Classification of ductal carcinoma in situ by gene expression profiling

Introduction Ductal carcinoma in situ (DCIS) is characterised by the intraductal proliferation of malignant epithelial cells. Several histological classification systems have been developed, but assessing the histological type/grade of DCIS lesions is still challenging, making treatment decisions based on these features difficult. To obtain insight in the molecular basis of the development of different types of DCIS and its progression to invasive breast cancer, we have studied differences in gene expression between different types of DCIS and between DCIS and invasive breast carcinomas. Methods Gene expression profiling using microarray analysis has been performed on 40 in situ and 40 invasive breast cancer cases. Results DCIS cases were classified as well- (n = 6), intermediately (n = 18), and poorly (n = 14) differentiated type. Of the 40 invasive breast cancer samples, five samples were grade I, 11 samples were grade II, and 24 samples were grade III. Using two-dimensional hierarchical clustering, the basal-like type, ERB-B2 type, and the luminal-type tumours originally described for invasive breast cancer could also be identified in DCIS. Conclusion Using supervised classification, we identified a gene expression classifier of 35 genes, which differed between DCIS and invasive breast cancer; a classifier of 43 genes could be identified separating between well- and poorly differentiated DCIS samples.


Introduction
Ductal carcinoma in situ (DCIS) of the breast represents a heterogeneous group of non-invasive breast tumours commonly detected in women undergoing screening mammography. DCIS is characterised by malignant epithelial cells accumulating in the ducts of the breast without invading through the basement membrane into the surrounding tissue. DCIS accounts for approximately 3% of symptomatic breast malignancies and for approximately 20% of breast malignancies in patients from population-based screening programs [1].
Different histological types of DCIS can be recognised, and a variety of classification systems have been developed [2]. Due to subjective interpretation of the morphology of the lesions, even experienced pathologists differ in their classification of DCIS [3]. Therefore, histological classification of DCIS may not be sufficient, and additional classification approaches could assist pathological classification.
It is assumed that most cases of DCIS will progress to invasive breast cancer. Because this progression may take many years and may not occur within the lifetime of a patient, elucidating the mechanisms of progression from in situ lesions to invasive disease and developing diagnostic tests would be of great clinical benefit.
Several models of the evolution of DCIS to invasive cancer have been suggested. One model suggests the linear progression from low-nuclear-grade DCIS to high-nuclear DCIS and DCIS = ductal carcinoma in situ; ER = oestrogen receptor; IDC = invasive ductal carcinoma; IHC = immunohistochemistry; LCIS = lobular carcinoma in situ; SNR = signal-to-noise ratio. the subsequent development of invasive cancer [4]. Based on specific genetic alterations found in the different types of DCIS, a more likely scenario is the evolution of well-, moderately, and poorly differentiated DCIS via distinct pathways. Following this idea, well-differentiated DCIS can give rise to lowgrade invasive carcinoma, whereas poorly differentiated DCIS can give rise to high-grade invasive breast cancer [5,6].
Several specific genetic alterations have been found in DCIS. HER2 gene amplification and protein overexpression are detected in up to 70% of poorly differentiated DCIS cases [7], and cyclin D1 is amplified and overexpressed in DCIS [8] in approximately 20% of the cases. Inactivating mutations of the E-cadherin gene are detected in almost all cases of lobular carcinoma in situ (LCIS) [9]. Several other genetic alterations in oncogenes (for example, C-MYC) and tumour suppressor genes (for example, p53) have been found in DCIS and are reviewed in Reis-Filho and colleagues [10] and Allred and colleagues [11].
Gene expression profiling has been shown to be a powerful tool for identifying profiles of tumour subtypes [12][13][14][15] and for correlating gene expression profiles with outcome in breast cancer [16][17][18]. The identification of specific gene expression patterns correlated with the different types of DCIS may help to elucidate the processes underlying the evolution of in situ carcinomas of the breast and also lead to a more reproducible classification of DCIS lesions.
To date, only a few studies of gene expression profiling of DCIS and a comparison with the gene expression pattern of invasive samples have been published and these are based on a small number of samples [19,20].
In the study presented here, gene expression profiling was performed on one LCIS and 39 DCIS samples to identify differentially expressed genes between well-, intermediately, and poorly differentiated DCIS. In addition, differences in gene expression between these cases of carcinoma in situ and 40 invasive breast carcinomas were studied.

Selection of samples
Cases of DCIS were selected from the tissue bank of the Netherlands Cancer Institute (Amsterdam, The Netherlands). These samples were obtained within 1 hour after surgery from patients who underwent wide local excision (n = 16) or mastectomy (n = 24). All samples were reviewed by two pathologists independently to determine the histological classification of the samples according to Holland and colleagues [21]; samples were classified as well, intermediately, or poorly differentiated. For analysis purposes, the intermediately differentiated DCIS cases were subclassified as those cases that were in part well differentiated (well to intermediately differentiated) and those that were in part poorly differentiated (mod-erately to poorly differentiated) in some areas. In cases in which there was a discrepancy in classification between the two pathologists, the histological slides were reviewed together to reach an agreement.
In addition, 40 cases of primary invasive breast cancer were selected; these were all cases of invasive ductal carcinoma (IDC) measuring between 1 and 5 cm and were graded as grade 1, 2, or 3 according to the method described by Elston and Ellis [22]. The study was approved by the medical ethical committee of the Netherlands Cancer Institute.

RNA isolation and amplification
RNA isolation and amplification were performed essentially as described by Weigelt and colleagues [23]. Thirty tissue sections of 30 μm of frozen material were cut. The first and the last tissue sections were 6 μm in thickness and were stained with haematoxylin and eosin to determine the percentage of tumour cells and to exclude invasive growth. Only samples with greater than or equal to50% of tumour cells were used for gene expression profiling.

Immunohistochemistry
The procedures applied are described in the supplementary information provided online [24].

Microarray hybridisation
Labeling of the amplified cRNA and microarray hybridisations were performed as previously described [25]. Equal amounts of amplified cRNAs of 100 invasive breast carcinomas were pooled and used as a reference. All hybridisations were performed on 18K human cDNA arrays (Central Microarray Facility, Netherlands Cancer Institute) [26].
Microarrays were scanned with the DNA Microarray Scanner G2565B (Agilent Technologies, Santa Clara, CA, USA). Selfself hybridisations were used to validate the quality of the hybridisations and as a negative control in the error model.

Processing of microarray data
Information on data processing is provided in the supplementary information [24].

Unsupervised hierarchical clustering
Two-dimensional unsupervised hierarchical clustering using Pearson correlation as distance function and complete linkage was performed using Genesis software (Technical University, Graz, Austria) [27,28].

Supervised classification
We performed supervised classification applying methods described previously [16,29,30]. Pathological features (histological type of the DCIS samples, histological grade of the invasive samples) were used to define groups for supervised classification. Genes were rank-ordered based on their signal-(page number not for citation purposes) to-noise statistic. Safe cutoffs were determined by comparing the signal-to-noise ratio (SNR) values with the results from 2,000 sample label permutations (Monte Carlo randomisation). For each group and a number of genes, a centroid is defined as the mean ratio per gene over all samples in that group. Correlation or Euclidean distance of each sample to those centroids determines their predicted group. Leave-out cross-validation was used to determine the optimal number of genes separating the groups. The number of left-out samples in this cross-validation procedure was dependent on the number of samples within the analysis set. SNR calculation, Monte Carlo randomisation, and cross-validation have been described previously [25].

Supplementary information
The microarray data, additional information on the methods, and the filtering results are provided as supplementary information [24].

Results
This study was performed to identify differences in gene expression (a) between DCIS and invasive breast cancer and (b) between different histological types of DCIS.

Tumour characteristics
Thirty-nine cases of DCIS of the breast were included in the analyses. By histological examination, they were assigned to the following groups: well differentiated (n = 6), intermediately differentiated (n = 18), and poorly differentiated (n = 14). For analysis purposes, the group of intermediately differentiated cases was further subdivided in well-intermediately (n = 10), true intermediately (n = 2), and intermediately-poorly (n = 6) differentiated type. One sample contains a mixture of well-and poorly differentiated DCIS components in the same tissue specimen. In addition, one case of LCIS was included.
To be able to compare DCIS with invasive breast cancer, 40 cases of invasive breast cancer were studied. Five tumours were histological grade 1, 11 samples were grade 2, and 24 samples were grade 3. Patient and tumour characteristics are summarised in Table 1.

Molecular subtypes of breast cancer
Several subtypes of breast cancer have been identified by gene expression profiling and have been correlated with clinical outcome [13,14]. This classification has been translated to classical immunohistochemistry (IHC): basal-type tumours are characterised by negative staining for oestrogen receptor (ER), progesterone receptor, and HER2 and are often positive for keratin 5/6; ERB-B2 tumours are HER2-positive, and luminal A and B tumours are ER-positive and HER2-negative. In our set of 40 in situ tumours, only two tumours are positive for CK5/6 by IHC. Both of them are poorly differentiated and negative for HER2 and ER by IHC. From the intrinsic gene set identified by Perou and colleagues [12], we could match 403 identifiers to our array platform. This set of genes was used to perform unsupervised hierarchical clustering of the 40 in situ samples. We clearly see a discrimination between tumours highly expressing genes of the luminal/ESR1 cluster and tumours negative for these genes, whereas the discrimination for the HER2-overexpressing groups was much less clear ( Figure 1 in the supplementary information [24]). We could not identify a large basal-type group, which is in agreement with the data obtained using IHC.

Unsupervised hierarchical clustering
Unsupervised hierarchical clustering of in situ and invasive samples First, the whole group of DCIS and invasive samples was clustered ( Figure 1a). As can be seen, the invasive samples cluster in three different groups (indicated as I, II, and III in Figure 1a). Ten out of 14 poorly differentiated DCIS samples cluster together in a fourth group, and a fifth group consists of 13 out of 18 cases of intermediately differentiated DCIS and four out of six of the well-differentiated in situ samples. The clustering seems not to be driven mainly by the ER status or the HER2 status of the samples. These results suggest that poorly differentiated DCIS samples show an overall gene expression profile other than that of the intermediately and well-differentiated DCIS samples.

Unsupervised hierarchical clustering of DCIS
We also performed unsupervised hierarchical cluster analysis to the series of DCIS cases only, resulting in two large groups. One group contains 10 poorly differentiated samples and only one well-differentiated sample, whereas 83% of the well-differentiated samples group in the other, second cluster. Most of the samples in this second group are ER-positive by IHC. In total, our sample set contains 18 cases with an intermediately differentiated component. Of these samples, 12 cluster in the arm of the well-differentiated samples. In accordance with the clustering results presented in Figure 1, these results also indicate that the overall gene expression profiles of in situ samples with an intermediately differentiated component are more similar to those of well-differentiated DCIS than to those of poorly differentiated DCIS. It is clear from these results that there are large differences in gene expression pattern between welland poorly differentiated DCIS.

Supervised classification
We performed supervised classification on different data sets to identify the genes differentially expressed between the groups of interest. These groups are (a) 40 in situ versus 40 invasive breast carcinomas, (b) 14 poorly differentiated DCIS cases versus 38 invasive grade 3 tumours, and (c) six cases of well-versus 14 cases of poorly differentiated DCIS.
(page number not for citation purposes)

Supervised classification of in situ versus invasive carcinomas
We investigated the differences in gene expression between in situ and invasive breast carcinoma samples. We therefore used the whole data set and assigned all 40 in situ samples to one group and all 40 invasive samples to a second group (analysis set 1). To obtain a profile taking into account the expression sets of both tumour types, significantly regulated genes were identified independently for both groups. The 1,706 overlapping genes were used for analysis. Monte Carlo randomisation revealed approximately 300 genes differentially expressed between in situ and invasive samples.
After cross-validation, classifier consisting of 35 genes resulted in a stable prediction of the differences between DCIS and invasive breast carcinomas, with an average performance of 91%. The gene list is provided in Table 2.

Supervised classification for poorly differentiated DCIS versus grade 3 invasive carcinoma
Because it is very likely that grade 3 invasive breast cancer arises from poorly differentiated DCIS [5,6], we applied the supervised classification procedure to the subset of poorly differentiated DCIS (n = 14) and grade 3 invasive tumours (n = 24) (analysis set 2). Again, the filtering procedure was applied to both groups independently. The overlapping fraction of this gene list contains 1,119 genes that were used to perform the analyses. Monte Carlo randomisation showed that 80 genes are differentially expressed between poorly differentiated DCIS and grade 3 invasive breast carcinoma samples. After cross-validation in 14 steps, the best performance of 93% is reached, when at least 50 genes are used to build the classifier. This performance remains stable with increasing numbers of genes. This means that 50 to 80 genes are able to discriminate between poorly differentiated DCIS and invasive grade 3 breast tumours (Figure 2a). These 80 genes are shown in Table 3. Between the 35-gene classifier of all DCIS and invasive samples and the subgroup classifier of 80 genes, 21 genes were present in both classifiers.

Supervised classification of well-versus poorly differentiated DCIS
We intended to find the most prominent differences between the well-and poorly differentiated DCIS samples. Sixfold cross-validation of six well-and 14 poorly differentiated in situ samples (analysis set 3) resulted in a set of 43 genes separating these groups with a performance of 90% ( Figure 3a, Table  4).
Because histological classification of intermediately differentiated DCIS versus well-or moderately differentiated DCIS is most challenging, we investigated whether gene expression profiling could be used to identify markers that could help in making this classification. We therefore included the cases classified as intermediately differentiated DCIS. Subsequently, we divided the sample set into one group of well/wellintermediately differentiated samples (n = 16) and a second group containing poorly/intermediately-poorly differentiated samples (n = 20). Supervised classification of these data revealed a set of 78 genes separating these two groups with an average performance of 89% (Table 5). We observed a separation of this data set in three distinct groups ( Figure 3). One group contains one intermediatelypoorly differentiated sample (17%) and 12 out of 14 poorly differentiated samples, and a second group all six well-differentiated samples and seven out of 10 well-intermediately differentiated samples. The third group shows no correlation with both profiles and consists of five out of six intermediatelypoorly and three out of 10 well-intermediately differentiated samples. This implies that this third group typifies mainly the intermediately-poorly differentiated samples. Well-intermediately differentiated samples are apparently very similar to welldifferentiated DCIS in their gene expression. These results are in accordance with the results of unsupervised hierarchical clustering of all in situ samples (Figure 4a).
Twenty-one genes are overlapping between the 43 genes of analysis set 3 and the 78 genes of analysis set 4. It is known that many poorly differentiated in situ breast carcinomas do not express the ER. In our data set, nine of all 14 poorly differentiated DCIS samples (64%) are negative for ER expression by IHC. There was a slight chance that our classifier would detect mainly the differences of ER-associated genes. We identified only one gene (LIV-1), beside the ER itself, directly ER-regulated in the classifier of 43 genes. Additionally, we compared the 43 genes with 2,460 ER-associated genes identified by van 't Veer and colleagues [16]. Thirteen genes, including the ER itself, have been found in both gene lists. So, most of the genes in this 43-gene classifier have not been correlated to ER expression so far, indicating that the differences between well-and poorly differentiated DCIS samples are not originating from the ER status of the samples.
Remarkably, completely different gene lists are found describing the differences in gene expression between different in

List of 80 genes able to discriminate between poorly differentiated DCIS and invasive grade 3 breast tumours
situ samples, on one hand, and DCIS and invasive samples on the other hand. These findings may indicate that gene regulation involved in progression from in situ to invasive breast cancer affects molecular mechanisms other than the mechanisms responsible for the development of the different types of DCIS.

Discussion
Although studies to identify gene expression signatures in DCIS are limited by difficulties in obtaining frozen material from DCIS, we were able to collect a relatively large series of DCIS cases for this purpose. It should be kept in mind that we did    We were able to show that well-(n = 6) and poorly (n = 14) differentiated DCIS show different gene expression profiles and can be distinguished by a classifier of 43 genes. Most of the genes differentially expressed between well-and poorly differentiated DCIS are involved in metabolism (for example, BTD, ETFA, GMFG, and PLAT) and cell communication (for example, ESR1, ACK1, CELSR2, and CCL19).
One of the top genes in the 43-gene classifier is BCL2. The mRNA expression of this anti-apoptotic protein is upregulated in the well-differentiated samples. In addition to its anti-apop- totic function, BCL2 has a suggested role in neuro-endocrine differentiation in colon carcinomas [31] and its downregulation is associated with poor prognosis in breast cancer [32].
Twenty-eight of the 43 genes are upregulated and 15 genes are downregulated in the well-differentiated samples ( Figure  3a). Whereas a number of the 28 upregulated genes are involved in DNA binding, no genes fulfilling this function are on the list of the 15 downregulated genes. Conversely, genes involved in phosphate metabolism (for example, GMFG, ACK1, and ATP5B) can be found within the 15 downregulated, but not in the 28 upregulated, genes.
It is known that HER2 is overexpressed in poorly differentiated DCIS in approximately 42% of the cases [7], and it has been suggested that HER2 overexpression is an early step in the evolution of a distinct type of breast carcinoma. In our data set of all in situ samples, we found a positive log2-ratio for HER mRNA expression in six of 14 poorly differentiated DCIS cases (43%) and in one case of intermediately-poorly differentiated DCIS. In all the other in situ samples, the log2-ratios of HER2 are negative. These results are in agreement with the hypothesis that HER2 overexpression is an early event in the development of poorly differentiated in situ breast carcinomas.
Supervised classification of well-, well-intermediately, intermediately-poorly, and poorly differentiated DCIS samples (analysis set 4) showed a separation of these samples in three groups: a 'good' group, a 'poor' group, and an 'intermediate' group containing mostly samples that were identified as intermediately-poorly differentiated samples by pathologists. This group also contains some samples pathologically classified as well-intermediately differentiated, whereas most of these samples fall in the 'good' group. These results indicate that welland well-intermediately differentiated DCIS are more similar to each other than poorly and intermediately-poorly differentiated DCIS are. Following this idea, well-and well-intermediately differentiated samples may be considered to be one group, whereas poorly and intermediately-poorly differentiated samples seem to be two distinct groups of DCIS. If these results can be validated in additional studies, this classification could help to decrease controversial classification of DCIS due to interobserver variability and to recognise well-differentiated DCIS with more accuracy.
Within the gene lists describing the differences between welland poorly differentiated DCIS, a number of genes refer to proteins for which antibodies are available. There is no single gene discriminating between the different types of DCIS, but it has to be investigated whether a combination of protein stainings in a patient's tissue can assist in better classification of DCIS. From the study presented here, potential candidates for such an approach are Bcl-2, Ack1, CCL19, and CELSR2, among others.  [19] studied six cases of DCIS with necrosis and four samples of DCIS without necrosis and identified a signature of 69 transcripts differentially expressed between these two groups. Ma and colleagues [33] used laser capture microdissection from paraffin-embedded material followed by gene expression profiling to identify molecular signatures in premalignant, preinvasive, and invasive stages of breast cancer. The results of their study suggested that tumour grade, rather than tumour stage, is associated with distinct gene expression patterns and that changes in gene expression required for invasive growth are already present in the DCIS stage [33]. In the study presented here, we compared the gene expression profiles of poorly differentiated DCIS lesions with those in grade 3 invasive breast tumours. In contrast to Ma and colleagues, we did not compare paired samples from the same patient but compared two groups of tumours. The 80-gene signature we identified is different from the signatures describing the differences between different grades of DCIS lesions. Schuetz and colleagues [34] identified gene expression signatures of in situ and invasive breast cancer by using 18 paired samples and combining laser capture microdissection and gene expression profiling on oligonucleotide microarrays. They showed that 546 probes were differentially expressed between DCIS and IDC. From the 18 genes they validated by real-time polymerase chain reaction, four (MMP11, PLAU, BGN, and FAP) are also present in our filtered data sets of significantly regulated probe sets comparing DCIS and invasive samples. They all show the same expression pattern as described by Schuetz and colleagues and are expressed at higher levels in the groups of invasive tumours. One of these genes (MMP11) is also part of the 35-gene and 80-gene classifiers. MMP11 and PLAU have already been correlated to invasion and poor prognosis [35,36]. FAP (seprase) is a membrane-bound protease that has been suggested to reduce the dependence of breast cancer cells on exogenous growth factors in vitro and thereby to facilitate tumour growth and metastasis [37]. Allinen and colleagues [38] identified comprehensive gene expression profiles of the different cell types in normal breast, DCIS, and invasive breast cancer tissue. These data show that dramatic gene expression changes occur between normal breast tissue and breast carcinomas and that these changes are already present at the DCIS stage. These results also suggest a role of the chemokines CXCL12 and CXCL14 in breast tumourigenesis. Neither chemokine is present on our array platform, but CXCR4, which is the receptor for CXCL12, is. CXCR4 does not appear in the set of significantly regulated genes, indicating that it does not play a crucial role in our series of tumours, which reflects the data of a mixed population of cells enriched for tumour cells, whereas Allinen and colleagues performed gene expression profiling on microdissected cell populations.
A recent study by Nagaraja and colleagues [39] describes gene expression patterns corresponding to normal breast, noninvasive breast cancer, and invasive breast cancer by using several cell lines. They identified genes involved in cellcell and cell-matrix interactions which were altered in their expression. A set of nine genes was sufficient to distinguish between invasive and non-invasive cell lines [39]. From this set of nine transcripts, six could be matched to our array platform. For three of them (cadherin 11, annexin A1, and vimentin), we observe the same expression pattern as published by Nagaraja and colleagues for the transition from in situ to invasive carcinoma. The other three transcripts (S100A8, claudin 3, and cadherin 1) are upregulated in the invasive cancer cell line in the data set of Nagaraja and colleagues, whereas we see a downregulation in the invasive grade 3 tumours compared with the group of poorly differentiated samples. This may be due to the fact that Nagaraja and colleagues generated in vitro data, which we compared with our human breast cancer data set.
Porter and colleagues [40] identified a subset of genes that are significantly regulated in DCIS or invasive carcinomas. They identified 26 genes that were differentially expressed between normal and DCIS samples or intermediate-and highgrade DCIS, respectively. From these, only XBP1 is present in one of our classifiers (78 genes). Porter and colleagues describe this transcript as tumour-specific, meaning upregulated in in situ and invasive tumours compared with their normal samples. We find that XBP1 is significantly more highly expressed in well-and well-intermediately differentiated DCIS samples than in poorly/intermediately-poorly differentiated ones.
Wulfkuhle and colleagues [41] performed proteomic analyses of six matched normal and DCIS samples of the human breast. They identified proteins that are more highly expressed in individual DCIS samples and that are involved in cytoskeletal regulation or vesicular trafficking or have chaperone activity. From the 15 proteins from which the expression has been validated by IHC, 12 are present as probes on our array platform. Three of those (profilin, stathmin, and prohibitin) are differentially regulated between DCIS and invasive samples, and all three show a higher expression in the invasive samples than in the DCIS samples. This is in line with the paper of Wulfkuhle and colleagues, which describes a higher expression of these proteins in the DCIS samples than in normal tissue. This indicates that changes in gene and protein expression observed in invasive tumours are already present in the transition from normal tissue to DCIS lesions.

Conclusion
We demonstrate here that gene expression profiling can distinguish between in situ breast cancer samples of well-versus poorly differentiated type. There appear to be a group of poorly differentiated samples, a group of well-and well-intermediately differentiated samples, and a third group containing mainly intermediately-poorly differentiated in situ cases. The quantitative differences in gene expression between these groups are mainly between twofold and fourfold. These differences are difficult to detect by classical IHC, because this technique is not very accurate in the quantification of small differences in protein expression. So far, there are no single markers that distinguish between the different types of DCIS, but the possibility of identifying a manageable panel of markers to distinguish the different types of DCIS lesions has to be further investigated.