Identification of Molecular Subtypes of Breast Cancer Using Hierarchical Clustering: Analysis of Inter-Observer Agreement.
Alan Mackay, Britta Weigelt, Anita Grigoriadis, Bas Kreike, Rachael Natrajan, Roger A'Hern, David S Tan, Mitch Dowsett, Alan Ashworth, Jorge S Reis-Filho. Institute of Cancer Research, London, United Kingdom; CRUK London Research Institute, United Kingdom; KCL Breakthrough Breast Cancer Research Unit, London, United Kingdom; Institute for Radiation Oncology, Arnhem, Netherlands
Background: Hierarchical clustering of microarray data using 'intrinsic' gene sets has been employed to classify breast cancers into the molecular subtypes: basal-like, HER2, luminal A, B and C, and normal breast-like. Over the past decade, five different 'intrinsic' gene sets composed of varying numbers of genes have been described for molecular subtype identification of breast cancers by hierarchical cluster analysis. The aim of this study was to determine the agreement between observers in the assignment of breast cancers into the molecular subtypes by hierarchical clustering.
Design: Microarray data from three publicly available breast cancer datasets (total n=779) were subjected to hierarchical cluster analysis using the five distinct 'intrinsic' gene lists. Five observers analysed the dendrograms obtained by hierarchical clustering and classified the breast cancers in each dataset into the molecular subtypes according to the description in each of the five original publications. Inter-observer agreement between the five breast cancer researchers was determined using the free-marginal Kappa score. The results were analysed for the whole classification and for each molecular subtype separately according to each 'intrinsic' gene list for each breast cancer dataset.
Results: The inter-observer agreement was substantial (Kappa scores ≥0.61) in all three datasets when four molecular subtypes had to be identified (i.e. basal-like, HER2, luminal, normal breast-like). With the introduction of the subdivision of luminal cancers into either luminal A, B and C or luminal A and B, none of the classification systems produced substantial agreement between the five observers in all datasets analysed. Analysis of agreement for each subtype separately revealed that only the basal-like and HER2 cancers could be reproducibly identified by independent observers through inspection of hierarchical clustering dendrograms (Kappa scores ≥0.81).
Conclusions: The assignment of breast cancers into the molecular subtype classes based on the interpretation of hierarchical clustering dendrograms is subjective and only shows a modest inter-observer agreement, in particular when luminal cancers are subclassified into two or three groups. For the implementation of a molecular breast cancer taxonomy, objective definitions of each molecular subtype and standardised methods for their identification are essential.
Tuesday, March 1, 2011 8:00 AM
Platform Session: Section C, Tuesday Morning