Using the discovery set of 997 breast cancers, we next asked whether novel biological subgroups could be found by joint clustering of copy number and gene expression data. On the basis of our finding that cis-acting CNAs dominated the expression landscape, the top 1,000 cis-associated genes across all subtypes (Supplementary Table 30) were used as features for input to a joint latent variable framework for integrative clustering20 (see Methods). Cluster analysis suggested 10 groups (based on Dunn’s index) (see Methods and Supplementary Figs 22 and 23), but for completeness, this result was compared with the results for alternative numbers of clusters and clustering schemes (see Methods, Supplementary Figs 23–27 and Supplementary Tables 31–33). The 10 integrative clusters (labelled IntClust 1–10) were typified by well-defined copy number aberrations (Fig. 4, Supplementary Figs 22, 28–30 and Supplementary Tables 34–39), and split many of the intrinsic subtypes (Supplementary Figs 31–33). Kaplan–Meier plots of disease-specific survival and Cox proportional hazards models indicate subgroups with distinct clinical outcomes (Fig. 5, Supplementary Figs 34, 35 and Supplementary Tables 40 and 41). To validate these results, we trained a classifier (754 features) for the integrative subtypes in the discovery set using the nearest shrunken centroids approach21 (see Methods and Supplementary Tables 42 and 43), and then classified the independent validation set of 995 cases into the 10 groups (Supplementary Table 44). The reproducibility of the clusters in the validation set is shown in three ways. First, classification of the validation set resulted in the assignment of a similar proportion of cases to the 10 subgroups, each of which exhibited nearly identical copy number profiles (Fig. 4). Second, the groups have substantially similar hazard ratios (Fig. 5b, Supplementary Fig. 35 and Supplementary Table 40). Third, the quality of the clusters in the validation set is emphasized by the in-group proportions (IGP) measure22 (Fig. 4).

Figure 4: The integrative subgroups have distinct copy number profiles. Genome-wide frequencies (F, proportion of cases) of somatic CNAs (y-axis, upper plot) and the subtype-specific association (–log 10 P-value) of aberrations (y-axis, bottom plot) based on a χ2 test of independence are shown for each of the 10 integrative clusters. Regions of copy number gain are indicated in red and regions of loss in blue in the frequency plot (upper plot). Subgroups were ordered by hierarchical clustering of their copy number profiles in the discovery cohort (n = 997). For the validation cohort (n = 995), samples were classified into each of the integrative clusters as described in the text. The number of cases in each subgroup (n) is indicated as is the in-group proportion (IGP) and associated P-value, as well as the distribution of PAM50 subtypes within each cluster. Full size image Download PowerPoint slide

Figure 5: The integrative subgroups have distinct clinical outcomes. a, Kaplan–Meier plot of disease-specific survival (truncated at 15 years) for the integrative subgroups in the discovery cohort. For each cluster, the number of samples at risk is indicated as well as the total number of deaths (in parentheses). b, 95% confidence intervals for the Cox proportional hazard ratios are illustrated for the discovery and validation cohort for selected values of key covariates, where each subgroup was compared against IntClust 3. Full size image Download PowerPoint slide

Among the integrative clusters, we first note an ER-positive subgroup composed of 11q13/14 cis-acting luminal tumours (IntClust 2, n = 45) that harbour other common alterations. This subgroup exhibited a steep mortality trajectory with elevated hazard ratios (discovery set: 3.620, 95% confidence interval (1.905–6.878); validation set: 3.353, 95% confidence interval (1.381–8.141)), indicating that it represents a particularly high-risk subgroup. Several known and putative driver genes reside in this region, namely CCND1 (11q13.3), EMSY (11q13.5), PAK1 (11q14.1) and RSF1 (11q14.1), which have been previously linked to breast13,23 or ovarian cancer24. Both the copy number (Fig. 4) and expression outlier landscapes (Fig. 2) suggest at least two separate amplicons at 11q13/14, one at CCND1 (11q13.3) and a separate peak from 11q13.5-11q14.1 spanning UVRAG–GAB2, centred around PAK1, RSF1, C11orf67 and INTS4, where it is more challenging to distinguish the driver24. Notably, the expression outlier profiles for this region are enriched for samples belonging to IntClust 2 (Fig. 2, inset region 23) and all 45 members of this subgroup harboured amplifications of these genes, with high frequencies of amplification also observed for CCND1 (n = 39) and EMSY (n = 34). In light of these observations, the 11q13/14 amplicon may be driven by a cassette of genes rather than a single oncogene.

Second, we note the existence of two subgroups marked by a paucity of copy number and cis-acting alterations. These subgroups cannot be explained by low cellularity tumours (see Methods). One subgroup (IntClust3, n = 156) with low genomic instability (Fig. 4 and Supplementary Fig. 22) was composed predominantly of luminal A cases, and was enriched for histotypes that typically have good prognosis, including invasive lobular and tubular carcinomas. The other subgroup (IntClust 4, n = 167) was also composed of favourable outcome cases, but included both ER-positive and ER-negative cases and varied intrinsic subtypes, and had an essentially flat copy number landscape, hence termed the ‘CNA-devoid’ subgroup. A significant proportion of cases within this subgroup exhibit extensive lymphocytic infiltration (Supplementary Table 45).

Third, several intermediate prognosis groups of predominantly ER-positive cancers were identified, including a 17q23/20q cis-acting luminal B subgroup (IntClust 1, n = 76), an 8p12 cis-acting luminal subgroup (IntClust 6, n = 44), as well as an 8q cis-acting/20q-amplified mixed subgroup (IntClust 9, n = 67). Two luminal A subgroups with similar CNA profiles and favourable outcome were noted. One subgroup is characterized by the classical 1q gain/16q loss (IntClust 8, n = 143), which corresponds to a common translocation event25, and the other lacks the 1q alteration, while maintaining the 16p gain/16q loss with higher frequencies of 8q amplification (IntClust 7, n = 109). We also noted that the majority of basal-like tumours formed a stable, mostly high-genomic instability subgroup (IntClust 10, n = 96). This subgroup had relatively good long-term outcomes (after 5 years), consistent with ref. 26, and characteristic cis-acting alterations (5 loss/8q gain/10p gain/12p gain).

The ERBB2-amplified cancers composed of HER2-enriched (ER-negative) cases and luminal (ER-positive) cases appear as IntClust 5 (n = 94), thus refining the ERBB2 intrinsic subtype by grouping additional patients that might benefit from targeted therapy. Patients in this study were enrolled before the general availability of trastuzumab, and as expected this subgroup exhibits the worst disease-specific survival at both 5 and 15 years and elevated hazard ratios (discovery set: 3.899, 95% confidence interval (2.234–6.804); validation set: 4.447, 95% confidence interval (2.284–8.661)).