able to markedly change the prognostic information yielded by PAM50 ROR.

In conclusion, our findings indicate that high CAV1 gene expression is associated with a particularly poor prognosis in patients with ROR high tumors. As CAV1 can mediate between malignant cells and the TME it may also be a promising therapeutic target. The underlying CAV1 expression markedly modified the prognostic information provided by PAM50 ROR. We have shown in two independent datasets that PAM50 ROR was only prognostic in tumors with high CAV1 expression. Thus, CAV1 expression could be a useful biomarker that may enhance and further refine PAM50 ROR risk stratification for patients with ROR high tumors.

Materials and methods

SCAN-B

The Swedish Cancerome Analysis Network—Breast (SCAN-B: ClinicalTrials.gov ID NCT02306096) is an ongoing population-based study that have enrolled breast cancer patients at seven hospitals in South Sweden and two additional hospitals (Uppsala and Jönköping)32,33. The enrollment of patients is integrated in clinical routine33 and all patients with newly diagnosed or suspected breast cancer are invited to participate. The Swedish National Quality Registry for Breast Cancer is used for collection of clinicopathological data, treatment information, and follow-up32,33,48.

Sample collection followed established SCAN-B procedures and protocols32,33. In brief, the remaining fresh collected tumor samples from surgical specimens were preserved in RNAlater (Qiagen, Hilden, Germany). Core needle biopsies were taken before neoadjuvant treatment and preserved in RNAlater. Gene expression profiling of the tumors was performed by massive parallel paired-end sequencing of mRNA (RNA-seq) using a custom SCAN-B workflow32,48. Details on library preparation, quality control, the analysis pipeline, and software used are described elsewhere32,33,48.

All clinicopathological data and gene expression data for SCAN-B patients used here were downloaded from the Supplementary Information and Data from Staaf et al.48. Expression levels were expressed in fragments per kilobase of exon per million mapped reads (FPKM) in an expression matrix48. To all FPKM data an offset of + 0.1 was added, and then the data was log2 transformed.

Patients were enrolled between September 1, 2010, and May 31, 2018, and followed until November 202148. From the beginning 7743 patients were included with a total of 8350 gene expression profiles (GEXs), as previously described48. After exclusion of GEXs from noninvasive cancer or lymph nodes, a total of 7142 patients remained with 7650 GEXs in the current study. In case multiple gene expression profiles from a single tumor passed quality control, the profile with the highest RNA concentration measured by NanoDrop spectrophotometry was chosen, as previously described48.This procedure left one GEX per patient for analysis. Further, patients with bilateral cancer or no available follow-up for distant metastasis were excluded, Fig. 1. After exclusions, GEX profiles from a total of 5326 patients were available for analysis. Information on PAM50 subtype and ROR category was obtained from Staaf et al.48 who assigned these categories using single sample predictors48.

Metabric

Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) is collection of clinically annotated primary fresh-frozen breast cancer specimens from five tumor banks in the UK and Canada34. The patients were diagnosed with non-metastatic breast cancer between 1977 and 200534. Manual curation and basic quality control of the clinicopathological data including treatment information was performed31. None of the HER2+ patients received trastuzumab. For a subset of 1980 patients, known as the METABRIC molecular dataset, gene-expression data from microarrays is available31,34,35. Details on sample handling, gene expression profiling and workflow are described elsewhere34,35. The METABRIC molecular dataset was downloaded from https://www.cbioportal.org/study/summary?id=brca_metabric and corresponding clinical data from Rueda et al.31. The genefu package49 was used to assign PAM50 subtype using nearest centroid correlation22 and calculate the PAM50 ROR score based on centroid correlations, tumor size and proliferation score according to the ROR equation with nodal status dependent cut-offs to assign categories, as described7,50,51.

In both SCAN-B and METABRIC, eight gene expression modules representing different biological functions in breast cancer were calculated as previously described52.

Statistical analysis

Differences in CAV1 mRNA expression depending on PAM50 subtype were evaluated using Kruskal–Wallis test. Correlations between CAV1 expression, ROR category and the eight gene modules52 were assessed using Pearson’s correlation (r). Pearson’s correlation was also used to assess correlations between CAV1 mRNA expression and mRNA expression of the PAM50 genes. Logistic regression was used to test whether CAV1 mRNA expression was independently associated with ROR category after adjusting for potential confounders (age at diagnosis, axillary lymph node status (pN1/2/3), tumor size (pT2/3/4), Grade (III vs I or II), ER+, PR+, HER2+, PAM50 subtype (Luminal A as reference), and (neo)adjuvant treatments.

Endpoints used for survival analysis were recurrence-free interval (RFI), distant metastasis-free interval (DMFI), and overall survival (OS) for both SCAN-B and METABRIC, as previously defined31,34,48. Breast cancer-specific survival (BCSS) was used as additional endpoint for METABRIC31,34.

For the survival analysis, log2 transformed CAV1 mRNA expression were categorized into tertiles, tertile 1 (T1), tertile 2 (T2), and tertile 3 (T3) to allow for non-linear effects. The lowest expression of CAV1 (T1) was used as reference. For survival analyses, the R packages ‘survival’ and ‘survminer’ were used.

Univariable survival analyses were performed using the Kaplan–Meier method and the Log-rank test. Cox proportional hazards models were used to obtain crude and adjusted Hazard ratios (HRs) with 95% confidence intervals (CI). The multivariable models were adjusted for age (binned in 5-year intervals for SCAN-B or continuous for METABRIC), tumor characteristics; axillary lymph node status (pN1/2/3), tumor size (pT2/3/4), Grade (III vs I or II), ER+, PR+, HER2+, PAM50 subtype (Luminal A as reference), PAM50 ROR category (High vs Low/Intermediate); and (neo)adjuvant treatments (endocrine treatment and chemotherapy for both SCAN-B and METABRIC and trastuzmab for SCAN-B only).

Schoenfeld’s residuals were used to test and graphically examine the proportional hazard assumption for the CAV1 tertiles in the adjusted model. To investigate effect modifications between the CAV1 tertiles and PAM50 ROR category, two-way interaction analyses on multiplicative and additive scales were performed in the multivariable model using the ‘interactionR’ package53.

Differential gene expression (DGE) analysis was conducted in SCAN-B using the ‘Limma-Voom’ package54 to find differentially expressed genes (DEGs) between the highest tertile (T3) and the lowest tertile (T1) of CAV1 expression. The criteria used to define DEGs is a false discovery rate (FDR) of ≤ 0.05 and log2 fold change (log2FC) ≥ 1.5 for up-regulated genes and log2FC ≤  − 1.5 for down-regulated genes. To correct for batch effects, batch was included in the Limma models. Gene set enrichment analysis (GSEA) was performed in ‘clusterprofiler’55 to find the statistically significant, concordant gene sets that differed between the highest tertile (T3) and the lowest tertile (T1) of CAV1 expression. Gene sets were grouped according to Gene Ontology (GO) and Hallmark Signature annotations56,57.

All statistical analyses were conducted in R version 4.2.2. P-values < 0.05 was considered statistically significant. All P-values were two-tailed. This study followed the Reporting Recommendations for Tumor Marker Prognostic Studies (REMARK) criteria58.

Ethics approval and consent to participate

Ethical approvals for the cohorts studied (SCAN-B and METABRIC) were obtained in relation to the primary projects and publications31,32,33,34,35,48. The SCAN-B study was approved by the Lund University ethics committee32,33,48. The METABRIC study was approved by the ethics committees at the University of Cambridge and the British Columbia Cancer Research Centre 31,34,35. All participants signed written informed consent. No separate approval was obtained for this specific study since it is based on previously published data. The study was conducted in accordance with the ethical principles of the Declaration of Helsinki.