BluePrint breast cancer molecular subtyping recognizes single and dual subtype tumors with implications for therapeutic guidance

Purpose BluePrint (BP) is an 80-gene molecular subtyping test that classifies early-stage breast cancer (EBC) into Basal, Luminal, and HER2 subtypes. In most cases, breast tumors have one dominant subtype, representative of a single activated pathway. However, some tumors show a statistically equal representation of more than one subtype, referred to as dual subtype. This study aims to identify and examine dual subtype tumors by BP to understand their biology and possible implications for treatment guidance. Methods The BP scores of over 15,000 tumor samples from EBC patients were analyzed, and the differences between the highest and the lowest scoring subtypes were calculated. Based upon the distribution of the differences between BP scores, a threshold was determined for each subtype to identify dual versus single subtypes. Results Approximately 97% of samples had one single activated BluePrint molecular subtype, whereas ~ 3% of samples were classified as BP dual subtype. The most frequently occurring dual subtypes were the Luminal-Basal-type and Luminal-HER2-type. Luminal-Basal-type displays a distinct biology from the Luminal single type and Basal single type. Burstein’s classification of the single and dual Basal samples showed that the Luminal-Basal-type is mostly classified as ‘luminal androgen receptor’ and ‘mesenchymal’ subtypes, supporting molecular evidence of AR activation in the Luminal-Basal-type tumors. Tumors classified as Luminal-HER2-type resemble features of both Luminal-single-type and HER2-single-type. However, patients with dual Luminal-HER2-type have a lower pathological complete response after receiving HER2-targeted therapies in addition to chemotherapy in comparison with patients with a HER2-single-type. Conclusion This study demonstrates that BP identifies tumors with two active functional pathways (dual subtype) with specific transcriptional characteristics and highlights the added value of distinguishing BP dual from single subtypes as evidenced by distinct treatment response rates. Supplementary Information The online version contains supplementary material available at 10.1007/s10549-022-06698-x.


Introduction
Breast cancer (BC) is a heterogenous disease with respect to clinical, histopathological, and molecular features. Based on clinical behavior and genomic characteristics, multiple methods have been utilized to categorize BC into distinct subgroups, be it with clinical subtyping for hormone receptor (HR) protein status or more recently with molecular subtyping based on RNA assays [1][2][3][4].
Clinical subtyping relies on well-established immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH) staining that determines estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) status [5,6]. The BluePrint (BP) 80-gene subtyping assay was developed to bridge clinical pathology and molecular subtyping, by using IHC-based receptor status and mRNA expression, resulting in a molecular diagnostic array with predictive value [1,7]. Each of the three subtypes determined by BP (Basal-type, Luminal-type, and HER2-type) is scored according to their respective gene signatures (consisting of 28, 58, and 4 genes, respectively) reflecting specific functional pathways, with the highest score determining the subtype [1,7]. In most cases, the highest score is significantly higher than the score of the other two subtypes, indicating a strong dominance of a single pathway activation in the tumor (so-called single subtype). However, in rare instances, the difference between the highest score and the second-highest score is statistically indiscernible, indicating that these tumors might be characterized by multiple activated pathways (dual subtype). Having a deeper understanding of which pathways are activated may help understanding the specific biology of BP dual subtypes that distinguish them from the single subtypes.
In addition to the standard BP subtypes (Basal-type, Luminal-type, and HER2-type), other studies have identified expression-based subtypes, which include normal-like, claudin-low, triple positive, and triple-negative [3,[8][9][10][11] types. Among others, Burstein and colleagues further classified the clinical triple-negative breast cancer (TNBC) subtype into basal-like immuno-activated (BLIA), basal-like immunosuppressed (BLIS), luminal androgen receptor (LAR), and mesenchymal-like (MES) [12]. Therefore, assessing further the differences between the BP scores may help identifying additional subtypes previously not detected by standard BP. Also, understanding the biological characteristics of BP dual subtypes may help in guiding more effective treatment plans.

Data
For this study, only data and no samples were collected, and all patient data were fully anonymized according to the 'General Data Protection Regulation' (GDPR) and the 'Health Insurance Portability and Accountability Act' (HIPAA) and are in compliance with the 'Data Protection Act.' This study was a retrospective analysis of (internal) studies between 2015 and 2020. These studies included those previously described in Beumer et al. [13], the FLEX registry trial (NCT03053193), the Neoadjuvant Breast Registry Symphony Trial (NBRST) (NCT01479101), and the Multi-Institutional Neo-adjuvant Therapy MammaPrint Project I (MINT) trial (NCT01501487). Most samples comply with MammaPrint (MP) eligibility criteria [14,15], stage I, II, or operable stage III breast cancer, tumor diameter ≤ 5 cm , and up to three positive lymph nodes, with any ER/PR/HER2 status. Microarray processing was performed following standard procedure at Agendia [7] (Supplementary methods). Agendia's customized diagnostic arrays were either a targeted array or a full genome array, as previously described [7,13,16].
Of the 15,580 samples analyzed with BP, 7985 had full-genome expression data available of which 1978 with clinicopathological information (Table 1). All samples analyzed with the targeted array had clinicopathological information available (Table 1b).
The Neoadjuvant Breast Registry Symphony Trial (NBRST) [17][18][19] classified BC patients according to MP and BP and compared it with conventional IHC/FISH subtyping to predict treatment sensitivity. From the entire NBRST trial dataset (n = 1060), a subset that received HER2-targeted therapy (n = 289) was used to evaluate the association between the dual subtypes and response to HER2-targeted therapy. The NBRST trial protocol was approved by Institutional Review Boards at all participating sites (ClinicalTrials.gov NCT01479101). All patients consented to participation in the study and clinical data collection. Part of the anonymized data (BP results and IHC) used in this study was generated from early-stage BC patients collected from standard diagnostic testing and was only used to identify potential dual subtypes and not for any gene expression analysis. The data from studies can be shared by the authors upon reasonable request.

BluePrint single and dual-subtype classification
Standard BP scores of 15,580 samples were calculated followed by dual-subtype classification, which was based on bootstrap technique [20], and multi-modality detection. Details on the procedure can be found in the Supplementary methods and Figure S1 (Fig. S1a).

Burstein classification
An algorithm published by Burstein et al., stratifies TNBCs into different subtypes by gene expression analyses of 80 signature genes. This algorithm was used to classify the Basal-single-type and Luminal-Basal-type samples into BLIA, BLIS, LAR, and MES [12].
Hallmark and Oncogenic gene sets from the Molecular Signatures Database v7.2 were used for gene set enrichment analysis (GSEA) [22]. Genes were ranked based on the effect size ratio using the Cohen's D effect size [23]. Differentially expressed genes (DEG) were considered significant with a p value ≤ 0.05 and a log 2 fold change ≥ 1.
Computational analysis and visualization were performed using R (v3.6.1) [24]. Principal component analysis (PCA) was performed using the "prcomp" package [25] (v3.6.2) and visualized using "ggplot" (v3.3.2) [26]. Unpaired, two-sample t-tests were used to measure if the means of ER, PR, and Ki-67 positivity were significantly different between single and dual subtypes. Chi-square test of Independence was used to test for differences of categorical variables within the Burstein classification (BLIA, BLIS, LAR, and MES) and a multivariate logistic regression analysis for response to therapy (pathological complete response, pCR) between single and dual subtypes. Molecular subtype classification algorithms were used from the "Genefu" package [27].

BluePrint single and dual subtype classification
Molecular subtyping of patient tumors (n = 15,580) was performed at Agendia using the BP 80-gene assay as previously described [1,7]. We applied the dual subtype classification method (see "Methods" for details) to assess the presence of multiple activated pathways. Most tumors were classified as single subtype (n = 15,087, 96.8%), followed by 449 (2.9%) tumors classified as dual subtype, and 44 (0.3%) tumors as triple subtype (Table 1). The most common dual subtypes in this dataset were the Luminal-Basal-type and the Luminal-HER2-type. These had sufficient numbers for downstream analyses while HER2-Basal-type and Luminal-HER2-Basaltype were not sufficient in size [28] and not further analyzed (Table 1a).

Principal component analyses using BluePrint reveals similarities between subtypes
To understand the similarities between single and dual subtypes, we performed PCA based on the BP gene expression signatures ( Fig. 1a-e). We observed a clear distinction of single subtypes shown in the first two principal components ( Fig. 1a-c). Luminal-Basal-type cluster separately from both Basal-single-type ( Fig. 1d) and Luminal-single-type ( Fig. 1e), conversely, Luminal-HER2-type (Fig. 1e, f) are more closely related with both Luminal-single-type (Fig. 1b) and HER2-single type (Fig. 1c).

Differential gene expression analysis highlights differences between BluePrint dual and single subtypes
Differential expression analysis using full-genome data (n = 7985) was performed to evaluate global transcriptional differences between single and dual subtypes. As expected from the PCA, when compared with their corresponding single subtypes, more DEGs were found for the Luminal-Basal-type (446 DEGs) (Fig. 2a, b) than for the Luminal-HER2-type (151 DEGs) (Fig. 2e, f).
Among the up-regulated genes in Luminal-Basal-type (vs. both Basal-single-type and Luminal-single-type) were present MUCL1, a known tumor suppressor gene [30], and CLCA2, a negative regulator of cancer cell migration and invasion [31].
Among the most up-regulated genes in Luminal-HER2type compared with Luminal-single-type tumors, we found GRB7, TCAP, and ERBB2 which belong to the HER2 amplicon and are known to be overexpressed in pathologically confirmed HER2 tumors [32]. Indeed, these genes were also up-regulated in HER2-single-type tumors (Fig. 2e). When comparing Luminal-HER2-type with HER2-single-type, ESR1 was found to be upregulated, similarly as in Luminal-single-type tumors. Additionally, Luminal-HER2-type tumors were mainly classified as either Luminal B (n = 34/99, 34%) or HER2 enriched (n = 44/99, 44%) using the intrinsic subtype classifier of the"Genefu" [27,33]. Together, these data suggest that both ER and HER2 are activated in Luminal-HER2-type tumors.

Differences between BluePrint single and dual subtypes may impact therapy response pathways
A better understanding of the underlying biological characteristics of the dual subtypes may come from analyzing gene pathway regulation.
Comparison of Luminal-Basal-type with Basal-singletype revealed upregulation of two estrogen response (ESR) and one androgen response (AR)-related gene sets (Fig. 2c). Same ESR gene sets were downregulated in Luminal-Basaltype versus Luminal-single-type, indicating that Luminal-Basal-type has intermediate ER levels. Conversely, AR was upregulated in Luminal-Basal-type, versus both the Basalsingle-type and Luminal-single-type. G2M and E2F pathways [34,35] were either downregulated or upregulated in Luminal-Basal-type compared with Basal-single-type and Luminal-single-type, respectively, indicating that Luminal-Basal-type are less proliferative than Basal-single-type, but more proliferative than Luminal-single-type tumors. Taken together, Luminal-Basal-type tumors show a distinct biology from their single counterparts with decreased proliferation than Basal-single-type and AR activation.
Compared with single HER2-single-type tumors, a Luminal-HER2 type shows downregulation of MAPK (MEK and RAF) signaling pathways and ER activation (Fig. 2g). Clinical characteristics of the single and dual BP subtypes and their response to therapy may confirm these hypotheses and provide additional insights.

BluePrint dual subtypes present clear clinicopathological differences from single subtypes
Standard BP Luminal-, HER2−, and Basal-type tumors were further stratified using the single-dual subtyping classification (Fig. 3a). Additionally, conventional clinical subtypes (based on IHC HR staining (ER, PR) and HER2 status) were further classified into BP single subtypes or dual subtypes (Fig. 3b, c).

Burstein LAR and MES subtypes are identified using BluePrint dual subtype classification
Since Luminal-Basal-type displays different transcriptional characteristics than Luminal-single-type and Basal-singletype, we classified them using the Burstein classifier to better understand their biology. Indeed, we found a significant association between BP single/dual subtypes and the Burstein BLIA, BLIS, LAR, and MES subtypes [12] (p value < 0.001) with the Basal-single-type classified mostly as BLIA or BLIS, whereas the Luminal-Basal-type as LAR or MES (Fig. 4a), irrespective of their standard BP subtype (Fig. 4b).

BluePrint dual subtype classification of the NBRST dataset shows refined prediction to therapy
Our findings indicate that the Luminal-HER2-type shares clinical and genomic features with Luminal-single-type and HER2-single-type and previous studies suggest that HR and HER2 co-expression is associated with endocrine and HER2-targeted therapy resistance [37,38]. Therefore, to better understand how Luminal-HER2-type relates to HER2-targeted therapy response, we analyzed the NBRST dataset (see Methods for details) [18] and selected only pathologically confirmed HER2+ tumors (n = 289) with gene expression and HER2-targeted therapy response data available [either Trastuzumab (T) only or with Pertuzumab (P)]. Patient tumors were stratified using the BluePrint dual subtype classification (Fig. 5).
BP HER2-single-type showed higher pCR rate to chemotherapy (C) + T compared to Luminal-HER2-type (61.3% vs. 23.8%, p = 0.032) (Fig. 5a). Although not significant (also due to lower numerosity of Luminal-HER2-type), this trend remained for patients that received additional P (Fig. 5b). Response rates of Luminal-HER2-type tumors was higher, but not significantly different than for Luminal-single-type. Instead, a significant higher response rate was observed for    Table S2).

Discussion
Molecular subtyping using the standard BP 80-gene assay enables to discern the tumor subtype by the underlying functional pathways and not merely by HR and HER2 status [1,7]. In most cases, the assay identifies a single, dominant activated pathway distinctive of a Luminal-, Basal-, or HER2-type tumor. This information often confirms the pathologically defined subtype but in many cases further classifies tumors from their initial clinical subtype into a different molecular subtype. This phenomenon has clinical implications for the treatment of patients, perhaps most notably in the ER+/Basal and HER2+/Luminal subtypes which have been previously described [18,19,39,40]. The vast majority of the breast cancer tumors analyzed in this study using the BP test show a single activated pathway (i.e., single BP subtype) (97%); however, less frequently, they exhibit multiple activated pathways (i.e., dual or triple subtype) (3%), as we showed in a preliminary analysis [41]. Notably, this dataset mostly reflects a HR + population but upon sampling the data based on observed frequencies of clinical subtypes, such a percentage raises to ~ 5%. Importantly, the single and dual assessment performed on the NBRST dataset and also reported for the TRAIN2 [42] and APHINITY [43] patient cohorts show a higher number of dual subtypes, ranging from 11 to 30%, indicating that the dual subtype classification might have a greater clinical impact on a HER2+ population and that the potential clinical utility should be found in specific subgroups rather than in the entire EBC population. The analysis on the NBRST dataset was performed on limited numbers of dual subtypes (n = 32); however, the size was sufficient to generate statistically powerful results.
Overall, in this manuscript, we aimed to provide a better understanding of the biological diversity of EBC and these results should be taken with caution with respect to any immediate change in clinical management.
Next, by analyzing whole-transcriptomic data, we set out to understand if and how dual subtypes were distinct from single subtypes. For the analysis, we focused on the Luminal-Basal-type and Luminal-HER2-type tumors as the other dual subtypes were limited in size.
Neither the Basal nor the Luminal BP template genes were able to fully capture the biology of the Luminal-Basaltype tumors. The majority of tumors expressing typical Basal gene patterns are TNBC by pathology [44], and it is known that there is a large overlap between BP Basal subtypes and TNBCs. Therefore, we applied the TNBC Burstein classifier on the Basal-single-type and Luminal-Basal-type. Basal-single-type tumors were mostly classified as BLIA and BLIS while Luminal-Basal-type tumors were more likely to be either LAR or MES. Genes described by Burstein et al. to be up-regulated in the LAR subtype, such as DHRS2, AGR2, FOXA1, AR, and MUCL1, were indeed higher expressed in Luminal-Basal-type compared with the Basal-single-type samples. Since the majority of Luminal-Basal-type tumors were classified as LAR, and according to Burstein et al., those patients derive benefit from traditional anti-estrogen or anti-androgen therapy, we could speculate that Luminal-Basal-type cancers would benefit from such treatment as well. Furthermore, ADH1B and FABP4 genes were up-regulated in Luminal-Basal-type samples compared with Basal-single-type samples. The upregulation of these genes is typical of the MES subtype, which is characterized by the dysregulation of cell cycle and DNA damage repair pathways. On the contrary, BLIS subtype-specific genes, HORMAD1, SOX10, SERPINB5, and FOXC1, were up-regulated in Basal-single-type samples compared with Luminal-Basal-type samples. Therefore, we could hypothesize that among the Basal-single-type samples, two subgroups are present which are indiscernible with the current dual subtype classification, but might have a different prognosis according to Burstein et al. and require additional analyses. Notably, majority of the Luminal-Basal-type showed a Ki67 positivity below 30% which might indicate that they share features with the TNLP tumors recently described by Bhargava and colleagues [36]. Additionally, no large agreement was found between any of the dual subtypes and the normal-like [3,4] (Table S3) or claudin-low classifications [9,27] (data not shown). Conversely, BluePrint Basal-, Luminal-, and HER2single type classifications were largely concordant with the intrinsic subtypes (> 90%) (see Table S3). Interestingly, and perhaps unexpectedly, the Luminal-Basal-type tumors were mostly classified as HER2-e intrinsic subtype, possibly due to the absence of Luminal-and Basal-type biology in the BP Luminal-Basal-type.
Luminal-HER2-type samples consistently showed patterns of both ER and HER2 activation (by expression and IHC/FISH), which may suggest similarities to the clinically triple-positive tumors [10]. Expression of both ER and HER2 may lead to receptor crosstalk which has often been associated with resistance to both endocrine and HER2targeted therapies [45]. However, down-regulation of the MAPK-related gene sets MEK and RAF may indicate no downstream activation of the HER2 pathway. Therefore, Luminal-HER2-type tumors are unlikely fueled through the HER2 pathway alone and HER2-targeted therapies might not be as effective as in the HER2-single-type tumors. This suggestion is strengthened by the observation in the NBRST data that Luminal-HER2-type tumors have a significantly lower pCR rate to neoadjuvant chemotherapy including HER2-targeted agents compared with HER2-single-type tumors (p-value < 0.032). This is supported by preliminary subanalysis of the TRAIN2 [42,46] and APHINITY [43,47,48] trial datasets, suggesting that BluePrint HER2-single-type tumors derive the most benefit from HER2 dualtargeted treatment [43].
It has been suggested that clinically triple-positive tumors develop endocrine resistance as downstream-activated MAPK inhibits ER transcription and phosphorylates ER [38]; however, in this study, Luminal-HER2-type tumors may be only driven by the ER pathway, as MAPK is downregulated compared with HER2-single-type tumors and not significantly different from that of Luminal-single-type tumors. Further analysis on Luminal-HER2-type samples treated with endocrine therapy is required to investigate and confirm this hypothesis.

Conclusion
Our study showed that by further dissecting the BP scores, it is possible to identify a small proportion of EBCs that have dual-activated BP pathways. These dual subtypes display specific transcriptional and clinicopathological features supporting the idea that they represent a different biological subgroup than their single counterparts. Most dual BP subtypes are either Luminal-Basal-type or Luminal-HER2-type.
The Luminal-Basal-type shows lower proliferation levels compared with the Basal-single-type and AR activation. Interestingly, using the Burstein classification, Luminal-Basal tumors are mostly classified as LAR and MES subtypes.
The Luminal-HER2-type resembles features of both the Luminal-single-type and HER2-single-type. However, patients with Luminal-HER2-type tumors have a lower pCR rate after receiving HER2-targeted therapies in addition to chemotherapy compared with patients with a HER2-single-type.
Taken together, BP dual classification shows potential clinical utility in helping treatment decision for a limited, but still relevant, fraction of EBC patients with dual subtypes that may benefit from additional or alternative targeted therapies. Even though molecular subtyping is not yet standardly used in routine clinical diagnostics, increasing number of evidences are emerging indicating that molecular subtypes should become part of breast cancer management [49]. In this light, results presented here further support the need toward such transition and implementation.
Future work will be focused on further confirming and prospectively validating the findings described here in additional independent datasets.