March 15, 2006 (Vol. 26, No. 6)
Using 3-D Modeling to Optimize Small Molecule Drug Discovery
Structure-based drug designers identify and optimize small molecules that affect the activity of target proteins, such as kinases, by understanding and modeling 3-D molecular interactions. Techniques have improved significantly, and those like x-ray crystallography, NMR, and computational modeling are now considered indispensable tools for all stages of drug discovery.
Novartis (www.novartis.com) Gleevec, an inhibitor of wild-type Bcr-Abl kinase, was a breakthrough discovery, according to Tomi K. Sawyer, Ph.D., senior vp, drug discovery, Ariad Pharmaceuticals (www.ariad.com). It established and validated the fact that you can make a small molecule inhibitor of a protein kinase and that it can work, in fact, quite dramatically.
Although Gleevec is a blockbuster drug for chronic myelogenous leukemia, clinical resistance can occur because the drug is ineffective against mutant forms of Bcr-Abl that patients may develop. Ariad is one of the few companies using structure-based drug design to identify next-generation Gleevec drugs that can inhibit these mutants.
It has been most challenging for companies to find small molecule inhibitors of a particular Bcr-Abl mutant, T315I, according to Dr. Sawyer. Threonine-315 resides near the ATP binding site and is thought of as a gateway residue that sits next to a fairly well-sized hydrophobic pocket, he explains. The change to isoleucine is just enough to compromise the ability of many inhibitors to optimally exploit their functional group binding in the hydrophobic pocket.
Ariad uses 3-D structural information, including x-ray crystallography and computational chemistry methodologies to identify clinically relevant inhibitors of Bcr-Abl. An early lead molecule, AP23464, showed potent inhibition of both oncogenic protein kinases Src and Bcr-Abl with the exception of the T315I mutation. A subsequent analog, AP23846, demonstrated improved pharmacology relative to AP23464.
W. Patrick Walters, Ph.D., senior research fellow, computational chemistry and molecular modeling, takes an integrated approach to library design at Vertex Pharmaceuticals (www.vrtx.com). While docking programs are essential, finding a small molecule that fits inside the active site is only 10% of the solution, he stresses. At the end of the day, you dont want just an inhibitor that fits in an active site, you want a drug that has an appropriate set of physical properties.
Vertex leverages its knowledge base to identify new molecules with desirable properties, says Dr. Walters. For example, scientists can probe hundreds of crystal structures available in company databases. The crystal structure of one or more lead compound, co-complexed with a protein of interest, can then be used to guide the placement of new molecules within the same active site. The company has developed software that utilizes this structure-guided docking approach to evaluate and prioritize large compound libraries.
Dr. Walters believes that combining its internally developed docking software with chemistry-based enumeration, filtering, and structure-based evaluations sets Vertex apart. The integrated software platform at Vertex allows chemists to go very rapidly from a chemical reaction to a model of a small molecule library in a protein active site.
Starting with a chemical reaction and a set of reagents, the software enumerates and filters the library based on a set of physical properties, such as the presence of reactive or potentially toxic functionality, aqueous solubility, or potential metabolic problems. The program generates 3-D structures for molecules surviving the filters and flexibly fits them into the protein active site for binding affinity prediction. Chemists then select a set of compounds to synthesize and test in biological assays, and the process is repeated until a compound with the desired property and activity profile is obtained.
Computational Model Using Torsional Space
Robert D. Clark, Ph.D., senior director, software research development at Tripos (www.tripos.com), believes the companys Galahad, a genetic algorithm that aligns a set of molecules and develops pharmacophore hypotheses, improves pharmacophore modeling. Launched in 2005, Galahad was created through a series of collaborations with Novo Nordisk, University of Sheffield, and Biovitrum.
To build a program that not only identifies critical similarities between molecules but also allows some partial matches, it required a fundamental change in the way programs work because its an inductive problem instead of a deductive problem, explains Dr. Clark. Galahad programmers took the problem out of Cartesian space, which focuses on xyz coordinates for each atom and put it into torsional space, otherwise known as internal coordinate space.
The Galahad program relies on measurements of pharmacophoric and steric tuplets, or triangles, within each potential conformation of each molecule. The corners of the tuplets represent different steric and pharmacophore features; for example, hydrogen bond donor, hydrogen bond acceptor, hydrophobe, positive, negative. The sides represent the relative distance between them. In that way, Dr. Clark says that Galahad creates a pharmacostere hypothesis, a molecular conformation that shares pharmacophore features, as well as similar shape.
Galahad employs Pareto multiobjective optimization to simultaneously balance steric, pharmacophoric, and energy information to calculate multiple pharmacophore answers for a particular ligand set. This means the program provides a mixture of answers, explains Dr. Clark, some of which may have the best energy, while others may have better steric or pharmacophore overlap.
Quantum Pharmaceuticals (www.q-pharm.com) protein-ligand modeling software employs first principle physics rather than statistics to calculate binding affinities, says Peter Fedichev, Ph.D., corporate scientific officer. Most software programs that calculate affinity constants take a statistical approach, whereby a training set of experimental data is used to create a predictive function for IC50s for new molecules.
Our approach is 100% not statistical, says Dr. Fedichev. A statistical model could miss strong binders that dont share enough features with the training set, he stresses. So, instead of using a statistical approach, we decided to build up a physical model that is based on quantum mechanics, on the physics of water, and on thermodynamics, to solve the problem from the other end.
Quantum created computational models for each component of protein-ligand interactions, tested them independently, and then combined them all together to make binding affinity predictions. For example, since proteins and small molecules interact in water and water produces important forces, Quantum modeled solvation effects. Instead of using known binding affinities as training sets, experimental data is used to test the accuracy of Quantums software.
When it comes to protein-protein interaction targets, our classic assays are insensitive, says Markus Schade, Ph.D., vp of NMR drug discovery, Combinature Biopharm (www.combinature.com).
NMR-based screening offers advantages for targeting shallow binding pockets, such as protein-protein interaction sites, and in the discovery and optimization of fragments, which are synthetic low molecular weight molecules, typically 300 Da and smaller. Although pharmaceutical companies are interested in fragments for many reasons, the smaller the molecule, the weaker the interaction, and the more challenging it is to detect this interaction.
NMR has three advantages for detecting weak interactions, according to Dr. Schade. It provides high sensitivity, direct structural information, and sufficient throughput to screen a mid-size library. He estimates that NMR can screen up to 1,000 compounds a day, while x-ray crystallography processes significantly fewer compounds.
As an example of Combinatures success with NMR fragment screening, Dr. Schade cites work identifying small molecule modulators of the PDZ domain, a protein-protein interaction domain that may play a role in cancer biology. In collaboration with H. Oschkinat, D. Phil, of the Research Institute for Molecular Pharmacology in Berlin, Dr. Schade identified three novel, chemically distinct classes of fragment hits by protein-detected NMR screening. He says that the 3-D structure of the complex between the PDZ domain and the best derivative confirmed binding to the active site and guides further lead optimization.
Structure-based drug design played a large role in the discovery and optimization of an inhibitor of the anti-apoptotic proto-oncogene Bcl that is planned to begin Phase I studies sometime in 2006, says Jonathan Greer, Ph.D., director, structural biology, advanced technology, global pharmaceutical R&D, Abbott Labs.
The company employed its SARbyNMR technology, screening approximately 10,000 fragments (250 Da or smaller) by NMR to identify weak binders to a desirable protein target, explains Dr. Greer. When two fragments are identified that bind to appropriate, adjacent sites on the target, each one can be optimized separately and then linked together. Once connected, the two fragments get the benefit of binding to two adjacent sites and become a potent, viable lead.
The rationale for SARbyNMR is that screeners can tap into vast chemical diversity without physically screening several million compounds. Structure-based drug design was also instrumental in optimizing other properties beyond potency in Abbotts Bcl-2 inhibitor, says Dr. Greer.
An earlier molecule in the series displayed high-binding to human serum albumin. Because of this specific interaction, less of the compound reached Bcl, the desired target, in vivo. After comparing the 3-D structures of the compound bound to the Bcl and albumin binding sites using crystallography and NMR, Abbott scientists added a substituent group to the compound that resulted in decreasing binding to albumin without affecting its potency against Bcl.