G Protein-Coupled Receptors (GPCRs) are integral membrane proteins that play important role in communicating diverse extracellular signals to second messenger systems within the cells. Due to their importance in regulating key physiological functions, these proteins are targets of about 50% of recently launched drugs. High-resolution experimental structures are available for only one GPCR, bovine rhodopsin. As a result, structure-based drug design efforts must rely on in silico models. There are currently two approaches that have been applied to build such models: (a) ab initio methods, which use only primary sequence data of the receptor; and (b) homology modeling, which uses the rhodopsin structure as a template. Here we describe Gmodel, a new hybrid method for building 3D atomic models of GPCRs. Gmodel starts by building a homology model, which is subsequently refined by optimizing side-chain as well as backbone conformations, typically in the presence of bound ligand. Mutation and ligand binding data are used to locate and orient ligands in the GPCR binding pocket. Two unique features of Gmodel are: (a) it incorporates a profile-based alignment in building the initial homology model and (b) it uses low-frequency normal modes to optimize the backbone conformation of the receptor. Gmodel was used to generate receptor models for the antagonist forms of 5HT1A and H3. The validity of this method was demonstrated by its ability to produce highly enriched hit lists upon screening a library of random compounds seeded with known actives. Enrichment factors nearly 100 fold better than random screening were obtained, demonstrating the usefulness of Gmodel in generating structures for lead discovery.