Graph Database Management Systems (GDBMS) are rapidly emerging as an effective and efficient solution to the management of very large data sets in scenarios where data are naturally represented as a graph and data accesses mainly rely on traversing this graph. Currently, the design of graph databases is based on best practices, usually suited only for a specific GDBMS. In this paper, we propose a model-driven, system-independent methodology for the design of graph databases. Starting from a conceptual representation of the domain of interest expressed in the Entity-Relationship model, we propose a strategy for devising a graph database in which the data accesses for answering queries are minimized. Intuitively, this is achieved by aggregating in the same node data that are likely to occur together in query results. Our methodology relies a logical model for graph databases, which makes the approach suitable for different GDBMSs. We also show, with a number of experimental results over different GDBMSs, the effectiveness of the proposed methodology.