Stromal signature identifies basal breast cancers

Finak et al. reply:

In our previous report in Nature Medicine¹, hierarchical clustering of breast tumor stromal-specific gene expression data revealed a group of subjects enriched for poor clinical outcome. From 163 genes differentially expressed between this cluster and the remaining samples, we derived a predictor consisting of 26 (SDPP) and showed that it predicted poor clinical outcome in multiple breast tumor data sets. The SDPP was explicitly constructed to bias against outcome-linked genes that strongly associated with histopathological subtypes (defined by immunohistochemistry) rather than as a measure of which stroma cluster a tumor belongs to. We clarify the points raised by Wennmalm et al.² In contrast to our approach, Wennmalm et al.² use the 26 SDPP genes to train a classifier to predict membership in our poor-outcome stroma cluster rather than to predict clinical outcome, a fundamentally different goal than that of our report¹ and one for which the SDPP genes are unlikely to be an optimal choice. Consistent with our stated results (Table 2 in ref. 1), Wennmalm et al.² observe that their poor-outcome cluster is enriched for basal tumors. However, their use of the Sorlie et al.³ centroids to assign tumor subtypes in our stroma gene expression data is, in our opinion, incorrect, as many of the genes in these centroids are epithelial specific. Consequently, Wennmalm et al.² assigned subtypes that are in disagreement with the immunohistopathology data, as their results predict five HER2-positive and eight basal tumors, whereas we reported ten HER2-positive and six triple-negative (basal) tumors. Because the stroma generally does not possess the estrogen receptor or HER2 signatures, it is likely to be confounded with the triple-negative (basal) subtype owing to the lack of ER or HER2 expression, thus increasing the apparent percentage of these tumors. We reported a similar observation for normal breast tissue⁴.