Mining large drug-oriented knowledge graphs enables predicting Adverse Drug Reactions (ADRs). Indeed, these graphs encompass knowledge elements about the molecular mechanism of drugs (e.g. drug targets, Gene Ontology annotations, gene variations, pathways). However, only few works explored further these graphs in the search for mechanistic explanation for this type of events. We assume that features documenting molecular mechanisms that take part in the prediction are particularly interesting features, since they may provide novel knowledge for the mechanism that may be underlying an ADR. We propose to explore PGxLOD, a knowledge graph built around drugs and pharmacogenomic processes in which they are involved, through the lens of several ADR datasets, each focusing on a particular type of ADRs. Particularly, we propose to use features resulting from the exploration of PGxLOD in a prediction task where best predictive features will be considered as potential elements of explanation.