Molecular characterisation of formalin-fixed paraffin-embedded (FFPE) breast tumour specimens using a custom 512-gene breast cancer bead array-based platform

Abstract

Background: Formalin-fixed, paraffin-embedded (FFPE) tumour tissue represents an immense but mainly untapped resource with respect to molecular profiling. The DASL (cDNA-mediated Annealing, Selection, extension, and Ligation) assay is a recently described, RT-PCR-based, highly multiplexed high-throughput gene expression platform developed by Illumina specifically for fragmented RNA typically obtained from FFPE specimens, which enables expression profiling. In order to extend the utility of the DASL assay for breast cancer, we have custom designed and validated a 512-gene human breast cancer panel.

Methods: The RNA from FFPE breast tumour specimens were analysed using the DASL assay. Breast cancer subtype was defined from pathology immunohistochemical (IHC) staining. Differentially expressed genes between the IHC-defined subtypes were assessed by prediction analysis of microarrays (PAM) and then used in the analysis of two published data sets with clinical outcome data.

Results: Gene expression signatures on our custom breast cancer panel were very reproducible between replicates (average Pearson's R²=0.962) and the 152 genes common to both the standard cancer DASL panel (Illumina) and our breast cancer DASL panel were similarly expressed for samples run on both panels (average R²=0.877). Moreover, expression of ESR1, PGR and ERBB2 corresponded well with their respective pathology-defined IHC status. A 30-gene set indicative of IHC-defined breast cancer subtypes was found to segregate samples based on their subtype in our data sets and published data sets. Furthermore, several of these genes were significantly associated with overall survival (OS) and relapse-free survival (RFS) in these previously published data sets, indicating that they are biomarkers of the different breast cancer subtypes and the prognostic outcomes associated with these subtypes.

Conclusion: We have demonstrated the ability to expression profile degraded RNA transcripts derived from FFPE tissues on the DASL platform. Importantly, we have identified a 30-biomarker gene set that can classify breast cancer into subtypes and have shown that a subset of these markers is prognostic of OS and RFS.

Figure 1

Expression of ESR1, PGR and ERBB2 correspond to their respective pathology IHC staining status for ER, PR and HER2. Stripcharts showing the expression of (A) ESR1, (B) PGR and (C) ERBB2 segregated by their respective negative and positive IHC categories. Gene-level data are displayed with the mean for each category denoted by the horizontal black bar and significance is assessed by a two-sided Welch's t-test.

Figure 2

Montreal cohort of 87 patients hierarchically clustered across 30 genes predicative of immunohistochemical (IHC) breast cancer subtype. Prediction analysis of microarrays (PAM)-determined expression from 30 genes were indicative of IHC breast cancer subtype. Hierarchical clustering of patients (columns) and genes (rows) tends to segregate triple-negative (TN; indicated in black), HER2+ (indicated in yellow) and HR+ (indicated in blue) tumours. Red indicates upregulation and green downregulation of transcripts for genes labelled on the right. Gene transcript expression levels are Z-score normalised with a colour key indicated in the top left corner.

Figure 3

Hierarchical clustering of expression data from UNCCC cohort using probes from the 30 genes indicative of immunohistochemical (IHC) subtype in the Montreal cohort. Probes (rows) mapping to the 30 genes indicative of IHC subtype were used to hierarchically cluster patients (columns) in the Parker et al (2009) published microarray data (GEO data set GSE10886). These 30 genes segregate patients by ‘intrinsic subtype’. Red indicates upregulation and green downregulation of transcripts for probes labelled on the right. Probe transcript expression levels are Z-score normalised with a colour key indicated in the top left corner. Hierarchical clustering was conducted in R using the heatmap.2 package, with a dissimilarity metric based on Euclidean distance and an average algorithm for clustering.