Title: The challenges and pitfalls of delivering breast cancer multi-gene diagnostic assays to the clinic.

Background: Breast cancer is a heterogeneous group of tumors with highly diverse clinical outcomes. Recently a number of groups have reported multi-gene based classifiers that aim to more accurately predict outcome than standard parameters such as estrogen receptor and HER2 status, node status, tumor size and grade. These efforts have largely focused on supervised analysis approaches whereby patient outcomes are used to train data sets to create gene signatures that are associated with clinical outcomes. In contrast our group has focused on the alternative approach of unsupervised clustering where the expression patterns are examined for the purposes of disease classification. These efforts have produced a five class classification termed Luminal A (good prognosis ER+), Luminal B (poor prognosis ER+), HER2- ER-, Basal-type and normal. These subtypes have quite distinct clinical features and treatment paradigms. With support from the NCI our group has undertaken a collaborative program to developed a multi-gene qRT-PCR assay for identifying biological subtypes from formalin-fixed paraffin-embedded tissues. A prototype diagnostic assay was compared to subtype classification made by microarray from fresh tissue and also compared to standard clinical markers.

Methods: Publicly available microarray data were used to develop a training set for classifying tumors into five different previously defined molecular subtypes of Luminal A, Luminal B, HER2, Basal-like, and Normal-like. Sample class predictors were developed from hierarchical clustering of the microarray data using two different centroid-based algorithms: Prediction Analysis of Microarray and a Single Sample Predictor. The training set data was applied to predicting sample class on an independent test set of 121 unique breast tumors procured as both fresh-frozen and formalin-fixed, paraffin embedded tissues (242 samples). Classification of the test set samples was determined from microarray data using a large 1300 gene set (Hu et al, BMC Genomics 2006), and using minimized ‘intrinsic’ gene sets for the qRT-PCR assay. In total, 212 genes (9 housekeepers and 203 classifiers) were tested by qRT-PCR on FFPE tissues and 131 genes were selected to perform the ‘intrinsic’ subtype classification.

Results: There were 6 out of 121 (5%) samples that were classified as Normal-like by either microarray or PCR and were excluded from further analyses. We found 83% concordance in classification between the microarray and qRT-PCR assays. By qRT-PCR alone, there was 85% concordance in subtype assignment between the two centroid-based methods. Classification assigned by microrray (1300 gene set) and by qRT-PCR had 93% agreement with ER-status by IHC. In 9 out of 12 FISH confirmed HER2-positive samples the multi-gene assay assigned the HER2 subtype. There was 80% sensitivity and 95% specificity between the qRT-PCR subtype classification and FISH. Single gene analyses between qRT-PCR and IHC status using receiver operator curves to optimize cut-offs showed 92% sensitivity and 89% specificity for ER and 69% sensitivity and 91% specificity for HER2.

Conclusions: Determining agreement in classification between platforms and procurement methods requires a variety of methods. We have shown that centroid-based algorithms are robust classifiers for breast cancer subtype assignment across platforms (microarray and qRT-PCR data) and procurement conditions (fresh-frozen and formalin-fixed, paraffin-embedded tissues). Our strategy for primer set validation and classification have applications in routine clinical practice for risk-stratifying breast cancers. We will discuss our plans for further development of intrinsic subtype testing, and the barriers presented by a diagnostic platform that requires accurate assessment of multiple genes in a robust manner.

98th AACR Annual Meeting-- Apr 14-18, 2007; Los Angeles, CA