Comprehensive Genomic Analysis Identifies Novel Subtypes and Targets of Triple-negative Breast Cancer

Abstract

Purpose

Genomic profiling studies suggest triple-negative breast cancer (TNBC) is a heterogeneous disease. In this study we sought to define TNBC subtypes and identify subtype-specific markers and targets.

Patients and Methods

RNA and DNA profiling analyses were conducted on 198 TNBC tumors (ER-negativity defined as Allred Scale value ≤2) with >50% cellularity (discovery set: n=84; validation set: n=114) collected at Baylor College of Medicine. An external data set of 7 publically-accessible TNBC studies was used to confirm results. DNA copy number, disease-free survival (DFS) and disease-specific survival (DSS) were analyzed independently using these datasets.

Results

We identified and confirmed four distinct TNBC subtypes: (1) Luminal-AR (LAR); 2) Mesenchymal (MES); 3) Basal-Like Immune-Suppressed (BLIS), and 4) Basal-Like Immune-Activated (BLIA). Of these, prognosis is worst for BLIS tumors and best for BLIA tumors for both DFS (logrank test p=0.042 and 0.041, respectively) and DSS (logrank test p=0.039 and 0.029, respectively). DNA copy number analysis produced two major groups (LAR and MES/BLIS/BLIA), and suggested gene amplification drives gene expression in some cases (FGFR2 (BLIS)). Putative subtype-specific targets were identified: 1) LAR: androgen receptor and the cell surface mucin MUC1; 2) MES: growth factor receptors (PDGF receptor A; c-Kit); 3) BLIS: an immune suppressing molecule (VTCN1); and 4) BLIA: Stat signal transduction molecules and cytokines.

Conclusion

There are four stable TNBC subtypes characterized by the expression of distinct molecular profiles that have distinct prognoses. These studies identify novel subtype-specific targets that can be targeted in the future for effective treatment of TNBCs.

INTRODUCTION

Recent studies have demonstrated that breast cancer heterogeneity extends beyond the classic immunohistochemistry (IHC)-based divisions of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (Her2)¹. 10-20% of primary breast cancers are triple-negative breast cancers (TNBCs)², which lack expression of ER, PR and Her2, present with higher grade, often contain mutations in TP53³, and have a poor prognosis⁴. Molecularly-targeted therapy has shown limited benefit so far in TNBCs, and although PARP inhibitors in the BRCA-mutant setting are promising^5,6, new strategies for classifying and treating women affected by this aggressive disease are urgently needed.

The intrinsic subtyping of breast cancer by gene expression analyses⁷ was recently supported by The Cancer Genome Atlas (TCGA) Program through mRNA, miRNA, DNA, and epigenetic analyses⁸. The basal-like subtype, traditionally defined by RNA profiling or cytokeratin expression⁹, account for 10-25% of all invasive breast cancers¹⁰. In addition, basal-like breast cancers account for 47-88% of all TNBCs^8,11,12. Tumors of the “claudin-low” (CL) subtype^13,14 have particularly poor prognoses compared to hormone-sensitive tumors¹⁵. The results from an aggregate analysis of publically available expression data sets performed by Lehmann et al.¹² suggested that TNBCs are more heterogeneous than previously described, and identified 6 subtypes: 1) androgen receptor positive; 2) claudin-low-enriched mesenchymal; 3) mesenchymal stem-like; 4) immune response; and two cell cycle-disrupted basal subtypes 5) BL-1 and 6) BL-2. However, immunohistochemical (IHC) detection of ER, PR, and HER2 protein is the clinical standard used to define TNBC. In the study by Lehmann et al., when tumors with IHC-confirmed ER, PR, and HER2 protein expression were analyzed, only 5 of the 6 described subtypes were observed (see supplemental figures 4 and 5 in Lehmann et al.¹²). Therefore, while previous genomic studies have advanced our understanding of TNBCs, stable subtypes, as well as subtype-specific molecular targets still need to be identified.

In this study, we investigated 198 previously uncharacterized TNBCs using mRNA expression and DNA profiling, and identified 4 stable TNBC subtypes: 1) Luminal/Androgen Receptor (LAR), 2) Mesenchymal (MES), 3) Basal-Like Immune Suppressed (BLIS), and 4) Basal-Like Immune Activated (BLIA). Using independent TNBC datasets, we show that BLIS and BLIA tumors have the worst and best prognoses, respectively (independently of other known prognostic factors), compared to the other subtypes. Our DNA studies demonstrate unique subtype-specific gene amplification, with CCND1, EGFR, FGFR2, and CDK1 amplified in the LAR, MES, BLIS and BLIA subtypes, respectively. Collectively, our RNA and DNA genomic results identify stable, reproducible TNBC subtypes characterized by specific RNA and DNA markers, and identify potential targets for more effective treatment of TNBCs.

MATERIALS AND METHODS

Patients and study recruitment

278 anonymized tissues collected from multiple U.S. and European sites were obtained from the Lester and Sue Smith Breast Cancer Tumor Bank at Baylor College of Medicine (BCM), diagnosis-confirmed and flash frozen. BCM purchased these tumors (with clinical information, including: age, menopausal status, histology, AJCC stage, tumor grade) from Asterand USA. No treatment or outcome data was available for these tumors. Tissues were managed by the BCM Breast Center's Human Tissue Acquisition and Pathology (HTAP) shared resource. Cellularity, histology, and IHC ER, PR, and HER2 receptor status in discovery and validation samples were assessed by Breast Center pathologists. Only tumors exhibiting >50% tumor cellularity were used. ER-receptor negativity is defined as Allred Scale ≤2.

RNA/DNA extraction and array experiments

For extraction and quality control details, see supplemental material. Briefly, tumors were profiled using the Affymetrix U133 Plus 2.0 gene expression array and affy¹⁶ package in R¹⁷. Discovery and validation set SNP experiments were performed on Illumina 610K and 660K platforms, respectively. Common SNPs were analyzed after independent processing in Illumina Genome Studio v2011 Genotyping Module 1.9.4.

PAM50, TNBCType, and ERSig

TNBCs were assigned to previously described subtypes using the TNBCType tool¹⁸. Intrinsic subtypes were established with the PAM50 Breast Cancer Intrinsic Classifier¹⁹, and compared to 67 non-TNBC randomly sampled tumors representing 80% of the assigned sample (confirmed by Pearson Correlation). This comparison was used to create a 32-gene centroid signature (derived from Williams et al.'s estrogen receptor 1 (ESR1) downstream targets gene list²⁰, accessed via the Molecular Signatures Database (MSigDB)²¹) in order to correlate TNBCs with ER activation (“ERSig”).

Gene selection, NMF clustering, differential expression, and centroid signatures

Genes were sorted by aggregate rank of median absolute deviations (MADs) across all samples and the MAD across each of the two most predominant clusters (approximating basal-like versus the remaining intrinsic subtypes) for the discovery set using R package Differential Expression via Distance Summary (DEDS)²². The top 1000 median-centered genes were utilized for clustering and split into 2000 positive input features²³. The ideal rank basis and factorization algorithm was determined using the R package Non-negative Matrix Factorization (NMF)²⁴ before taking the 1000-iteration consensus for a final clustering basis of 4.

Genes were sorted by DEDS using: 1) Goeman's Global Test (GGT)²⁵ applied to each set individually for all 18,209 genes, using a Benjamini-Hochberg (BH) False Discovery Rate (FDR) multi-test correction; and 2) computed log2(Fold Change) (“FC”) values. The top 20 unique genes by p-value and log2(FC) became a classifier comprising 80 genes and representing the median quantiles of all 80 genes for each discover set cluster, with cases assigned by minimum average Euclidean distances of quantile gene expression data. Non-significant p-values (p>0.05 by 10,000 permutations) or deviations from any centroid >0.25 were left unclassified.

Preprocessing and assignment of expression data for publically-accessible cases

Normalization and quality control procedures identical to the primary study sets (but using the Partek Genomics Suite program²⁶ to perform ANOVA-based batch correction across the 221 arrays prior to summarization of probe set data) were performed on 7 publically-accessible studies in Gene Expression Omnibus (GEO) with TNBCs (by IHC) profiled on the Affymetrix U133 Plus 2.0 array (“external set”). Series GEO matrices and accompanying TNBC tumor clinical data from the Sabatier²⁷ (also included in external set) and Curtis¹¹ studies were assigned using gene-centric representation of array data.

Ingenuity pathways analysis

Significant genes (BH correction p-value<0.001 from GGT) for each dataset group were uploaded independently into Ingenuity Systems’ Interactive Pathway Analysis (IPA) software (www.ingenuity.com). A 0.05 significance threshold was used for pathway enrichment. Molecules, chemicals, or groups with regulatory function(s) were analyzed by IPA to produce final gene lists.

Copy number segmentation and analysis

Allele-Specific Piecewise Constant Fitting (ASPCF) analysis and Allele-Specific Copy Number (CN) Analysis of Tumors (ASCAT, default values)²⁸ of 84 discovery and 58 validation set tumors yielded 62 and 46 samples, respectively, with assigned reliable DNA ploidy- and tumor percentage-corrected integer CNs. These segments were uploaded collectively and individually by assigned expression-based subtypes to Genomic Identification of Significant Targets in Cancer (GISTIC) 2.0²⁹ (default settings, with a 0.5 linear margin for gains and losses).

Survival analyses

Survival curves were constructed using the Kaplan-Meier product limit method and compared between subtypes with the log-rank test using publically available datasets for which disease-free survival and disease-specific survival results are available; however, no treatment information was available for these datasets. Cox proportional hazard regression model adjusted for available prognostic clinical covariates was performed to calculate subtype-specific hazard ratios, 95% confidence intervals, and disease-free and overall survival (DFS and OS, respectively). Survival analyses were performed using the R package survival.

RESULTS

Patient population

198 TNBCs were assigned to discovery (n=84) or validation (n=114) sets based on chronological acquisition of tissue. Subjects were predominantly postmenopausal, Caucasian, and of mean and median age of 53 (Table 1). 95% of TNBCs were invasive ductal carcinomas, predominantly Stages I-III (1% were metastatic breast cancers), and >75% of tumors were >2cm at diagnosis.

Table 1.

Clinical characteristics of the patients and tumor samples used in study.

Characteristic	Both Sets		Discovery Set		Validation Set		p-value
	n	%	n	%	n	%
Number of Tumors	198		84	42	114	58
Age	192		84		108		0.02
<50 yrs	81	42	43	51	38	35
≥50 yrs	111	58	41	49	70	65
Missing	6		0		6
Race	194		80		114		0.10
Caucasian	187	96	75	94	112	98
Asian/Pacific Islander	7	4	5	6	2	2
Missing	4		4		0
Menopausal status	167		71		96		0.24
Premenopausal	62	37	31	44	31	32
Menopausal	11	7	3	4	8	8
Postmenopausal	94	56	37	52	57	59
Missing	31		13		18
Body mass index	166		65		101		0.65
Underweight (< 18.5)	3	2	2	3	1	1
Normal (18.5-24.9)	46	28	17	26	29	29
Overweight (25-29.9)	61	37	26	40	35	35
Obese (≥30)	56	33	20	31	36	35
Missing	32		19		13
Tumor size	195		83		112		0.01
<2 cm	35	18	10	12	25	22
2-5 cm	139	71	60	72	79	71
>5 cm	12	6	10	12	2	2
Any size with direct extension	9	5	3	4	6	5
Cannot be assessed	3		1		2
Positive lymph nodes	150		66		84		0.14
0	74	49	29	44	45	54
1-3	49	33	28	42	21	25
4-9	17	11	6	9	11	13
>10	10	7	3	5	7	8
Unknown	48		18		30
Metastases	146		64		82		0.86
No metastases	144	99	63	98	81	99
Metastases found	2	1	1	2	1	1
Unknown	52		20		32
Histology	198		84		114
Infiltrating ductal carcinoma (IDC)	188	95	82	98	106	93	0.41
Infiltrative lobular carcinoma (ILC)	1	0.5	0	0	1	1
Adenocarcinoma/carcinoma, not otherwise specified	7	3.5	2	2	5	4
Medullary carcinoma	2	1	0	0	2	2

mRNA profiling of TNBCs reveals four stable molecular phenotypes

Using RNA gene expression profiling, we explored TNBC molecular phenotypes. NMF was performed on 1000 discovery set genes selected to maximize separation across and within conventional intrinsic subtypes. These tumors were most stably divided into 4 clusters by cophenetic, dispersion, silhouette, and Statistical Significance of Clustering (SigClust)³⁰ metrics, in addition to visual inspection of the consensus heat map (Figures 1A-B, S1). This quadrilateral division of data was also observed in the validation set tumors using the same input features (Figures 1D-E, S2). ER-, PR- and Her2-negativity was IHC-confirmed by our participating pathologist, Dr. Contreras (Figures S3). Differentially-expressed genes (BH-adjusted p-value<0.001 from GGT) were significantly enriched only within corresponding discovery and validation set clusters (Fisher Exact test p=4.01E-30, 3.47E-17, 2.88E-46, and 3.61E-10, respectively, Tables S1-5), independently confirming the 4 molecular phenotypes observed. Additionally, significant enrichment of discovery set IPA results in the validation set also support the four cluster separation (Tables S6-10).

Figure 1. Classification of TNBCs by mRNA profiling reveals four stable molecular phenotypes.

84 (discovery set) and 114 TN breast tumors (validation set) both demonstrate 4 stable clusters by NMF of mRNA expression across the top 1000 genes (IQR summarized) selected by DEDS aggregate rank of median absolute deviations (see complete methods) of the discovery set. A & D. Cophenetic and dispersion metrics for NMF across 2-10 clusters with 50 runs suggest 4 stable clusters. Full metrics are available for each set in Supplemental Figures 1 and 2. B & E. Silhouette analyses and consensus plots for rank basis 4 NMF clusters (1000 runs, nsNMF factorization). Average silhouette widths worsened with increasing clusters beyond the 4 shown. SigClust was significant for all pairwise comparisons with this feature set. C & F. PAM50 intrinsic subtypes and TNBCType distributions by 4 NMF clusters.

Comparison of our NMF results to Perou's “PAM50” TNBC molecular classification (luminal A, luminal B, HER-2-positive, basal-like and normal-like subtypes)⁹ shows clusters 3 and 4 to be entirely basal-like, containing 86% and 74% of all PAM50 basal-like tumors in the discovery and validation sets, respectively (Figure 1C). Conversely, cluster 1 contains all luminal A, luminal B, and Her2-positive PAM50 tumors, and cluster 2 contains basal-like and normal-like PAM50 tumors.

We then compared our NMF results with the Lehmann/Pietenpol “TNBC Type” molecular classification (basal-like-1, basal-like-2, immunomodulatory, luminal androgen receptor (LAR), mesenchymal, and mesenchymal stem-like subtypes)¹², in which “claudin-low” tumors are split between the mesenchymal and mesenchymal stem-like subtypes. Our results show cluster 1 contains all of Lehmann's LAR tumors, and cluster 2 contains most of Lehmann's mesenchymal stem-like and some claudin-low mesenchymal tumors (Figures 1F, S4B, S5). Conversely, our TNBC clustering did not separate Lehmann's¹² “basal-like 1” and “basal-like 2” types even when utilizing all six subtype signatures described in Lehmann et al.¹² in a semi-supervised NMF (2188 genes; Figure S4). Instead, Lehmann's basal-like-1 and basal-like-2 tumors are split between clusters 3 and 4 (Figure S4). Finally, Lehmann's remaining claudin-low mesenchymal tumors reside in cluster 3, while the immunomodulatory tumors are distributed across clusters 2 and 4, which express common signaling pathways (Figures S4-5).

Gene signatures define four prognostically-distinct TNBC subtypes

Using the discovery and validation sets we developed and confirmed an 80-gene signature for these clusters (Figure 2A, Tables S11-16). This analysis was repeated using an independent set of 221 publically accessible TNBCs with IHC data (“external set”, Tables S17-18, Figure 2B), and other publically accessible datasets with available clinical data (Tables S19-20). Comparisons of group assignment against existing NMF clusters demonstrated strong reproducibility, with Rand indices of 0.94 (p<0.0001) and 0.82 (p<0.0001), respectively (Tables S21-22).

Figure 2. Gene signature defines four subtypes of TNBC with prognostic differences.

Discovery, validation, and external sets tumors with intermediate grade, high ESR1, PGR, and ERBB2 expression, activated ER downstream targets, and luminal A/B subtypes are enriched in subtype 1. A. The four assigned subtypes in both the discovery (84/84) and validation sets (114/114). B. Gene signature applied successfully to 220 of 221 external set TNBCs. C. Clinical outcomes from independent sets classified by the discovery set-based signature. Subtype 4 has a better prognosis for both disease-free and disease-specific survival.

Clinical outcome data was available for this publically available “external set” of TNBCs. However, treatment information for the “external set” data is not available. Analysis of disease-free survival (DFS) and disease-specific survival (DSS) showed that subtype 3 has the worst prognosis of all 4 subtypes, while subtype 4 has a relatively good prognosis for DFS (logrank test p=0.042 and 0.041, respectively) and disease-specific survival (DSS; logrank test p=0.039 and 0.029, respectively) (Tables S23-24, Figure 2C). The associations between subtypes 3 and 4 and DFS and DSS remained significant in multivariate models adjusted for available prognostic clinical covariates..

TNBC subtype-specific enrichment of molecular pathways

Differentially expressed genes from each subtype (BH-adjusted p-value<0.001 from GGT) were analyzed for pathway enrichment. Results from the validation and external sets significantly overlapped the discovery set, with predicted regulator activation and inhibition patterns stable across the three datasets but distinct between subtypes (Tables S25-29, Figure 3).

Figure 3. Molecular pathways enriched in the four identified subtypes of TNBCs.

Significant pathways from the discovery set also found in validation and external sets are listed for the LAR, MES, BLIS and BLIA subtypes.

Subtype 1 tumors exhibit androgen receptor, ER, prolactin, and ErbB4 signaling (Figure 3), but ER-alpha-negative IHC staining. Gene expression profiling demonstrates expression of ESR1 (the gene encoding ERα; Figure S6), and other estrogen-regulated genes (PGR, FOXA, XBP1, GATA3). Thus, these “ER-negative” tumors demonstrate molecular evidence of ER activation. This may be because 1% of these tumor cells express low levels of ER protein, defining them as “ER-negative” by IHC analysis. These observations suggest subtype 1 tumors may respond to traditional anti-estrogen therapies as well as to anti-androgens, as previously suggested¹². To be consistent with previous studies¹², we termed Subtype 1 the Luminal/Androgen Receptor (LAR) subtype.

Subtype 2 is characterized by pathways known to be regulated in breast cancer, including cell cycle, mismatch repair and DNA damage networks, and hereditary breast cancer signaling pathways (Figure 3). Additionally, genes normally exclusive to osteocytes (OGN) and adipocytes (ADIPOQ, PLIN1), and important growth factors (IGF-1) are highly expressed in this subtype, previously described as “mesenchymal stem-like” or “claudin-low” (Figure S7). Therefore, we named Subtype 2 the Mesenchymal (MES) subtype.

Subtype 3 is one of two basal-like clusters, and exhibits downregulation of B cell, T cell, and natural killer cell immune-regulating pathways and cytokine pathways (Figure 3). This subtype has the worst DFS and DSS, and low expression of molecules controlling antigen presentation, immune cell differentiation, and innate and adaptive immune cell communication. However, this cluster uniquely expresses multiple SOX family transcription factors. We termed Subtype 3 the Basal-Like Immune Suppressed (BLIS) subtype.

Immune regulation pathways are upregulated in Subtype 4, the other basal-like cluster (Figure 3). Contrary to BLIS, Subtype 4 tumors display upregulation of genes controlling B cell, T cell, and natural killer cell functions. This subtype has the best prognosis, exhibits activation of STAT transcription factor-mediated pathways, and has high expression of STAT genes. To contrast BLIS tumors, we termed Subtype 4 the Basal-Like Immune Activated (BLIA) subtype.

DNA copy number analysis identifies TNBC subtype-specific focal changes

We next investigated TNBC subtype-defined CN variation (CNV) by ploidy- and tumor percentage-correcting 62 discovery and 46 validation set TNBCs, before analyzing them together in GISTIC 2.0. Overall, genomes were very unstable and exhibited common TNBC chromosomal arm gains and deletions (Tables S30-35, Figure 4A, S7-8). Focal variations present in all 4 TNBC subtypes include: 1) focal gains on 8q23.3 (CSMD3), 3q26.1 (BCHE), and 1q31.2 (FAM5C), which are the greatest gains and characterize >84% of all tumors; and 2) focal losses on 9p21.3 (CDKN2A/B), 10q23.31 (PTEN), and 8p23.2 (CSMD1) (Figure 4B).

Figure 4. DNA Copy Number Analysis identifies focal changes in TNBC subtypes.

DNA copy number changes observed in each subtype are listed. A. Focal gains (red) and losses (blue) detected by GISTIC 2.0 are plotted by log10(q-value) and reported by cytoband. Adjacent numbers are percentages of subtype specific cases (n = 24, 17, 33, 34, respectively) with this focal aberration. Presence of a colored square demonstrates this region was detected by subtype-specific GISTIC 2.0 analysis as well. All structural events for each subtype and set are available in the supplemental. B. Broad copy number events distinguish the LAR subtype from all others. Gains (red) and losses (blue) are plotted along the genome, with darker colors representing a region enriched to the displayed subtype by Fisher Exact Test.

Subtype-specific variation is greatest between LAR and the remaining 3 subtypes (Figure 4). LAR tumors have focal gains twice as frequently on 11q13.3 (CCND1, FGF family) and 14q21.3 (MDGA2), but 1/3 as frequently on 12p13.2 (MAGOHB, KLR subfamilies) and 6p22.3 (E2F3, CDKAL1) compared to MES, BLIS, and BLIA tumors (Figure 4). The LAR subtype also has more frequent deletions of 6q, lacks arm-wide deletions across 5q, 14q, and 15q, and has significantly fewer focal deletions on 5q13.2 (RAD17, ERBB2IP), 12q13.13 (CCNT1, ERBB3), 14q21.2 (FOXA1), and 15q11.2 (HERC2) (Figures 4, S8). MES and BLIA tumors, which exhibit increased normal (diploid) immune cell infiltration, are characterized by lower aberrant cell fractions than LAR and BLIS tumors (Figure S9). Additional subtype-specific gene overexpression includes: 1) LAR: AR, MUC1; 2) MES: IGF-1, ADRB2, EDNRB, PTGER3/4, PTGFR, PTGFRA; 3) BLIS: VTCN1; 4) BLIA: CTLA4 (Tables 2, S36-39).

Table 2.

Selected genes from pathway analysis with significant relative overexpression (>2-fold, BH p≤0.05) in discovery and validation sets.

TNBC Subtype	Symbol	Description	Discovery Fold-Change	Druggable	CNV Seen
1: Luminal AR (LAR)	DHRS2	dehydrogenase/reductase (SDR family) member 2	68.6
	PIP	prolactin-induced protein	21.1
	AGR2	anterior gradient 2 homolog (Xenopus laevis)	17.1		Yes
	FOXA1	forkhead box A1	17.1		Yes
	ESR1	estrogen receptor 1	13.9	Yes
	ERBB4	v-erb-a erythroblastic leukemia viral oncogene homolog 4 (avian)	11.3	Yes
	CA12	carbonic anhydrase XII	11.3	Yes
	AR	androgen receptor	9.8	Yes
	TOX3	TOX high mobility group box family member 3	7.5		Yes
	KRT18	keratin 18	4.3		Yes
	MUC1	mucin 1, cell surface associated	4.3	Yes
	PGR	progesterone receptor	3.5	Yes
	ERBB3	v-erb-b2 erythroblastic leukemia viral oncogene homolog 3 (avian)	3		Yes
	RET	ret proto-oncogene	2.5	Yes
	ITGB5	integrin, beta 5	2.1	Yes
2: Mesenchymal (MES)	ADH1B	alcohol dehydrogenase 1B (class I), beta polypeptide	42.2	Yes
	ADIPOQ	adiponectin, C1Q and collagen domain containing	32
	OGN	osteoglycin	16
	FABP4	fatty acid binding protein 4, adipocyte	14.9
	CD36	CD36 molecule (thrombospondin receptor)	14.9
	NTRK2	neurotrophic tyrosine kinase, receptor, type 2	6.1	Yes
	EDNRB	endothelin receptor type B	5.7	Yes
	GHR	growth hormone receptor	4.9	Yes
	ADRA2A	adrenoceptor alpha 2A	4.6	Yes
	PLA2G2A	phospholipase A2, group IIA (platelets, synovial fluid)	4.6	Yes
	PPARG	peroxisome proliferator-activated receptor gamma	4	Yes
	ADRB2	adrenoceptor beta 2, surface	3.5	Yes
	PTGER3	prostaglandin E receptor 3 (subtype EP3)	3.2	Yes
	IL1R1	interleukin 1 receptor, type I	3	Yes
	TEK	TEK tyrosine kinase, endothelial	2.8	Yes
3: Basal-like Immune Suppressed (BLIS)	ELF5	E74-like factor 5 (ets domain transcription factor)	7
	HORMAD1	HORMA domain containing 1	5.7		Yes
	SOX10	SRY (sex determining region Y)-box 10	4.9		Yes
	SERPINB5 serpin peptidase inhibitor, clade B (ovalbumin), member 5		4.6
	FOXC1	forkhead box C1	4.6
	SOX8	SRY (sex determining region Y)-box 8	4.3
	TUBB2B	tubulin, beta 2B class IIb	3.2	Yes
	VTCN1	V-set domain containing T cell activation inhibitor 1	3
	SOX6	SRY (sex determining region Y)-box 6	3
	KIT	v-kit Hardy-Zuckerman 4 feline sarcoma viral oncogene homolog	2.5	Yes
	FGFR2	fibroblast growth factor receptor 2	2	Yes	Yes
4: Basal Immune Activated (BLIA)	CXCL9	chemokine (C-X-C motif) ligand 9	5.3
	IDO1	indoleamine 2,3-dioxygenase 1	4.9
	CXCL11	chemokine (C-X-C motif) ligand 11	4.9
	RARRES1	retinoic acid receptor responder (tazarotene induced) 1	4		Yes
	GBP5	guanylate binding protein 5	4.3		Yes
	CXCL10	chemokine (C-X-C motif) ligand 10	4.3	Yes
	CXCL13	chemokine (C-X-C motif) ligand 13	4.3
	LAMP3	lysosomal-associated membrane protein 3	3.7		Yes
	STAT1	signal transducer and activator of transcription 1, 91kDa	3
	PSMB9	proteasome (prosome, macropain) subunit, beta type, 9	2.8	Yes
	CD2	CD2 molecule	2.5	Yes
	CTLA4	cytotoxic T-lymphocyte-associated protein 4	2.5	Yes
	TOP2A	topoisomerase (DNA) II alpha 170kDa	2.1	Yes	Yes
	LCK	lymphocyte-specific protein tyrosine kinase	2.1	Yes

DISCUSSION

Using RNA and DNA profiling, we identified four stable, molecularly-defined TNBC subtypes, LAR, MES, BLIS, and BLIA, characterized by distinct clinical prognoses, with BLIS tumors having the worst and BLIA tumors having the best outcome. DNA analysis demonstrated subtype-specific gene amplifications, suggesting the possibility of using in situ hybridization techniques to identify these TNBC subsets. Our results also demonstrate subtype-specific molecular expression, thereby enabling TNBC subtype classification based on molecules they do express as opposed to molecules they do not express.

Many highly expressed molecules in specific TNBC subtypes can be targeted using available drugs (Tables 2, S36-39). Our results suggest that AR antagonists¹² and MUC1 vaccines may prove effective for the treatment of AR- and MUC1-overexpressing LAR tumors, while beta-blockers, IGF inhibitors, or PDGFR inhibitors may be useful therapies for MES tumors. Conversely, immune-based strategies (e.g., PD1 or VTCN1 antibodies) may be useful treatments for BLIS tumors, whereas STAT inhibitors, cytokine or cytokine receptor antibodies, or the recently FDA-approved CTLA4 inhibitor, ipilumimab³¹ may be effective treatments for BLIA tumors. Thus, these studies have identified novel TNBC subtype-specific markers that distinguish prognostically distinct TNBC subtypes and may be targeted for more effective treatment of TNBCs.

Lehmann's TNBC-subtyping study identified six TNBC subtypes through the combined analysis of 14 RNA profiling datasets (“discovery dataset”)¹². Assignment to these subtypes was confirmed using a second dataset comprised of 7 other publically-available datasets, however all six subtypes were not detected when subtyping was limited to only those tumors with ER, PR, and HER2 IHC data. In addition, basal-like-1 and basal-like-2 tumors are not readily distinguishable by hierarchical clustering of public TNBC data sets using Lehmann's gene signatures³², despite demonstration of molecular heterogeneity beyond the classic intrinsic subtypes. In Lehmann's study, TNBCs strongly segregated into stromal, immune, and basal gene modules, partially supporting our model. Additional studies have also demonstrated that an immune signature is an important clinical predictor for ER-negative tumors^33,27,34. The large set of ER-, PR-, and HER2-characterized tumors used in our study enabled us to further separate TNBCs into LAR, MES (including “claudin-low”), BLIS, and BLIA subtypes, and define the clinical outcome of each subtype.

Previous genomic profiling studies have not demonstrated this degree of heterogeneity in basal-like breast tumors. Profiling of TCGA data across miRNA, DNA, and methylation data supported the intrinsic subtypes of breast cancer and grouped all basal-like tumors⁸. In the Curtis dataset¹¹, unsupervised clustering by CNV-driven gene expression did not identify multiple basal-like subtypes, confirming that CNV alone does not distinguish these tumor subtypes. However, our integrated DNA and mRNA data demonstrate that gene amplification drives several subtype-specific genes. The CCND1 and FGFR2 genes are amplified in LAR tumors, while MAGOHB is more commonly amplified in MES, BLIS and BLIA tumors. Conversely, CDK1 is amplified in all 4 TNBC subtypes (most highly in BLIA tumors) and thus represents a potential target. While broad and focal CNs differentiate LAR tumors from the remaining subtypes, they cannot dissociate BLIS and BLIA tumors.

All LAR and most mesenchymal stem-like tumors identified by the Pietenpol group¹² fall within our LAR and MES subtypes. However, our study splits the remaining proposed subtypes, including Lehmann's basal-like-1 and basal-like-2 tumors into distinct BLIS and BLIA subtypes based on immune signaling. Furthermore, stratification of our subtypes is based on a few broad biological functions. LAR and MES tumors downregulate cell cycle regulators and DNA repair genes, while MES and BLIA tumors upregulate immune signaling and immune-related death pathways (Table S36-39). Conversely, our BLIS and BLIA subtypes show a relative lack of P53-dependent gene activation (P53 mutations characterize most TNBC tumors), and BLIA tumors highly express and activate STAT genes. Both our current study and the study by Lehmann et al. used RNA-based gene profiling to subtype TNBCs. Until more TNBC datasets are analyzed, it will not be clear which specific subgrouping will ultimately be most clinically useful. The study by Lehmann et al. subdivided TNBCs into 6 subtypes while this manuscript describes subgrouping of TNBCs into 4 distinct subtypes, 2 of which overlap with Lehmann et al. (LAR & MES), while our other 2 subtypes (BLIS & BLIA) contain mixtures of the other 4 Lehmann subgroups (see Figure 1 C&F). Our attempt at reproducing the 6 Lehmann et al. subgroups by clustering our data using their gene signatures was unsuccessful (n = 198, Figure S5). The exact subdivision of these TNBC subtypes, while important, is less important than the clinical prognosis defined by each subtype, and most importantly, the specific molecular targets identified within the subtypes. To this point, the identification of specific targets that modulate the immune system in the BLIA and BLIS subtypes is one of the most important and unique findings in this study.

In summary, using RNA profiling we have defined 4 stable, clinically-relevant subtypes of TNBC characterized by distinct molecular signatures. Our results uniquely define TNBCs by the molecules that are expressed in each subtype as opposed to molecules that are not expressed. Furthermore, these newly defined subtypes are biologically diverse, activate distinct molecular pathways, have unique DNA CNVs, and exhibit distinct clinical outcomes. By identifying molecules highly expressed in each TNBC subtype, this study provides the foundation for future TNBC subtype-specific molecularly-targeted and/or immune-based strategies for more effective treatment of these aggressive tumors.

Statement of Translational Relevance.

This study describes the results of RNA and DNA genomic profiling of a large set of triple-negative breast cancers. We identified four stable triple-negative breast cancer subgroups with distinct clinical outcomes defined by specific over-expressed or amplified genes. The four subgroups have been named the “Luminal / Androgen Receptor (LAR)”, “mesenchymal (MES)”, “basal-like / immune-suppressed (BLIS)”, and “basal-like / immune activated (BLIA)” groups. We also identified specific molecules that define each subgroup, serving as subgroup-specific biomarkers, as well as potential targets for the treatment of these aggressive breast cancers. Specific biomarkers and targets include the androgen receptor, MUC-1, and several estrogen-regulated genes for the LAR subgroup; IGF-1, prostaglandin F receptor for the MES subgroup; SOX transcription factors and the immune regulatory molecule VTCN1 for the BLIS subgroup; and STAT transcription factors for the BLIA group. Thus, these studies form the basis to develop molecularly targeted therapy for triple-negative breast cancers.