Identification and validation of the cellular senescence-related molecular subtypes of triple negative breast cancer via integrating bulk and single-cell RNA sequencing data

Abstract

Patients with triple-negative breast cancer (TNBC) reportedly benefit from immune checkpoint blockade (ICB) therapy. However, the subtype-specific vulnerabilities of ICB in TNBC remain unclear. As the complex interplay between cellular senescence and anti-tumor immunity has been previously discussed, we aimed to identify markers related to cellular senescence that may serve as potential predictors of response to ICB in TNBC. We used three transcriptomic datasets derived from ICB-treated breast cancer samples at both scRNA-seq and bulk-RNA-seq levels to define the subtype-specific vulnerabilities of ICB in TNBC. Differences in the molecular features and immune cell infiltration among the different TNBC subtypes were further explored using two scRNA-seq, three bulk-RNA-seq, and two proteomic datasets. 18 TNBC samples were collected and utilized to verify the association between gene expression and immune cell infiltration by multiplex immunohistochemistry (mIHC). A specific type of cellular senescence was found to be significantly associated with response to ICB in TNBC. We employed the expression of four senescence-related genes, namely CDKN2A, CXCL10, CCND1, and IGF1R, to define a distinct senescence-related classifier using the non-negative matrix factorization approach. Two clusters were identified, namely the senescence-enriching cluster (C1; CDKN2A ^high CXCL10 ^high CCND1 ^low IGF1R ^low) and proliferating-enriching cluster (C2; CDKN2A ^low CXCL10 ^low CCND1 ^high IGF1R ^high). Our results indicated that the C1 cluster responds better to ICB and behaves with higher CD8⁺ T cell infiltration than the C2 cluster. Altogether, in this study, we developed a robust cellular senescence-related classifier of TNBC based on the expression of CDKN2A, CXCL10, CCND1, and IGF1R. This classifier act as a potential predictor of clinical outcomes and response to ICB.

Introduction

Immune checkpoint blockade (ICB) has led to great success in treating patients with several types of solid tumors, such as melanoma [1] and lung cancer [2]. However, only about 10-30% of patients benefit from it [3-5]. This observation highlights the importance of understanding molecular heterogeneity and biomarker research for identifying phenotypes that exhibit an improved response to ICB.

Combining ICB with chemotherapy has increased the response rates to therapy in patients with triple-negative breast cancer (TNBC) [6,7]. TNBC, with a high tumor mutational burden and high immune infiltration, responds better to ICB than luminal breast cancer [8,9]. Nevertheless, the benefits of ICB in TNBC are limited, and which TNBC subgroup may benefit the most from ICB has not yet been determined. Thus, precisely identifying biomarkers for patients with TNBC who respond well to ICB is crucial.

Cellular senescence is a permanent state of cell cycle arrest that plays a dual role in tumor immunity [10,11]. On the one hand, senescent cells exhibit a protective effect against tumorigenesis by enhancing immune clearance and tissue remodeling [12]. On the other hand, the senescence-associated secretory phenotype (SASP) factors released by the senescent cells suppress tumor immunity [13,14]. Therefore, further comprehensively analyzing the associations between cellular senescence and tumor immunity in TNBC is essential.

Infiltration of CD3⁺ T cells and CD8⁺ T cells, PD-1/PD-L1 expression, and tumor mutation burden are important determinants of response to ICB. SASP could be utilized by senescent tumor cells to influence the effect of ICB. The combination of ICB and senolytic has been used to increase immune surveillance in some special cancer types such as KRAS-mutant pancreatic ductal adenocarcinoma [15], CDK4/CDK6 inhibitor responsive melanoma [16], and brain metastatic breast cancer [17]. However, the relationship of senescence phenotype and response to ICB in TNBC is still unclear. As a result, exploring specific senescence-related biomarkers in TNBC patients may accurately identify patients who will benefit from ICB.

In this study, we aimed to define a distinct cellular senescence-related classifier and identify potential subtype-specific vulnerabilities of ICB in TNBC by integrated analysis of three scRNA-seq cohorts and multiple bulk RNA sequencing cohorts. The findings of our study may give aid to elucidate the senescence-related heterogeneity in TNBC at both the single-cell and bulk levels.

Materials and methods

ICB cohorts

To investigate and verify the association between cellular senescence and tumor immunity, breast cancer (BC) cohorts with ICB response or T cell expansion information and transcriptomic data were analyzed at bulk and single-cell levels. Data of the single-cell cohort (Bassez cohort 1) [18] were publicly available at http://biokey.lambrechtslab.org (Table 1). Data of the bulk cohorts were accessed through GEO accession numbers GSE173839 [19] and GSE124821 [20] (Table 1).

Table 1.

Information of datasets used in this study

Datasets	Available omics	ICB	Source
METABRIC cohort	Transcriptomics	No	Curtis et al., 2012
TCGA-BRCA cohort	Transcriptomics	No	Cancer Genome Atlas Network, 2012
GSE173839 cohort	Transcriptomics	Yes	Pusztai et al., 2021
GSE58812 cohort	Transcriptomics	No	Jezequel et al., 2015
GSE124821 cohort	Transcriptomics	Yes	Hollern et al., 2019
Bassez cohort 1	ScRNA-seq	Yes	Bassez et al., 2021
GSE75688 cohort	ScRNA-seq	No	Chung et al., 2017
GSE176078 cohort	ScRNA-seq	No	Wu et al., 2021
PDC000120 cohort	Proteomics	No	Krug et al., 2020
PDC000173 cohort	Proteomics	No	Mertins et al., 2016

ICB, immune checkpoint blockade.

Non-ICB cohorts

To compare the immune cell infiltration and senescence heterogeneity between two clusters, multiple cohorts without ICB information were analyzed at transcriptomic and proteomic levels. Transcriptomic data of The Cancer Genome Atlas-BRCA (TCGA-BRCA) cohort were downloaded from the UCSC Xena data portal (https://xenabrowser.net) [21], while those of the METABRIC cohort were downloaded from cBioPortal (https://www.cbioportal.org) [22-24] (Table 1). Data of the other transcriptomic cohorts at the bulk and single-cell levels were accessed through GEO accession numbers GSE58812 (bulk) [25], GSE75688 (single-cell) [26], and GSE176078 (single-cell) [27] (Table 1). Data from two proteomic cohorts were downloaded from the National Cancer Institute’s Clinical Proteomic Tumor Analysis Consortium (CPTAC) portal (https://proteomic.datacommons.cancer.gov/pdc/) through accession numbers PDC000120 [28] and PDC000173 [29] (Table 1). Two proteomic cohorts were merged as an independent proteomic cohort, and “combat” was used to remove the batch effect.

scRNA-seq data analysis

Bassez cohort 1 and GSE75688 datasets contain annotated cell types of each sample. Data of cancer cells from patients with TNBC were analyzed in the Bassez cohort 1 dataset, whereas those of cancer cells from all patients with BC were analyzed in the GSE75688 dataset. In the GSE176078 dataset, data from patients with TNBC were analyzed using the Seurat package [30]. Epithelial cells could be recognized based on the following characteristics: KRT19+ and PTPRC-. In contrast, CD4⁺ T cells, CD8⁺ T cells, naïve T cells, and Treg cells were classified and recognized by Uniform Manifold Approximation and Projection (UMAP) [31] based on CD4, CD8A, CCR7, and FOXP3 expression, respectively. The “harmony” package was used to remove the batch effect [32]. UCell method was used to calculate the subtype score based on specific genes [33].

Transcriptomic and proteomic data analysis at the bulk level

The “limma” package [34] was used to identify differentially expressed genes (DEGs) between ICB-treated patients with pathological complete remission (pCR) and those without pCR in the GSE173839 dataset. Genes with |log fold change| ≥ 0.85 and false discovery rate < 0.05 were regarded as DEGs. We used DEGs as input to perform the Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis via Metascape online tools [35] (https://metascape.org/gp/index.html#/main/step1) with minimum overlap = 5, minimum enrichment = 3, and P-value < 0.05. The senescence-related gene list comprised genes from hsa04218, GO: 0090398, and a core panel of frailty biomarkers [36] (Table S1). Transcriptomic data of the GSE173839 dataset was used to define distinct senescence-related classifiers by the non-negative matrix factorization (NMF) approach [37]. Furthermore, transcriptomic data of the TCGA-TNBC, METABRIC-TNBC, GSE58812, GSE75688, GSE124821 datasets, and proteomic data were used to verify the classifiers. The “IOBR” package [38] was used to evaluate immune cell infiltration in the GSE173839, TCGA-TNBC, METABRIC-TNBC, GSE58812, and GSE75688 datasets by performing CIBERSORT [39], ESTIMATE [40], Xcell [41], MCPCounter [42], quanTIseq [43], EPIC [41], and TIMER [44] analysis. Gene Set Enrichment Analysis (GSEA) analysis [45] was employed to identify the cluster exhibiting senescence-enriched phenotypes in the TCGA-TNBC, METABRIC-TNBC, GSE58812, and proteomic datasets with the list of senescence-related gene set downloaded from KEGG website (Table S1: hsa04218). The immunophenoscores (IPS) of the TCGA-TNBC cohort used to predict the efficacy of ICB were downloaded from The Cancer Immunome Atlas database (https://tcia.at/home) [46,47].

Patients and tissue microarray specimens

Formalin-fixed paraffin-embedded tissue blocks were collected from the Affiliated Hospital of Jining Medical University, China. Tissues were collected from patients with TNBC who received no chemotherapy or radiotherapy. For tissue microarray construction, all specimens were re-evaluated using hematoxylin and eosin staining, and the representative areas were selected and constructed into 2.0 mm tissue cores. In this study, a total of 18 cases were analyzed (Table S2). All of the research was reviewed and approved by the Ethics Committee of the Affiliated Hospital of Jining Medical University (approval number: 2021-08-C015). Informed consent was obtained from all subjects involved in the study.

mIHC assay development and staining

Tissue microarray slides were dewaxed in xylene and rehydrated by graded ethanol solutions, followed by 1-hour baking at 65°C. Antigen retrieval was performed using a high-pressure heat method with sodium citrate solution (pH = 6). After cooling down to room temperature, the slides were blocked and incubated with the IGF1R antibody (Abcam, ab263903, 1:600) at room temperature for 1 h. Subsequently, secondary antibodies and Opal 520 fluorophores (NEL811001KT, Akoya) were incubated at room temperature for 10 min after washing with TBST thrice. Slides stained using IGF1R-Opal 520 were retrieved by microwave oven with sodium citrate solution (pH = 6). After cooling down to room temperature, the slides were blocked and incubated with the other three antibodies and Opal fluorophores (CXCL10-Opal 690, PTG, 10937-1-AP, 1:300; CD3-570, PTG, 17617-1-AP, 1:1000; CD8-620, PTG, 66868-1-Ig, 1:6000). Staining was performed using DAPI as per the standard procedure. mIHC Slides were scanned by Vectra Polaris and pictures were captured by Vectra Polaris 1.0.10. The ratio of marker-positive cells was calculated by QuPath software [48].

Statistical analysis

Statistical analyses were generated using R v4.1.3 (https://www.r-project.org) or GraphPad Prism 9 (https://www.graphpad.com/). A comparison of immune cell infiltration scores and expression of SASP factors between two groups was analyzed using two-sided Wilcoxon tests. Categorical variables were compared between two groups or more than two groups using the Chi-square test or Fisher’s exact test. The receiver operating characteristic (ROC) curve was used to evaluate predictive performance. Kaplan-Meier plots of OS, progression-free interval (PFI), and metastasis-free survival (MFS) were performed using GraphPad Prism 9. Values of P < 0.05 using the Gehan-Breslow-Wilcoxon test were used to define differences in survival time.

Results

Identifying core senescence-related genes that correlated with response to ICB

To investigate the biomarkers for ICB, we identified DEGs between data of patients who responded to ICB and those who did not respond to ICB in the GSE173839 dataset (Figure 1A). For analysis of pathway enrichment using DEGs as input, pathways such as IL17 signaling pathway, p53 signaling pathway, chemokine signaling pathway, and cellular senescence that had been reported to regulate the immune system and inflammatory response in the progressive of various diseases were enriched (Figure 1B). As reported before, cellular senescence was mediated by the p53 signal pathway [49,50] and SASP was mediated by the IL17 signaling pathway [51,52] and the chemokine signaling pathway [53,54]. Although the senescence-related pathway is not the most significantly enriched in the list, the other pathways are connected with the cellular senescence signal pathway closely. As a result, the association between cellular senescence and ICB response was further explored. The six most significant senescence-related DEGs (Figure 1C) were selected for further validation in the scRNA-seq cohort (Figure 1D). Among them, the expression of CDKN2A, CXCL10, CCND1, and IGF1R in the TNBC cancer cells was associated with T cell expansion (Figure 1E). UCell score based on the four genes of each cancer cell was remarkablely different between the TNBC samples with (E group) and without (NE group) T cell expansion (Figure 1F). This result demonstrated that the E group had a higher four-gene score than the NE group (Figure 1G).

Figure 1.

Identification of genes correlated with immune checkpoint blockade response. A. Volcano plot depicted the differentially expressed genes (DEGs) between patients in the pathological complete regression (pCR) group and that in the non-pCR group using transcriptomics data from the GSE173839 dataset. B. Bubble plot of Kyoto Encyclopedia of Genes and Genomes analysis using DEGs as input showed that cellular senescence pathway was enriched. C. Six senescence-related genes, including CDKN2A, CXCL10, CCND1, IGF1R, CXCL8, and CCNE, were identified as DEGs in the GSE173839 dataset. D. Uniform Manifold Approximation and Projection (UMAP) plot of cancer cells between triple-negative breast cancer (TNBC) patients with T cell expansion (E) and TNBC patients without T cell expansion (NE) using scRNA-seq data from Bassez cohort 1. E. UMAP plot showed that the expressions of CDKN2A, CXCL10, CCND1, and IGF1R were significant difference between cancer cells in the E group and cancer cells in the NE group. F. UMAP plot of UCell score based on the expressions of CDKN2A, CXCL10, CCND1, and IGF1R in cancer cells derived from TNBC patients. G. Cancer cells in the E group had stronger UCell scores based on the expressions of CDKN2A, CXCL10, CCND1, and IGF1R than cancer cells in the NE group. DEGs, differentially expressed genes; pCR, pathological complete regression; Uniform Manifold Approximation and Projection, UMAP; triple-negative breast cancer, TNBC; ****P < 0.0001.

Defining and verifying a classifier of TNBC based on CDKN2A, CXCL10, CCND1, and IGF1R expression

Consensus clustering analysis of the NMF algorithm was used to identify distinct senescence pattern clusters based on the expression of CDKN2A, CXCL10, CCND1, and IGF1R in the GSE173839 dataset. Two clusters were identified as k = 2 when the magnitude of the cophenetic correlation coefficient began to decrease, including 41 cases in cluster C1 and 30 cases in cluster C2 (Figure 2A). The heatmap plot exhibited the consensus matrix of NMF clustering results using transcriptomic data in the GSE173839 dataset (Figure 2B). The C1 cluster exhibited higher expression of CDKN2A and CXCL10 and lower expression of CCND1 and IGF1R than the C2 cluster (Figure 2C). Meanwhile, the proportion of patients with pCR was higher in the C1 cluster than in the C2 cluster (Figure 2D). The immune infiltration score calculated using ESTIMATE (Figure 2E) and xCELL (Figure 2F) indicated that the C1 cluster had more immune activity than the C2 cluster. For further validation, three transcriptomic data samples of the TNBC cohorts were used to verify the senescence pattern clusters constructed based on the GSE173839 cohort. The results indicated that the C1 cluster had higher expression of CDKN2A and CXCL10 and lower expression of CCND1 and IGF1R than the C2 cluster in protein level in the TCGA-TNBC, GSE58821, and METABRIC-TNBC cohorts (Figure 3A-C). Similarly, the C1 cluster displayed higher immune infiltration levels than the C2 cluster in three cohorts (Figure 3D-I). Moreover, patients in the C1 cluster had better OS (Figure 3J) and PFI (Figure 3K) than those in the C2 cluster in the TCGA-TNBC cohort. Furthermore, patients in the C1 cluster had better OS (Figure 3L) and MFS (Figure 3M) than those in the C2 cluster in the GSE58821 cohort.

Figure 2.

Development of senescence-related classifier by Non-negative Matrix Factorization Approach (NMF) using data from the GSE173839 dataset. A. NMF rank survey indicated the number of clusters should be equal to 2. B. The heatmap plot exhibited the consensus matrix of NMF clustering results. C. The heatmap plot exhibited the expressions of CDKN2A, CXCL10, CCND1, and IGF1R between C1 and C2. D. The bar plot showed the percent of patients with pathological complete regression (pCR) in C1 was higher than that in C2. E. The bar plot showed patients in C1 had higher ESTIMATE immune scores than that in C2. F. The bar plot showed patients in C1 had higher xCell immune scores than that in C2. NMF, Non-negative Matrix Factorization Approach; pCR, pathological complete regression; ***P < 0.001.

Figure 3.

Validation of the senescence-related classifier using data from three transcriptomics datasets at the bulk level. (A-C) The heatmap plot exhibited the expressions of CDKN2A, CXCL10, CCND1, and IGF1R between C1 and C2 using data from the TCGA-TNBC (A), GSE58812 (B), and METABRIC-TNBC (C) dataset, respectively. (D, E) The bar plot showed patients in C1 had higher ESTIMATE immune scores (D) and xCell immune scores (E) than that in C2 using data from the TCGA-TNBC dataset. (F, G) The bar plot showed patients in C1 had higher ESTIMATE immune scores (F) and xCell immune scores (G) than that in C2 using data from the GSE58812 dataset. (H, I) The bar plot showed patients in C1 had higher ESTIMATE immune scores (H) and xCell immune score (I) than that in C2 using data from the METABRIC-TNBC dataset. (J, K) The Kaplan-Meier curve plot showed patients in C1 had longer overall survival (OS) (J) and progress free survival (K) than that in C2 using data from the TCGA-TNBC dataset. (L, M) The Kaplan-Meier curve plot showed patients in C1 had longer OS (L) and metastasis free survival (M) than that in C2 using data from the GSE58812 dataset. TNBC, triple-negative breast cancer; OS, overall survival.

Verifying the classifier of TNBC at the single-cell level

Eleven primary BC tumor samples in the GSE75688 cohort were subjected to bulk RNA-seq and scRNA-seq. Firstly, we continuously divided these 11 patients into C1 and C2 clusters by NMF using data at the bulk level (Figure 4A). The C1 cluster was also found to contain a higher expression of CDKN2A and CXCL10, a higher immune score, and a lower expression of CCND1 and IGF1R than those in the C2 cluster (Figure 4A and 4B). Secondly, we evaluated the expression of four genes and the four-gene score of the same patients using data at the single-cell level (Figure 4C-E). The four-gene score was calculated by UCell algorithm which is the more appropriate method for evaluating the pathway or multiple-gene score when using data at the single-cell level. The patients with a higher four senescence-related gene score may be closer to patients with a senescence phenotype as well as patients in the C1 cluster. Our results demonstrated that cancer cells in the C1 cluster have a higher four-gene score than those in the C2 cluster at the single-cell level (Figure 4F), which further validated the robustness of our classifier.

Figure 4.

Validation of the senescence-related classifier using transcriptomics data at the single-cell level. A. The heatmap plot exhibited the expressions of CDKN2A, CXCL10, CCND1, and IGF1R between C1 and C2 using bulk-RNA-seq data from the GSE75688. B. The bar plot showed patients in C1 had higher ESTIMATE immune scores than that in C2 using bulk-RNA-seq data from the GSE75688. C. Uniform Manifold Approximation and Projection (UMAP) plot of cancer cells between C1 cluster and C2 cluster using scRNA-seq data from the GSE75688. D. UMAP plot showed the expressions of CDKN2A, CXCL10, CCND1, and IGF1R in cancer cells. E. UMAP plot of UCell score based on the expressions of CDKN2A, CXCL10, CCND1, and IGF1R in cancer cells. F. Cancer cells in C1 cluster had stronger UCell scores based on the expressions of CDKN2A, CXCL10, CCND1, and IGF1R than cancer cells in C2 cluster. UMAP, Uniform Manifold Approximation and Projection; ****P < 0.0001.

CD8⁺ T cell infiltration difference between patients with TNBC in C1 and C2 clusters

CIBERSORT, MCPCounter, quanTIseq, EPIC, and TIMER analysis were used to evaluate the infiltration abundance of CD8⁺ T cells in the TCGA-TNBC, GSE58812, and METABRIC-TNBC cohorts at the bulk level. Our results indicated that the C1 cluster had a higher proportion of CD8⁺ T cells than the C2 cluster in the TCGA-TNBC (Figure 5A), GSE58812 (Figure 5B), and METABRIC-TNBC (Figure 5C) cohorts. TNBC data in the GSE176078 cohort was analyzed at the single-cell level (Figure 6A). We first calculated the four-gene score of each PTPRC-KRT19+ epithelial cell based on the expression of CDKN2A, CXCL10, CCND1, and IGF1R by the UCell algorithm (Figure 6A-C). The cells with a four-gene score > 0 were regarded as senescence-positive cells and patients with a proportion of senescence-positive cells in total epithelial cells > 50% were categorized to the C1 cluster, whereas others were categorized to the C2 cluster (Figure 6C). We then classified and recognized CD4⁺ T cells, CD8⁺ T cells, naïve T cells, and Treg cells from total CD3⁺ T cells of patients in the C1 and C2 clusters (Figure 6D and 6E). The results indicated that patients with TNBC in the C1 cluster had a higher infiltration abundance of CD8⁺ T cells than those in the C2 cluster at the single-cell level (Figure 6F).

Figure 5.

Exploration of the infiltration of CD8⁺ T cells between C1 and C2 using bulk-RNA-seq data by MCPcounter, quanTiseq, CIBERSORT, EPIC and TIMER methods. A. The violin plot showed samples in C1 had higher infiltration level of CD8⁺ T cells than that in C2 using data from the TCGA-TNBC dataset. B. The violin plot showed samples in C1 had higher infiltration level of CD8⁺ T cells than that in C2 using data from the GSE58812 dataset. C. The violin plot showed samples in C1 had higher infiltration level of CD8⁺ T cells than that in C2 using data from the METABRIC-TNBC dataset. TNBC, triple-negative breast cancer.

Figure 6.

Exploration of the infiltration of CD8⁺ T cells between C1 and C2 using scRNA-seq data from the GSE176078 dataset. A. Uniform Manifold Approximation and Projection (UMAP) plot of epithelial cells derived from TNBC patients. B. UMAP plot of UCell scores based on the expressions of CDKN2A, CXCL10, CCND1, and IGF1R in epithelial cells. C. The TNBC patients were divided as C1 and C2 clusters according to the percent of UCell-score-positive cells. D. UMPA plot of CD3⁺ T cells derived from TNBC patients in C1 cluster. E. UMPA plot of CD3⁺ T cells derived from TNBC patients in C2 cluster. F. The bar plot showed TNBC patients in C1 cluster had higher infiltration level of CD8⁺ T cells than that in C2 using data at the single-cell level. UMAP, Uniform Manifold Approximation and Projection; TNBC, triple-negative breast cancer.

Essentially, mIHC staining was conducted to verify the co-expression of CXCL10 and IGF1R with CD3 and CD8 in 18 TNBC PPFE samples. Here, CXCL10 was marked red, IGF1R was marked green, CD3 was marked orange, CD8 was marked yellow, and DAPI was marked blue (Figure 7A). Results indicated that the patient with CXCL10^highIGF1R^low had stronger staining intensity of CD3 and CD8 than the patient with CXCL10^lowIGF1R^high (Figure 7A). In addition, we found the ratio of CXCL10-positives cells is positively correlated with the ratio of CD3-positive cells (Figure 7B) and CD8-positive cells (Figure 7C). However, the ratio of IGF1R-positive cells is negatively correlated with the ratio of CD3-positive cells (Figure 7D) and CD8-positive cells (Figure 7E).

Figure 7.

Multiplex immunofluorescence staining analysis of CXCL10 and IGF1R co-expression with CD3 and CD8 in 18 triple-negative breast cancer samples. A. Two cases of multiplex immunofluorescence staining of CXCL10, IGF1R, CD3 and CD8. B. Correlation of CD3 with CXCL10. C. Correlation of CD8 with CXCL10. D. Correlation of CD3 with IGF1R. E. Correlation of CD8 with IGF1R.

Senescence-enriching phenotype of the C1 cluster

We identified biomarkers between epithelial cells from patients in the C1 cluster and epithelial cells from patients in the C2 clusters in the GSE176078-TNBC cohort (Figure 8A). KEGG analysis using these markers as input showed that the cellular senescence pathway was enriched at the single-cell level (Figure 8B). The GSEA analysis using data from the TCGA-TNBC, GSE58812, and METABRIC-TNBC cohorts revealed that the C1 cluster exhibited a senescence-enriching phenotype at the bulk level (Figure 8C). Similarly, data from one merged proteomic cohort were analyzed. We found that the C1 cluster exhibited a senescence-enriching phenotype and higher expression of CDKN2A and CXCL10 and lower expression of CCND1 and IGF1R than the C2 cluster at the protein level (Figure 8D).

Figure 8.

Identification of C1 cluster as a senescence-enriching phenotype. A. Uniform Manifold Approximation and Projection (UMAP) plot and volcano plot depicted the differentially expressed genes (DEGs) between epithelial cells in C1 cluster and epithelial cells in C2 cluster using scRNA-seq data from the GSE176078 dataset. B. Bubble plot of Kyoto Encyclopedia of Genes and Genomes analysis using DEGs as input also showed that cellular senescence pathway was enriched. C. Gene Set Enrichment Analysis (GSEA) plot showed cellular senescence pathway was enriched in C1 cluster using bulk-RNA-seq data from the TCGA-TNBC, GSE58812, and METABRIC-TNBC dataset, respectively. D. The heatmap plot exhibited the proteomics expression levels of CDKN2A, CXCL10, CCND1, and IGF1R between C1 and C2, and GSEA plot showed cellular senescence pathway was enriched in C1 cluster at the protein level. E. GSEA plot showed cellular senescence pathway was enriched in C1 cluster using data from the merged CPTAC BRCA dataset. F. The heatmap plot exhibited that the transcriptomics expression levels of some senescence-associated secretory phenotype (SASP) factors, such as CCL5, CCL8, IL1B, IL7, MMP7, ICAM1, ICAM3, TNFRSF1B etc., was higher in C1 cluster than that in C2 cluster using data from the TCGA-TNBC and GSE58812 dataset, respectively. G. The violin plot showed the proteomics expression levels of some SASP factors, such as CXCL10, IL18, ICAM1, and MMP12, was higher in C1 cluster than that in C2 cluster using data from the merged CPTAC BRCA dataset. UMAP, Manifold Approximation and Projection; DEGs, differentially expressed genes; GSEA, Gene Set Enrichment Analysis; TNBC, triple-negative breast cancer; SASP, senescence-associated secretory phenotype.

MMP12 were higher in the C1 cluster than in the C2 cluster (Figure 8F). Meanwhile, the expression of the growth factor IGFBP4 was lower in the C1 cluster than in the C2 cluster at the transcriptomic and proteomic levels (Figure 8E and 8F).

Sensitivity of CDKN2A^highCXCL10^highCCND1^lowIGF1R^low patients with TNBC to ICB

To validate the reliability of our classifier in predicting ICB efficacy, the IPS score of TCGA-TNBC samples and a mouse transcriptomic dataset (GSE124821) containing information on ICB treatment were analyzed. Our results showed that samples in the C1 cluster had higher IPS scores than those in the C2 cluster using data from the TCGA-TNBC cohort, indicating that the C1 cluster patients received more benefits from ICB than the C2 cluster patients (Figure 9A). Moreover, we found that the ICB-sensitive samples exhibited higher expression of Cdkn2a and Cxcl10 and lower expression of Ccnd1 and Igf1r than the ICB-resistant samples (Figure 9B) in the GSE124821 cohort. Besides, we divided samples into the C1 cluster (Cdkn2a ^high Cxcl10 ^high) and the C2 cluster (Ccnd1 ^high Igf1r^high) by NMF based on the mRNA expression of the four senescence-related genes in the GSE124821 cohort (Figure 9C). Our result demonstrated that a higher proportion of samples sensitive to ICB was seen in the C1 cluster than in the C2 cluster (Figure 9D). Then we performed ROC analysis using “pROC” package. In detail, the senescence-related cluster status (C1 or C2) and status of response to ICB (sensitive or resistant) of these samples were used for performing ROC analysis and calculating the area under the curve (AUC) value. The ROC analysis indicated that the classifier based on four senescence-related genes predicts the efficacy of ICB remarkably well (area under curve value = 0.79, 95% confidence interval: 0.84-0.74; Figure 9E). Similarly, the C1 cluster exhibited a senescence-enriching phenotype and higher expression of SASP factors, such as IL1a, IL1b, IL6, CCL5, and CCL8 than the C2 cluster in the GSE124821 cohort (Figure 9F and 9G).

Figure 9.

Identification of C1 cluster as the subtype with well response to ICB. A. The bar plot showed samples in C1 cluster had higher immunophenoscores (IPS) than that in C2 cluster using data from the TCGA-TNBC. B. The heatmap plot showed that samples sensitive to immune checkpoint blockade (ICB) had a higher expression of Cdkn2a, Cxcl10 and a lower expression of Ccnd1, Igf1r than samples resistant to ICB in the GSE124821 dataset. C. The heatmap plot exhibited the expressions of Cdkn2a, Cxcl10, Ccnd1, and Igf1r between C1 and C2 using data from the GSE124821 dataset. D. The bar plot showed C1 cluster had a higher percent of samples sensitive to ICB than C2 cluster. E. The receiver operating characteristic curve plot showed an area under curve value equal to 0.79 (95% CI: 0.84-0.74), which meant a great performance in predicting immune checkpoint blockade response. F. Gene Set Enrichment Analysis plot showed cellular senescence pathway was enriched in C1 cluster using data from the GSE124821 dataset. G. The heatmap plot exhibited that the transcriptomics expression levels of some senescence-associated secretory phenotype factors, such as Ccl1, Ccl5, Ccl8, Il1a, Il1b, Il6, Mmp3, Mmp10, Mmp12, Tnfrsf1b etc., was higher in C1 cluster than that in C2 cluster using data from the GSE124821 dataset. IPS, immunophenoscores; TNBC, triple-negative breast cancer; ICB, immune checkpoint blockade.

Discussion

Novel biomarkers identified or validated using gene expression data at single-cell resolution may perform better than those using the traditional RNA-seq data at the bulk level. Our study focused on identifying cellular senescence-related biomarkers for TNBC with different ICB responses at both single-cell and bulk levels. On the one hand, cellular senescence promotes immunosuppression and decreases the efficacy of immunotherapy in glioblastoma [55]. On the other hand, interferon-dependent and cytokine-induced senescence lead to self-sustaining senescence surveillance of melanoma, and patients with metastatic melanoma that lost senescence-inducing genes and amplificated senescence inhibitors progressed rapidly after receiving ICB therapy, suggesting that senescence may play a critical role in killing cancer cells that escape from ICB therapy [56,57]. Although the interaction between cellular senescence and anti-tumor immunity is complex [10,11,58], mounting evidence has highlighted the necessity of understanding the senescent heterogeneity in cancer, which may help to identify biomarkers of ICB [58,59].

In this study, we developed a robust cellular senescence-related classifier that divided TNBC patients into two clusters (C1 cluster vs C2 cluster) based on the expression of CDKN2A, CXCL10, IGF1R, and CCND1 using non-negative matrix factorization (NMF) approach. Patients in the C1 cluster have high expression levels of CDKN2A and CXCL10 and low expression levels of IGF1R and CCND1, which was a senescence phenotype and predicted a sensitive response to ICB. On the contrary, patients in the C2 cluster expressed low levels of CDKN2A and CXCL10 and high levels of IGF1R and CCND1, which was a proliferation phenotype and predicted a resistant response to ICB.

P16, a tumor suppressor and classical mediator of cellular senescence coded by CDKN2A, was associated with a better prognosis in human BC [60]. Loss of p16(Ink4a) has been shown to render BC resistant to endocrine-based therapies [61,62]. Our study showed that DEGs between ICB responder and non-responder groups (Figure 1B) or between patients in cluster C1 and cluster C2 (Figure 8B) were enriched in KEGG-endocrine-resistant pathways. This finding indicates that there is complex crosstalk between endocrine-based therapy and ICB. However, research on the association of CDKN2A and anti-tumor immunity is limited. Our results demonstrated that both CDKN2A and CXCL10, as SASP factors, were positively correlated with the infiltration abundance of CD8⁺ T cells, which were the main cells involved in anti-tumor immunity.

CCND1 is a key regulator of cell cycle and proliferation, and overexpression of CCND1 promotes tumorigenesis, cell proliferation, tamoxifen resistance, and recurrence of BC [63-65]. IGF1/IGF1R signaling regulates cell growth and promotes growth effects in TNBC cells [66]. IGF1R inhibition enhances the effects of chemoimmunotherapy combined with ICB by initiating autophagy and enhancing CD8⁺ T cell infiltration [67]. In this study, cluster C2 exhibited a high expression of CCND1 and IGF1R, low-immune activity, rare CD8⁺ T cell infiltration and worse prognosis, which indicated that IGF1R inhibition combined with chemoimmunotherapy, and ICB and might help prolong the OS of TNBC patients.

Our study found that cluster C1 which exhibited a senescence-enriching phenotype and was characterized by high expression of CDKN2A and CXCX10 had high CD8⁺ T cell infiltration and responsiveness to ICB in multiple datasets at both single-cell and bulk levels. We found that cluster C1 significantly increased the expression of inflammatory factors, including IL-1 and IL-6, and the expression of chemokines, including CCL5 and CCL8, and decreased the expression of growth regulators such as IGFBP4. The activity of IL-1 links innate and adaptive immunity and can therefore be clinically translated into the context of preventive and therapeutic strategies by promoting T cell immunity [68]. Similarly, CCL5 and CCL8 were associated with T cell infiltration and response to ICB [69,70]. The abnormal expression of these regulators in tumor tissues presumably determines the difference in both prognosis and response to ICB between patients in clusters C1 and C2.

However, our study has some limitations. First, data of all patients with BC in the GSE75688 cohort were analyzed regardless of molecular subtype owing to a limited number of samples and single tumor cells. Second, although we performed a multi-immunohistochemistry assay in human TNBC specimens and found the expression of CDKN2A and CXCL10 to be positively associated with CD3 and CD8, the underlying molecular mechanism remains unclear and needs further investigation.

Conclusion

As the subtype-specific vulnerabilities of ICB in TNBC are unclear, we aimed to identify markers related to cellular senescence that could potentially serve as predictors of ICB response in TNBC. Here, we succeeded in developing a robust cellular senescence-related classifier of TNBC based on the expression of CDKN2A, CXCL10, CCND1, and IGF1R by analyzing data at both single-cell and bulk levels. The expression of these genes was found to be relevant to clinical outcomes and response to ICB across multiple cohorts.

Disclosure of conflict of interest

None.

Abbreviations

TNBC

Triple negative breast cancer

ICB

Immune checkpoint blockade

SASP

Senescence-associated secretory phenotype

UAMP

Uniform manifold approximation and projection

DEGs

Differentially expressed genes

pCR

pathological complete remission

NMF

Non-negative matrix factorization

IPS

Immunophenoscores