Docking ligands onto binding site representations derived from proteins built by homology modelling.

Due to the abundant sequence information available from genome projects, an increasing number of structurally unknown proteins, homologous to examples of known 3D structure, will be discovered as new targets for drug design. Since homology models do not provide sufficient accuracy to apply common drug design tools, a new approach, DragHome, has been developed to dock ligands into such approximate protein models. DragHome combines information from homology modelling with ligand data, used by and derived from 3D quantitative structure-activity relationships (QSAR). The binding-site of a model-built protein is analysed in terms of putative ligand interaction sites and translated via Gaussian functions into a functional binding-site description represented by physico-chemical properties. Ligands to be docked onto these binding-site representations are similarly translated into a description based on Gaussian functions. The docking is computed by optimising the overlap between the functional description of the binding site and the ligand, generating multiple solutions. For a set of different ligands, these solutions are ranked according to the internal similarity consistance among the various ligands in the binding modes obtained from docking. DragHome has been validated at examples for which crystal structures are available: structurally distinct thrombin inhibitors were docked onto models of thrombin generated from serine proteases of 28 to 40 % sequence identity, yielding ligand binding modes with an average RMS deviation of 1.4 A. Mostly the near-native solutions are ranked best. Molecular flexibility of ligands can be considered in terms of pre-calculated multiple conformers. DragHome has been used to automatically generate an alignment of 88 thrombin inhibitors, for which a significant 3D QSAR model could be derived. The contribution maps resulting from this analysis can be interpreted with respect to the surrounding protein model. They highlight inconsistencies and deficiencies present in the model. In future developments, this information could be fed back into a subsequent modelling step to improve the protein model.