Adverse drug reactions are an unresolved issue that can result in mortality, morbidity and substantial healthcare costs. Many conventional machine learning methods have been used for predicting post-marketing drug side-effects. However, owing to the complex chemical structures of certain drugs and the nonlinear and imbalanced nature of biological data, some side-effects might not be detected. Motivated by the drug discovery research studies that have shown that deep learning outperformed machine learning methods over prediction tasks, we proposed: (i) to exploit the unsupervised deep learning approaches to predict ADRs; (ii) to use a two-stage framework to predict personalized ADRs and repurpose the drugs. This work demonstrates that the proposed framework shows promise in providing more-accurate prediction of side-effects and drug repurposing.