All human tumour samples were collected from fresh frozen primary breast tumours using protocols approved by the Institutional Review Board and were profiled as described previously using Agilent (Agilent Technologies, United States) oligo microarrays 9141516. The primary microarray data for the 232 sample data set is available in the Gene Expression Omnibus (GEO) [GEO:GSE3165]. The data set containing only tumours with informative LOH status can be found in GEO [GEO:GSE10884], with 13 new samples in this study.

Patient DNA from lymphocytes, normal breast tissues and breast tumours was isolated using the DNeasy kit (Qiagen, Germany). We used two polymorphic markers; a variable number tandem repeat (VNTR) in intron 20 and D13S153; a microsatellite marker for RB1 LOH analyses. The primers were previously published for intron 20 17. The primers for D13S153 [AFM058xd6a, AFM058xd6m] were obtained from the Genome Data Bank (Johns Hopkins University, Baltimore, MD) 18. The PCR products were run on the Agilent Bioanalyzer using DNA 1000 kit (Agilent, United States). The patient was called informative when there were two alleles present in their normal DNA. LOH was called when there was at least a 50% loss of an allele/band in the tumour for at least one of the two polymorphic markers.

The chi-square test and Fisher-Freeman-Halton exact test were used to examine correlations between RB1 LOH status, immunostaining and tumour subtype using SAS 9.1 (Cary, NC). An analysis of variance (ANOVA) and unpaired Student's t-test were performed and a box plot graph plotted to compare RB1 LOH status or immunostaining with gene expression using web based 'Statistics to Use' 19. A Significance Analysis of Microarrays (SAM) was performed to identify genes that were significantly differentially expressed between tumours with RB1 LOH-positive tumours compared with LOH-negative tumours 20. Expression analysis systemic explorer was used to identify gene ontology categories overrepresented in the RB LOH gene list compared with the genes present on the array.

Whole genome RVista was used to determine the known transcription factor binding sites overrepresented in the 1 kb upstream region in the lists of genes examined compared with the rest of the RefSeq genes in the whole human genome 21. Hypergeometric mean analysis was performed as described by Chung and colleagues 22; this comparison gives the likelihood of finding co-occurrences between these gene sets by chance. The simulation was performed independently for each pair of gene sets analysed.

Formalin-fixed, paraffin-embedded tissue sections (approximately 5 μm) were processed using standard immunostaining methods. Following deparaffinisation in xylenes, slides were rehydrated through a graded series of alcohol and rinsed in PBS. Endogenous peroxidase activity was blocked with 3% hydrogen peroxidase. Samples were steamed for antigen retrieval with 10 mM citrate buffer (pH 6.0) for 30 minutes. Slides were then incubated for 20 minutes with diluted normal blocking serum. The sections were incubated for 60 minutes at room temperature with primary antibody pRb (Visionbiosystems Novocastra, NCL-L-RB-358 clone 13A10, 1:50 dilution) or p16 (Santa Cruz, H-156, 1:50 dilution). The slides were incubated for 45 minutes with diluted biotinylated secondary antibody (1:250 dilution) and 30 minutes with Vectastain Elite ABC reagent (Vector Laboratories, United States). Sections were incubated in peroxidase substrate solution for visualisation. Slides were counterstained with haematoxylin and examined by light microscopy. Tumour immunoreactivity was scored as: 0 = negative, 1 = weak positive, 2 = moderate positive and 3 = strong positive. The evaluation of p16^INK4a and RB1 protein staining was performed by two blinded independent researchers.

Tumours of the basal-like subtype have been shown in several studies to have a high mitotic rate and to highly express a proliferation gene signature 1112. For this reason, we postulated that there might be a defect in the RB pathway in these tumours. We first examined the expression levels of the core components and regulators of the RB pathway in a previously published microarray data set 914 that contained 232 microarrays consisting of 184 primary breast tumour samples and nine normal breast samples (Figure 1). On average, RB1 was expressed at the lowest levels in basal-like tumours and at the highest levels in luminal A tumours (Figures 1 and 2a), while the converse was observed for the average expression of a previously defined proliferation gene signature 9 (Figure 2b). Tumours of the basal-like subtype also frequently showed high levels of p16^INK4a (Figures 1 and 2c) and E2F1 (Figure 1 and data not shown). Cyclin D1 levels, on the other hand, were elevated mainly in luminal tumours including noticeably higher expression in many luminal B tumours (Figures 1 and 2d); in addition, it was recently reported that there are often high level gains of the 11q13 locus that includes cyclin D1 in luminal tumours 23, therefore suggesting that cyclin D1 high expression can be considered a 'luminal event'. Similar findings concerning RB1 mRNA expression and the basal-like subtype have also recently been reported suggesting that this is a reproducible expression feature of basal-like tumours 24.

Due to the low expression of RB1 message in basal-like tumours we decided to examine these breast carcinomas for LOH at the RB1 locus. We investigated 88 paired primary human breast carcinomas and normal tissue genomic DNA samples to assess the frequency of LOH in RB1. We used two polymorphic markers located at the RB1 locus (13q14); a variable number tandem repeat (VNTR) in intron 20 and D13S153, that is a microsatellite marker located within intron 2. There were 67 cases that were informative for at least one of these two markers (see Figure 3a as an example LOH). In total, 26 tumours showed RB1 LOH for at least one marker (26 of 67, 38.8%). This is consistent with the frequency of RB1 LOH seen in previous studies of breast cancer (26 to 47%) 5625.

Next, using a previously defined proliferation gene signature 9, we clustered the gene expression data for this signature using just the 67 LOH informative patients (Figure 3b). This analysis was able to sort the samples into two groups with the right most group containing most of the RB1 LOH positive tumours (21 of 27, 77.8%). The left most group contained 29 of 40 (72.5%) of the RB1 LOH-normal tumours. It is important to note, however, that many of the LOH positive tumours that did cluster in the left group were located on the outer nodes and still showed higher expression of the proliferation markers than the tumours in the centre. Overall, an ANOVA showed that RB1 LOH was highly correlated with the high expression of the proliferation gene cluster (p = 0.0001) (Figure 3c).

We next classified the tumours that were assayed for RB1 LOH analysis according to intrinsic subtype using the five-class single sample predictor 9. The frequency of RB1 LOH varied by molecular subtype (p = 0.0002) (Table 1). The lowest LOH frequencies were observed in the luminal A (3 of 20, 15%) and normal-like (0 of 7, 0%) subtypes, while the HER2-enriched subtype had a frequency near the breast cancer average (3 of 9, 33.3%). The highest frequency of RB1 LOH was observed in tumours of the basal-like (13 of 18, 72.2%) and luminal B subtypes (8 of 13, 61.5%), both of which are known to be highly proliferative tumour subtypes 91011.

RB1 is expressed ubiquitously in mammary epithelial cells and typically shows a nuclear staining pattern in normal human breast tissue (Figure 4a). RB1 immunostaining was statistically correlated with RB1 message levels (p = 0.0081) but, as has been described before, RB1 protein expression did not correlate with RB1 LOH (p = 0.5) 6; however, a trend for low expression of RB1 message to occur with RB1 LOH was observed (p = 0.11). RB1 protein expression also tended to be low in basal-like tumours (Figure 4b and Table 1), but this relationship was only near statistical significance (p = 0.064).

High p16^INK4a staining (3+), which is a hallmark of lost RB1 function 2627, was seen in 32 of 119 tumours assayed and often included both nuclear and cytoplasmic staining (Figure 4c). p16^INK4a immunostaining was statistically correlated with p16^INK4a message levels (p = 0.013), especially when high staining was observed (data not shown). p16^INK4a immunostaining was also associated with intrinsic tumour subtype (p = 1.41E-05) with 22 of 33 (66.7%) of basal-like tumours showing 3+ staining (Figure 4c, Table 1). The correlation between RB1 LOH status and p16^INK4a expression levels was statistically significant (p = 0.01) (Figure 4d, Table 2). In addition, p16^INK4a was similarly highly expressed in transgenic murine mammary tumours with loss of RB function driven by SV40 large T-antigen or T121 (Additional File 1) 14. When considered together, these data suggest that RB1 LOH in basal-like tumours cause a loss of RB1 protein function, which the cells attempt to compensate for by increasing p16^INK4a gene and protein expression levels.

To determine if there was a gene expression signature related to RB1 LOH, a two-class SAM 20 was performed of RB1 LOH positive tumours vs. LOH-normal tumours. In total there were 452 genes that varied with RB1 LOH status with a false discovery rate of 0.94%, as compared with 11 genes identified at a false discovery rate of 16.7% when a similar analysis was performed using RB immunostaining data (0 vs. 1, 2 and 3). Of these genes, 423 were highly expressed in tumours with RB1 LOH and an analysis of the gene ontologies associated with this gene set showed that cell cycle, cell division, DNA metabolism, spindle organisation and biogenesis, and response to DNA damage were the top biological processes discovered when using Bonferroni-corrected scores. Interestingly, E2F1, E2F3 and E2F5 were present on this supervised gene list and highly expressed in tumours with RB1 LOH. Also present in this list was RB1CC1, a regulator of RB1 expression that has been shown to contain truncating mutations in breast cancers 28. We used whole genome RVista to calculate which transcription factor binding sites might be present within the 1000 bp upstream regions of these genes 21 and determined that the top three transcription factor binding sites with p < 0.005 were E2F4:DP1, E2F1:DP1:RB and E2F4:DP2, showing that a majority of these genes are likely to be E2F-regulated; other statistically significant transcription factor binding sites were HIF1:ARHN. Only 29 genes were significantly negatively/downregulated from the RB1 LOH SAM analysis, and there were no significant gene ontology categories enriched in this list.

Recently, an RB-loss gene expression signature was derived using mouse fibroblasts with either acute or chronic knockout of RB1 using conventional knockout as well as Cre-Lox technology 29. As might be expected, this RB-loss signature was highly expressed in basal-like tumours (Additional File 2). This RB-loss signature significantly overlapped with the proliferation signature (29 of 139 genes of the RB-loss signature are contained in the 140 gene human proliferation signature used here, hypergeometric mean p < 0.001), thus serving as further evidence that the proliferation signature contains many RB1-E2F regulated genes. There was statistically significant overlap among all three RB-pathway signatures studied here (ie, RB-LOH, RB-loss and proliferation) as determined by hypergeometric mean analysis p < 0.001, and thus all three signatures are probably tracking a common biology that is RB1-E2F dependent. There were 20 genes that overlapped between all three gene lists (Figure 5), which included cell cycle-related genes including the spindle assembly checkpoint proteins BUB1 and MAD2, and many commonly used chemotherapeutic drug targets including TOP2A (doxorubicin, etoposide), thymidylate synthetase (5-FU), ribonucleotide reductase M2 (hydroxyurea) and CDC2 (flavopiridol, staurosporine).

All four of these RB-pathway associated lists (RB-LOH, RB-loss, proliferation signature and the 20 common genes) were highly predictive of breast cancer patient outcomes when using a two-class or three-class, average value rank order expression cutoff and when tested on the NKI295 patient data set (Figure 6, shown using RB-LOH list and data not shown for the other three lists that give very similar results) 30, and on a previously described 251 patient data set (data not shown) 31. It should be noted, however, that several groups have independently identified different gene lists that contain a large number of so-called proliferation/RB-pathway genes 1332333435 and it was expected that the RB-LOH signature would be a strong prognostic profile.

The RB-pathway associated signatures were predictive of poor prognosis in breast cancers. But are they predictive of response to therapy? To answer this question, we determined if the RB-LOH signature correlated with pathological complete response in a well-annotated data set of neoadjuvant chemotherapy-treated patients 36. The study by Hess and colleagues included 133 patients with stage I to III breast cancer that were treated with preoperative weekly paclitaxel followed by fluorouracil-doxorubicin-cyclophosphamide chemotherapy. High expression of each of the four RB-pathway signatures was associated with pathological complete response in the breast and regional lymph nodes (Additional File 3). Thus, high expression of the RB-LOH and proliferation signatures is associated with a good response to neoadjuvant chemotherapy in breast cancers.