Finak and colleagues used laser capture microdissection to gather samples of invasive breast tumour stroma, which were used for gene expression profiling. They found that the breast tumour stroma could be separated into three clusters according to gene expression: one correlated with significantly reduced recurrence rate and increased relapse-free survival, a second cluster was associated with significantly increased recurrence rate, and the third was associated with a mixture of outcomes. They showed that the clusters were independent of other prognostic markers, such as lymph node status, tumour grade, and oestrogen receptor and ERBB2 expression, indicating that the tumour stroma can be divided into clinically useful subtypes. Gene ontology analysis of the 163 genes that showed the highest differential expression between clusters revealed that poor outcome was mainly associated with the expression of genes involved in angiogenesis and responses to hypoxia, whereas good outcome was mainly associated with increased expression of genes involved in TH1 immune responses.
Next, based on the 163-gene profile, the authors produced a prognostic classifier comprising 26 genes — the stroma-derived prognostic predictor (SDPP) — that could accurately predict patient outcome using tumour stromal samples, but not samples from isolated tumour epithelium. Importantly, they showed that the SDPP accurately predicted outcome using data sets of whole breast cancer samples (containing epithelium and stroma). Numerous prognostic classifiers have been developed — including the 70-gene signature (MammaPrint) that was approved by the US Food and Drug Administration last year — and so Finak and colleagues compared the accuracy of the SDPP. Using the same cohort from the Netherlands on which MammaPrint was developed, they found that the SDPP was 5.96 times more likely to identify a 'true' poor outcome patient with ERBB2+ breast cancer than was MammaPrint. The only classifier with comparable accuracy was the SFT/DTF signature developed using fibroblastic tumours. However, the SFT/DFT (comprising 656 genes) and SDPP have only three genes in common, indicating that the SDPP detects unique biological processes in the stroma. Consistently, they showed that combining the output of the SDPP with that of four prognostic classifiers improved the prediction accuracy.
This is a preview of subscription content, access via your institution