Targetable ERBB2 mutation status is an independent marker of adverse prognosis in estrogen receptor positive, ERBB2 non-amplified primary lobular breast carcinoma: a retrospective in silico analysis of public datasets

Genomic and clinical outcome data associated with tumor samples from patients with primary breast cancer in TCGA 2015 (N = 817), METABRIC 2012 and 2016 (N = 2509), and MSK-IMPACT 2018 (N = 918) were accessed online via CBioportal [25‐27]. From these datasets, ER+ and HER2− cases of stage I–III ILC and IDC with both clinical outcome and mutational data called from next-generation sequencing (NGS) analyses were selected (N = 1580). Cases of mixed or non-ILC/IDC histology, ER-negative/undetermined, HER2+/undetermined, carcinoma in situ, and stage 4/undetermined were excluded. For TCGA and MSK datasets, HER2 status was determined by IHC (positive/negative) or, where IHC was indeterminate, by ISH assessment of ERBB2 amplification, in line with standard clinical practice [11]. For METABRIC cases, HER2 status was determined using the Affymetrix SNP6 copy number inference pipeline.

The primary outcome measure, available in all datasets, was OS. Variables included ERBB2 and ERBB3 mutational status. ERBB2mut status was subcategorized as oncogenic or uncharacterized by cross-reference with existing data that identified mutations targetable by HER2-inhibition: G309A/E, S310F, L755S, del755-759, S760A, D769H, D769Y, V777L, P780ins, V842I, and R896C [13]. Cases were denominated oncERBB2mut if tumors harbored at least one oncogenic ERBB2 mutation.

Clinical and NGS mutation data were integrated with clinicopathological features including histological subtype, lymph node (LN) status, and tumor size and grade. Normalized gene expression data were publicly available for METABRIC (Illumina HT12 microarray) and TCGA (RNA-seq) datasets.

For analysis of binary somatic mutation status, combined cohort analysis was performed. Enrichment for cases with mutations in candidate genes (ERBB2/3mut) by histological subtype (ILC vs. IDC) was determined by χ² test for association between categorical variables. For ILC and IDC separately, Kaplan-Meier (KM) survival curves stratified by mutation status were compared using logrank and generalized Wilcoxon tests. Multivariate analysis of OS was performed using a Cox regression model. Covariates included classic prognostic markers tumor grade, size (< 20 mm or ≥ 20 mm), and LN status (positive or negative). Adjustment was made for age at diagnosis (< 50 or ≥ 50 years). Tumor grade was classified as low (grade 1–2) or high (grade 3).

To derive a novel gene expression signature of HER2 activity that accounted for the effect of potentially targetable ERBB2mut in ERBB2 non-amplified tumors, we applied a weighted average difference (WAD) method to gene expression data in cases from the METABRIC 2012 (N = 1980) and TCGA 2015 (N = 817) cohorts [25, 28, 29]. Gene expression in ERBB2mut cases (N = 38, selected by ERBB2 non-amplified status and patient age > 50) was compared with the same order of magnitude of ERBB2 wild-type cases (N = 79, selected by ERBB2 non-amplified status, grade > 1, stage > I, patient age > 50). This was repeated for oncERBB2mut cases (N = 23) using the same comparator and selection criteria. To incorporate the effect of HER2 activity via ERBB2 amplification, the overlap of differentially expressed genes DEGs shared by both comparisons (ERBB2mut and oncERBB2mut vs. ERBB2 wild-type) with DEGs from a further comparison of ERBB2 amplified (N = 247) vs. non-amplified (N = 1733) cases in METABRIC was calculated. Finally, to incorporate the downstream phenotype (HER2 status), the overlap of this list with DEGs from a comparison of clinical HER2+ vs. HER2− cases in TCGA was calculated. In contrast to ERBB2mut cases, matching was not performed for ERBB2 amplified or HER2+ cases because numbers were higher and within an order of magnitude across groups, such that similar variation in gene expression could reasonably be assumed.

Multiple gene expression signatures of HER2 activity have been derived using cell line models and patient tumors [30‐33]. We compared our novel gene signature with the HER2 activity signature established by Desmedt et al [31] with respect to its ability to detect potentially targetable ERBB2mut cases in our ILC/IDC dataset. This was achieved by multivariate regression modeling of response to neratinib for each gene signature using breast cancer cell line pharmacogenomic data from the BROAD Institute, accessed online via the CellMinerCDB portal [34, 35]. The Pearson coefficient for each significantly correlated signature gene was used to calculate normalized signature scores for each METABRIC case in the current dataset. To validate the prognostic effect of ERBB2 in ILC, cases were then stratified by gene signature score (upper vs. lower quartiles) and the signatures compared by histological subtype using a Cox regression model of a 10-year OS.

All cases of primary (stage I–III) ER+ and HER2− ILC and IDC from METABRIC, TCGA, and MSK-IMPACT cohorts were selected (total N = 1580, see Table 1). Baseline clinicopathological characteristics by individual cohort are summarized in Table 1.

To detect prognostic effects of low-frequency mutations and add to the existing body of literature on ILC-specific mutational drivers, a combined cohort of three public datasets was collated. We first evaluated the implications of combining cases of ILC and IDC from potentially disparate datasets.

In the combined cohort, long-term follow-up of patients is dominated by the largest dataset (METABRIC, N = 702: mean follow-up 133 months) while early events are enriched by the two smaller datasets (TCGA and MSK, N = 878: mean follow-up 33 months) (see Supplemental Figure S1(A) in Additional file 1 for KM plot). Compared to the smaller datasets with shorter follow-up (TCGA and MSK), patients in METABRIC were more likely to be over 50 years of age and have T1 tumors (< 20 mm diameter), but no significant difference in grade or LN status was found (see Table 2). As the principle skew in the combined dataset is towards longer follow-up in METABRIC cases, OS analysis was limited to 10 years—thus providing a clinically meaningful endpoint for all patients, irrespective of age. Comparison of METABRIC with combined TCGA and MSK-IMPACT cohorts revealed no significant difference in a 10-year OS (logrank p = 0.225, see Supplemental Figure S1(B) in Additional file 1).

We next assessed the prevalence of ERBB2/3mut in our dataset of ILC and IDC (N = 1580). Overall prevalence of ERBB2mut was 2.2% (N = 34). ERBB2mut was enriched in ILC, with prevalence of 5.7% (N = 16) vs. 1.4% in IDC (N = 18) (p < 0.0001). In contrast, prevalence of ERBB3mut was lower (1.6% overall, N = 26), and there was no enrichment of ERBB3mut in ILC vs. IDC (1.1%, N = 3 vs. 1.8%, N = 23; p = 0.604). Due to the small number of cases in ILC, further analysis of ERBB3mut was not performed.

In ILC, ERBB2mut clustered in the tyrosine kinase domain of HER2 (15 out of 16 cases; 93.8%). Of these, the majority have been characterized as oncogenic and potentially targetable using neratinib (oncERBB2mut, 11 out of 15 kinase domain mutations in ILC; 73.3%). In IDC, all kinase domain mutations were known oncogenic (N = 8) and non-characterized mutations were distributed evenly across protein domains (see Fig. 1). The most frequently occurring ERBB3mut coded for E928G, which lies in the tyrosine kinase domain of HER3 (N = 7 out of 26; 27%).

Using our meta-cohort of 1580 cases, we next assessed ERBB2mut as a prognostic marker of OS in ILC (N = 279) and IDC (N = 1301). Median duration of patient follow-up was 50 months (range 0–351 months). In patients with ILC, median OS was significantly shorter if tumors were ERBB2mut positive (inclusive of oncogenic and uncharacterized mutations) vs. ERBB2mut negative (66 vs. 211 months, p = 0.0001). In contrast, there was no significant difference in OS for ERBB2mut cases of IDC (159 vs. 166 months, p = 0.733) (see Fig. 2).

ERBB3mut status was not a significant prognostic indicator of OS in cases of IDC (HR = 1.46, 95% CI 0.65–3.28; p = 0.359). There were no events (patient deaths) at 10 years of follow-up in the three cases of ERBB3mut ILC.

To test the effect of therapeutically actionable ERBB2 mutations on a clinically relevant endpoint, we stratified 10-year OS by oncERBB2mut status. In ER+, ERBB2 non-amplified ILC, oncERBB2mut status was prognostic of 10-year OS independently of LN status, tumor grade, and size (HR 3.65, 95% CI 1.21–11.00; p = 0.021, see Fig. 3). The independent prognostic value of oncERBBmut status on 10-year OS was retained after further adjusting for age at diagnosis (HR 3.19, 95% CI 1.06–9.62; p = 0.039).

Unselected ERBB2mut status (including oncogenic and uncharacterized mutations) was also adversely prognostic of 10-year OS in univariate analysis (HR 3.66, 95% CI 1.54–8.73; p = 0.003), but in contrast to characterized oncERBB2mut, this prognostic effect was not independent of LN stage, tumor grade, and/or size. The independent prognostic value of oncERBB2mut was further verified at 5 years of follow-up (HR 3.668, 95% CI 1.096–12.275; p = 0.035).

Since existing gene signatures of HER activity derive from HER2 status (by IHC and/or ISH), we generated a novel gene signature incorporating DEGs in ERBB2mut and oncERBB2mut cases, as described in the “Patients and methods” section and outlined in Fig. 4. Our aim was to establish a novel signature that reflects HER2 pathway activity more completely than existing signatures, whether induced by ERBB2 mutations or amplifications, and to be able to apply it in a wider range of patients. A list of up and downregulated ERBB2 “mutant” DEGs (N = 20) was generated by combining the overlap between DEGs for METABRIC ERBB2 amplified vs. non-amplified, ERBB2 mutated vs. wild-type (ERBB2mut and oncERBB2mut separately), and TCGA HER2+ vs. HER2− (Fig. 4a, and see Additional file 2 for Supplemental Tables S1–3 with all gene lists).

As shown in Fig. 4b, ERBB2mut and oncERBB2mut clustered in the upper quartile of the novel gene signature score in cases of ER+, HER2− ILC/IDC from METABRIC. In contrast, ERBB2mut and oncERBB2mut cases did not cluster when scored by the established gene signature of HER2 pathway activation but were evenly distributed across the cohort. Both the novel and established HER2 pathway gene signatures were predictive of response to neratinib in multivariate analysis using a pharmacogenomic breast cancer cell line model (Pearson R = 0.9, p < 1e⁻¹⁴ for both; see Supplemental Figure S2 in Additional file 3 for the original plots from the CellMinerCDB online portal). The regression model used incorporates drug response and gene expression data from N = 36 breast cancer cell lines [35, 36]. Our findings using this model verify that the novel signature derived from DEGs in ERBB2mut vs. wild-type cases incorporates oncogenic HER2 activity due to ERBB2mut. This suggests that a gene signature may have value in predicting targetable ERBB2 alterations including ERBB2mut in HER2− breast cancer. Finally, Fig. 4c demonstrates that the novel gene signature score (stratified into upper vs. lower quartiles) was adversely associated with 10-year OS (HR = 2.3, 95% CI 1.04–5.05; p = 0.040), thus providing mRNA level validation of ERBB2mut as an adverse prognostic marker.

Using the same dataset (CellMinerCDB), it was not possible to associate ERBB2mut status with neratinib response as only one breast cancer cell line (MDA-MB-468) was ERBB2mut positive. Response to neratinib for MDA-MB-468 was in the upper tertile of breast cancer cell lines (data in the public domain via CellMinerCDB online portal).