To identify MRI-based radiomic features that could be obtained automatically by a deep learning (DL) model and could predict the clinical characteristics of breast cancer (BC). Also, to explain the potential underlying genomic mechanisms of the predictive radiomic features. A denoising autoencoder (DA) was developed to retrospectively extract 4,096 phenotypes from the MRI of 110 BC patients collected by The Cancer Imaging Archive (TCIA). The associations of these phenotypes with genomic features (commercialized gene signatures, expression of risk genes, and biological pathways activities extracted from the same patients’ mRNA expression collected by The Cancer Genome Atlas (TCGA)) were tested based on linear mixed effect (LME) models. A least absolute shrinkage and selection operator (LASSO) model was used to identify the most predictive MRI phenotypes for each clinical phenotype (tumor size (T), lymph node metastasis(N), status of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2)). More than 1,000 of the 4,096 MRI phenotypes were associated with the activities of risk genes, gene signatures, and biological pathways (adjusted P-value < 0.05). High performances are obtained in the prediction of the status of T, N, ER, PR, HER2 (AUC>0.9). These identified MRI phenotypes also show significant power to stratify the BC tumors. DL based auto MRI features performed very well in predicting clinical characteristics of BC and these phenotypes were identified to have genomic significance.