Drug development is a costly and time-intensive process. However, promising strategies such as drug repositioning and side effect prediction can help to overcome these challenges. Repurposing approved drugs can significantly reduce the time and resources required for preclinical and clinical trials. Furthermore, early detection of potential safety issues is crucial for both drug development programs and the wider healthcare system. For both goals, drug repositioning and side effect prediction, existing machine learning (ML) approaches mainly rely on data collected in preclinical phases, which is not necessarily representative of the real-world situation faced by patients. In this chapter, we construct a knowledge graph based on diagnoses, prescriptions and diagnostic procedures found in large-scale electronic health records, as well as secondary information from different databases, such as drug side effects and chemical compound structure. We show that modern Graph Neural Networks (GNNs) allow for an accurate and interpretable prediction of novel drug-indication and drug-side effect associations in the knowledge graph. Altogether, our work demonstrates the potential of GNNs for knowledge-informed ML in healthcare.