Cellular and molecular determinants of breast cancer sensitivity to all-trans retinoic acid: Identification of a gene expression fingerprint predicting responsiveness

All-trans retinoic acid (ATRA) and derived natural as well as synthetic retinoids are promising agents in the treatment and chemoprevention of various types of neoplasia, including mammary tumors. ATRA is an important component of the therapeutic schemes used for the treatment of a rare form of Acute Myelogenous Leukemia known as Acute Promyelocytic Leukemia. A rational use of the paradigmatic retinoid, ATRA, in a heterogeneous disease, like breast cancer, requires the definition of the cellular and molecular determinants of sensitivity to the agent. The major aim of the study was the definition of a predictive gene expression fingerprint that can be used for the selection of patients who may benefit from treatment protocols containing ATRA. To this purpose, we selected 45 breast cancer cell lines characterized for the constitutive whole genome gene expression profiles. The sensitivity of 30 cell lines (training set) to the anti-proliferative action of ATRA was defined after challenge with increasing concentrations of the retinoid for 3, 6 and 9 days. This analysis established that Luminal and ER-positive cell lines are enriched within the ATRA sensitive group. In contrast, cell lines characterized by a Basal-like phenotype, according to the PAM50 gene expression signature, are generally refractory to the growth inhibitory action of ATRA. The sensitivity of Luminal-A and Luminal-B and the general refractoriness of Basal-like tumors to ATRA was validated in short-term tissue slice cultures of surgical breast cancer specimens. The training set was used to define a gene-expression fingerprint consisting of approximately 50 genes significantly associated with ATRA sensitivity. The fingerprint was generated by reprocessing the RNA sequencing data contained in the CCLE (Cancer Cell line Encyclopedia) of the Broad Institute and it was built from approximately 60,000 coding and non-coding loci. The approach involved the use of general linear models (machine learning algorithm). The identified gene-expression fingerprint was subsequently used to successfully predict ATRA sensitivity in a test set consisting of the remaining 15 cell lines. As a first step towards the use of the fingerprint for the stratification of patients, we evaluated the proportion of predicted ATRA sensitive breast tumors in the TCGA dataset. In accordance with the cell line and primary tumor data, approximately 30% of the Luminal tumors present with a high similarity score to the identified gene expression fingerprint associated with ATRA sensitivity. In contrast, only 5% of the Basal-like or Triple-negative mammary tumors are characterized by the same high similarity score. Curiously, the ATRA sensitivity signature seems to be tumor context independent, as it correctly identifies the 20 Acute Promyelocytic Leukemia patients present in the 198 Acute Myelogenous Leukemia patients present in the TCGA datase