Development of computer aided drug discovery methods based on machine learning techniques and application to the dopamine D1 receptor

• Main Author: Oloff, Scott H
• Format: Dissertation
• Language: English
• Subjects: health sciences, pharmacology
• ISBN: 054234047X; 9780542340475

The overall objective of this work was to develop and test novel computational drug discovery approaches. Although the methods I developed have applicability to any target protein, dopamine D1 receptor-ligand information was used as the validation system because of the pharmacological significance of this receptor and the available experimental resources. I first developed a QSAR method based on the Support Vector Machine algorithm, improving upon the overall process of model development and virtual screening. Using these advancements, I virtually screened databases with more than 750,000 compounds, and identified 54 candidate D1 ligands, only five of which had been characterized previously as D1 ligands. Two commercially available compounds were obtained for experimental testing, and found to have D 1 binding affinity similar to dopamine. Although the predicted affinities of these compounds were higher than found experimentally, the results suggest that compound predictions made by a high percentage of QSAR-based models are more reliable in database screening.
The second aim was to develop a dynamic structural model of the D1 receptor using a de novo molecular dynamics approach following the work of Goddard's group, and compare that model to a semi-empirical homology-based model. The active site of the ab initio model provides a feasible docking orientation for D 1 ligands, and is in agreement with site-directed mutagenesis data. The model predicts the second extracellular loop, which has low sequence similarity to the D5 receptor, folds over the binding pocket suggesting possible interactions for a D1 versus D5 selective ligand. Differences between the two modeling approaches have provided hypotheses for future work in both the experimental and computational arenas.
In my third aim, I developed a novel approach that predicts "Complementary Ligands Based on Receptor Information" (CoLiBRI). This approach represents both known receptor active sites and their ligands in a universal space of chemical descriptors. The latter permits the identification of probable ligands for a test receptor, as well as predicting probable target receptors for test ligands. This tool is exceptionally accurate for its speed and is best applied as a prescreening tool for large chemical databases that would be unfeasible to process with traditional docking.