Automated chemical hypothesis generation and database searching with Catalyst^®

Summary

Simulation technology is a time-tested approach for lowering research costs in industries where trial-and-error techniques dominate. Research in pharmaceutical drug discovery is such an industry, but development of suitable simulation methods has lagged behind because of the difficulty in effectively modeling molecular behavior. The Catalyst program represents a new approach that focuses on modeling the drug-receptor interaction from the point of view of the receptor, using information derived only from the drug. Molecules are described as collections of chemical functions arranged in three-dimensional space. Conformational flexibility is modeled by creating multiple conformers, judiciously prepared to emphasize representative coverage over a specified energy range. A training set, consisting of approximately 20 molecules ranging in activity over four orders of magnitude, becomes the information set used to derive a hypothesis, a minimal collection of chemical features common across the set that explains the observed activity. A hypothesis can serve as an electronic query for searching 3D databases of structures for compounds that, fitting its constraints, are potential leads for further development. The method was applied in different ways to four problems in medicinal chemistry, i.e., angiotensin converting enzyme (ACE), protein farnesyl transferase (PFT), human immunodeficiency virus (HIV) protease (PHIV), and HIV reverse transcriptase (RTHIV) inhibition. From a training set of 20 small peptides with ACE inhibitory activity, a five-featured hypothesis was found that is predictive for inhibitors outside of the set. From a training set of 20 CaaX box tetrapeptides, a hypothesis for PFT inhibition was computed that explains the activities of non-peptide inhibitors reasonably well. Representative compounds from three structurally different classes of RTHIV inhibitors were used to generate a simple four-featured hypothesis that suggests how all three classes could bind to the same receptor. Finally, an enzyme-bound conformation of SB-203386, an inhibitor of PHIV, determined by X-ray crystallography was used to manually construct a hypothesis. This was then used to search 3D databases for structures of interest as inhibitors of the enzyme.