Breast cancer subtype predictors revisited: from consensus to concordance?

Background: At the molecular level breast cancer comprises a heterogeneous set of subtypes associated with clear differences in gene expression and clinical outcomes. Single sample predictors (SSPs) are built via a two-stage approach consisting of clustering and subtype predictor construction based on the cluster labels of individual cases. SSPs have been criticized because their subtype assignments for the same samples were only moderately concordant (Cohen's κ<0.6).

Methods: We propose a semi-supervised approach where for five datasets, consensus sets were constructed consisting of those samples that were concordantly subtyped by a number of different predictors. Next, nine subtype predictors - three SSPs, three subtype classification models (SCMs) and three novel rule-based predictors based on the St. Gallen surrogate intrinsic subtype definitions (STGs) - were constructed on the five consensus sets and their associated consensus subtype labels. The predictors were validated on a compendium of over 4,000 uniformly preprocessed Affymetrix microarrays. Concordance between subtype predictors was assessed using Cohen's kappa statistic.

Results: In this standardized setup, subtype predictors of the same type (either SCM, SSP, or STG) but with a different gene list and/or consensus training set were associated with almost perfect levels of agreement (median κ>0.8). Interestingly, for a given predictor type a change in consensus set led to higher concordance than a change to another gene list. The more challenging scenario where the predictor type, gene list and training set were all different resulted in predictors with only substantial levels of concordance (median κ=0.74) on independent validation data.

Conclusions: Our results demonstrate that for a given subtype predictor type stringent standardization of the preprocessing stage, combined with carefully devised consensus training sets, leads to predictors that show almost perfect levels of concordance. However, predictors of a different type are only substantially concordant, despite reaching almost perfect levels of concordance on training data.