CVC Seminar: Wed June 1 1:30-3pm ACES 4.304

Maysam Moussalem

A Reduced Flexibility Model and Improved Scoring Scheme For Lead Optimization In Computational Drug Design

Virtual drug screening and design offer a much faster and cheaper alternative to the the process of testing tens of thousands of drugs, one by one, in the laboratory. Computational drug design can be summarized as the following sequence of phases: a target first needs to be identified; this is generally a protein which plays a central role in the causes and symptoms of the disease. Once a suitable target has been determined, structural and physicochemical information on it is gathered, using a range of techniques selected from crystallography, genomics, proteomics and bioinformatics. A database of potential drug candidates is then compiled and the compounds are tested for their binding affinity with the target using structure-based drug design tools such as high throughput screening. The goal of this process is to come up with the best 0.1% of compounds from a database of more than a million potential candidates. Once a suitable lead (or small set of leads) has been identified, optimization techniques are applied for improving binding affinity and pharmacokinetic properties. These include structural refinement via methods such as molecular docking and energetics-based re-ranking methods, and combinatorial chemistry. Finally, top leads are tested for cross-reactivity, side-effects, non-toxicity, bioavailability, absorption, metabolism and half-life. This process is both very lengthy and very costly.

In this research, we focus on lead optimization, in which a target-drug complex returned as the result of a docking experiment must go through iterative chemical optimization and structural refinement, with the goal of improving the complex’s binding affinity as well as its pharmacokinetic properties, making it a stronger candidate for subsequent stages of the drug development process. The two biggest unsolved challenges in computational drug design which tend to limit its scope of application are: (1) the need for accurate and efficient modeling and exploration of proteins’ large number of degrees of freedom, which would result in better pose prediction, and (2) improved accurate and efficient scoring functions which are capable of distinguishing correct binding poses from false ones for large varieties of target-drug combinations, as well as provide an accurate estimate of binding affinities. In order to develop efficient methods, certain assumptions and simplifications are often made, such as the rigid body assumption used when modeling larger proteins, and the development of all-purpose empirical and knowledge-based scoring schemes based on experimentally determined interaction pattern preferences. While these approaches allow for increased efficiency, they often do not yield optimal results.

The goal of the proposed research is to develop a computational approach to lead optimization based on an efficient search algorithm which samples relevant conformational and positional space with the help of a reduced protein flexibility model, and which is guided by a scoring function applicable and tunable to a wide array of diverse targets belonging to different protein families. In order to support such an accurate scoring function while maintaining computational efficiency, a parallel algorithm will be developed and implemented for a hybrid mixture of distributed as well as shared memory.