Relapse

All statistical analyses were carried out in SPSS version 16.0 (SPSS, Inc., Chicago, IL) and R version 2.6.0 (www.r-project.org). For breast cancer subtype prediction with qRT-PCR data, the 50-gene PAM50 classifier, as described in detail by Parker et al. (25 (link)), was used to assign breast cancer subtypes to the 357 training samples. The algorithm maps the gene expression in each specimen to centroids (a multidimensional average expression of the 50 discriminatory genes) that were previously constructed from prototypical examples of the five breast cancer subtypes (luminal A, luminal B, HER2-enriched, basal-like, and normal breast-like) (25 (link)). We assigned a subtype to each tumor specimen tested by calculating the distances to each of the subtype centroids with the Spearman rank correlation test. Tumors for which the difference between Spearman rank correlation coefficients for the luminal A and B centroids was less than 0.1 were considered borderline.
Expression of ER and ER-associated genes is a characteristic of luminal breast cancers as defined by microarray expression profiling (1 (link)–3 (link)). Approximately 30% of tumors in the luminal B cluster expressed HER2 and associated genes, and in this study, we defined tumors that expressed hormone receptor proteins (ER or PR) and were positive for HER2 as being of the luminal–HER2-positive subtype. However, the remaining 70% of luminal B tumors primarily differed from better prognosis luminal A tumors by virtue of higher expression of proliferation genes (15 (link),18 (link)). We investigated whether addition of the proliferation marker Ki67 to the immunopanel of ER, PR, and HER2 could distinguish luminal B tumors (ie, hormone receptor–positive, HER2-negative, and Ki67-high tumors) from luminal A tumors (ie, hormone receptor–positive, HER2-negative, and Ki67-low tumors). In the UBC-WashU series, we used the 50-gene PAM50 classifier to identify tumors as being either luminal A or luminal B, to determine the optimal cut-point value for the Ki67 index. We then compared quantitative data from visual assessment of Ki67 immunohistochemical labeling against these gene expression profile–based assignments for hormone receptor–positive, HER2-negative tumors. The optimal cutoff value for Ki67 was selected by use of the receiver operating characteristic (ROC) method, by minimizing the sum of the observed false-positive and false-negative errors with bootstrapping methodology (33 ). In this fashion, the cutoff value was selected against the gold standard of gene expression profiling, as opposed to assigning a cut point against clinical outcome (which can be difficult to extrapolate to other patient populations with differences in treatment and risk).
The immunopanel thereby defined (ie, ER, PR, HER2, and Ki67) was used to assign tumors of the BCCA validation series to breast cancer subtypes and to assess clinicopathological characteristics and the relation to patient outcome. We estimated 95% confidence intervals (CIs) with bootstrapping methodology (32 (link)) for the reported percentages of luminal subtypes as defined by the immunopanel. Differences in clinicopathological characteristics, including age, tumor grade, tumor size, and lymph node status, among breast cancer subtypes were examined by use of χ² tests. For univariate survival analysis, relapse-free survival and breast cancer–specific survival were estimated by use of Kaplan–Meier curves (34 (link)), and the statistical significance of survival differences was assessed with a log-rank test (35 (link)). For relapse-free survival, survival time was censored at death if the cause was not breast cancer or if the patient was alive without relapse on June 30, 2004. For breast cancer–specific survival, survival time was censored at death if the cause was not breast cancer or if the patient was still alive on June 30, 2004 (the date for outcome data collection). Patients with unknown cause of death were excluded from breast cancer–specific survival analysis. For multivariable survival analyses, Cox regression models (36 ) were used to estimate the association between the Ki67 index and breast cancer subtypes, with adjustment for with standard clinicopathological variables, including age at diagnosis (as a continuous variable), histological grade (grade 3 vs grade 2 or 1), tumor size (>2 vs ≥2 cm), lymphovascular invasion (positive vs negative), and number of positive axillary lymph nodes as a percentage of the total numbers examined (coded in three categories, in which 0%–25% was compared with 0%, and >25% was compared with 0%). We classified patients with breast cancer in the British Columbia population by using the percentage of positive lymph nodes as a continuous variable in the Cox model because this variable was shown to be more prognostic than a categorical variable of one to three positive lymph nodes vs four or more than positive lymph nodes (37). Only patients with information for all the covariates were included in the Cox regression analyses. Smoothed plots of weighted Schoenfeld residuals were used to test proportional hazard assumptions (38 ), and no evidence that these assumptions were invalid was observed. All statistical tests were two-sided, and P values of less than .05 were considered statistically significant. The data were assembled to provide more than 80% power for testing hypotheses regarding the biomarkers in all patients combined and for patient subgroups that were defined by the adjuvant therapies received.