Magnetic resonance imaging (MRI) captures the three dimensional way in which tumors are organized in the breast, defined as imaging phenotypes in the I-SPY 1 trial. We developed a gene set based on the way in which breast epithelial cells aggregate and organize in three dimensional cultures. We investigated whether these organizational genes correspond to the imaging phenotypes.

MRI phenotypes have been shown to correspond to the degree of response to neoadjuvant chemotherapy, and are used to predict the ability to achieve breast conservation treatment (Mukhtar et al, Ann Surg Oncol, 20: 3823–3830, 2013). We hypothesized that the molecular profile accompanying phenotypic changes occurring during the organization process of non-malignant acini may explain the molecular basis of MRI tumor phenotypes.

We have developed prediction models for MRI phenotypes based on expression profiles identified during the organization process of non-malignant breast epithelial cells in three-dimensional laminin-rich extracellular matrix (Fournier et al. Cancer Res, 66, 7095-102, 2006). We analyzed a subset of 324 organizational genes in 147 samples from stage II-III breast cancer in the I-SPY 1 trial cohort with Agilent microarrays and MRI annotation (CALGB 150007/150012; ACRIN 6657). MRI phenotype of the index lesion was assessed by the site radiologists using the following radiographic criteria: A) well defined unicentric mass; B) well-defined multilobulated mass; C) area enhancement with nodularity; D) area enhancement without nodularity; E) Septal spreading. The distribution of phenotypes in I-SPY 1 was: A: 16%, B: 33%, C: 30%, D: 15%, E: 7%. We developed predictors for MRI phenotypes dichotomized as either "well-defined" (A and B) or "non-well-defined" (C, D and E). Using patent-pending algorithms we selected a subset of the organizational genes with the greatest predictive power to identify MRI "well-defined" phenotypes in the I-SPY 1 cohort. Then prediction models were developed using the 50 top ranking genes and logistic regression methods. We followed Monte-Carlo cross validation method to make sure that the performance of a model is not a result of over-fitting the model to the sample data using separate datasets to support the modeling. The samples were randomly partitioned 10,000 times in a 85% training, and 15% test ratio to test models performance. The resulting performance of predictive models in the test set had an average classification accuracy of greater than 80% (ROC statistics AUC>0.8) using gene expression measurements from between 16 to 20 genes. For random lists of 16-20 genes, the accuracy was approximately 50% (ROC AUC∼0.5). Amongst the MRI models were several genes that are known to regulate key cellular pathways such as cell division, metabolism, and migration using MetaCore pathway analysis. Taken together, the results suggest that the MRI phenotypes may be a manifestation of the organization genes that determine behavior in three dimensional culture. Future research includes confirming these results using an independent dataset, defining the potential drivers of MRI phenotypes, and determining if putative drivers provide a key contribution to tumor subtypes.