Optimiertes Design kombinatorischer Verbindungsbibliotheken durch Genetische Algorithmen und deren Bewertung anhand wissensbasierter Protein-Ligand Bindungsprofile

In dieser Arbeit sind die zwei neuen Computer-Methoden DrugScore Fingerprint (DrugScoreFP) und GARLig in ihrer Theorie und Funktionsweise vorgestellt und validiert worden. DrugScoreFP ist ein neuartiger Ansatz zur Bewertung von computergenerierten Bindemodi potentieller Liganden für eine bestimmte...

Full description

Saved in:
Bibliographic Details
Main Author: Pfeffer, Patrick
Contributors: Klebe, Gerhard (Prof. Dr.) (Thesis advisor)
Format: Doctoral Thesis
Published: Philipps-Universität Marburg 2009
Online Access:PDF Full Text
Tags: Add Tag
No Tags, Be the first to tag this record!

This thesis describes the theory, development, application and validation of two new chemoinformatic tools to find new drugs named DrugScoreFP and GARLig. DrugScoreFP is an extension of the well-known scoring function DrugScoreCSD. The method can be used for rescoring docking solutions by adding structural information from a user-defined set of protein-ligand complexes resulting in a tailor-made protein-specific scoring function. The new method demonstrates significant improvements in finding near-native poses in comparison to 12 established scoring functions. Using the in-house investigated protein structures trypsin, thermolysin and tRNA-guanin transglycosylase (TGT), we identified six fragment-sized molecules which were found to inhibit these targets and one thermolysin crystal structure in complex with one of the predicted fragments. GARLig is a library design tool based on docking and a self-adaptive genetic algorithm for structure-based sidechain-optimization of small molecule skeletons. Multiple scoring functions such as AutoDock4 Score, GOLDScore and DrugScoreCSD can be applied to the search procedure as possible decision criteria for potential library candidates. The program has been validated on trypsin, thrombin, factor Xa, plasmin and cathepsin D. GARLig was able to find natural substrates and known binders by validating less than 8% of large combinatorial libraries, meanwhile outperforming other search strategies such as a random search and a Monte Carlo Sampling.