Multi-relational graphs contain various types of relations that usually come with variable frequency and have different importance for the problem at hand. Existing graph sampling approaches ignore the multi-relational nature of such graphs. We propose an approach to modeling the importance of relation types for sampling and show that we can learn the right balance: relation-type probabilities that reflect both frequency and importance. We use relation-dependent sampling to develop a scalable graph neural network and apply it for multi-relational link prediction. Our experiments specifically consider drug-drug interaction (DDI) prediction, an important task during drug development. In that context, we further show the benefit of considering a special relation type, negative edges. Synergistic drug combinations (e.g., drugs that synergize by improved efficacy and reduced side effects) can be regarded as negative evidence for DDIs (which often indicate adverse reactions). We add drug synergy data to provide extra expert knowledge which can be easily integrated into our model and yields improved performance.