The ability to integrate a wealth of human-curated knowledge from scientific datasets and ontologies can benefit drug-target interaction prediction. The hypothesis is that similar drugs interact with the same targets, and similar targets interact with the same drugs. The similarities between drugs reflect a chemical semantic space, while similarities between targets reflect a genomic semantic space. In this paper, we present a novel method that combines a data mining framework for link prediction, semantic knowledge (similarities) from ontologies or semantic spaces, and an algorithmic approach to partition the edges of a heterogeneous graph that includes drug-target interaction edges, and drug-drug and target-target similarity edges. Our semantics based edge partitioning approach, semEP, has the advantages of edge based community detection which allows a node to participate in more than one cluster or community. The semEP problem is to create a minimal partitioning of the edges such that the cluster density of each subset of edges is maximal. We use semantic knowledge (similarities) to specify edge constraints, i.e., specific drug-target interaction edges that should not participate in the same cluster. Using a well-known dataset of drug-target interactions, we demonstrate the benefits of using semEP predictions to improve the performance of a range of state-of-the-art machine learning based prediction methods. Validation of the novel best predicted interactions of semEP against the STITCH interaction resource reflect both accurate and diverse predictions.