Automated Classification of Breast Cancer Across the Spectrum of ERBB2 Expression Focusing on Heterogeneous Tumors With Low Human Epidermal Growth Factor Receptor 2 Expression

Abstract

PURPOSE Although pharmaceutical companies conduct clinical trials of novel human epidermal growth factor receptor 2 (HER2)-low–directed drugs, diagnosing HER2-low cancer by immunohistochemistry (IHC) and in situ hybridization (ISH) remains challenging. This study investigates the performance of first-in-kind computerized intelligence to classify samples across gene expression levels and differentiate HER2-low tumors. MATERIALS AND METHODS We classified 251 samples: 142 primary invasive breast cancers (IBCs), 75 ductal carcinomas in situ (DCIS), and 34 mammaplasties (reference) using mRNA expression data from the QuantiGene Plex 2.0 assay. We used g3mclass probabilistic software to assess the number of classes in the assay data, the mean and the variance in each class, diagnostic cutoffs, and the prevalence of each class in the study population. RESULTS HER2-low (IHC score of 1+ or 2+/ISH–) accounted for 31% of IBC. First, we discovered that HER2-low tumors were represented by cases with normal ERBB2 transcript levels that were expected to produce physiologic levels of HER2 (70%) and cases with abnormally upregulated unamplified ERBB2 (30%). We termed the latter cancers ERBB2-up as they do not meet the standard definitions for ERBB2 overexpression and amplification. Second, HER2-low IBC classified as ERBB2-up had not only abnormally increased luminal growth and adhesion markers ( ERBB2, ESR1, PGR, IGF1R, VAV2, VAV3, KRT8, CDH1) but also downregulated myoepithelial marker ( KRT5). The vascularization ( RAP1 and C3G), immune cell infiltration ( VAV1), and mesenchymal transition ( CDH2) markers were dysregulated. Finally, in the independent cohort of DCIS, 40% of HER2-low DCIS shared similar traits with HER2-low IBC except for rare downregulation of KRT5 and no change in C3G, VAV1, and CDH2. CONCLUSION We demonstrated how innovative bioinformatic tools could help diagnose cancer across the spectrum of ERBB2 expression to aid decision making for HER2-low.