The US public spends billions of dollars on prescription drugs every year; however striking variations exist in drug therapy as characterized by adverse drug reactions (ADRs) or drug side-effects. ADRs are defined as those unintended and undesired responses to drugs beyond their anticipated therapeutic effects during clinical use at normal doses. It is one of the major causes of failure in drug development process. Severe ADRs that go undetected until the post-marketing phase of a drug often lead to patient morbidity, as exemplified by numerous drug withdrawals with Rofecoxib (Vioxx) and Cerivastatin (Baycol) being the most prominent examples. Therefore, accurate prediction of potential ADRs is required in the entire life cycle of a drug, including early stages of drug development, different phases of clinical trials, and post-marketing surveillance. In an effort to address this problem, we explored the use of laboratory and retrospective medication order data of 1.9 million individuals in the Vanderbilt University Medical Center (VUMC) electronic medical record (EMR) system to ascertain ADRs. Using laboratory signals for drug safety studies may reduce errors caused by biased reporting, which has the potential to support active clinical drug surveillance. Experimental results have shown that the overall recall and precision were 80% and 44%, respectively. We also proposed to predict ADRs by integrating three different types of drug information from public databases: (1) chemical structures, (2) biological properties (i.e., protein targets, pathways), and (3) phenotypic characteristics (i.e., indications, other known side effects). Five different machine learning algorithms were compared and the best AUC was 0.9524 achieved by the Support Vector Machine (SVM). The data fusion approach is promising for large-scale ADR predictions in both preclinical and post-marketing phases.