Identification of three subtypes of triple-negative breast cancer with potential therapeutic implications

For both internal and external cohorts, raw data were MAS5-normalized in the Affymetrix® Expression Console (v1.3.1) and then log2-transformed. Genes were then median-centered and scaled in each cohort separately. All microarrays complied with quality criteria. Microarray and patient clinical data have been deposited in the GEO under the GSE83937 accession number. Five publicly available datasets were pooled for a total of 257 TNBC (Additional file 1).

To annotate the clusters, we used clinicopathologic characteristics, 54 gene-expression signatures (GES), Gene Ontology enrichment analysis (GOEA), and non-TNBC data. Depending on the nature of data, the following methods helped to explore differences among the clusters: one-way analysis of variance (ANOVA) with Tukey post hoc test, Fisher’s exact test, Cox regression model, Kaplan-Meier curves with log-rank test, and Pearson’s correlation coefficient.

Fifty-four GES were selected for functional annotation of breast cancer tumors. Twelve GES were used for breast cancer molecular subtyping: 4-TNBC, androgen receptor (AR), basal-like, tumor identity card (CIT), claudin-CD24, claudin-low, ER, ER-negative, ERBB2, molecular apocrine, PAM50, and TNBCtype. Eleven were used for immune response dissection: B-cell, Cell type Identification By Estimating Relative Subsets Of known RNA Transcripts (CIBERSORT) [n = 22] (http://​cibersort.​stanford.​edu/​), cytolytic activity (CYT), interferon (type I IFN), interleukin-8 (IL-8), M2/M1 macrophage ratios (M2/M1, M2/M1 [Becker]), MHC-1, MHC-2, STAT1, and T cell [10]). Three relate to microenvironment cells: epithelial cells, fibroblasts, and neurons (http://​xcell.​ucsf.​edu/​). Three were linked to metabolism evaluation: adipocytes, glycolysis, and iron (IRGS); and 22 to critical biological pathways in cancer: AKT, β-catenin, chromosomal instability (CIN), E2F3, EGFR, HOXA, mitochondrial oxidative phosphorylation (MITO/OXPHOS), MYC, p53, PIK3CA, perineural invasion (PNI), prolactin (PRL), proliferation, PTEN loss, RAS, reactive stroma, SRC, Stroma-CD10, TGFβ, VEGF, wound response, and YAP1-WWTR1. Finally, three prognostic GES were also used: 38-GES, van’t Veer 70-GES, and genomic grade index (GGI). Complete GES list, methods, and references are described in Additional file 2.

Functional annotation of each cluster through GO biological process analysis was performed using the ToppGene web tool [11]. Two methods were used to select genes differentially expressed across the clusters. SAM method was performed to obtain lists of genes with significantly different expression between clusters (one versus one and one versus the others): genes for which all corresponding probe sets had a q value of 0% were retained. In addition to SAM method, expression of the 5% most variable probe sets was represented on a heatmap, with patients ordered according to the clusters; hierarchical clustering (centered Pearson correlation distance, Ward’s method) was performed on the probe sets in order to visually detect groups of genes with high expression patterns corresponding to specific cluster(s) of patients. Visually identified overexpressed gene sets in heatmap were designated by “H” followed by a number, which referred to the number of the corresponding cluster (1, 2, or 3). If sets of genes contained different subsets, this number was followed by “a,” “b,” or “c.”

Eighty-seven cases of archival formalin-fixed, paraffin-embedded tissues of the internal cohort were evaluated on tissue microarrays, containing a median of 3 replicate 0.6 mm cores per case. Tissue microarrays (TMA) construction was described elsewhere [6]. Furthermore, 42 full sections of these surgical specimens were available for histological evaluation. Immunohistochemistry was performed using the Ventana BenchMark Ultra platform (Ventana Medical Systems, Tucson, AZ). Details of the antigen retrieval technique and dilution of primary antibodies (CD20, CD21, CD138, MECA79, UCHL1/PGP9.5, and S100) are described in Additional file 3.

Tumor-infiltrating lymphocytes (TILs), especially CD138-positive plasma cells and CD20-positive B lymphocytes, were assessed according to recommendations of an international working group [12, 13]. The presence of tertiary lymphoid structure (TLS) and lymphoid clusters, with and without germinal centers, respectively, was evaluated in the 42 specimens with available full sections, due to their typical localization in the surrounding area of the tumors. Furthermore, the number of CD21-positive follicular dendritic network was assessed by counting positive structures by IHC on a representative full section per tumor [14]. The presence of high endothelial venules (HEV) was assessed on full sections by counting the number of MECA79-positive vessels in five 400× magnification hotspots per tumor [15]. Cases were classified as nerve fibers positive versus nerve fibers negative using UCHL1/PGP9.5 (neuronal marker) and S100 (pan-specific Schwann cell marker). Pathologists were blinded for TNBC cluster assignment.

Non-TNBC external data (n = 894) were used to refine non-basal-like TNBC cluster (Additional file 1).

We considered a two-sided P value of less than 5% to be statistically significant. For SAM method, 0% q values were retained. All statistical analyses and figures’ generations were performed using R [16] and the packages affy 1.50.0, amap 0.8.14, cluster 2.0.4, citbcmst 1.0.4, fpc 2.1.10, and samr 2.0.

Three clinicopathologic characteristics were differentially represented in function of clusters (Table 1). Patients belonging to C1 were older than C3 patients (Tukey, P = 0.0066). Histological grade and Nottingham prognostic index were higher in C2 compared to C1. Contrary to our previous work based on 107 patients of this larger TNBC cohort, no prognostic difference was found between the three clusters: overall survival, P = 0.54; metastasis-free survival, P = 0.64; and event-free survival (EFS), P = 0.41. EFS analysis conducted on internal and external cohorts showed no significant differences (P = 0.20 and 0.07, respectively); however, pooling both cohorts (n = 427) showed that C3 patients have a better prognosis compared to C2 patients (P = 0.0480) (Additional file 6).

First principal component (PC1) separated C1 tumors from C2 and C3 tumors, while second principal component (PC2) separated C2 from C3 (Additional file 7). The 20 probe sets of each of the two components with the highest absolute weights were selected. A total of 20 unique genes (15 for PC1 and five for PC2) could be identified and were retained for bibliographic analysis (Additional file 8). Eight PC1 genes were known to be linked to molecular apocrine subtype. Seven belonged to a set of genes upregulated in PIK3CA-mutated ERalpha-positive breast cancer cells. Association between high AR level and PIK3CA mutations was consistent with Lehmann’s work, which showed that mutations in PIK3CA exon 9 and 20 were more frequent in luminal androgen receptor TNBC subtype [17]. The five well-identified PC2 genes were specific of B cell lineage differentiation. In conclusion, PCA showed that the two biological key features that distinguished the three TNBC clusters were molecular apocrine phenotype for C1 compared to C2 and C3, and B cell lineage differentiation for C3 compared to C2.

Various GES indicated that C1 clustered molecular apocrine (luminal androgen receptor) breast tumors (Figs. 1 and 2 and Additional files 9 and 10). Molecular apocrine subtype identification by means of GES was concordant with previous IHC results (AR, FOXA1) [6]. This subtype is known to be characterized by a molecular profile common to ER-positive breast cancer and an enrichment in ERBB2-positive tumors [18‐21]. These two characteristics were demonstrated by means of ER and ER-negative GES, and ERBB2 GES, respectively.

Basal-like features were clearly displayed by C2 and C3 based on different GES (CIT, PAM50, 4-TNBC, basal-like) (Additional files 9 and 10).

Numerous GES (CIN, EGFR, HOXA, MYC, p53, PNI, proliferation, wound response) showed that the biological aggressiveness of TNBC was high in C2 and C3 compared to C1 and more pronounced (i.e., mesenchymal phenotype) in C2 compared to C3 (E2F3, PNI, VEGF, YAP1-WWTR1) (Additional file 10).

We focused on immune response in C2 and C3 because intratumoral immune response is known to play an important role in basal-like subtypes and has prognostic significance. Two immune GES, relative to M2/M1 macrophage ratios, displayed high scores in C2 compared to C3 (P <  0.0001), a finding which was confirmed by CIBERSORT analysis (Table 2). Since M2 macrophages are considered as immunosuppressive cells, this result underlined an immune pro-tumorigenic signal in C2 compared to C3.

High TGFβ GES scores confirmed C2 pro-tumorigenic status. Indeed, in late-stage tumors, TGFβ acts as a pro-tumorigenic cytokine produced by tumor cells and tumor-infiltrating immune cells [22].

All other immune GES (B-cell, IFN type I, CXCL8 [IL8], MHC-1, MHC-2, STAT1, T cell, and notably CYT GES based on the expression of GZMA [granzyme A] and PRF1 [perforin 1]) showed high scores in C3, which suggested a pronounced adaptive immune response.

CIBERSORT analysis showed that 14 out of 22 immune cells were differentially distributed among C2 and C3 (Table 2). According to relationships between these hematopoietic subsets and survival in breast cancer, C2 tumors were highly infiltrated by seven immune cells associated with adverse outcomes (hereafter named “pro-tumorigenic” immune cells) [23]. On the contrary, C3 tumors were highly infiltrated by six immune cells associated with favorable outcomes (hereafter named “anti-tumorigenic” immune cells). Contrary to Gentles’ results and according to numerous studies, macrophages M1 were considered as anti-tumorigenic immune cells [24]. Out of 14 immune subsets, only dendritic cells activated were discordant with this interpretation.

In brief, two opposite immunophenotypes have been identified in basal-like TNBC; the first one characterized by immune cells known to stimulate breast cancer growth (pro-tumorigenic) and the second one by immune cells known to inhibit it (anti-tumorigenic). Furthermore, prognostic analyses based on immune continuous score GES in internal and external cohorts showed that C2 and C3 patients with high CYT, MHC-2, and STAT1 scores, reflecting an anti-tumorigenic immune response, have a better prognosis (Additional file 11).

Gene expression related to iron metabolism clearly decreased from C2 to C3 and from C3 to C1 (Additional file 10). Glycolysis seemed more elevated in C2 and C3 compared to C1. Considering biological knowledge, high iron metabolism and glycolysis are associated with aggressive tumors [25, 26].

Scores calculated according to the three prognostic GES were significantly lower in C1 compared to C2 and C3 (Additional file 10). The 70-GES prognostic score kinetics pattern separated the three clusters in the following order: C2 > C3 > C1. This finding was not corroborated by prognostic analysis. Therefore, we can only conclude that it more likely underlines biological aggressiveness of C2 and C3 tumors.

Four clusters named H1, H2a, H2b, and H3 of highly expressed genes, belonging to C1, C2, and C3, were visually individualized from heatmap composed of the 5% most variable probe sets (n = 1843). Furthermore, eight gene lists of most differentially expressed genes between the three clusters were obtained by means of SAM. All these gene lists were submitted to the ToppGene web tool (Additional files 12 and 13). C1 was characterized by three luminal hallmarks (hormone metabolic process, lipid metabolic process, and oxidation-reduction process) on one hand, and angiogenesis and developmental processes, on the other hand. Mitotic cell cycle process, which is considered as a basal-like hallmark, was overrepresented in C2 and C3 compared to C1. Genes highly expressed in C2 vs C3 were notably linked to central nervous system development process, here considered as neurogenesis, extracellular matrix disassembly, and collagen metabolic process, i.e., invasion and progression. To our knowledge, this is the first time that high neurogenesis-related gene expression is linked to a breast cancer subtype. This finding was associated with UCHL1/PGP9.5 neuronal marker and S100 pan-specific Schwann cell marker enrichment in C2 compared to C3: P = 0.0249 and P = 0.0205, respectively (Fig. 3). Neuron GES corroborated this difference. Finally, high PNI scores in C2 (C2 > C3 > C1) could result from high nerve density in this subtype (Additional file 10). Further investigations are needed to unravel the role of nerve cells and de-differentiated Schwann cells in cancer progression and invasion [27, 28]. Immune response activation, and notably B cell activation, was preponderant in C3.

Three gene subclusters (H3a, H3b, and H3c) could be distinguished in H3. GenomicScape was used to refine B lymphocyte subsets in these three gene subclusters [29]. H3c was strongly enriched in genes highly expressed in plasma cells (Additional file 14). On the contrary, H3b genes focused our attention on the early phase of B cell differentiation. MS4A1, which encodes CD20, a B lymphocyte surface molecule absent in plasma cells, was found overexpressed in this subcluster. No clear information emerged from H3a GenomicScape atlas screening. However, H3a GOEA underlined the following immune response processes: CXCR3 signaling pathway (CXCL9, CXCL10, CXCL11), TCR signaling pathway (TRA, ICOS, ZAP70, MAP4K1, TARTP [CD3G], UBASH3A), and T cell cytotoxicity (GZMB, GSMH, GNLY). To conclude, high expression of genes encoding for lymphorganogenic chemokines (CCL19, CCL21, CXCL13) and other TLS markers (ICAM1, ICAM2, ICAM3, VCAM1, CCL22) were observed in C3 versus C2 (P <  0.0001) [30, 31]. In brief, the presence of different B lymphocyte subsets in C3 suggested that complete B cell differentiation characterized this TNBC cluster and might be orchestrated with Th-1 cells and Tfh in TLS located in close proximity to tumors.

In TLS, intratumoral favorable immune phenotype relies on coordinated activation of pro-inflammatory genes, which include Th-1 transcripts (e.g., IFNG, STAT1, IL12A, IL12B, IRF1, TBX21, CD8B), Th-1 chemokine genes (e.g., CXCL9, CXCL10, CCL5), Tfh chemokine gene (CXCL13) and immune effector genes (e.g., GNLY, GZMB, PRF1), and B cells genes (e.g., CD19, IGKC) [29]. This immune process is accompanied by a counter activation of immunosuppressive mechanisms (e.g., CD274 [PD-L1], CTLA4, FOXP3, IDO1, PDCD1), which will be discussed further [3, 32]. In our study, differential gene-expression analyses displayed a significant higher level of all except two genes (FOXP3, CD8B) in C3 compared to C2.

CD274 (PD-L1), which is considered as an adaptive immune resistance marker in basal-like breast cancers, was overexpressed in C3 compared to C2 in internal and external cohorts (P <  0.0001) [33]. This information has to be pointed out because it may offer targeted therapeutic intervention for C3 patients.

Histological evaluation of TLS and lymphoid clusters showed that they were more frequent in C3 (10/12) than in C2 (9/22) (P = 0.0297) and in C1 (2/8) (P = 0.0194) (Fig. 3). No difference was found between C2 and C1. This result was corroborated by the presence of HEV (MECA79-positive) and follicular dendritic cell networks (CD21-positive) in C2 compared to C3: C2 (n = 22) vs C3 (n = 12), P <  0.0001 and P = 0.0017, respectively (Fig. 3). TILs displayed the following profile: C3 > C2 = C1 (P = 0.0014). Furthermore, plasma cell (CD138-positive cells) and B lymphocyte (CD20-positive cells) infiltrates were significantly higher in C3 compared to C2, P = 0.0297, and P = 0.0001, respectively (Fig. 3).

To explore the possible existence of biological gradients between C2 and C3, we calculated correlation coefficients between C2 and C3 cluster probability orthogonal projection on C2-C3 axis and GES scores for each tumor (Additional file 5). Ten coefficients had absolute value superior to 0.5 (P <  0.0001) for internal cohort (Additional file 15). Nine of these correlations related to immune response GES. These results showed that anti-tumorigenic immune response decreases from C3 to C2 and, on the contrary, pro-tumorigenic immune response increases from C3 to C2.

Expressions of 47 immune checkpoint genes were compared between C2 and C3 in internal and external cohorts [34]. Only significant and concordant results in both cohorts are discussed thereafter. Thirty-four immune checkpoints, including CD274 (PD-L1), CTLA4, and PDCD1 (PD1), displayed a C2 < C3 profile, five a C2 ≈ C3 profile, and one (VTCN1) a C2 > C3 profile (Additional file 16). Higher expression of a large majority (85%) of immune checkpoint genes was observed in C3 compared to C2, which was characterized by a marked immune response. This finding should be interpreted as a consequence of an immune response against tumor cells. In other words, immune inhibitory response by means of immune checkpoint upregulation could be triggered by a high anti-tumorigenic immune response to fine-tune and limit global immune response, i.e., to maintain immune self-tolerance. In regard to these results, we can hypothesize that immune checkpoints do not represent the main effectors of C2 immunosuppression and that they might contribute to the inefficacy of C3 anti-tumor immune attack.

Fuzzy clustering separated external TNBC patients into three clusters, named C’1 (n = 61; 23.8%), C’2 (n = 97; 37.7%), and C’3 (n = 99; 38.5%). Naming of external TNBC clusters was based on biological correspondence with internal TNBC clusters (e.g., C’1 for C1). Distribution of patients among the three clusters was similar (Additional file 17).

C’1 TNBC patients were older than C’2 and C’3 TNBC patients (P = 0.0002). SBR1-2 histological grades were overrepresented in C’1 compared to C’2 and C’3 (P = 0.0134) (Additional file 18). Prognosis of TNBC patients belonging to C’3 showed no difference compared to C’2 patients (P = 0.10). No statistical difference was found for cluster and SBR distributions and age according to the internal and external cohort as a whole or split into the three clusters (Additional files 18 and 19).

All categorical GES distributions, except for PAM50, were quite similar (Fig. 1 and Additional files 9 and 20). The main difference regarding PAM50 was a high proportion of HER2E subtype in C’1 (47.5%) contrary to C1 (0%) and a low proportion of Luminal B subtype in C’1 (6.5%) contrary to C1 (54.5%). In fact, C1 is mostly composed of molecular apocrine subtypes, which are known to be characterized by luminal and HER2 features [18‐21]. Furthermore, in C1, correlation coefficients with PAM50 Luminal B and HER2E centroids were close (data not shown). Here, subtype assignation should rather be “luminal B and HER2E,” i.e., molecular apocrine. Our conclusion is that PAM50 subtype assignment should be considered cautiously when used for TNBC subtyping, in contrast to CIT or TNBCtype GES which seem more appropriate.

Thirty-two out of 47 (68%) continuous score GES displayed exactly the same profile between internal and external cohort (Fig. 2 and Additional files 10 and 21). When selecting GES with two similarities, 40 out of 47 (85%) GES gave the same information.

CIBERSORT analysis showed that seven out of 22 immune cells were differentially distributed among C’2 and C’3 (Additional file 22). These profiles were the same as those observed in our internal cohort. C’2 tumors were infiltrated by three pro-tumorigenic immune cells and C’3 by four anti-tumorigenic immune cells.

When investigating the presence of an immune response gradient between C’2 and C’3, as was done on internal cohort, results corroborated what had been found: a continuous decrease of anti-tumorigenic immune response and continuous increase of pro-tumorigenic immune response from C’3 to C’2 (Additional file 23).

The same GOEA process was applied to the external cohort. Four clusters named H’1, H’2a, H’2b, and H’3 of highly expressed genes were visually individualized from heatmap composed of the 5% most variable probe sets. GOEA result comparison showed that the three clusters of the internal and external cohort were very similar (Additional files 12 and 13).

In conclusion, independent fuzzy clustering and functional annotation of the three clusters by means of clinicopathologic characteristics, GES, and GOEA indicated that the same molecular entities defined the different clusters in both the internal and external cohorts.

In regard to previous results, C1 can be considered as the less aggressive TNBC subtype. The question is now: What are the differences between C1 subtype and non-TNBC, from a biological point of view? In order to answer this question, non-TNBC external data were used.

PCA of TNBC (n = 257) and non-TNBC (n = 894) external data showed that C’1 patients were close to non-TNBC patients in the first plane (Additional file 24). PC1 separated non-TNBC and C’1 from basal-like clusters (C’2 and C’3) and PC2 separated C’2 and C’3 basal-like clusters as also observed for internal TNBC cohort. Interpretation of PCA result leads us to think that non-TNBC and non-basal-like TNBC probably share some biological similarities.

GES functional annotation comparison of C’1 (TNBC) and non-TNBC tumors showed both similarities and differences (Additional files 25, 26, 27 and 28). Among the main differences were intermediate prognostic scores for the three prognostic GES for C’1, highest for C’2 and C’3, and lowest for non-TNBC. Other results (ER, EGFR, and p53 GES scores) pointed out that aggressiveness of C’1 breast tumors was intermediate between non-TNBC and basal-like tumors. In conclusion, external data analysis showed that C’1 patients were closer to non-TNBC patients and that aggressiveness of C’1 tumors was intermediate between non-TNBC and basal-like tumors (C’2 and C’3).