Knowledge graphs encode semantic knowledge that can be exploited to enhance different data management tasks, e.g., query answering, ranking, or data mining. We tackle the problem of predicting interactions between drugs and targets, and propose esDSG, an unsupervised approach able to predict links from subgraphs that are not only highly dense, but that comprise both similar drugs and targets. The esDSG approach extends a state-of-the-art approximate densest subgraph algorithm with knowledge about the semantic similarity of the nodes in the original graph, and then predicts potential novel interactions from the computed dense subgraph. We have conducted an initial experimental study on a benchmark of drug-target interactions. Our observed results suggest that esDSG is able to identify interactions in graphs where existing approaches cannot perform equality well. Further, a large number of esDSG predictions can be validated using external databases as STITCH and Kegg. These results, although initial, reveal how semantics in conjunction with topological information of the knowledge graph may have a great impact on pattern discovery tasks.