Clinical implementation of the intrinsic subtypes of breast cancer.

Our group has reported several intrinsic gene sets important for identifying subtypes of breast cancer with clinical signifi cance. 1–3 In these studies we have explored and published methods for sample classifi cation across diff erent genomic platforms and tissue qualities. In 2009, we suggested the use of a standardised gene set (PAM50) for subtype classifi cation to improve the classifi cation concordance reported by investigators. 3 However, a standardised gene set does not completely resolve discrepancies between researchers since the genes might be quantitatively measured using diff erent platforms and normalisation methods. Weigelt and co-workers 4 applied three diff erent intrinsic gene sets to four data sets using one prediction method and showed a range of agreement. Because of the level of discordance that was reported, they concluded that identifi cation of the intrinsic subtypes is not ready for clinical implementation. We disagree with this interpretation. Careful examination of Weigelt and co-workers' analyses 4 revealed bioinformatics-based technical limitations that reduced the accuracy in subtype predictions and concordance of these three predictors. These limitations are highlighted in the accompanying letters, but we emphasise here the importance of dataset to dataset normalisation. The webappendix shows the relationship between the four datasets when they are not normalised (as done by Weigelt and co-workers 4) and when row centring and column standardisation is done (as advocated by ourselves and others 1–3,5,6). Additionally, diff erences in the composition of datasets (ie, proportion of oestrogen receptor-positive [ER+] tumours) can aff ect sample classifi cation. 6 Nonetheless, all three gene sets were signifi cant predictors of outcomes in univariate and multivariate testing, which suggests that this is a robust classifi cation method. Many clinical assays begin in the research setting and are refi ned over time until ready for clinical use. Clinical concordance testing rarely has perfect agreement even under the best of circumstances, such as measuring a single analyte with a locked-down protocol across CLIA (Clinical Laboratory Improvement Amendments) laboratories. There is little value, and potential harm, to draw conclusions about the robustness and utility of a test based on research data from independent laboratories not intended for concordance testing, as Weigelt and colleagues 4 and the accompanying commentary 7 interpreted their fi ndings. The interpretation of Weigelt and co-workers 4 is also based on the hypothesis that training sets with diff erent tumours and genes should result in high agreement in subtype classifi cation. …