Abstract P4-09-01: Retrospective evaluation of precision of gene-expression-based signatures of prognosis and tumor biology in replicate surgical biospecimens from patients with breast cancer
Barry, WT; Marcom, PK; Geradts, J; Datto, MB
Published in: Cancer Research
Background: Numerous gene-expression signatures have been developed for breast cancer. However, assessments of their validity continue to largely ignore the impact of intratumor heterogeneity and technical variation on clinical utility. Here, the collection of replicate specimen in Barry et al (2010) is used to evaluate a broad collection of gene-expression signatures of prognosis and tumor biology.Methods: Eighteen patients with multiple frozen cores (one patient with quadruplicate, twelve with triplicate, and five with doublet samples) were previously identified in the Duke Breast SPORE tissue repository. Cores were assessed for percent invasive cancer cellularity, and tumor size, grade, and ER/PR status. RNA was extracted and hybridized to AffymetrixH133Plus2.0 microarrays. Expression signatures of prognosis and tumor biology were developed or translated to the Affymetrix platform using routines by Prat et al. (2012) and Haibe-Keins et al. (2012). Association between signatures, and precision among replicates are evaluated using Pearson and intraclass correlation. Coefficients are reported with 95% confidence intervals.Results: Among prognostic signatures, an imputed 70-gene signature of MammaPrint had the highest level of precision (ICC = 0.96, 0.91–0.98); strong concordance is seen with Affymetrix-based PAM50 risk-of-relapse (ICC = 0.82, 0.65–0.92); 16 of 18 patients had constant subtype calls. Substantially lower concordance was seen in the GENIUS prognostic model (ICC = 0.62, 0.35–0.82). In ER+ patients (n = 13), two algorithms to impute the 21-gene model of OncotypeDX recurrence score were concordant (r = 0.98, ICC= 0.90 and 0.83) and distinct from the GENIUS ER+ score (average r = 0.74, ICC=0.78). Imputed models for the Rotterdam 76-gene model (+ER status), GGI (+grade), and ROR-T (+tumor size) showed lower levels of correlation (0.47<rho<0.64), suggesting independent prognostic information is conveyed by the unique combinations of clinical and genomic factors. High correlation is seen between models of ER status ranging from single-gene data (ICC = 0.90) to signatures from archived specimen (ICC = 0.96) and cell lines (ICC = 0.85), with almost complete agreement to IHC results. Two signatures of PI3K from specimen with mutation data were partially correlated and concordant (r = 0.52, ICC= 0.96 and 0.91), while two signatures from overexpression systems were uncorrelated, and one with poor precision (r = 0.12, ICC= 0.92 and 0.46). Signatures of PR, EGFR and p53 status will also be reported.Conclusions: The number of genomic signatures in breast cancer from archived specimen and cell-line experiments continues to grow, but there are limited resources for validating their prognostic/predictive value in patient populations. Reproducibility across biological replicates is a critical component in establishing clinical utility of a signature that is distinct from using technical replicates for the repeatability of analytes on the array platform. We demonstrate how archived specimen can confirm reproducibility in the ‘Test Validation Phase’ of biomarker development, as advocated by the Institute of Medicine, and inform trial designs to prospectively test clinical utility.Citation Information: Cancer Res 2012;72(24 Suppl):Abstract nr P4-09-01.
