Abstract PD1-03: Multivariate Analysis of Subtype and Gene Expression Signatures Predictive of Pathologic Complete Response (pcr) in Triple-Negative Breast Cancer (TNBC): CALGB 40603 (Alliance)

KA Hoadley,T Hyslop,C Fan,DA Berry,O Hahn,SM Tolaney,WM Sikov,CM Perou,LA Carey

DOI: https://doi.org/10.1158/1538-7445.sabcs16-pd1-03

IF: 11.2

2017-01-01

Cancer Research

Abstract:Abstract Support: U10CA180821, U10CA180882 Background: The addition of either carboplatin (Cb) or bevacizumab (Bev) to standard neoadjuvant chemotherapy (NACT) increases pCR rates in TNBC overall and in the dominant subset of basal-like cancers (Sikov et al, JCO 2015; Sikov et al, SABCS 2014). Multigene expression signatures more accurately reflect tumor biology for response prediction and prognosis than individual gene expression. We evaluated the ability of multivariate analysis of gene expression signatures to create predictive models for achievement of pCR in TNBC. Methods: RNA sequencing was successful on 389 pretreatment samples from patients with available pCR data, and used to assign PAM50 subtype and calculate gene signatures scores for 489 published expression signatures. Elastic net, a penalized regression model for high dimensional variable selection, was used to select features associated with pCR in all TNBC and in the basal-like subset. Models were derived in a training set (2/3 of samples) and validated in a separate test set (1/3). A separate model was derived using 196 TNBC samples from patients treated only on the standard NACT +/- Cb arms for application to external TNBC neoadjuvant data sets not treated with Bev. Results: Consistent with our prior partial data set, 343 (88%) of the cancers were classified basal-like, in whom the in breast pCR rate was 54%; the remainder were classified normal-like (n=32) or HER2-enriched (n=14) with a non-basal pCR rate of 56%. Elastic Net analysis in all TNBC generated a model of 23 signatures and treatment assignment with 68% sensitivity and 64% specificity. The area under the curve was 0.64 (p-value=0.0019). Nineteen modules, including immune cell signatures (Th1, NK, IgG), immunoglobulin variable region expression, addition of Cb and Bev and expression of genes at regions 15q25, 17p11.2-13.3, and 8p22 were positively associated with response. The latter two regions are associated with aggressive breast cancer, and while not part of the 17p13 signature, this region contains TP53, a gene important in TNBC. Six modules were associated with resistance, including luminal progenitor, TGFB, NOTCH, FOS/JUN, 8p amplicon, and eosinophil signatures. When limited to basal-like samples, a model including 32 modules and addition of Cb and Bev was generated, with 62.3% sensitivity and 59.1% specificity. Seventeen features were selected in both models. Omitting Bev-treated patients, a model using just the gene expression signatures was developed. The predictive value of this model will be assessed using an external cohort of TNBC patients treated with neoadjuvant docetaxel and Cb (NCT01560663) and results presented. Conclusions: Multivariate analysis of gene expression signatures derived from pretreatment samples enabled the construction of models to predict achievement of pCR in TNBC. These models performed well on our test set, and will be assessed for their predictive ability in other TNBC data sets. If validated by future analyses, this could help us identify patients likely to achieve pCR with standard NACT and may benefit from the addition of agents such as Cb or Bev. ClinicalTrials.gov Identifier: NCT00861705. Citation Format: Hoadley KA, Hyslop T, Fan C, Berry DA, Hahn O, Tolaney SM, Sikov WM, Perou CM, Carey LA. Multivariate analysis of subtype and gene expression signatures predictive of pathologic complete response (pCR) in triple-negative breast cancer (TNBC): CALGB 40603 (Alliance) [abstract]. In: Proceedings of the 2016 San Antonio Breast Cancer Symposium; 2016 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2017;77(4 Suppl):Abstract nr PD1-03.