Signatures from 3D mammary acini predict survival in two large independent datasets of breast cancer patients.

A40 Predicting clinical outcome and therapy responses of breast cancer patients is challenging and several groups have used transcriptional profiling of “training sets” (i.e. sets of patients with known outcome) to develop expression signatures. The supervised approach has been criticized as it is strongly dependent on the patient set used and is subject to inadequate validation. We previously described a novel unsupervised strategy to identify predictive signatures for breast cancer using 3D culture models of differentiating non-malignant human mammary epithelial cells (HMECs). These cells proliferate for several days before forming growth-arrested, polarized mammary acini. A panel of 22 genes was identified that distinguishes proliferating from quiescent HMECs in 3D culture. Breast tumors expressing profiles most like the proliferating cells had the poorest outcome when validated using the van de Vijver set of 295 breast tumors. To evaluate the prognostic capacity of this 3D signature in additional breast cancer patients we analyzed the Wang data set, which includes 286 node-negative invasive ductal carcinoma biopsies and the Stanford-Norway data set, which includes 121 breast tissue samples of a range of histology subtypes including 4 normal breast, 2 DCIS, 100 invasive ductal carcinoma, 3 fibroadenoma, 8 lobular carcinoma, and 1 each of mucinous, papillary, pleomorphic, and undifferentiated carcinomas. Each data set used a different microarray platform. We used our 22 gene signature and unsupervised hierarchical clustering to group the tumors into subclasses. Each dataset was separated into two groups that accurately predicted overall survival when analyzed by the method of Kaplan-Meier (p=0.000013 and 0.045, for the Wang and Stanford datasets, respectively). Tumors with signatures most similar to differentiated acini had the best prognosis; in contrast to those with patterns like proliferating cells. This study underscores the utility of the unsupervised approach, as well as the value and potential clinical relevance of 3D culture models.