Report for 2^nd quarter of second year (November 2005 January 2006)

TArget Specific Scoring FUNctions (TASSFUN)

In the TASSFUN project, the structure-based COMparative BINding Energy (COMBINE) method is being applied to generate target-specific tailor-made scoring functions. As a 3D QSAR tool, these scoring functions will be used in virtual screening to select specific ligands of different proteases related to the blood coagulation cascade. The project is supervised by Dr. Niklas Blomberg (GSI CompChem AZ Mlndal) and Dr. Rebecca Wade (EML Research, Heidelberg).

Crystal structures of protein-ligand-complexes of trypsin (15) and urokinase (30) were selected to construct training sets for building target-specific COMBINE models. From these structures, one template protein structure was selected for building the COMBINE models. Outliers, with a structure-bioactivity-relationship inconsistent with the other data, were removed. The models reached correlation coefficients of r², q², and standard error of prediction (kcal/mol) of 0.83, 0.62, and 0.92 for urokinase and of 0.96, 0.83, and 0.40 for trypsin. The preparation of binding data and structures for the training set for building a COMBINE model for thrombin is in progress.

For using the COMBINE models in virtual screening with large data sets, the programmed and implemented software tools were changed to allow the transition from a semi- to a fully-automatic procedure suitable for running in parallel on a compute-cluster. Now, it is possible, starting from a receptor model and a SD file containing hundreds of small molecule ligands, to dock these by the program GOLD and to minimize them afterwards by molecular mechanics calculations in AMBER8. Subsequently, the electrostatic and van der Waals interaction energies are calculated automatically and are combined with electrostatic solvation energy terms in an input file for GOLPE. The interaction and electrostatic solvation energy terms are calculated by the program ANAL and by solving the Poisson-Boltzmann equation with UHBD, respectively.

As a first test set, the ligands used for model building were docked ten times into the receptor models with GOLD and their binding affinities were predicted with COMBINE with an overall accuracy for all docked solutions of better than 1.4 kcal/mol. In the next step, ligands with known inhibitor constants but unknown binding conformation were selected from prepared ligand data files. For urokinase, a training set of 180 structures and for trypsin a training set of 88 structures (data set of Bhm, Strzebecher and Klebe, J. Med. Chem. 1999) were used. From preliminary analysis of these predictions, the best ranked docking solution has an average difference between predicted and experimental binding affinity of around 1.4kcal/mol in the COMBINE models.

In the next time period, further docking experiments will be performed with a larger amount of inhibitors and the prediction of binding affinity values by the existing COMBINE models will be tested in more detail. In addition, the method will be extended to further proteases, like thrombin.