HER2 is not a cancer subtype but rather a pan-cancer event and is highly enriched in AR-driven breast tumors

Genomic data are from TCGA [13], Metabric [2], and the USO1062 phase III trial [14]. A flow chart and metadata for these three cohorts are available in Additional file 1.

We assessed the performance of three HER2A classification schemes: four or more ploidy-corrected copies of HER2, five or more total copies of HER2, and four or more centromere-corrected copies of HER2 (i.e. ratio of HER2 to chromosome 17 centromere copy number, estimated as the average number of copies for 431 genes on 17p) (Additional file 2). Normal mixture modeling (R package mclust) was used to define a bimodal threshold for HER2 overexpression from RNAseq data (TCGA), Illumina microarray data (Metabric), and Nanostring data (USO1062), and for HER2 and phospho-HER2 protein abundance from RPPA data (TCGA). Voom + limma in R was used for differential gene expression analysis. PAM50 subtype and a measure of chromosomal instability and breakage (CIN) were included in the model when assessing differentially expressed genes between HER2A and non-HER2A breast tumors. CIN is a persistently high rate of loss and gain of chromosomes or chromosomal segments, and can confound differential gene expression when its prevalence is different between HER2A and non-HER2A tumors. We calculated CIN as the total number of segments on the autosomal chromosomes with distinct copy levels. ER expression, PR expression and proliferation score as calculated by the PAM50 algorithm were included in the model when assessing genes differentially expressed between HER2E and non-HER2E breast tumors. Gene set enrichment analysis was done using the camera function in the Bioconductor package limma. Camera is a gene set enrichment test that accounts for correlation between genes that belong to the same gene set [21]. We set the inter-gene correlation value to 0.05, to obtain fewer significant hits that are more biologically interpretable. We assessed the enrichment of the C2 collection from the Molecular Signature Database (MSigDB) [22]. Pathways from the Kyoto Encyclopedia of Genes and Genomes (KEGG) were not considered due to licensing restrictions. Gene sets were filtered by p value corrected for multiple testing (false discovery rate (FDR)) [23]. The Cox proportional hazards model was used for survival analysis and to generate hazard ratios (HR), using the survival package in R. We censored data for patients who had not had an event of progressive disease or death at the date of their last tumor assessment. Human reference genome hg19 was used in all analyses.

We built a signature of 14 genes induced by androgen or reflective of active AR signaling (positive signature genes, present in at least two out of three MSigDB C2 gene sets: Doane breast cancer classes up, Doane response to androgen up, and Farmer breast cancer cluster 7), and 31 genes suppressed by androgen (negative signature genes, from gene set Doane breast cancer classes down). We did not include genes located on chromosome 17, to not confound the detection of AR-driven tumors with HER2 amplification. The expression of each signature gene was z-score-normalized across the ER- tumors per cohort (i.e. normalized to a mean of 0 and standard deviation of 1). The AR-ness score for an ER- breast tumor was then defined as the average z-scored expression of 14 positive signature genes, minus the average z-scored expression of 31 negative signature genes. Out of 45 signature genes, 42 were available for Metabric. Only nine signature genes were present on the Nanostring platform used for USO1062. These clinical trial tumors were therefore not scored for AR-ness.

We associated HER2 amplification with subtype, for those cancers with well-established subtypes. For gastric, bladder, ovarian and head and neck cancer, we used molecular subtypes derived by TCGA, based on either expression data (bladder, ovary, head and neck) or a diverse panel of molecular data (gastric) [24‐27]. For colon cancer, we used the consensus subtypes derived from six independent classification systems [28]. The subtype calls (CMS final network and random forest classifier for non-consensus samples) provided by Guinney and colleagues were used [28].

We defined activating HER2 mutations as those with increased tyrosine kinase activity and cellular signaling, that increase cellular transformation and tumor formation, and/or sensitize tumor cells to HER2-targeted therapies in at least one of the referenced studies [29‐31]. This set consists of G309A, S310F, L755S, D769H, D769Y, V777L, V842I, and T862A. Other HER2 mutations include those with no functional effect (e.g. R678Q) or that have not been tested (e.g. L755W).

We defined HER2-amplified (HER2A) tumors as having a ploidy-corrected copy number for HER2 ≥ 4 (i.e. ratio of copy number to ploidy ≥2). Ploidy corrects for the extensive background amplifications seen in breast tumors. This threshold maximized concordance with HER2 over-expression, clinical HER2 status, HER2 protein abundance, and phosphorylated HER2 protein abundance in TCGA and Metabric (see Additional file 2). This definition covers 12.3% (106/864) of TCGA breast tumors (Fig. 1a), and 12% (133/1107) of Metabric tumors (Fig. 1b). Ploidy-corrected HER2A status is 96–98% concordant with HER2 overexpression in the two cohorts, and improves precision by 9–36% compared to two other SNP6-based measures of HER2 amplification (five or more total copies of HER2, or ratio of HER2 to chromosome 17 centromere copy number ≥2) (Additional file 2, Additional file 3A, D-F). Among TCGA HER2A tumors, 78% have elevated HER2 protein levels, compared to 49–70% with alternative measures of HER2A, and 71% have elevated phosphorylated HER2, indicative of activation, in contrast to 43–63% (Additional file 3B-D). Ploidy-corrected HER2A status is also concordant with clinical measures of HER2 copy number and protein levels (HER2+) in both cohorts: 90% of HER2A tumors are clinically HER2+, while only 4.8% of non-HER2A tumors are HER2+ (Additional file 3D, F). Alternative HER2A measures predict only 57–85% of HER2+ tumors (Additional file 3D, F, H). For the USO1062 trial, we considered the 8% (79/987) of tumors with total HER2 copy number ≥5 as HER2A due to unavailability of tumor ploidy (Fig. 1c, Additional file 3H).

Concordance between HER2A and the PAM50 HER2E subtype was remarkably weak: only 47% of HER2A tumors are HER2E while 18% are luminal A, 24% luminal B, and 11% basal-like across the three cohorts combined (Table 1, Additional file 3D, F, H). This genomically confirms the prior observation that half of clinical HER2+ tumors fall in the HER2E subtype, while the rest are observed predominantly in the luminal subtypes [13]. The non-HER2E tumors that are HER2A are very clearly classified as luminal or basal-like by PAM50 despite HER2 amplification, with their PAM50 subtype scores comparable to those of non-amplified tumors (Additional file 4A-B). HER2A is thus a genomic event found across all PAM50 subtypes, while the HER2E subtype, of which 46% are non-HER2A tumors, may be driven by additional factors. Given this strong discrepancy, we set out to understand the genomic correlates of HER2 amplification across all subtypes, and to understand the additional factors beyond HER2A driving the HER2E subtype.

We assessed the genomic correlates of HER2 amplification by comparing amplified and non-amplified tumors across all subtypes. We focused on mutations and copy number alterations in 43 genes previously shown by TCGA to be frequently altered in breast cancer [13] (Additional file 5). The mutation and copy number profile of HER2A tumors largely reflect those of the underlying subtype, rather than those driven by HER2 amplification (Additional file 4C). Only three mutations and copy number alterations have a significant association with HER2A status in an individual subtype: PIK3CA mutations are more prevalent in basal-like HER2A compared to non-HER2A tumors (40.9% (n = 9) vs. 10% (24)); GATA binding protein 3 (GATA3) mutations are more prevalent in HER2E tumors without HER2 amplification (1.8% (2) vs. 13.3% (11)); and BRCA1 deletions are more prevalent in HER2E tumors with HER2 amplification (21.8% (n = 26) vs. 4.5% (4)) (Additional file 4C). HER2A also shows no evidence of being a transcriptional subtype. Only 43 protein-coding genes are differentially expressed between HER2A and non-HER2A breast tumors, when accounting for PAM50 subtype and chromosomal instability (Additional file 4D, Additional file 6A). Twenty-nine of these are neighbors of HER2 on 17q12-21 and can be explained by co-amplification. Expression of genes outside of this region is moderately impacted (<3 times altered in any cohort), with the exception of two secretoglobins: mammaglobin A (SCGB2A2) and lipophilin B (SCGB1D2). These are chromosomal neighbors on 11q13, form a covalent complex [32], and are 4–8 times more highly expressed in HER2A compared to non-HER2A TCGA tumors (Additional file 4E). Other reported HER2 target genes did not validate when correcting for subtype and chromosomal instability [33, 34]. HER2 amplification thus shows minor association with transcriptional changes outside 17q12-21, consistent with previous findings based on clinical HER2+ status [7, 15]. Taken together, we found that HER2 amplification is a discrete event on top of a luminal or basal transcriptional and mutational state.

While HER2E tumors are believed to be HER2-driven, only half of HER2E tumors are HER2A. This brought into question whether HER2E is a consistent subtype. We set out to better understand the molecular composition of the HER2E subtype, other than increased transcription of HER2 and GRB7, the two amplicon genes on the PAM50 panel. We performed gene set enrichment analysis between HER2E and non-HER2E TCGA tumors, accounting for ER expression, PR expression and PAM50 proliferation score, and omitting genes on chromosome 17 to reduce interference with HER2 amplification (Additional file 7A-B). The most significant gene set enriched in HER2E tumors (Doane breast cancer classes up; Additional file 8A) is composed of genes upregulated in a subset of ER-/PR- tumors but that, paradoxically, are direct targets of ER, responsive to estrogen, and/or typically expressed in ER+ breast cancer [35]. Genes downregulated in those ER-/PR- tumors are concordantly lower expressed in the HER2E tumors (hit #7; Additional file 8A).

The answer to this puzzling expression pattern may be in androgen receptor (AR), another steroid hormone receptor that is highly expressed in some breast tumors, has overlapping target genes with ER, and is on average eight times more highly expressed in HER2E compared to non-HER2E tumors (Additional file 7A). Indeed, MDA-MB-453, a TNBC cell line that expresses those paradoxical genes and lacks typical basal-like cytokeratins, was shown to respond to androgen in an AR-dependent and ER-independent manner, and its expression profile is, at least in part, AR-induced [35]. This was confirmed independently in vivo for AR antagonist bicalutamide [36]. Additional evidence indicative of androgen signaling in HER2E tumors is the enrichment of gene set Farmer breast cancer cluster 7 (hit #6; Additional file 8A). These genes were found to be highly expressed in breast tumors considered molecular apocrine, based on their active AR signaling (i.e., expression of genes induced by androgen in LNCaP prostate cancer cells), weak ER signaling, and morphological hallmarks of apocrine tumors such as abundant eosinophilic cytoplasm and prominent nucleoli [6]. Beyond these sets, three additional gene sets in the top 10 (hits #3, 4 and 9; Additional file 8A) indicate that HER2E tumors display a transcriptional profile that is more similar to ER+ breast tumors than to the basal subtype, despite the shared ER- status [6, 35, 37]. This is supported by expression of luminal cytokeratins (KRT7, KRT8, KRT18), luminal markers FOXA1 and XBP1, and lack of basal-like cytokeratins such as KRT5, KRT6A, and KRT81 (Additional file 7A). Furthermore, HER2E tumors not only selectively express genes that were previously observed in ER-/AR+ or molecular apocrine breast tumors, but also genes induced by androgen [35] (Additional file 7B). The minimal gene overlap between these sets increases our confidence that they independently support overlap between HER2E and androgen signaling (Additional file 8B). These gene set enrichment results suggest that AR regulates the transcriptional program of HER2E tumors.

Other subtypes besides HER2E may also contain AR-driven tumors. We therefore derived a signature of AR-ness that is agnostic to subtype. We selected 14 genes included in at least two of three gene sets reflective of active AR signaling (hits #1, 6 and 69; Additional file 7B), and 31 AR-repressed genes (hit #7). This resulted in a 45-gene AR-ness signature (Additional file 7C). Because AR is co-expressed with ER in up to 90% of ER+ breast cancer [38] and AR can recapitulate the ER-mediated transcriptional program seen in luminal breast cancers [39], we applied the AR-ness signature to tumors that are ER- by IHC, to identify ER- tumors with apocrine features, active AR signaling, and/or expressing androgen-induced genes. As shown in Fig. 1d and Additional file 8C, basal-like tumors score in general low for the AR-ness signature, that is, androgen-induced genes are on average lower expressed than genes reflective of inactive AR signaling (TCGA: 116/126, 92%; Metabric: 100/121, 83%). HER2E ER- tumors on the other hand score high for AR-ness (TCGA: 39/39, 100%; Metabric: 62/67, 93%). In the context of the IntClusts, most of the AR-driven Metabric tumors reside in IntClust4-, which consists of ER- tumors with favorable outcome, and in IntClust5, which captures most of HER2-amplified cancers regardless of ER status [4] (Fig. 1e). ER- TCGA tumors with a positive AR-ness score have concordant higher abundance of AR protein levels (t test, p = 3e-7; Additional file 8D). Across subtypes, we found that 34% (61/178) of ER- TCGA tumors and 48% (98/205) of ER- Metabric tumors have a positive AR-ness score. These prevalence rates are similar to those observed by others [6, 35, 36].

Around half of AR-driven ER- tumors are HER2A (TCGA 46%, Metabric 58%). Subdivision by subtype revealed that two thirds of HER2E tumors are HER2A, and AR-ness score is consistent across HER2A and non-HER2A tumors of this subtype, showcasing that HER2E is a consistent subtype independent of HER2A status (Fig. 1d, Additional file 8C). Results for basal-like tumors differed by cohort. In TCGA with only 10 AR-driven basal-like tumors, AR-ness scores did not differ by HER2A status (Fig. 1d). In Metabric, AR-ness scores were significantly higher in HER2A basal-like tumors (t test, p = 3.2e-5), and concordantly the 21 AR-driven basal-like tumors were enriched for HER2A (52% vs. 3% of AR-inactive tumors; Fisher’s exact test, p = 8e-8) (Additional file 8C). These findings suggest crosstalk between HER2 amplification and AR activity in certain contexts.

The finding that ER- and/or HER2A tumors are frequently AR-driven suggests that those tumors may benefit from treatment regimens including AR antagonists. TN tumors with a luminal AR-driven (LAR) profile, defined as expressing AR and downstream AR targets and co-activators [36], have been shown to benefit less from standard neoadjuvant chemotherapy compared to other TN breast tumors [40]. Eight out of nine LAR-specific genes are significantly higher expressed in HER2E tumors, independent of HER2A status (Additional file 7A). Our signature thus not only identifies TN LAR tumors, covering basal-like, HER2E and IntClust4-, but also HER2A tumors. While AR antagonists may be beneficial for this subset of TN and HER2A tumors, this may not be the case for luminal tumors. The luminal A and B tumors with high AR-ness score but that were ER- by IHC had ESR1 expression levels comparable to ER+ luminal tumors, suggesting that endocrine therapy could suffice.

To support this hypothesis, we assessed prognosis in TCGA and Metabric. In TCGA, HER2E ER+ tumors were associated with the worst and luminal A tumors with the best overall survival, and this was the case in both HER2A and non-HER2A tumors (multivariate Cox proportional hazards model with HER2A status and subtype based on PAM50, ER IHC and AR-ness score; hazard ratio (HR) ER+ luminal A vs. ER+ HER2E, 0.301, 95% CI 0.109–0.833) (Fig. 1f). In Metabric, where women were enrolled before the general availability of HER2-targeted agents, both HER2E and basal tumors were associated with significantly worse survival than luminal A tumors, independent of AR activity and HER2A status (Additional file 8E).

Specifically for ER- tumors, there is a trend in TCGA towards worse survival in patients with ER- tumors with active AR signaling compared to ER- tumors with inactive AR signaling in both HER2A and non-HER2A tumors (multivariate HR 1.853, 95% CI 0.745–4.608) (Fig. 1g). Though not conclusive, this suggests that patients with ER- tumors with active AR signaling do worse on chemotherapy, the current standard of care for TNBC, and could potentially benefit from alternative treatment options including AR antagonists. AR activity was not prognostic in Metabric (Additional file 8F), possibly impacted by lack of exposure of HER2A Metabric tumors to HER2-targeted agents [4].

Next, we explored the extent of amplification near HER2 and its impact on gene expression, to identify putative HER2-cooperating oncogenic drivers. HER2A breast tumors are more chromosomally instable than non-HER2A tumors (Additional file 9A), which in the past led to the hypothesis that HER2 amplification drives the selection of additional copy number aberrations [41, 42]. The HER2 amplicon has no conserved breakpoints, but does have a minimal core of genes amplified in most tumors (Fig. 2a, Additional file 9B-C). Ten genes within the core HER2 amplicon, spanning 237 kb, are amplified to five or more total copies in at least 92% of HER2A TCGA tumors. A broad HER2 amplicon with genes amplified in at least 60% of HER2A TCGA tumors covers 1.14 Mb and 37 genes from LRRC37A11P to CASC3. We confirmed the HER2 amplicon boundaries in the Metabric cohort (Additional file 9C), concordant with previous publications [42, 43]. Four regions on chromosome 17 outside of the broad HER2 amplicon are significantly co-amplified with HER2 in the combined TCGA and Metabric cohorts when taking chromosomal instability into account (Fig. 2b, Table 2). One of these regions is adjacent to the HER2 amplicon. Genes in the other three regions, including 11 cancer genes (as reported [44, 45]), are amplified in 3.4% to 27.6% of HER2A tumors and in a maximum 5.6% of non-HER2A tumors (Additional file 10). Taken together, chromosome 17 and in particular the HER2 amplicon are likely to include HER2-cooperating drivers such as GRB7, STARD3, MIEN1, and LASP1 [43, 46, 47]. Most tumors have focal HER2A amplification on top of a largely diploid 17q, but 19% of TCGA HER2A tumors have arm-level gain, and 4% are defined as HER2A due to 17q amplification without further focal HER2 amplification (Fig. 2b, Additional file 9D).

Outside of chromosome 17, we identified 10 significantly co-amplified regions containing 116 genes (Table 2, Additional file 9E, Additional file 10). These genes, though often co-amplified with HER2, are not more highly expressed in HER2A than non-HER2A tumors (Additional file 6A), and only one of these is a known cancer gene, ubiquitin ligase CCNB1IP1. Genes SCGB1D2 and SCGB2A2 that are differentially expressed by HER2A status (FC >4) are not significantly co-amplified with HER2. We thus find no evidence for co-operating copy number drivers with HER2 outside chromosome 17.

HER2 amplification is prevalent in several other cancers. We found 1.8% of all primary non-breast TCGA tumors to be HER2A (Table 3). An even higher rate of 3.4% was seen in another cohort of ~ 7300 solid tumors, encompassing primary, locally recurrent, and metastatic tumors [48]. To date, anti-HER2 therapies are indicated for the treatment of breast and metastatic gastric cancer [49]. Based on TCGA HER2A prevalence, these two cancers are estimated to annually account for ~ 31000 new cases in the USA (Table 3) [50]. Other cancers may account for another ~ 14750 new HER2A cases. We explored the genomics of these HER2A tumors to better understand similarities and differences to HER2A breast cancer, and potential for therapeutic intervention.

HER2-targeted therapy may be an opportunity for HER2A bladder, endometrial, and ovarian cancer. HER2 expression, protein, and phospho-protein levels are higher in HER2A compared to non-HER2A tumors of these types in addition to gastric cancer (Fig. 3a-c). HER2 transcript levels are also higher in HER2A colon, lung, and head and neck carcinoma, but this does not translate to increased HER2 protein and/or pHER2 levels. Levels of downstream phosphorylation markers for ERBB3, pan-AKT, and ERK1/2 are not affected by HER2 amplification (Additional file 11A-C). HER2 may also carry mutations in the kinase and extracellular domains, some implicated in tumorigenesis [29, 51]. Three percent of breast tumors in both the TCGA and Metabric cohort carry an HER2 protein-altering mutation (Additional file 11D). Prevalence in non-breast cancers varies from 0.4% (ovary) to 8.6% (bladder) (Additional file 11D), consistent with previous studies [48, 51]. These mutations are mutually exclusive with HER2 amplification in 92% of mutated tumors (Additional file 11D). In the absence of HER2 amplification, HER2 mutations do not increase HER2 or pHER2 levels (Fig. 3b-c).

HER2A tumors in these cancers share certain similarities with breast cancer. We observed the same pattern in 17q arm-level gain and amplification (Additional file 12A): 51% (n = 50) of non-breast HER2A tumors have focal HER2 amplification on top of a largely diploid 17q, 48% (n = 47) show additional arm-level gain, and one uterine tumor only has arm-level amplification. As in breast, HER2 amplification is found in multiple transcriptional subtypes of bladder, colon, ovarian, and head and neck cancer (Additional file 12B). Non-breast HER2A tumors show focality of HER2 and closest neighboring genes. The core HER2 amplicon in non-breast cancers (amplified in at least 80% of HER2A) covers 79 kb and six genes, from PNMT to GRB7 (Additional file 12C). The broad pan-cancer HER2 amplicon shared by at least 60% of HER2A tumors spans 532 kb, contains 13 additional genes from CDK12 to PSMD3, and is narrower compared to breast. Genes in the broad amplicon have on average 0.85 fewer copies in non-breast compared to breast HER2A tumors (Fig. 2a, Additional file 12C). Expression levels of amplicon genes also vary by cancer (Fig. 3d). Relative expression of core amplicon genes normalized to levels in HER2-diploid tumors are similar in HER2A breast, bladder, head and neck, endometrial, cervical, ovarian, and colon cancers, and are significantly lower in HER2A gastric, lung, kidney, and uterine tumors (one-sided t test in comparison to HER2A breast, FDR p = 0.013, 0.001, 0.003, and 0.025, respectively). Relative expression levels of broad amplicon genes outside of the core are only lower in lung adenocarcinoma and kidney HER2A tumors in comparison to breast (both FDR p = 0.021) (Fig. 3d).

HER2 amplification does not induce transcriptional changes outside of the amplicon in a coherent pan-cancer manner. Only 14 protein-coding genes, all located on 17q12-21, consistently distinguish pan-cancer HER2A from non-HER2A tumors (Additional file 6B). Concordantly, of the 43 genes differentially expressed by HER2A status in breast cancer (Additional file 6A), only genes on 17q12-21 show consistent higher expression in HER2A non-breast tumors (Additional file 12D). The non-amplicon SCGB targets are not consistently higher in HER2A tumors (Additional file 12E).

We also discovered a set of tumors that express HER2 at levels found in HER2A tumors, but which are not amplified (31 breast tumors, Fig. 1a-c; and 16 other TCGA tumors, Fig. 3d, bottom). These tumors, other than cervix and bladder, consistently upregulate the closest HER2-neighboring genes PGAP3, MIEN1, and GRB7, at levels higher than observed in HER2A tumors (Additional file 13A–H). More distal HER2 neighbors MED1, CDK12, NR1D1, and TOP2A are not overexpressed. Overexpression of HER2 and closest HER2-neighbors in non-HER2A tumors is not driven by amplification of those genes (Additional file 13I). This suggests that there is a regional control of gene expression. We assessed epigenetic changes and found that high expression of HER2 and closest HER2-neighbors in those tumors is associated with CpG hypomethylation in the bodies or maximum 2 kb upstream of these genes (Additional file 13J-M). The two non-HER2A bladder tumors that overexpress HER2 but lack transcriptional regional control (Additional file 13E) concordantly did not undergo reduced methylation (Additional file 13M). HER2-neighboring genes have been suggested to contribute to HER2A cancer [43, 46]. The co-expression supports the model that co-amplified genes near HER2 contribute to oncogenesis. This small pan-cancer population of non-HER2A tumors with substantial overexpression of HER2 and neighboring genes may also benefit from HER2-targeted treatment.