Details concerning the aCGH profiling of the samples and subsequent normalization can be found in Materials and Methods. Detailed clinical data of the breast cancer cohort profiled is available in Additional Data File 1, while the raw and normalized aCGH data for tumors and cell lines is available from NCBI's Gene Expression Omnibus (GEO) 161718 under the series accession number GSE8757. Briefly, after segmentation of the mode-normalized data using the CBS algorithm 19, we applied the method described in 5 to define thresholds for gain and loss. We observed that because the cellularity of samples varied widely (mean cellularity, expressed as percentage, was 69% with a standard deviation of 19%), the genome instability index (GII; defined as the fraction of genome altered) was highly correlated with cellularity (Additional Data File 2, panel A). To correct for this unwanted effect without sacrificing a considerable number of samples, thresholds were redefined separately for each sample using a cellularity correction model similar to the model described in 20 (see also Materials and Methods). After correction, the GII became independent of cellularity (Additional Data File 2, panel B), thus validating the approach we adopted. The choice of thresholds was further validated with the help of breast tumor cell lines with known gains and losses. Thresholds for amplification were initially defined for cell-lines with known amplicons and rescaled for primary tumors using the cellularity correction (see Materials and Methods).

To test our normalization and segmentation further, we evaluated the concordance of alteration patterns between the oligo array and a genosensor BAC array, on which 126 of the 171 breast tumors had been previously profiled 10 (see also Materials and Methods). After matching the locations of the oligos to the 281 BACs representing cancer-related loci, we found a strong concordance between both types of copy number data (28 of the 34 matched regions, 82%, showed strong agreement with a Fisher-exact test p < 0.05; see Additional Data File 2, panel C). A similar degree of good concordance between BAC and oligo data was recently observed across a panel of 19 prostate cancers 21.

Genomewide patterns of gain and loss showed a significant number of highly recurrent altered regions (Figure 1 and Additional Data File 3). The patterns for tumors and cell lines were remarkably similar to each other and in concordance with previously published studies 16713. Interestingly, the pattern was also similar to that reported for lung cancer 22. In brief, chromosomal regions that were most commonly gained in both tumors and cell lines were 1q21.1-qtel, 5ptel-5p13.3, 8p12-8q24.3, 17q12, 17q21-17q25.1 and 20q11-qtel. Chromosomal regions that were most commonly lost in both tumors and cell lines were 8ptel-8p12, 11q14-qtel, 13q21-qtel and 17ptel-17p11.2. However, there were also notable differences between tumors and cell lines. Specifically, cell lines showed a higher frequency of losses on chromosomes 9, 18 and X, and a lower frequency of losses on 16q, as compared with tumors. On the other hand, tumors showed a higher frequency of gains on 16p. In agreement with 6 we observed regions of recurrent high-level amplification on chromosomes 8, 11, 12, 17 and 20 (Figure 1a) bounding well-known breast cancer oncogenes (e.g. BRF2, ASH2L, CCND1, EMSY, ERBB2, NCOA3, MYBL2, STK6) 102324, although amplification frequencies were much lower on chromosomes 12 and 20 as compared with those reported in 6. In contrast, cell lines did show amplification frequencies on chromosomes 12 and 20 that were more in line with those observed in 6 (Figure 1b). We found homozygous deletion (HD) to be a rare event in primary tumors and only found evidence of HD in two cell lines and one tumor on chromosome 13q14 where the retinoblastoma gene (RB-1) resides.

To perform dimensional reduction we developed an extension (CRalg) of the minimal regions algorithm of Rouveirol (MRalg) 25, which, in contrast to MRalg, identifies common regions of alteration (CRA) (see Materials and Methods). Using CRalg we achieved a substantial dimensional reduction (from 27695 oligos to 5914 CRA that showed at least 5% changes across tumors) without losing any information in the process (note that the MRalg and CRalg algorithms will work unchanged if instead of using 1 and -1 to indicate gain and loss, we used the precise segment values; thus, CRalg achieves a dimensional reduction without further information loss), automatically including gains and losses in the same matrix. However, a drawback of CRalg was the relatively larger number of variables (5914 CRA compared with 1134 minimal regions of alteration (MRA)) and the high degree of redundancy/correlation since many adjacent CRA only differed in value in one sample. In order to reduce the redundancy of the CRA matrix, we applied an algorithm that merged together adjacent regions that differed in only a single sample (see Materials and Methods). This gave a reduced matrix of 1063 merged CRA (mCRA) over 171 breast tumors.

While standard hierarchical clustering algorithms have been successfully applied to BAC-derived continuous log-ratio data, we explored the possibility of incorporating the inherent discreteness of copy number data into the unsupervised classification analysis. Specifically, we performed (complete linkage) hierarchical clustering over the matrix of mCRA using the number of copy number state differences as a distance metric. This revealed a complex pattern of gains and loss across the cohort (Figure 2). Using the methodology implemented in the R-package pvclust 2627 for testing the robustness of the clusters, we found that only one reasonably sized cluster of 26 samples was reliable with a robustness index larger than 90% (Figure 2 and Additional Data File 4). This cluster was characterized by a very low GII (average of 0.036 ± 0.035) relative to the rest of samples (average of 0.22 ± 0.12), which was highly significant (Wilcoxon test p < 10^-13). We verified that this result was independent of cellularity by showing that this cluster did not have a significantly lower cellularity than the rest of samples (Wilcoxon rank sum test p = 0.69). The 26-sample cluster was made up of proportionally more ER-negative (15) than ER-positive tumors (11) (Fisher-exact test p = 0.007) as well as more basal (6) than luminal tumors (5) (Fisher-exact test p = 0.01), but was equally distributed in terms of histological grade (3 grade I, 10 grade II and 13 grade III, p = 0.28), the immunohistochemical markers ERBB2, PGR, AR and p53, and p53 mutation status (21 samples with no p53 mutation and 4 with p53 mutation, p = 0.79). Among the 26 samples there were 8 with gains of ERBB2 and 5 of these had a high-level ERBB2 amplification. This confirms the observation made in 10 that a proportion of ERBB2-amplifier tumors show little overall genomic instability.

Two further, yet much smaller, clusters with robustness indices greater than 90% and of relatively high GII were also identified (Figure 2). The cluster with the highest GII was made up of 9 samples and was mainly characterized by gains of 1q, 8q, telomeric end of 17q and 20, and unaltered chromosome 16. Most of the samples were ER negative (6 ER- versus 3 ER+) and of high grade (7 grade III, 1 grade II and 1 grade I). Another robust cluster of 12 samples and intermediate GII was characterized mainly by loss of chromosome 17, loss of 16q and gain of 8p. This cluster was made up of 7 ER+ and 5 ER- samples and was also mostly high grade (7 grade III, 3 grade II and 2 grade I). The rest of samples could not be characterized as members of large stable clusters.

The identification of a subclass of breast tumors of low genomic instability that was proportionally enriched in terms of ER- and basal tumors was striking and suggested to us that, in contrast to present belief, there is a subtype of ER- tumors of relatively low genomic instability and which includes a subset of ERBB2-amplifier tumors. Further evidence for this came from a Wilcoxon rank sum test comparing the GII distributions of ER- and ER+ samples, which showed that the GII of ER- samples was not significantly higher than that for ER+ samples (Figure 3a, p = 0.35). Importantly, among the 15 ER-samples within the 26 sample low-GII subgroup, 10 were of high grade, 4 of intermediate grade and only 1 of low grade, which was proportionally similar to the distribution in the rest of the ER- cohort (30 high grade, 9 intermediate grade and 3 low grade tumors, p = 0.88). This showed that the ER- samples in the low GII cluster were not necessarily of lower grade.

To obtain further evidence for the existence of a low GII ER- subtype, we sought independent validation in three external breast cancer cohorts 6911 for which copy number data was available. Specifically, we computed the GII in these external cohorts as described in Materials and Methods and tested, using a one-sided Wilcoxon rank sum test, whether there was a substantial number of ER- samples of relatively low GII (Figure 3b, c and 3d). Lending further support to the existence of this low-GII subtype, in two of these external cohorts 69 we did not find the GII of ER- samples to be significantly higher than that for ER+ samples.

In terms of the intrinsic subtype classification 282930, for which a single sample predictor (SSP) was recently derived and validated in external cohorts 31, we found that the 26-sample low-GII subgroup was made up of 6 basal, 3 HER2+, 4 luminal-A, 4 normal and 1 luminal-B tumors (8 samples could not be classified owing to missing gene expression information). As before, when taking into account all samples, the basal subtype did not have a significantly higher GII than the luminal-A subtype (p = 0.44) (Figure 3e). We interpreted this result as further evidence for the existence of a low-GII basal subtype. The only statistically significant differences between the GII distributions of the various intrinsic subtypes were between the normal subtype and all others (p < 0.05 for all comparisons) and between the luminal-A and luminal-B subtypes (p = 0.009). We observed a similar GII distribution in another cohort for which expression data was available 6 (Figure 3f). Specifically, in this cohort as well, the basal subtype did not have a significantly higher GII than the luminal-A subtype (p = 0.26), while the luminal-B subtype did (p = 0.03).

To further characterize the identified low-GII subgroup, we attempted to derive an associated transcriptomic signature from the 113 samples for which additional gene expression information was available. To this end we used a multiple logistic regression model and ranked genes according to the difference of their model Akaike information criterion (AIC) score 32 with respect to a null model AIC score that only included ER status (see Materials and Methods). The null distribution for AIC scores was obtained by performing 10000 random permutations of the sample expression values. Hence, this method allowed us to rank the genes according to how well they discriminated between the 26-sample low-GII cluster and the rest of the cohort, independently of ER status. To correct for multiple testing we converted the p values into q-values 33, which provided us with an estimate of the false discovery rate (FDR). This showed that, for example, among the top-50 genes we would expect on average about 10 false positives, thus confirming the existence of an expression signature associated with this subclass.

To derive a classifier based on this gene signature we decided on a linear discriminant classifier where class assignment is determined by a nearest centroid criterion using an euclidean distance metric. The centroids were constructed using the top-37 genes (Additional Data File 5), yielding an average of 7 false positives. To test this classifier we first applied it to the 135 NCH samples with gene expression information 15. This classified 15 ER- and 9 ER+ into the putatively low-GII subgroup (25 ER- and 84 ER+ were classified into the other group), which we verified had a lower GII than the rest of the samples (Wilcoxon test p < 10^-4). It is striking that even though the classifier was derived independently of ER status, classifying for this particular subgroup of low GII predetermined samples to be more likely ER- than ER+ (Fisher-exact test p = 0.0003). Interestingly, applying the classifier to four additional breast cancer cohorts with expression profiles 6343536 showed that the corresponding putatively low-GII subgroup in these cohorts was also enriched for ER- tumors (Additional Data File 6). In a further data set 30 where only 85 samples (18 ER- and 56 ER+) were available the predicted low-GII subclass had only 4 ER+ and 3 ER- samples (and 3 samples had missing ER information), which did not reach statistical significance, but suggested to us that it possibly would if more samples were available.

Fortunately, the tumors in 30 were profiled recently at the copy number level 12. This allowed us to validate the hypothesis that our gene expression classifier selects a particular subtype of low GII in breast cancer. Using the GII for the samples in this cohort we compared the GII of the predicted low-GII subgroup, as determined by our expression classifier, with the rest of samples in that cohort (Figure 4a and 4B). In spite of only 10 samples being classified into the predicted low-GII subgroup, we could verify that it was characterized by a lower GII when compared with the rest of the samples (Wilcoxon test p = 0.001). Moreover, among these 10 samples, 6 were of high grade, 2 of intermediate grade and none were of low grade (2 samples had missing information). For the other external cohort for which both copy number and expression data was available 6, the predicted low-GII subgroup had a lower median GII than the rest of samples, but did not reach statistical significance (Figure 4c and 4D).

To better understand the nature of the expression classifier we performed both gene ontology (GO) analysis using GOTM 37 and pathway analysis using MSigDB 38. GOTM on the 37 genes making up the classifier showed enrichment of inflammatory and defense response genes (CXCL1, CXCL2, XCR1, LY96, NMI, TLR2, uncorrected p < 10^-5), which were generally upregulated in the low-GII subgroup, and marginal enrichment of signal transduction (RASSF2, SNX4, CASP1, MKNK1, RHPN1, INPP5D, uncorrected p = 0.002) and apopotosis genes (BCL2A1, MRPS30, CASP1, CASP4, TLR2, uncorrected p = 0.004). Pathway analysis using MSigDB confirmed the involvement of the caspase, cell death, TNF-α-NF-κβ, inflammatory response and signalling pathways, although these statistical associations were lost on correction for multiple testing (data not shown).

Of the 171 breast tumors, 113 were also profiled on Agilent gene expression arrays 15. This allowed us to evaluate the contribution of gene-dosage levels to gene expression (Additional Data File 7). Of the 5914 CRA, 4551 (77%) contained at least one Agilent probe. Of these 4551 CRA, 2407 harbored at least one Agilent probe for which there was at least 10 (~5%) expression values in the altered (i.e. gained or lost) group of samples (note that owing to missing values in the gene expression data, p values could not be reliably computed for many probes). Thus, for 2407 CRA at least one reliable p value (Wilcoxon test) could be computed (see Materials and Methods) to evaluate the significance of the association between copy number and aberrant expression. We found that from the 2407 CRA, there were 806 CRA for which there was at least one probe with significant association (p < 0.05) between gain and overexpression, and 412 for which there was at least one probe with significant association between loss and underexpression. On average about 34% of probes in regions that were gained in at least 5% of samples were significantly overexpressed relative to the samples that showed no copy number alteration. Similarly, about 29% of probes in regions that were lost in at least 5% of the samples were significantly underexpressed relative to the samples that showed no copy number alteration. This confirms the finding reported elsewhere 1 that a significant proportion of gene expression variation is caused by underlying copy number alterations.

To find the CRA showing the strongest associations between copy number and expression we first tabulated those CRA with at least 10% gains or losses and which showed a significant association with expression (p < 0.05; see Additional Data File 8). To narrow this down to a smaller set of the most significant regions ('hotspots') we next selected those CRA with an association index (AI) value larger than or equal to 0.5 and a most significant p value of less than 0.001, where the AI was defined as the fraction of probes within the CRA that had significant p values (see Methods). This yielded 196 and 63 hotspots that showed significant association with overexpression and underexpression, respectively (Additional Data File 9). In the case of loss and associated underexpression this table included the well-known tumor suppressors RB1, CDH1, MBD2 and EP300, while in the case of gain and overexpression it included many well-known and potentially novel oncogenes such as MUC1 on 1q21.3, ASH2L, BRF2, LSM1 on 8p12, FADD on 11q13, ERBB2, PNMT, GRB7 on 17q12, TOP2A, THRA, NR1D1 on 17q21, and NCOA6, YWHAB, UBE2C on 20q13. Of these candidate oncogenes, several, notably TOP2A, PNMT and UBE2C, have appeared in prognostic gene expression signatures 394041, thus re-emphasizing their important role in breast cancer. Among the hotspots that were gained, we provide a further selection of those that also showed frequent amplifications and which are therefore likely to harbor candidate oncogenes (Table 2).

As the identified 196 and 63 hotspots represent the regions of strongest association between copy number and coordinate aberrant expression, it was natural to investigate whether any of these regions also showed association with clinical outcome. To this end we performed univariate Cox proportional hazard regressions comparing the HRs for samples with gain (loss) of a hotspot with samples without altered hotspots for three different outcome endpoints (overall survival (OS), disease free interval (DFI) and time to distant metastasis (TTDM); see Additional Data File 9). In addition, we performed Cox regressions for those hotspots with at least 10 amplifications and estimated the HR for samples with and without amplification (Additional Data File 9). This analysis showed that there were three cytoband regions of frequent amplification (8q22.3, 8q24.11-8q24.13 and 11q14) and associated with either OS or TTDM (log-rank test p < 0.05; see Table 3). We verified that for all of these regions samples with amplification had approximately a twofold risk increase of poor outcome compared with samples without the amplification (Table 3). Interestingly, 37 tumors had amplifications in any one of these hotspot regions and this subgroup had also a much higher risk of distant metastasis and a poorer survival rate compared with samples without amplification (TTDM HR = 3.0 (1.6-5.8) p < 10^-3; see Table 3). Genes located in these hotspots, which also showed coordinate overexpression, included EDD1, WDSOF1 (8q22.3), THRAP6 (8q24.11), DCC1 (8q24.12), SQLE, KIAA0196 (8q24.13) and NDUFC2, ALG8, USP35 (11q14.1) (Tables 2 and 3 and Additional Data File 9). Importantly, multivariate Cox regression analysis in a model that included the NPI (one of the strongest prognostic factors in univariate analysis), showed that amplification of any of these five regions was a strong prognostic factor independently of NPI (Table 4).