We discuss the development of a coarse-grained (CG) model for molecular dynamics (MD) simulation of a hydrophobic dipeptide, diphenylalanine, in aqueous solution. The peptide backbone is described with two CG beads per amino acid, the side groups and charged end groups are each described with one CG bead. In the derivation of interaction functions between CG beads we follow a bottom-up strategy where we devise potentials such that the resulting CG simulation reproduces the conformational sampling and the intermolecular interactions observed in an atomistic simulation of the same peptide. In the CG model, conformational flexibility of the peptide is accounted for through a set of intra-molecular (bonded) potentials. The approach followed to obtain the bonded potentials is discussed in detail. The CG potentials for nonbonded interactions are based on potentials of mean force obtained by atomistic simulations in aqueous solution. Following this approach, solvent mediation effects are included in the effective bead-bead nonbonded interactions and computationally very efficient (solvent-free) simulations of self-assembly processes can be performed. We show that the conformational properties of the all-atom dipeptide in explicit solvent can be accurately reproduced with the CG model. Moreover, preliminary simulations of peptide self-assembly performed with the CG model illustrate good agreement with results obtained from all-atom, explicit solvent simulations.