Classical pathology and mutational load of breast cancer – integration of two worlds

Abstract

Breast cancer is a complex molecular disease comprising several biological subtypes. However, daily routine diagnosis is still based on a small set of well‐characterized clinico‐pathological variables. Here, we try to link the two worlds of surgical pathology and multilayered molecular profiling by analyzing the relationships between clinico‐pathological phenotypes and mutational loads of breast cancer. We evaluated the number of mutated genes with somatic non‐silent mutations in different subgroups of breast cancer based on clinico‐pathological, including immunohistochemical and tumour characteristics. The analysis was performed for a cohort of 687 primary breast cancer patients with mutational profiling, gene expression and clinico‐pathological data available from The Cancer Genome Atlas (TCGA) project. The number of mutated genes was strongly positively associated with higher tumour grade (p = 1.4e−14) and with the different immunohistochemical and PAM50 molecular subtypes of breast cancer (p = 1.4e−10 and p = 4.3e−10, respectively). We observed significant associations (|R| > 0.4) between the abundance of mutated genes and expression levels of genes related to proliferation in the overall cohort and hormone receptor positive cohort, including the Recurrence Score gene signature (e.g., MYBL2 and BIRC5). Specific mutated genes (TP53, NCOR1, NF1, PTPRD and RB1) were highly significantly associated with high loads of mutated genes. Multivariate analysis for overall survival (OS) revealed a worse survival for patients with high numbers of mutated genes (hazard ratio = 4.6, 95% CI: 1.0 – 20.0, p = 0.044). Here, we report a strong association of the number of mutated genes with immunohistochemical and PAM50 subtypes and tumour grade in breast cancer. We provide evidence that specific levels of the mutational load underlie different morphological and biological phenotypes, which collectively constitute the current basis of pathological diagnosis. Our study is a step towards genomics‐informed breast pathology and will provide a basis for future studies in this field bridging the gap between morphology, tumour biology and medical oncology.

Introduction

Breast cancer is a complex molecular disease 1, 2 and the most common cancer in women worldwide with each 8th to 9th woman affected during lifetime in the United States and Western Europe. While mortality is decreasing due to advances in both diagnostic and therapeutic approaches, breast cancer is still the second most common cause of death from cancer in women 3.

As demonstrated by a multitude of high‐throughput molecular profiling studies 4, 5, 6, 7 over the last 15 years spearheaded by the pioneering work of Perou et al 4, breast cancer is a heterogeneous disease comprising several biologically defined subtypes that are associated with distinct clinical behaviour and each require specific therapeutic strategies 8. Currently, four subtypes (luminal A and luminal B, HER2‐enriched and basal like/triple negative) are firmly established but more recent data even suggest a finer molecular classification in ten subtypes 9 whose clinical implications, however, still need to be explored. Complementing these advances, several multigene predictors have been developed to aid prediction of response towards adjuvant therapy and risk stratification of recurrence 10, 11, 12.

The current histomorphological typing system 13 still plays a significant role in the detection of rare subtypes but yields only limited clinically exploitable information for the most prevalent histological breast cancer type, invasive carcinoma of no special type (NST). With a shift to smaller tumours due to screening programs, tumour staging per se has only limited prognostic power 13, 14 with recent studies suggesting that a refinement of the current staging system through inclusion of molecular profiles might prove beneficial 15, 16. The Elston–Ellis modification 17, 18 of the Scarff–Bloom–Richardson grading system 19 is widely used to estimate outcomes but the semi‐quantitative evaluation of morphological features rather than quantitative assessment of genetic parameters appears to introduce a considerable bias leaving grading results alone with a comparably limited impact on clinical decision making 13, 14, 20.

These developments and subsequent implementation of molecular classifications into routine diagnostics led to guidelines 21 which attribute a less important prognostic and particularly predictive power to classic morphology‐based parameters, that is, typing, grading and staging 13, 14, 20, 21, 22.

This being said, it should be noted that, in principle, conventional morphology‐based parameters, when applied in combination and evaluated by trained experts contribute to a reliable, valid and robust estimation of the course of disease in many cancers including breast cancer 20. Very surprisingly, however, studies investigating the actual genetic underpinnings of phenotype‐based prognosticators and comparing these results with established molecular subtypes are still rare. A landmark study by Sotiriou et al 23 demonstrated histological grading to be associated with different gene expression signatures which also allowed stratifying the large set of grade 2 tumours into subgroups with different clinical outcome.

Interestingly, recent genomic profiling studies 24, 25, 26, 27, 28, 29, 30, 31 suggested a substantial degree of intra‐ and inter‐tumour heterogeneity fuelling selection processes during evolution of breast cancer 32, 33, which may complicate prognostication as well as prediction and impede cancer precision medicine approaches 34. In this context, it is worth recalling that just as the molecular phenotype, the morphological phenotype of the tumour, including tumour grade and tumour size is essentially a result of accumulated genetic aberrations over time, thereby reflecting tumour evolution at the phenotypic level.

Availability of a large dataset from The Cancer Genome Atlas (TCGA) enabled us to narrow in on these associations. Specifically, here, we investigated relations between the abundance of mutated genes and the morphological as well as biological phenotype of breast cancers.

Material and methods

Dataset

The study cohort (for details, see Table 1) included 687 breast cancer patients with both tumour exome sequencing and tumour grade data available. The primary variable under investigation was the number of mutated genes defined as genes carrying at least one somatic non‐synonymous mutation in the coding region.

Table 1.

Clinico‐pathological characteristics of the study cohort including 687 primary breast cancer patients

Characteristics	Group	Number of patients
Total	All patients	687 (100%)
Age	Median (q1‐q3)	59 (49–68)
Size	Median (q1‐q3)	2.5 (1.9–3.8)
N	N0	324 (47.2%)
	N+	351 (51.1%)
	NA	12 (1.8%)
G	G1	74 (10.8%)
	G2	319 (46.4%)
	G3	294 (42.8%)
Histopathology	Lobular	107 (15.6%)
	NST/other	579 (84.3%)
	NA	1 (0.2%)
Hormone receptor	HR−	148 (21.5%)
	HR+	532 (77.4%)
	NA	7 (1.0%)
HER2	HER2−	517 (75.3%)
	HER2+	113 (16.5%)
	NA	57 (8.3%)
Subtype	HR+/HER2−	405 (59.0%)
	HR+/HER2+	90 (13.1%)
	HR−/HER2−	111 (16.2%)
	HR−/HER2+	23 (3.4%)
	NA	58 (8.4%)
TP53	Mut	207 (30.1%)
	wt	480 (69.9%)
PIK3CA	Mut	212 (30.9%)
	wt	475 (69.1%)

q1 = first quartile, q3 = third quartile, NA = missing data

To obtain this quantity, we counted the number of genes with simple somatic mutations (substitutions, small insertions or deletions) in the coding region that led to a change in the amino acid sequence. Mutation data, clinico‐pathological data and RNAseq data were downloaded from TCGA at https://tcga-data.nci.nih.gov/tcga/dataAccessMatrix.htm. Overall survival (OS) data with a limited mean follow‐up time of 3.0 years and a limited median follow‐up time of 2.0 years were avaiable. Seventy‐four (10.8%) out of the 687 patients were followed until an event. Additionally, data on tumour grading and on tumours size (in cm) that were not available from TCGA were manually extracted from the corresponding pathology reports.

In the current study, we only used data sets that were published previously and labelled ‘No restrictions; all data available without limitations’ according to the TCGA publication guidelines (see http://cancergenome.nih.gov/publications/publicationguidelines, for details). We also extracted the PAM50 data supplied as supplementary data in 2.

With respect to the somatic mutation data, non‐synonymous mutations comprising insertions, deletions and substitutions in the coding region of genes were included. Structural rearrangements, aneuploidy, translocations or duplications as well as silent mutations were excluded from the analysis. For details visit https://wiki.nci.nih.gov/display/TCGA/Mutation+Annotation+Format.

Statistics

All statistical analyses and visualizations were performed using the statistical environment R including the R‐packages clinfun, beeswarm, survival and binom 35, 36, 37, 38, 39.

The number of mutated genes was correlated with patient age and tumour size using Spearman correlations, with tumour grade using the Jonckheere–Terpstra test, with histopathological subtype, hormone receptor (HR) status, HER2 status and molecular subtype using the exact Fisher test. p‐values < 0.05 were considered statistically significant. Robust linear regression was executed using Huber's M‐estimator as it is implemented in the function rlm of the R package MASS 40. Correlations of the number of mutated genes with the mutation status of 23 recurrently mutated breast cancer genes 2 were investigated using Wilcoxon's rank test. p‐values < 0.05 after Bonferroni‐correction for the 23 genes were considered statistically significant.

Genome‐wide gene expression data (generated by RNAseq) of breast cancer, HR+ breast cancer and HR− breast cancer were correlated with the number of mutated genes using the Spearman correlation coefficient. p‐values < 0.05 after Bonferroni correlation for the total number of 20 531 genes were considered statistical significant. For visualization, fold changes (FC) between highly mutated tumours (22 or more mutations) and lowly mutated tumours (21 or less mutations) were calculated. Gene ontology (GO) categories were obtained from GO2MSIG 41.

Univariate and multivariate survival analysis was executed using the function coxph from the R package survival. p‐values < 0.05 were considered significant. Prognostic evaluation of cut‐off points for the number of mutated genes was executed as it is implemented in the Cutoff Finder software 42. The mutation status of each of the 23 recurrently mutated breast cancer genes was correlated with patient age and tumour size using the Wilcoxon test, with tumour grade using the chi‐squared test for trends, with histopathological subtype, HR status, HER2 status and molecular subtype using the exact Fisher test. p‐values <0.05 after Bonferroni‐correction for the 23 genes were considered statistically significant.

Results

We performed an integrated analysis of somatic mutations data, clinico‐pathological data and RNAseq data that were publicly available from the TCGA project. Characteristics of the cohort comprising 687 patients are provided in Table 1. The primary variable under investigation was the number of genes carrying at least one non‐silent somatic mutation (hereinafter simply termed ‘mutated genes’).

Associations of clinico‐pathological parameters with mutational load

First, we correlated the number of mutated genes with the clinico‐pathological characteristics of the tumour (Table 2, Figure 1). The number of mutated genes was strongly associated with tumour grade increasing from a median of 23 mutated genes in G1 tumors via 27 in G2 tumors to 43 in G3 tumours (p = 1.4e−14). When analysing the three components of tumour grade, nuclear pleomorphism (atypia) showed the strongest association with the number of mutated genes (p = 5.5e−11), but mitotic count and the degree of tubule formation also contributed significantly (p = 4.8e−08 and p = 1.4e−03, respectively). In addition, the number of mutated genes was strongly associated with the molecular subtype of breast cancer as determined by routine immunohistochemistry and the PAM50 classifier (p = 1.4e−10 and p = 4.3e−10, respectively). As shown in Table 2 and Figure 1, HR−/HER− breast cancer and HR−/HER2+ breast cancer harboured much more mutated genes compared with HR+/HER2− breast cancer. In contrast, we observed no statistically significant differences between the mutational loads of the two main histological subtypes, invasive lobular carcinoma (ILC) and NST (27 versus 23 mutated genes (median), p = 0.23). To narrow in on relations between histological tumour types and mutational load, we analyzed the abundance of mutated genes in all ILC (of any grade) and compared the results with the mutational load of NST carcinomas stratified according to grade (supplementary material Figure S1). While ILC harboured a mean of 27 mutated genes, G1–G3 NST displayed 23, 27 and 43 mutated genes, respectively (p = 0.027, p = 0.66, p = 0.0011). These data showed that ILC of any grade are fairly comparable with G2 NST with respect to their mutational load and ILC, G1 and G2 NST are in the lower range of mutational loads compared with G3 NST (p = 9.12e−12). Interestingly, ILC of any grade shared a high burden of mutated PIK3CA with G1–G3 NST. Highly mutated genes in ILC were PIK3CA (43.9%), CDH1 (56.1%), MLL3 (7.5%), FOXA1 and TP53 (6.5% each), as well as MAP3K1 and GATA3 (4.7% each). 54.6% of NST G1 showed mutated PIK3CA, followed by MAP3K1 (16.4%), GATA3 (14.6%), MAP2K4 (10.4%), and FOXP1 and TP53 (3.6% each). With higher tumour grades, NST tumours with mutated MAP2K4 (G2: 5.0%, G3:1.8%), MAP3K1 (G2: 10.5%, G3: 3.2%) and GATA3 (G2: 14.2%, G3: 5.6%) decreased while cases with TP53 mutations strongly increased (G2: 19.7%, G3: 52.6%).

Table 2.

Association of the number of mutated genes (non‐silent somatic mutation) with clinico‐pathological characteristics of breast cancer

Variable	Number of mutated genes (median)	p‐value
Age	R = 0.12, slope = 2.1, intercept = 26.0	0.0016
1. Tumour size	R = 0.17, slope = 1.6 intercept = 33.1	0.00018
2. Nodal status	N0 = 34, N+ = 30	0.0048
3. Histopathology	ILC = 27, NST = 33	0.23
4. Tumour grade	G1 = 23, G2 = 27, G3 = 43	1.0e−14
5. Hormone receptor (HR)	HR+ = 28, HR− = 46	1.0e−10
6. HER2	HER2− = 31, HER2+ = 40	0.0066
7. Subtype (based on IHC)	HR+/HER2− = 27, HR+/Her2+ = 40, HR−/HER2− = 49, HR−/HER2+ = 41	1.0e−10
8. Subtype (based on PAM50)	Luminal A = 27, Luminal B = 39, Basal‐like = 46, HER2‐enriched = 50, Normal‐like = 70	4.3e−10

The number of mutated genes correlated significantly with patient age, tumour size, nodal status, tumour grade, as well as immunohistochemical and PAM50 molecular subtypes. Associations with tumour grade, hormone receptor (HR) status and molecular subtype were much stronger compared with the associations with the other clinico‐pathological parameters. Associations with patient age and tumour size were quantified using Spearman correlations and robust linear modelling. For patient age, the slope describes the increase in mutated genes per 10 years. For the tumour size, the slope describes the increase in mutated genes per 1cm increase in the maximum tumour dimension

Figure 1.

Association of the number of mutated genes (non‐silent somatic mutation) with clinico‐pathological characteristics of breast cancer. (A) In the beeswarm plot, each coloured dot represents a tumour. The bands indicate the first quartile, the median and the third quartile (n = median of mutated genes). The number of mutated genes was strongly associated with tumour grade, molecular subtyping by immunohistochemistry and SISH for ER, PR and HER2 as well as molecular subtyping by PAM50. (B) Scatterplots showing weak, but significant association of the number of mutations with patient age and tumour size in cm. The red curves represent the results of robust linear fitting. Not that the results of linear modeling appear as curves (not straight lines), because of the logarithmic scale of the y‐axis.

Further, we detected significant associations of the number of mutated genes with patient age (p = 0.0016), tumour size (p = 0.00018), and nodal stage (p = 0.0048). The median number of mutated genes in node positive (N+) tumours was 30 compared with 34 in N0 tumours. To quantify the strength of association with the continuous parameters, we executed robust linear regression of the number of mutated genes against patient age and against tumour size. Only 2.1 mutated genes were added per 10 years of the patient's life (slope, p = 0.0012), while 26.0 mutated genes (intercept, p = 4.2e−09) occurred independent of the age at diagnosis. Further, only 1.6 mutated genes were added when the tumour's maximum dimension extended by 1 cm (slope, p = 0.0015), while 33.1 mutated genes (intercept, p = 5.8e−44) occurred independent of the tumour size.

Finally, we executed the aforementioned analyses separately in HR+ breast cancer (n = 532) and in HR− breast cancer (n = 148). The number of mutated genes significantly correlated with tumour grade and HER2 status in HR+ breast cancer (p = 4.8e−09 and p = 0.00039), but not in HR− breast cancer. Interestingly, the correlation with patient age was stronger in HR+ breast cancer (R = 0.2) compared with HR− breast cancer (R = 0.14). Using estimates from robust linear regression, this corresponded to 3.4 mutated genes per 10 years life time occurring in HR+ tumours compared with 2.5 mutated genes per 10 years life time occurring in HR− tumours.

Associations of recurrently mutated genes with mutational load

Second, we correlated the number of mutated genes with the mutation status of each of the 23 genes that were reported by TCGA 2 as recurrently mutated in breast cancer (Figure 2, supplementary material Table S1). It turned out that the number of mutated genes was strongly associated with the mutation status of the tumour suppressor gene TP53 (median of the number of mutated genes 47 in TP53‐mutated tumours versus 27 in TP53 wild‐type tumours; p = 3.8e−21), while it was not associated with the mutation status of the oncogene PIK3CA (p = 0.36). Additionally, a significantly enhanced number of mutated genes was observed in, NCOR1‐mutated, NF1‐mutated, PTPRD‐mutated and RB1‐mutated tumours (Figure 2). In addition, we analyzed the correlation of the total mutational load with the mutation status of BRCA1 and BRCA2. Note that, in this analysis, only somatic mutations were taken into account, as germline variant calling is currently not available for the TCGA BRCA data set. BRCA1 somatic mutated tumours tended (p = 0.058) to harbour an enhanced total number of mutated genes (n = 67) compared with tumours without somatic BRCA1 mutations (n = 32). BRCA2 somatic mutated tumours had a significantly (p = 0.0037) enhanced total number of mutated genes (n = 112) compared with tumours without somatic BRCA2 mutations (n = 32).

Figure 2.

Association of the number of mutated genes (non‐silent somatic mutation) with mutation status of genes frequently mutated in breast cancer. In the beeswarm plot, each coloured dot represents a tumour. The bands indicate the median including the first and third quartile (n = median of mutated genes). The number of mutated genes was not associated with PIK3CA mutation status, but strongly and highly significantly associated with mutated TP53. Moreover, the number of mutated genes was strongly and significantly associated with mutated NCOR1, NF1, PTPRD and RB1.

Associations of gene expression profiles with mutational load

Third, we analyzed the correlation of the gene expression landscape with the number of mutated genes (Figure 3, supplementary material Table S3). To this end, Spearman correlations were calculated between RNAseq data and the number of mutated genes. The number of genes whose expression showed a strong positive or negative correlation with the number of mutated genes (|R| > 0.4) was 44 in breast cancer (all p < 8.8e−28; Figure 4A), 11 in HR+ breast cancer (all p < 4.7e−22; Figure 4B) and 23 in HR− breast cancer (all; p < 5.1e−07; Figure 4D). In a functional analysis based on GO categories, we found the cellular process ‘mitotic cell cycle (GO:0000278)’ to be strongly and significantly enriched in the gene lists for breast cancer (20 genes, p = 3.3e−19) and for HR+ breast cancer (5 genes, p = 1.0e−05). When relaxing the correlation threshold to |R| > 0.3, larger numbers of 120 and 84 cell cycle genes belonged to the gene lists (significance of enrichment: p = 2.6e−56 and p = 2.8e−52, respectively). Moreover, we observed similar results for associations between the abundance of mutated genes and the five proliferation‐related genes from the Recurrence Score gene signature 11 in HR+ breast cancers (all correlation p < 2.4e−22; Figure 4C, supplementary material Table S2). Taken together, in the overall cohort as well as in HR+ ‐driven tumours, we found especially upregulated cell cycle and proliferation‐related genes to be associated with higher numbers of mutated genes. Specifically, CDC20, MYBL2 and BIRC5 (Survivin) as well as MKI67, which is routinely used for diagnostic work‐up of breast cancers ranged among the top genes. In the gene expression data of HR− breast cancer, functional analysis did not detect any enriched GO category and all correlating genes were downregulated.

Figure 3.

Gene expression FCs with 95% CIs between tumours with a high number of mutated genes (22 or more) and a low number of mutated genes (21 or less). Cell cycle genes [GO category 0000278: ‘mitotic cell cycle’; (40)] are colored in yellow. (A) Analysis of breast cancer: FCs of 44 genes whose expression strongly correlated with the number of mutated genes (|Spearman‐R|>0.4). (B) Analysis of HR+breast cancer: FCs of 11 genes whose expression strongly correlated with the number of mutated genes (|Spearman‐R|>0.4). (C) Analysis of HR−breast cancer: FCs of 23 genes whose expression strongly correlated with the number of mutated genes (|Spearman‐R|>0.4). (D) Analysis of HR+breast cancer: FCs of the 16 cancer genes included in the Recurrence Score assay.

Figure 4.

Association of the mutation status of specific genes with tumour grade and immunohistochemical as well as PAM50 subtypes (mutation rates including 95% confidence intervals). Genes with significant association after Bonferroni correction are shown (TP53, GATA3, CDKN1B, PIK3CA, CDH1, and MAP3K1). For example, the number of PIK3CA mutated tumours decreased from 54.3 via 36.7% to 18.7% with increasing tumour grade (G1, G2 and G3), while the number of TP53 mutated tumours increased from 5.4 via 15.4% to 52.4%. Mutation status of PIK3CA, TP53, GATA3, CDH1 and MAP3K1 correlated significantly with the molecular subtype. PIK3CA mutations were much less frequent in triple‐negative breast cancer (7.2%) compared with the other subtypes (all mutation rates >17.4%). TP53 mutation rate was much higher in triple‐negative (71.2%) and HR−/HER2+(69.6%) breast cancer compared with HR+/HER2−(16.3%) and HR+/HER2+(32.2%) breast cancer.

Associations of recurrently mutated genes with clinico‐pathological parameters

Next, we correlated the mutation status of the 23 genes that were described as recurrently mutated in breast cancer 2 with tumour grade, molecular subtype and other clinico‐pathological characteristics of breast cancer (Figure 4). Mutation status of PIK3CA, TP53, CDH1 and CDKN1B correlated significantly with tumour grade. Mutation status of PIK3CA, TP53, GATA3 and CDH1 correlated significantly with the molecular subtype. Further, we found a strong association between CDH1 mutation status and the histopathological subtype, where 56.1% of the lobular cancers had CHD1 mutations (G1: 57.9%, G2: 56.3%, G3: 50.0%) compared with only 2.1% of NST/other cancers (G1: 1.8%, G2: 2.1%, G3: 2.1%). Finally, we detected a moderate, but significant association of mutated MLL3 (p = 0.00046) and FOXA1 (p = 0.00097) with patient age. Median age of patients harbouring these mutations was 63 and 67 compared with a median age of 59 in patients without these mutations.

Further, we investigated the correlation of somatic BRCA1 and BRCA2 mutation status with the clinico‐pathological data. Mutations in BRCA1 were detected only in the basal‐like molecular subtype (in 3.9% of these tumours), but absent in all other PAM50 subtypes (p = 0.053). BRCA2 mutations were found in the HER2‐enriched subtype (8.3%) in the basal subtype (3.9%) and in the luminal A subtype (1.2%), but absent in the other PAM50 subtypes (p = 0.032).

Associations of overall survival with mutational load

Finally, we investigated the correlation of outcome with the number of mutated genes (Figure 5). In univariate analysis of overall survival, the number of mutated genes was a negative prognostic marker with a hazard ratio of 1.18 per 100 mutated genes (p = 0.00029). In a multivariate analysis including correction for patient age, tumour size, nodal status, histopathological type, tumour grade, HR status and HER2 status, the number of mutated genes was no longer significant. Next, we analyzed the prognostic power of a dichotomized version of the number of mutated genes. To this end, we made use of the Cutoff Finder 40 to study the impact of the choice of the cut‐off point on differences in survival. In multivariate analysis, we detected a borderline‐significant impact of the number of mutated genes on survival for cut‐off points between 18.5 and 27.5 (Figure 4). An optimal cut‐off point occurred when classifying tumours in those with 21 or less mutated genes and those with 22 or more mutated genes. This classification yielded a hazard ratio of 4.6 (95%CI: 1.0–20.0) and significance of p = 0.044 in multivariate analysis of overall survival.

Figure 5.

Association of overall survival with the total number of mutated genes (non‐silent somatic mutation) in breast cancer. Hazard ratios (haz. rat.) are shown for a dichotomized version of the number of mutated genes that was varied along the x‐axis. (A) In univariate analysis, no cut‐off point between 10 and 180 yielded a significant association with survival. (B) In multivariate analysis including correction for patient age, tumour size, nodal status, histopathological type, tumour grade, hormone receptor status and HER2 status, a high number of mutations were borderline‐significant associated with shorter survival for cut‐off points between 18.5 and 27.5. The most significant association was obtained for a cut‐off point of 21.5 with hazard ratio = 4.6 (1.0–10.0) and p = 0.044.

Discussion

Here, we show that the number of mutated genes, that is, the diversity of the gene pool putatively contributing to tumour evolution 34 is strongly associated with both morphological and biological features of breast cancer, which are commonly being used in routine diagnostics to guide clinical management and therapy worldwide. We also provide evidence for associations between the number of mutated genes and mutation status of specific genes as well as with expression levels of genes routinely used to assess the proliferative capacity of breast tumours (eg, Ki67, [43]) and to predict therapy response (eg, cancer gene signature of the Recurrence Score assay 11).

Gene expression profiling studies combined with survival analysis 4, 5, 6, 7, 8 revealed that hormone receptor (HR)‐positive (ie, luminal A and luminal B) and HR‐negative tumours (ie, HER2‐enriched and basal‐like) are distinct diseases in both molecular and clinical terms. The intrinsic molecular subtypes are largely distinguished by gene sets measuring epithelial and basal differentiation as well as proliferation and the HER2 pathway. The robustness of the stratifying signatures has been demonstrated across several platforms 2, 6, 7. In daily pathological diagnostics, however, immunohistochemistry is applied to facilitate approximation of these subtypes 43 and to stratify breast cancer patients into different risk groups and oncological therapies. A more recent combined analysis of genetic and RNA expression profiles of the largest breast cancer series to date comprising 2000 patients profiles by the METABRIC study group 9, even revealed 10 different subgroups associated with different clinical outcomes; however, this has not entered clinical decision making yet.

In line with these findings, we observed significant differences in the relative abundance of mutated genes across molecular subtypes determined by immunohistochemistry and the PAM50 classifier: the median of mutated genes increased from 27 in HR+/HER2− tumours to 40 in HR+/HER2+ tumours, 41 in HR−/HER2+ and 49 in HR−/HER2− tumours and from 27 in Luminal A to 39 in Luminal B, 46.5 in Basal‐like, 50.5 in HER2‐enriched and 70 in normal‐like tumours.

Recent large‐scale molecular profiling studies 2, 9 revealed considerable divergence of the main breast cancer subtypes on the genetic level. Interestingly, as shown by TCGA 2, luminal breast cancers are the most heterogeneous subtypes. Mutations frequently affected transcription factors of the ER pathway (GATA3, FOXA1 and RUNX1) and the PI3K cascade with mutations in PIK3CA as the most frequent genetic aberration followed by mutually exclusive mutations in PIK3R1, AKT and PTEN. Other frequent targets are the MAP‐kinases MAP3K1 and MAP2K4 and epigenetic modulators, for example, genes encoding methyltransferases (eg, MLL3), demethyltransferases (eg, KDM6A) as well as several adenine‐thymine‐rich interactive domain‐containing proteins (ARID1A and others). TP53 mutations occurred in approx. 25.8% of luminal B and 8.8% of luminal A tumours. In contrast, basal‐like breast cancer appears to harbour the most unique and distinctive genomic profile. They are primarily characterized by genomic instability that is putatively caused by the loss of function of three tumour suppressor genes and cell‐cycle regulators by mutations or deletions: TP53 (75%), RB1 (4%), and BRCA1/2 (7.9%). Similar to basal‐like breast cancers, HER2‐enriched subtypes showing a certain relation to luminal subtypes also harbour a high frequency of TP53 (63.9%) that is accompanied by PIK3CA mutations (38.9%).

In accord with these observations, extended analysis of the whole cohort showed a significantly high prevalence of mutated TP53 in TNBC and HER2‐enriched breast cancer, while mutated CDH1 and GATA3 occur significantly more frequently in the HR+ subtype and the HR+/HER2− subtype, respectively. We also noted that while PIK3CA and CDH1 mutations were significantly enriched in tumours with lower tumour grades, TP53 and CDKN1B were significantly associated with higher tumour grades. However, when analyzing CDH1 status within histological subtypes, we did not observe any significant correlations with tumour grade. As already reported by other investigators 2, 44, 45, mutated CDH1 occured almost exclusively in lobular breast cancer.

Interestingly, we did not observed significant associations between the abundance of mutated genes and several of the frequently mutated breast‐cancer related genes briefly summarized above. However, we identified five mutated genes (TP53, NCOR1, PTRD, NF1 and RB1), which showed strong associations to high numbers of mutated genes independent of the tumour subtype. As expected, non‐silent mutations in TP53 and RB1, which have been causally linked to genomic instability of tumours 46, 47, are strongly correlated with the number of mutated genes in breast cancer. Moreover, our data show that high numbers of mutated genes are strongly associated with mutant NCOR1, which is mutated in ∼4% of breast cancer cases 2. Although we cannot provide evidence for a causal relationship, it is of note that NCOR1 is a gene encoding a transcriptional co‐regulator serving as a linker for transcription factors and chromatin‐modifying enzymes and has been implicated in maintaining genomic stability via HDAC3 48. Pointing towards a similar direction, the mutant variant of PTPRD (genetic aberrations in ∼7% of breast cancers 2) encoding the protein tyrosine phosphatase receptor type delta was also strongly associated with high numbers of mutated genes in breast cancer. PTPRD has been shown to act as tumour suppressor 49 and impairs function of Aurora A 50, a protein that has been shown to modulate genomic stability 51. Last but not least, the mutated variant of the putative breast cancer oncogene NF1 (genetic alterations in ∼5% of breast cancers 2) correlated with a high abundance of mutated genes in our study. Neurofibromin1 acts as a negative regulator of RAS by stimulating GTPase activity of RAS 52 and impaired NF1 function has been linked to increased mutations frequencies in mice 53. Taken together, our observational data indicate that it may be worthwhile to investigate the specific functional relevance of these five genes in breast cancer biology and maintenance of genome integrity in future studies.

In addition, we sought to identify differentially expressed genes that are associated with the number of mutated genes in breast cancer. Interestingly, we found many upregulated cell cycle‐related proteins to be related to high numbers of mutated genes in the overall cohort. Among the top upregulated genes were MYBL2, BIRC5 and AURKA, which are also part of the cancer genes interrogated by the Recurrence Score assay 11. Narrowing in on this, we observed that particularly the proliferation gene signature of the Recurrence score shows a strong correlation with the mutation rate in HR+ breast cancer subtypes. A broader analysis of the HR+ cancers corroborated MYBL2 and identified other upregulated cell‐cycle associated genes, such as PLK1 and CDC20. These data suggest that the link between proliferation and abundant mutations is largely restricted to HR+ tumours and does not pertain to HR− tumours. It is tempting to speculate that this proliferation signature indicates that increased tumour cell replication and associated stochastically higher mutation rate in luminal tumours fuels the pool of mutated genes thereby advancing tumour heterogeneity and evolutionary selection processes 34.

Remarkably, we found that tumour grade, that is, phenotypes indicating the degree of cellular differentiation, but not the two main breast cancer types, that is, NST and ILC, are strongly associated with different numbers of mutated genes. Additionally, we identified a significant but comparably weaker correlation of the mutational load with two other measures reflecting the temporal development of the tumour, that is, age at diagnosis and tumour size. Specifically, the expected number of acquired mutated genes was estimated as 2.1 per 10 years and 1.6 per 1 cm, respectively. Hypothesizing that the abundance of mutated genes is a measure of genetic heterogeneity of the tumour 34, we here provide evidence that particularly pathological grading of breast cancer is in fact a microscopic read‐out of tumour heterogeneity, which indicates different biological and clinical behaviour of the tumour. Several studies either sequencing tumour bulks or single tumour cells have demonstrated that breast cancer subtypes exhibit considerable spatial and temporal heterogeneity on the genetic level within the primary, during metastatic progression and in patient‐derived xenografts 24, 25, 26, 27, 28, 29, 30, 31. This genomic diversity within breast cancers is a result of but also facilitates cellular evolution 34 allowing the tumour to dynamically adapt to external or internal stimuli as originally conceptualized by Nowell in 1976 54.

While we and others 31, 33, 34, 54 hypothesize that this diverse gene pool may confer survival advantage to the tumour, to our surprise and counterintuitive, local spread to lymph nodes was associated with slightly but significantly lower numbers of mutated genes compared with node negative tumours. Currently, we are unable explain these results satisfactorily and future studies are needed to investigate these relations further. Although data on follow‐up times of the cohort are rather short and limited, and hence, limit analysis as well as data interpretation, our survival data analysis indicates that increasing numbers of mutated genes per se are significantly associated with increased hazard ratios independently of other clinico‐pathological parameters. Hence, our data may suggest that the abundance of mutated genes has a certain relation with metastatic progression ultimately leading to tumour‐related death 55, 56, 57. Further studies are warranted to substantiate our findings and conclusions.

To sum up, we here provide evidence for significant associations between the genotype and morphological as well biological phenotypes of breast cancer. Specifically, while histological subtypes do not display specific numbers of non‐synonymous somatic mutations, immunohistochemically defined subtypes of breast cancer are associated with different amounts of mutated genes increasing from HR+/HER2− to HR+/HER2+, HR−/HER2+ and HR−/HER2− breast cancers. Moreover, we observed significant correlations with the expression level of genes involved in cellular proliferation as well as with specific breast‐cancer related genes. Interestingly, the increasing abundance of mutated genes displayed strong correlations with increasing tumour grades and also with tumour size and age at diagnosis albeit to a lower degree. The three main implications of these observations are: 1) a critical reappraisal of tumour grading that warrants further investigation as the tumour grade assessed by conventional light‐microscopy for prognostication of the course of disease may reflect genetic tumour heterogeneity and in turn possibly the evolutionary potential of tumour cells, 2) different molecular subtypes of breast cancer may harbour a different capacity for tumour evolution due to different abundances of mutated genes and 3) we provide evidence for a strong relation between the proliferative capacity as well as specific mutated genes and the number of genes harbouring somatic non‐silent mutations in breast cancer.

In conclusion, we believe that our study is a step further towards genomics‐informed breast pathology and will provide a basis for future studies in this field bridging the gap between morphology, tumour biology and medical oncology.

List of abbreviations

AFF2

AF4/FMR2 family member 2

ARID1A

AT‐rich interactive domain‐containing protein 1A

AURKA/STK15

Aurora A kinase

BIRC5

Survivin

CBFB

Core‐binding factor subunit beta

CDH1

Cadherin‐1

CDC20

cell‐division cycle protein 20

CTCF

CCCTC‐binding factor

ER

Estrogen receptor

FC

Fold change

GATA3

Transacting T‐cell‐specific transcription factor 3

HER2

Human epidermal growth factor receptor 2

HR

Hormone receptor

ILC

Invasive lobular carcinoma

MAP3K1

Mitogen‐activated protein kinase kinase kinase 1

MAP2K4

mitogen‐activated protein kinase kinase 4

MLL3

Histone‐lysine N‐methyltransferase 3

MYBL2

Myb‐related protein B

NCOR1

Nuclear receptor corepressor 1

NF1

Neurofibromin 1

NST

Invasive carcinoma of no special type

TNBC

Triple negative breast cancer

PIK3CA

phosphatidylinositol‐4,5‐bisphosphate 3‐kinase, catalytic subunit alpha

PLK1

Polo‐like kinase

PTEN

Phosphatase and tensin homolog

PTPRD

Receptor‐type tyrosine‐protein phosphatase delta

PR

Progesteron receptor

RB1

retinoblastoma 1

RUNX1

Runt‐related transcription factor 1

SF3B1

Splicing factor 3B subunit 1

SPC24

kinetochore protein SPc 24

TBX3

T‐box transcription factor

TP53

tumour suppressor 53

Author contributions

AS and JB conceived and designed the study and supervised experiments. JB, MB, FK, JKL, SB, GB, MD, CD, WW and AS analyzed sequencing data and clinico‐pathological parameters. AS and JB wrote the manuscript with contributions from MB, FK, SB, CD, JKL, WW, and SL. All authors read and approved the final manuscript.

Supporting information

Figure S1: Association of the number of mutated genes (non‐silent somatic mutation) with histopathological subtypes of breast cancer (ILC all grade, NST grade 1, NST grade 2, NST grade 3). In the beeswarm plot, each coloured dot represents a tumour. The bands indicate the first quartile, the median and the third quartile (n = median of mutated genes). Mutational load of ILC was slightly, but significantly, higher than in NST G1 tumours (p = 0.027). Mutational load of ILC did not show a significant difference compared with NST G2 tumours (p = 0.66). Mutational load was considerably lower in ILC compared with G3 NST tumours (p = 0.0011).

Table S1: Association of the number of somatically mutated genes (non‐silent mutation) with mutation status of genes frequently mutated in breast cancer. For comparison, the Wilcoxon test was applied and bold p‐values indicate statistical significance after Bonferroni adjustment. TSG: Tumour suppressor gene.

Table S2: Association of the number of somatically mutated genes (non‐silent mutation, median) with the expression levels of the 16 cancer genes from the OncotypeDX assay. * = p < 0.05, ** = p < 0.05 after Bonferroni‐correction.

Table S3 (separate Excel file): Detailed overview of the statistical results of the genes whose expression levels were found to be significantly associated with the abundance of mutated genes in all breast cancers, as well as in ER− and ER+ subtypes.