Concordance of immunohistochemistry based and gene expression-based subtyping in breast cancer

Background: Use of immunohistochemistry-based surrogates of molecular breast cancer subtypes is common in research and clinical practice, but information on their comparative validity and prognostic capacity is scarce. Methods: Data from 2 PAM50-subtyped Swedish breast cancer cohorts were used: Stockholm tamoxifen trial-3 with 561 patients diagnosed 19761990 and Clinseq with 237 patients diagnosed 2005-2012. We evaluated 3 surrogate classifications; the immunohistochemistry-3 surrogate classifier based on estrogen receptor, progesterone receptor, and HER2 and the St. Gallen and Prolif surrogate classifiers also including Ki-67. Accuracy, kappa, sensitivity, and specificity were computed as compared with PAM50. Alluvial diagrams of misclassification patterns were plotted. Distant recurrence-free survival was assessed using Kaplan-Meier plots, and tamoxifen treatment benefit for luminal subtypes was modeled using flexible parametric survival models. Results: The concordance with PAM50 ranged from poor to moderate (kappa = 0.36-0.57, accuracy = 0.54-0.75), with best performance for the Prolif surrogate classification in both cohorts. Good concordance was only achieved when luminal subgroups were collapsed (kappa = 0.71-0.69, accuracy = 0.90-0.91). The St. Gallen surrogate classification misclassified luminal A into luminal B; the reverse pattern was seen with the others. In distant recurrence-free survival, surrogates were more similar to each other than PAM50. The difference in tamoxifen treatment benefit between luminal A and B for PAM50 was not replicated with any surrogate classifier. Conclusions: All surrogate classifiers had limited ability to distinguish between PAM50 luminal A and B, but patterns of misclassifications differed. PAM50 subtyping appeared to yield larger separation of survival between luminal subtypes than any of the surrogate classifications.