Molecular profiling has been enthusiastically received as an exciting new technology by the scientific community. However, there were also some calls for caution which criticized the over-enthusiastic and uncritical embracing of a new method still lacking rigorous validation [44
]. First of all, the new molecular taxonomy was based on findings derived from different microarray platforms. To address the question of reproducibility of microarray technology, the US Food and Drug Administration (FDA) instigated the Microarray Quality Control (MAQC) project. The data collected in this extensive study showed that microarray measurements are highly reproducible within and across different microarray platforms. As a consequence, the FDA judged the microarray technology sufficiently reliable for clinical and regulatory purposes [46
]. However, it needs to be pointed out that different profiling gene sets cannot be transferred from one microarray platform to another without extensive modification. As a consequence, the results of different microarray platforms cannot be directly compared. Another point of critique concerns the samples used for microarray technologies. As the tissue samples are not microdissected before analysis, the gene profiles represent not only the tumor cells but also the peritumoral reaction and the host tissue. It seems probable that the subtype of normal-like breast carcinoma is an artifact of gene expression profiling, caused by the analysis of samples with a high content of normal breast epithelial cells and stromal cells [13
]. Furthermore, in the case of microinvasive carcinoma with extensive DCIS, there is no safe way of guaranteeing that the invasive carcinoma and not the in situ component are analyzed.
Beyond the level of methodology, the rather small sample sizes upon which the groundbreaking works of Perou and Sorlie were based were regarded with some uneasiness. Indeed, in his seminal work describing ‘molecular portraits of human breast tumors’, Perou analyzed a set of 65 surgical specimens derived from 42 different individuals [4
]. In the following 2 publications by the Stanford group that further outlined the concept of molecular profiling of breast carcinoma, Sorlie et al. based their analyses on 78 and 115 breast carcinomas, respectively [9
]. Considering that in these studies a whole new molecular profile-based taxonomy of breast carcinomas was developed, a greater sample size would have validated the new concept with greater statistical power.
As molecular profiling is often hailed as a largely unbiased analysis tool that allows scientists to avoid the subjectivity of immunohistochemistry and histopathological grading, it is often forgotten that the hierarchical clustering method, far from being a completely automated analysis, is a method that requires the input of a human observer for the final interpretation of the data. In a test encompassing the 5 major intrinsic gene lists, Mackay et al. [47
] demonstrated that none of the classification systems produced almost perfect interobserver agreement. The best interobserver agreement was documented for basal-type and HER2-enriched breast carcinomas, whereas poor interobserver agreement was found for luminaltype breast carcinomas.
It is also important to bear in mind that the analyzing tool of hierarchical clustering can only be applied retrospectively, to adequately sized collectives. The classical microarray-based hierarchical clustering method is therefore not suitable for assigning intrinsic subtypes to particular samples as the dendrogram changes with each additionally included sample. To circumvent this particular problem, single sample predictors (SSPs) have been devised [13
]. SSPs are based on the median expression patterns of the different intrinsic subtypes (i.e. centroids). The SSP allows for any given sample to be assigned to the intrinsic subtype most similar in its molecular profile. However, the reliability of SSPs has been discussed controversially. While some groups have reported reliable and reproducible intrinsic subgroup assignments through SSPs [40
], others have questioned the validity of this method and claimed that only basal-like breast carcinomas were safely identified by this method [32
]. Further studies are needed to ascertain whether SSPs represent a valid method for molecular subtyping.
It must be noted that, while the technique of RNA-based molecular subtyping allows fascinating insights into the molecular properties of breast cancer, it is not (yet) a very good tool for assessing individual samples.
Despite the scientific and medial hype, we are therefore not yet at a point where this new technology has a relevant impact on day-to-day clinical practice.