Molecular subtyping of breast cancer improves identification of both high and low risk patients

Abstract

Background: Transcriptome analysis enables classification of breast tumors into molecular subtypes that correlate with prognosis and effect of therapy. We evaluated the clinical benefits of molecular subtyping compared to our current diagnostic practice.

Materials and methods: Molecular subtyping was performed on a consecutive and unselected series of 524 tumors from women with primary breast cancer (n = 508). Tumors were classified by the 256 gene expression signature (CIT) and compared to conventional immunohistochemistry (IHC) procedures.

Results: More than 99% of tumors were eligible for molecular classification and final reports were available prior to the multidisciplinary conference. Using a prognostic standard mortality rate index (PSMRi) developed by the Danish Breast Cancer Group (DBCG) 39 patients were assigned with an intermediate risk and among these 16 (41%) were furthermore diagnosed by the multi-gene signature assigned with a luminal A tumor and consequently spared adjuvant chemotherapy. There was overall agreement between mRNA derived and IHC hormone receptor status, whereas IHC Ki67 protein proliferative index proved inaccurate, compared to the mRNA derived index. Forty-one patients with basal-like (basL) subtypes were screened for predisposing mutations regardless of clinical predisposition. Of those 17% carried pathogenic mutations.

Conclusion: Transcriptome based subtyping of breast tumors evidently reduces the need for adjuvant chemotherapy and improves identification of women with predisposing mutations. The results imply that transcriptome profiling should become an integrated part of current breast cancer management.

Introduction

Despite the early detection and improvement of treatment over the last decades, breast cancer is still the second leading cause of cancer-related death in women [1]. Transcriptomic and genomic profiling has enabled classification of breast cancer into intrinsic molecular subtypes and breast cancer is no longer considered a single disease [2–6]. The subtypes are biologically distinct entities with specific prognostic and therapeutic features. The pivotal study proposed five subclasses: (i) the ER-receptor positive and human epidermal growth factor receptor 2 (HER2)-receptor negative tumors i.e., luminal A (lumA), luminal B (lumB) and normal breast-like subclass, (ii) the HER2-receptor positive tumors: HER2-like subclass and (iii) the ER- and HER2-receptor negative tumors called the basal-like (basL)subclass [2–5]. Four of the subclasses can be distinguished by a 50-gene molecular classifier (PAM50) which has been developed as a commercial FDA approved platform (Prosigna^®) [7]. Recent taxonomies optimized the subclasses by applying integrative genomic analysis [8,9] and Guedj et al. [9] refined the subclasses by introducing six stable molecular subtypes based on genomic rearrangement and the expression of 256 transcripts. This taxonomy is remarkably robust and has been validated in nearly 3000 breast cancer samples, showing a high correlation between clinical characteristic and patient outcome [9]. The CIT-classifier is however not validated in randomized trials like, e.g., the PAM50/Nanostring and the MammaPrint signatures [10,11]. The PAM50 and the CIT signatures are based on RNA isolated from tissue stored under very different conditions and include an uneven number of subgroups, e.g., the PAM50 do not comprise the normal-like (normL) or the molecular apocrine (mApo) subgroups; thus a direct comparison of the PAM50 signature derived from microarray data is irrelevant. Based on survival data, other commercial molecular algorithms have emerged, that score the samples into low, medium or high clinical risk of recurrent disease [7,12–14]. However, a comprehensive genomic study integrating both genetic and epigenetic alterations concluded that breast cancers are biologically defined by five intrinsic subtypes and that clinical heterogeneity can be explained by subsets within the subtypes [15]. We chose to implement the CIT-classifier as a supplement to the existing diagnostic set-up, as it was built on an open-source and high-throughput platform (Affymetrix U133A gene expression microarrays) and included all of the intrinsic subtypes [16,17]. Moreover the transcriptomic data could provide any additional signatures, e.g., receptor status, without excess costs.

Still, in daily clinical practice the definition of molecular subtypes are only important if this knowledge improves the standard of care. It is well-established that the biological hallmark of luminal A subtype is low proliferation, high expression of the ESR1 gene and a favorable clinical outcome [4,5,18]. Since 2011, the St Gallen international expert consensus panel has recommended merely endocrine therapy in patients with luminal A disease [19,20]. Attempts have been made indirectly to approximate luminal A - like subtype by the use of IHC biomarkers - ER and/or PGR positive, HER2 negative and low Ki67 protein staining [20–22,]. Classification with only four biomarkers may however not entirely recapitulate the intrinsic subtype of breast cancer [18,23], so from 2014, we optimized our diagnostic work-up and implemented the six-class taxonomy on all patients undergoing breast cancer surgery at Rigshospitalet as a supplement to the existing procedures (Figure 1). Subsequently (in the current year), the Danish Breast Cancer Group (DBCG) introduced their revised guidelines recommending a PAM50 classifier for patients at intermediate risk.

Figure 1. Complete diagnostic work-up. A workflow illustrating the routine assessment of consecutive breast cancer patients enrolled in the complete diagnostic work-up including both standard histopathological evaluation and microarray analysis. In addition, blood samples were obtained for screening of germline predisposition, in case of a basal-like subtype, receptor negative profile or if patients were under the age of 40. A final report on molecular subtyping was available for clinical decision at the following multidisciplinary conference.

Here, we present the data from the first year of systematic molecular diagnostics of unselected breast cancer patients. The aim was to determine the feasibility of microarray-based transcriptomic profiling and to evaluate the reliability of molecular subtyping compared to our traditional stratification of the patients. Moreover, we analyzed the presumptive benefits of molecular subtyping for selection of patients eligible for genetic screening of six breast/ovarian cancer predisposing genes. Taken together we report that transcriptome based subtyping of breast tumors reduces the need for adjuvant chemotherapy and improves identification of women with predisposing mutations compared to the conventional work-up.

Material and methods

Patients and tumor samples

Over 11 months (2014–2015) consecutive female breast cancer patients (Stage I–III) were included in the study cohort provided they received breast and histopathological assessment at Rigshospitalet. The study was approved by The Danish data Protection Agency (jr. no.: 2012-58-0004) and DBCG, (jr. no.: DBCG-2015-14). Following surgical resection, fresh tumor specimens were evaluated by designated pathologists and tumor biopsies (≈ 100 mg) were stored in RNALater (Thermo Fisher Scientific Co., Waltham, MA,USA). A neighboring tumor section was sampled for verification of invasive tumor tissue in each case. Screening for genetic pre-disposing mutations followed the guidelines of DBCG (In short: breast cancer before the age of forty, both breast and ovary cancer, two first-degree relatives diagnosed before the age of fifty or three first-degree relatives with at least one before the age of fifty). Following implementation of molecular subgroups, patients with either a TNBC or a basL subtype were likewise tested for predisposing mutations.

Clinical risk assessment

According to the treatment algorithm of DBCG, patients 60 years or older with a node negative T1, ER positive and HER2 negative breast cancer are assigned to the low risk group and is not recommended adjuvant systemic treatment. Patients in the intermediate and high risk groups with ER positive breast cancer are recommended adjuvant endocrine therapy with tamoxifen or an aromatase inhibitor according to their menopausal status [24]. One year of trastuzumab combined with three weekly cycles of eprirubicin and cyclophosphamide followed by three weekly cycles of docetaxel were recommended to patients with HER2 overexpressed or amplified tumors [25]. In postmenopausal patients with intermediate or high risk ER positive breast cancer a prognostic standard mortality rate index (PSMRi) is used to allocate patients to adjuvant chemotherapy. The PSMRi was built by the DBCG, from the data of 6529 postmenopausal patients, post-surgery for ER positive breast cancer were allocated to five years of an aromatase inhibitor or tamoxifen [26]. Patients without excess mortality by PSMRi are not recommended for adjuvant chemotherapy.

Gene expression analysis

RNA was isolated using the AllPrep DNA/RNA purification kit (Qiagen, Hilden, Germany) and the QIACube workstation according to the manufacturer's instructions. The integrity of the RNA was measured using the Agilent RNA 6000 Nano Kit on an Agilent 2100 Bioanalyzer (Agilent Technologies, Inc., Santa Clara, CA, USA). The purified RNA was immediately analyzed on arrays. RNA was reverse transcribed and used for cRNA synthesis, labeling and hybridization with GeneChip® Human Genome U133 Plus 2.0 Array (Affymetrix, Santa Clara, CA, USA) according to the manufacturer’s protocol. The arrays were washed and stained with phycoerytrin conjugated streptavidin using the Affymetrix Fluidics Station 450 and the arrays were scanned in the Affymetrix GeneArray 3000 7G scanner to generate fluorescent images. Cell intensity files (.CEL files) were generated in the GeneChip Command Console Software (AGCC; Affymetrix). Raw intensity .CEL files were preprocessed by quantile normalization and gene summaries were extracted via robust multi-array average (RMA). The probe level data (.CEL files) were transformed into expression measures using R version 3.2.2 (https://www.R-project.org/). Expression values of each sample were combined in a batch with expression measures of 12 selected samples from the study. Subsequently, the values were batch corrected against CIT ‘core set’ using ‘ComBat’ implemented in ‘sva’ R package [27,28]. After such processing, breast cancer subtyping was performed using CIT predictor [9] incorporated in ‘citbcmst’ R package (http://CRAN.R-project.org/package=citbcmst). The subtypes are assigned to the closest centroids computed on 375 probe sets. To test the in-house settings, we used the CIT ‘core set’, and found compliance of the assigned subclasses. Microarray data are available as .CEL files in the online repository Array Express (accession number: E-MTAB-5724). It was noted, that the subtypes derived by the CIT classifier correlate to the ones from the PAM50 classifier, with few differences; CIT normal-like subgroup does not correspond to normal breast tissue but exhibits similar expression profiles to the luminal A subgroup. The CIT normal-like samples would classify as PAM50-luminal A. The CIT luminal C (lumC) and molecular apocrine subtypes include samples with overexpressed HER2; however the two subtypes represent patients with distinct ER status (positive and negative, respectively). CIT luminal A and B largely overlap with the PAM50 luminal A and B subtypes.

Proliferative index and receptor profile

All samples were assigned with a relative proliferative index (PI), comprising expression values of 79 genes encoding proliferative and cell cycle markers (100 probe sets). Previous in-house analysis revealed difference in PI between normal tissue and malignant tumor samples (data not shown) and normal tissue samples expressed a PI below five, thus a cutoff value was set at 5.5 and a PI ≥5.5 ensured that the tissue was malignant. Samples with a PI <5.5 were subjected to a cancer-type classifier consisting of 641 probe sets and more than 2400 samples were derived [29]. Samples classified as normal tissue by mRNA profile were excluded from further analysis. Confirmed tumor samples with PI <6.5 were assigned as low proliferative in the subsequent analysis. The expression value thresholds for receptor status were set by default as (intermediate-positive): HER2 (11.001–11.5001), estrogen receptor (ER) (9–9.5), and progesterone receptor (PGR) (6–7) For MKI67 expression status, the average expression value of two probe sets were used (212023_s_at and 212021_s_at).

Statistical analysis

Principal component analysis (PCA) analysis and visualization of the data were performed in the Qlucore Omics Explorer™ software (Qlucore AB, Lund, Sweden). Correlation calculations, plots and boxplots were generated in R. We applied the Anderson-Darling for normality-test of the data derived from the proliferative index (PI). Testing for overrepresentation of germline mutations among basL-subtypes was done by using Fisher’s exact test.

Histopathological diagnosis and ER, HER2 and Ki67 protein immunohistopathological analysis

Standard histopathological diagnosis of breast cancer samples was performed by a designated pathologist by light microscopy on glass slides from formalin fixed, paraffin embedded tissue blocks, according to the WHO-classification recommendations [30]. Analyses for ER, Ki67 protein and HER2 were performed by immunohistochemistry (IHC) using tissue micro array technique (TMA), with two cores of 2 mm from the invasive front of each tumor, as previously reported [31]. Staining for ER (SP1, diluted 1:25), Ki67 protein (MIB1; murine monoclonal antibody 1) and HER2 (4B5), all from Ventana Medical Systems, were carried out according to the manufacturer’s instructions. Scoring of ER and Ki67 protein were semi quantitative with a positive cutoff point of >1% for ER positive tumor. Scoring of HER2 was performed as described by Hansen et al. [32] following the national guidelines (www.dbcg.dk). Online available datasets for comparison of percentage of positive Ki67 protein-cells were downloaded from www.ebi.ac.uk/arrayexpress/ (accession numbers; E-GEOD- 43358, -76040 and -76250).

Germline mutation screening

Genomic DNA was isolated using the ReliaPrep Large Volume HT gDNA Isolation Kit (Promega, Madison, WI, USA) and a Tecan Freedom EVO HSM2.0 Workstation according to the manufacturer’s instructions. Germline mutation screening was performed using a gene panel consisting of six breast cancer-predisposing genes, including BRCA1, BRCA2, CDH1, PTEN, RAD51C and TP53 as described by Jonson et al. [33]. Sequencing was performed on a MiSeq (Illumina, San Diego, CA, USA) to an average depth of at least 100×. Sequencing data were analyzed using Sequence Pilot (JSI medical systems, Ettenheim, Germany), where variants were called if the non-reference base frequency was above 25%. Variants are numbered according to the following GenBank accession numbers: NM_007294 (BRCA1), NM_000059 (BRCA2), NM_004360 (CDH1), NM_000314 (PTEN), NM_058216 (RAD51C) and NM_000546 (TP53) using the guidelines from the Human Genome Variation Society (www.hgvs.org/mutnomen). All variants, except well-known polymorphisms and neutral variants were verified by Sanger sequencing on an ABI 3730 DNA analyzer using DNA purified from a second blood sample.

Results

Baseline characteristics

Clinical characteristics of the five hundred and eight patients and basic histopathological features of their tumor samples are summarized in Table 1. More than 80% of the patients were aged 50 or older and the majority were diagnosed with stage I and II breast cancer. In total 524 tumor samples were included in the study. Sixteen of the patients had two tumors at the time of diagnosis and one sample was incorrectly preserved for RNA-extraction and could not undergo molecular classification, resulting in a total of 523 tumor samples which were finally eligible for molecular analysis.

Table 1. Clinical characteristics of patients and basic histopathological features of primary tumor samples.

Patients^a	Number	%
Age at diagnosis (n = 508)
<50 years	93	18.3
≥50 years	415	81.7
Size of primary tumor (508)
<2 cm	303	59.6
2.1–5 cm	189	37.2
>5 cm	16	3.1
Positive Lymph Nodes (n = 508)
0	300	59.1
1–3	143	28.1
≥4	63	12.4
N/A	2	0.4
Malignancy grade^b (n = 524)
1	39	7.4
2	329	62.8
3	137	26.1
N/A	19	3.6
Treatment (n = 508)
No adjuvant therapy	29	5.7
CT^c+ET(+T, if HER2+)	224	44.1
ET	193	38.0
CT (+T, if HER2+)	55	10.8
N/A	7	1.4

^a Patients with two tumors: n = 16. Tumor samples: n = 524.

^b Malignancy grade is assigned for lobular and ductal carcinomas.

^c EC (Epirubicin + Cyklofosfamid) ×3 → Docetal ×3

CT: chemotherapy; ET: tamoxifen; T: trastuzumab; N/A: not applicable; HER2: human epidermal growth factor receptor 2

Molecular subtyping of breast cancer

Molecular subtyping was initiated on the day of the primary breast surgery and the diagnostic report was completed within six days and available prior to the clinical multidisciplinary conference (Figure 1). Only samples with proliferative index (PI) >5.5 were considered for subtype classification. Samples with PI <5.5 (n = 54) were classified by our cancer-type classifier [29] and samples classified as normal tissue were excluded from the analysis (n = 3), resulting in 520 tumor samples for molecular subtyping. The distribution of molecular subtypes in the study cohort is outlined in Table 2, together with treatment regimen for patients with ER-positive samples. Close to 15% of the samples could not be classified as a core-class and were consequently assigned as ‘mixed’, meaning that they consisted of two or three subtypes. Among the 76 mixed samples, 70 fell between the luminal subtypes where the lumA/normL were the most frequent (n = 29) combination. The remaining six mixed samples fell between either Basl/mApo or mApo/lumC. The spatial distribution of molecular subtypes and the WHO morphological classes is depicted in the principal component analysis (Figure 2(A)). Molecular subtypes were not associated with neither the WHO stage (Figure 2(B)) nor the histological subtype (invasive ductal (78%), invasive lobular (11%), mucinous, tubular and others) and molecular subtype (Figure 2(C)).

Figure 2. Principal component analysis of subtypes, stages and histopathological diagnosis. (A) A principal component analysis (PCA) showing the distribution of 520 breast cancer samples according to six molecular subtypes based on the expression profiles of the 375 probe sets. The subclasses cluster was assigned as separate entities with the samples classified as mixed located in the areas between the core clusters. The basal-like subtype, representing the double negative breast cancer samples, forms its own distinct cluster. (B) The distribution of patient stage from I–III in the spectra of subclasses clearly illustrates that assigned subclass is not dependent on patient stage. (C) The scattering of the assigned WHO histopathological diagnosis in the allocated molecular subclasses demonstrates that there is no association between molecular subgroup and histopathological diagnosis.

Table 2. Distribution of the molecular subtypes of 520 samples.^a

Subtype	Basal-like	Molecular apocrine	Luminal C	LuminalB	LuminalA	Normal-like	Mixed
All (n = 520)	50	28	40	76	182	68	76
	(9.6%)	(5.4%)	(7.7%)	(14.6%)	(35.0%)	(13.1%)	(14.6%)
ER positive^b (n = 461)	14	10	39	75	182	67	74
	(3.0%)	(2.2%)	(8.5%)	(16.3%)	(39.5%)	(14.5%)	(16.1%)
HER2 positive^c (n = 69)	0	17	18	17	5	2	10
	–	(24.6%)	(26.1%)	(24.6%)	(7.2%)	(2.9%)	(14.5%)
No adjuvant therapy	1	–	1	2	17	3	4
	(7.1%)		(2.6%)	(2.7%)	(9.3%)	(4.5%)	(5.4%)
ET	–	1	9	11	105	37	32
		(10.0%)	(23.1%)	(14.7%)	(57.7%)	(55.2%)	(43.2%)
CT^d (+T, if HER2+)	1	1	–	–	1	–	–
	(7.1%)	(10.0%)			(0.5%)
CT^d+ET (+T, if HER2+)	11	8	29	60	59	26	36
	(78.6%)	(80.0%)	(74.4%)	(80.0%)	(32.4%)	(38.8%)	(48.6%)
N/A	1	–	–	2	–	1	2
	(7.1%)			(2.7%)		(1.5%)	(2.7%)

^a In total 524 tumor samples were included in the cohort. One sample was incorrectly preserved for RNA-extraction and three samples were not classified as they consisted of normal tissue (0.8%).

^b ER + determined by IHC (n = 461).

^c HER2 status determined by immune histochemistry and fluorescent in situ hybridization.

^d EC (Epirubicin + Cyklofosfamid) × 3 → Docetal × 3.

CT: chemotherapy; N/A: not applicable; ET: tamoxifen; T: trastuzumab; ER: estrogen receptor; HER2: human epidermal growth factor receptor 2.

Receptor status defined by IHC and expression data

We compared the status of hormone receptors assigned by IHC and/or fluorescent in situ hybridization (FISH) with their microarray derived status (Supplementary Table 1). ER-positive breast cancer samples by IHC and those assigned by expression array overlapped to a large extent independently of the threshold levels (Supplementary Figure 1(A)). The algorithm identified 91% (420/461) of the IHC-positive samples with the 1% cutoff, whereas 33 (IHC 9%) and 41 (IHC 1%) samples, respectively, were not assigned as positive by the array. HER2 receptor status by IHC was scored according to international guidelines, where score 0 and 1 are considered negative, 2 is equivocal (considered positive by a HER2/centromere 17 ratio ≥2 by FISH) and 3 is considered as positive [34]. In total, 71 samples were HER2-positive by IHC and FISH and 42 samples were HER2-positive by array (Supplementary Table 1). There was a strong correlation between samples assigned as negative by array and scored 0–2 by IHC, whereas the correlation between HER2 positive samples was less profound (Supplementary Figure 1(B,C)). Twenty-seven out of 71 IHC/FISH positive samples (38%) did not exhibit elevated mRNA expression and were subsequently reevaluated by a designated pathologist. This provided an explanation in 70% of cases (n = 19) and were related to multiple tumors and intratumor heterogeneity or border-line assessments.

Proliferation index and Ki67 protein expression

Ki67 protein expression (IHC) is of some prognostic significance in breast cancer [35,36] and we compared the analytical validity of the analysis to a PI based on gene expression profiles [29]. In addition to our own data set, we examined three online available datasets (n = 368 samples in total) hereby confirming the distribution of the percentages of Ki67 protein positive cells, although identification of 60 and 70% nearly did not occur as the in-house opposed the on-line data (Figure 3(A)). In contrast to scoring of Ki67 by IHC, signature based assessment of proliferation is based on the expression levels of 79 transcripts and thus provides an objective, quantitative and reproducible proliferation value. The distribution of PI in our cohort was seen to approach that of a normal Gaussian distribution, as p = .04 by the applied test for normality (Figure 3(B)). Furthermore, there was a strong correlation, r = 0.91, between the PI levels (calculations) and the transcript levels of Ki67 protein (Figure 3(C)). On the contrary, the percentage of Ki67 protein positive cells showed a weak correlation with PI (r = 0.68;Figure 3(D)). There was discrepancy in 20 tumors that had 15% or less of Ki67 protein positive cells, but were scored with high PI ( >7). A review of the individual histopathological reports clarified four of the cases; one sample was due to fibroadenomatosis and three were due to multiple tumors. Two tumors had 80% Ki67 protein positive cells but low PI and the review of the discrepancy could only be clarified for the one, since it was due to multiple tumors, whereas a plausible explanation for the second sample remains unknown.

Figure 3. Correlation of Ki67 protein by immunohistochemistry (IHC) and expression profile. (A) The distribution of Ki67 protein positive cells from each representative tumor slide showing peaks from 5–30, 50 and 80% (in-house data to the left and online data to the right). (B) The distribution of the proliferative index (PI) extracted from the microarray analysis from each tumor specimen resembles that of a normal distribution. (C) The correlation of the PI and MKI67 protein encoding Ki67 shows a high correlation of r = 91. (D) The correlation of Ki67 protein positive cells by IHC and PI from array analysis shows a reduced correlation of just r = 0.68. Twenty-two samples (marked in square boxes) were found to have negative correlation and rendered further investigation.

Luminal A subtyping

To explore possible benefits of implementing the 256-gene expression signature for subtyping in a routine diagnostic work-up of breast cancer patients, we focused on the patient group who might benefit from omitting chemotherapy. Based on the prognostic standard mortality rate index (PSMRi), 39 women were annotated as being in the intermediate-low risk group. Of the 39 patients, 16 (41%) had a luminal A subtype and were treated solely with endocrine therapy. In our consecutive cohort, a total of 195 patients were treated with endocrine therapy alone (Table 2). Hence, as a direct consequence of implementing molecular subtyping, we showed an increase of 9% (16/179) in comparison to the original diagnostic set-up, in the total number of patients treated with endocrine therapy alone.

Identifying genetic predisposition by molecular subtype

As a part of the diagnostic work-up, screening for pathogenic germline mutations in breast/ovarian cancer predisposing genes (BRCA1, BRCA2, CDH1, PTEN, RAD51C and TP53) was performed. In total, 70 patients were screened, resulting in identification of eight pathogenic BRCA1/BRCA2 mutations (Supplementary Table 2). Noteworthy, seven of the eight pathogenic mutations were assigned to the basL-subtype and 17% (7/41) of the patients with a basL-subtype were carriers of a pathogenic BRCA1/BRCA2 mutation. Nine out of the 50 patients with a basL-like subtype did not deliver a blood-sample for genetic screening. In addition, 12 out of 23 receptor-negative tumors, predominated by mApo subtype, were screened resulting in the identification of a single pathogenic BRCA1 mutation. Finding of a pathogen germline mutation is significantly enriched in a basL-like subtype in comparison to a receptor-negative, non-basL-like subtype (p < .05). Moreover, 17 patients were screened due to young age (<40 years) among patients with a non- basL-like subtype (lumA (n = 4), lumB (n = 6), lumC (n = 3), mApo (n = 2), normL (n = 1) and mixed (n = 1)), did not carry any germline mutations in the six genes analyzed. Thus, the results indicate that basL-subtype is a predictor of patients at a greater risk of carrying a BRCA1/BRCA2 germline predisposing mutation.

Discussion

For more than a decade, several studies have shown that primary breast cancers can be classified according to specific molecular based signatures into intrinsic subtypes. Accordingly, breast cancer classification is moving towards the molecular classification based on whole-transcriptome profiling. Molecular taxonomies comprise important information regarding diagnostics, treatment and clinical outcome [12,37]. Accommodating this evolution, the microarray based molecular subtyping (CIT-classification) was implemented as a supplement to our routine diagnostic and clinical setting for all primary breast cancer patients undergoing surgery at Rigshospitalet, Copenhagen University Hospital, Denmark.

Strengths of our study include that RNA-yield and downstream gene expression analysis from fresh or frozen biospecimens are superior to formalin-fixed paraffin-embedded (FFPE) tissues and the reduced gene expression from FFPE stored samples may tend to underestimate expression of important biomarkers [38,39]. However, classifications by multi-gene signatures tend to perform reasonablywell on both FFPE and fresh frozen tissue [40]. A corner-stone in subtyping of breast cancer is a precise measurement of proliferation rate as well as ER and HER2-status. We found some concordance (91%) between ER positive samples comparing IHC and mRNA results; however expression array failed to detect a number of IHC assigned ER-positive samples. The advantage of IHC staining is the ability to capture a single ER-positive cell among 100 tumor cells, excluding any surrounding or normal breast cancer tissue in the examination. We tried to minimize the normal tumor contamination by letting the resection of tumor-biopsies for transcriptome analysis solely handled by our specialized breast cancer pathologist. However, the risk of normal tissue contamination, especially among small size tumors, are a possible limitation of our study and a small number of samples were not included in the study at all, when the pathologist had evaluated the tumor to be too minute to spare material or ensured high tumor cell content in the sampling.

Our result, as well as others, clearly depicts the limitations of Ki67 protein staining used for quantitative measurements, with a specific threshold, probably due to the heterogeneous expression on IHC [41–43]. In regard to the proliferation rate, the St. Gallen consensus rapport is complex, since clinic-pathological luminal A-like subtype definition is dependent on Ki67 protein staining with a cutoff that varies between laboratories [20]. We have shown that Ki67 protein expression is nearly distributed as normal Gaussian and well-correlated with the PI index, whereas Ki67 protein assessment by histopathological methods is biased. However, the poor correlation between expression levels and IHC in our study may be more prominent since Ki67 protein merely was evaluated based on two TMA cores as opposed to a whole slide section. Stressing the fact that these results are obtained from a routine diagnostic and clinical setting, expression array derived measurements for proliferation is superior to standard IHC.

The choice of method for luminal A subclass identification may depend on the available molecular platforms at-hand. As discussed in the above section, IHC for luminal Alike classification is suboptimal since it is dependent on the Ki67 protein index. Indeed, it is by now established that multi-gene signatures are superior to single-biomarker subtype classification and a subset of 50 genes was found to be the minimum number of genes in order to robustly identify the four basic intrinsic subtypes (luminal A, luminal B, HER2-enriched and basL) without compromising precision [7,18,23]. To ensure that the majority of the biological heterogeneity was contained in each consecutive sample, we chose a comprehensive taxonomy model for optimal distinction between the intrinsic subclasses. Moreover, the multi-gene platform enables an illustrative presentation of each sample according to the ‘reference cluster’. This may be in contrast to the smaller gene-panels, where edgy samples remain unrecognized. Indeed, this is relevant when considering the heterogeneity of the large ER+/PGR + groups.

An interesting group is a group of mixed samples. Close to 15% of our samples could not be assigned with a core class and might therefore represent a challenge when conveying research data into clinical practice. Similarly in the comparative French cohort, one third of samples were not assigned with the same subclass in all three classification algorithms and therefore represented a mixed group. Moreover, the fraction of non-tumor cells may dilute the intrinsic subclass-signal and questions have been raised, whether the norm-like subgroup is in fact an artifact due to high-contest of non-tumor cells [44]. Still, we cannot exclude that the fraction of normal tissue content in our samples results in the relatively large number of both mixed and normal-like samples, which remains a weaker point of the methodology of our study. To address this important issue, Guedj et al. [9] estimated the rate of non-diploid cells and their distribution within the subgroups by SNP array data and showed that normL ranked third and LumA and LumB consisted of the highest fraction of non-diploid cells. They validated the SNP-array data by histological estimates of non-tumor cell fraction and found that the SNP-array based assessments of non-diploid cells were in fact lower than pathological tumor cell content hereby substantiating that normL is a recognized subgroup [9]. It is reasonable to assume that low-risk patients suffering from a normL tumor also can be spared chemotherapy, all though long-term follow-up and additional samples are needed to test this hypothesis.

It is well-founded that the ‘triple-negative’ phenotype and/or the basL subtype are associated with the risk of a germline BRCA1 mutation [2,5,45]. As depicted in Figure 1, all patients under 40 years of age with a basL subtype or receptor negative tumor were eligible for screening of pathogenic germline mutations. We found an overrepresentation of BRCA1, and to some extent, BRCA2 mutations in the basL subtype, since 17% of the patients were genetically predisposed to breast cancer. The original retrospective study identified 11% BRCA1 germline mutation carriers among 268 breast cancer patients with a basL phenotype [2]. This is in agreement with our prospective cohort where 12% (5/41) of the basL subtype subsequently was found to harbor a BRCA1 germline mutation. While only half as many patients with a non-basL substype were screened (n = 23 vs. n = 41) it is remarkable, that we merely identified a single predisposing BRCA1 mutation in these patients. It is obvious to suggest, that the young patients with a non-basL subtype most likely harbor other predisposing genomic alterations other than the genes included in our NGS-panel (BRCA1, BRCA2, CDH1, PTEN, RAD51C and TP53). An ongoing large-scale screening and validation of candidate breast cancer predisposition genes aim to identify exactly which genes it will be [46].

In this prospective study, close to 100% of the samples were eligible for analysis demonstrating that microarray based classification is suitable for clinical practice. Moreover, time is an important parameter in cancer diagnostics and here we showed that an array-based signature is only a few-day procedure from surgery to clinical report.

To the best of our knowledge, this study is one of the first reporting the results of implementing a six subclass array-based classification into a diagnostic setting as a supplement to the existing diagnostic work-up. The cohort is increasing prospectively and ongoing retrospective evaluation of the results will be carried out with rational intermissions. This study has focused on the two extremes of the benefits of molecular subtyping; the identification of intermediate risk patients with not only a luminal A , but also with a basL tumor and a greater risk of being genetically predisposed. Future studies should unravel the clinical relevant characteristics of the remaining subgroups by subsequent mutational testing, prognostic outcome and treatment-regime, since this may possibly pinpoint the breast cancer patients that could benefit from an optimized personalized treatment and follow-up regimen.

Acknowledgments

We thank Susanne Smed, Maria Gushchina and Julie Fisker Nielsen for excellent laboratory assistance.

Disclosure statement

None of the authors have any potential conflicts of interest.

Additional information

Funding

The study was supported by the Capital Region of Denmark.