Predicting binding modes, binding affinities and 'hot spots' for protein-ligand complexes using a knowledge-based scoring function

The development of a new knowledge-based scoring function (DrugScore) and its power to recognize binding modes close to experiment, to predict binding affinities, and to identify `hot spots' in binding pockets is presented. Structural information is extracted from crystallographically determined protein-ligand complexes using ReLiBase and converted into distance-dependent pair-preferences and solvent-accessible surface (SAS) dependent singlet preferences of protein and ligand atoms. The sum of the pair preferences and the singlet preferences is calculated using the 3D structure of protein-ligand complexes either taken directly from the X-ray structure or generated by the docking tool FlexX. DrugScore discriminates efficiently between well-docked ligand binding modes (root-mean-square deviation < 2.0 Angstrom with respect to a crystallographically determined reference complex) and computer-generated ones largely deviating from the native structure. For two test sets (91 and 68 protein-ligand complexes, taken from the PDB) the calculated score recognizes poses deviating < 2 Angstrom from the crystal structure on rank 1 in three quarters of all possible cases. Compared to the scoring function in FlexX, this is a substantial improvement. For five test sets of crystallographically determined protein-ligand complexes as well as for two sets of ligand geometries generated by FlexX, the calculated score is correlated with experimentally determined binding affinities. For a set of 16 crystallographically determined serine protease inhibitor complexes, a R-2 value of 0.86 and a standard deviation of 0.95 log units is achieved as best result; for a set of 64 thrombin and trypsin inhibitors docked into their target proteins, a R-2 value of 0.48 and a standard deviation of 0.7 log units is calculated. DrugScore performs better than other state-of-the-art scoring functions. To assess DrugScore's capability to reproduce the geometry of directional interactions correctly, 'hot spots' are identified and visualized in terms of isocontour surfaces inside the binding pocket. A data set of 159 X-ray protein-ligand complexes is used to reproduce and highlight the actually observed ligand atom positions. In 74% of all cases, the actually observed atom type corresponds to an atom type predicted by the most favorable score at the nearest grid point. The prediction rate increases to 85% if at least an atom type of the same class of interaction is suggested. DrugScore is fast to compute and includes implicitly solvation and entropy contributions. Small deviations in the 3D structure are tolerated and, since only contacts to non-hydrogen atoms are regarded, it does not require any assumptions on protonation states.