Structure Profiling and Geometric Optimization of Protein-Ligand Complexes for the Scoring Function HYDE

Abstract

Computational drug design relies heavily on three-dimensional protein structures. They are the foundation for analyzing binding poses as well as developing new ideas for small molecule drugs. One goal in research is the prediction of binding affinity. Such predictions are made by assessing the non-covalent interactions between the small molecule and the protein as well as the change in rigidity of the overall system. Thus, a so called scoring function needs to be defined and validated on high quality protein-ligand complexes with known binding affinity data. In this thesis, the scoring function HYDE is equipped with an updated version of its geometric optimization function GeoHYDE. In the update, the Continuous Torsion Score was newly developed and the underlying Torsion Library of 2013 revised in terms of peaks as well as the torsion rule subset analysis. Since the aim of GeoHYDE is a local revision of the interaction profile to maximize the HYDE score in the given pose, deviations should be observed as a measure for its performance quality. The state of the art metric is the root mean squared deviation (RMSD) between initial and final atom coordinates in regards to the geometric optimization. But since the model of a protein structure is just the most preferred interpretation of electron density observations, slight alterations in the model's coordinates should be less relevant if still confirmed by electron density. Hence, the electron density score for individual atoms (EDIA) and molecular fragments (EDIAm) for any element in the periodic table is proposed in this thesis. With EDIAm, the missing piece in the automatic high quality structure data set assemblage is now present. Thus, the tool \texttt{StructureProfiler} was created and the data set ProtFlex18 consisting of 2386 pockets was extracted from the protein data base to consequently analyze the performance of GeoHYDE. As the final part of this thesis, GeoHYDE was parametrized and tested on the training and two test sets extracted from ProtFlex18. In 74 to 79% of all cases tested, EDIAm and RMSD both asses the geometrically optimized pose as very close to the crystallized one with a median HYDE score difference of 0.32 kJ. Including side chain flexibility in the pocket, the medians of final HYDE scores further improve but at the cost of at least four times rising computation time. HYDE in combination with GeoHYDE performs in the lower to middle third on the widely used validation data set CASF-2016 depending on the type of the test scheme.