CELIA M. HENRY, C&EN WASHINGTON

"We're in a very target-rich but lead-poor post-genomics era for drug discovery," notes Raymond Stevens, a profes-sor of molecular biology at Scripps Re- search Institute and cofounder of the San Diego-based biotech firm Syrrx. "The question is how are we going to handle all this target information. Structure-based biology or design has been a powerful tool, but it's been too slow and not reliable enough." Many companies—Syrrx included—are out to change that reputation.

	uxS—a
	DRUG TARGET L LuxS—a metalloenzyme involved in bacterial quorum sensing, a form of bacterial intercellular communication—is a target in Structural GenomiX’ antibacterial therapeutic program. It is shown bound to the small molecule methionine.metalloenzyme involved in bacterial
	COURTESY OF STRUCTURAL GENOMIX

In structure-based drug design, the three-dimensional structure of a drug target interacting with small molecules is used to guide drug discovery. "Structure-based drug design represents the idea that you can see exactly how your molecule interacts with its target protein," says Raymond Salemme, founder, president, and chief scientific officer of Three-Dimensional Pharmaceuticals in Exton, Pa. This structural information can be obtained with X-ray crystallography or nuclear magnetic resonance spectroscopy (NMR). Ideally, these two techniques complement one another. However, most companies that are specializing in structure-based drug design focus on only one method of structure determination, at least initially.

Originally, structure-based drug design was equated with de novo design or building a molecule from the ground up. The active site of the protein was a space to be filled with a molecule that complemented it in terms of shape, charge, and other binding components.

"For a variety of reasons, that turned out to be a difficult task, not the least of which was the fact that you were relying on imperfect computational docking routines and energy calculations such that the molecules that you came up with often did not conform to the expectation in either rank or potency when you made them," says Richard Ogden, senior director of scientific development at Agouron Pharmaceuticals in San Diego, now a subsidiary of Pfizer.

 "THE INITIAL expectation of structure-based drug design—that you were going to be able to design molecules and they were going to work right out of the box—was unrealistic," Three-Dimensional's Salemme says. "We didn't understand the thermodynamics well enough. We weren't really making molecules that had all the properties of drugs. They were good inhibitors, but they weren't drugs. People got disenchanted with structure-based drug design. They flipped over to the other side and said, 'We don't have to think about this. We can just make a lot of molecules with combinatorial chemistry.' I think that was rapidly shown to be an ineffectual strategy, mainly because no matter how many molecules you can make with combinatorial chemistry, it is still infinitesimally small compared to the possibilities."

However, combinatorial chemistry is the best thing that ever happened to structure-based drug design, Salemme believes. "The problem with combinatorial chemistry in the early days was that you were making huge numbers of compounds, many of which were irrelevant," he says. Purely random combinatorial chemistry has evolved into focused combinatorial libraries, which can be considered an "abstract form of structure-based drug design," Salemme says. Researchers realized that specific knowledge of the target could guide the power of combinatorial chemistry to rapidly make many compounds. "We recognized this eight or nine years ago and built the entire architecture of our company around that idea," Salemme says.

AUTOMATION Cofounder Raymond Stevens is shown here with Syrrx's crystallization robot, which helps speed the process.
PHOTO BY ANNE HAMERSKY

Structure-based drug design is at its most powerful when coupled with combinatorial techniques. "The big revolution has been to use combinatorial chemistry as a way of parallelizing the synthesis of molecules that can then be empirically checked through some sort of activity or binding assay to rapidly sort through all of these good ideas to find out which ones really work. Then you can go back and get the crystal structures of those and further refine the ideas," Salemme says.

"Clearly, having a lot of synthetic material in the form of combinatorial libraries in hand allows you to get into the experimental realm much more quickly," Agouron's Ogden says. "I think most companies that do this incorporate high-throughput screening as initial steps in a structure-based design program."

One of the driving forces behind structure-based drug design is lead optimization, says Michael Milburn, vice president of structural biology at Structural Genomix, based in San Diego. "Structure is a really good way of quickly getting a handle on how the lead compound binds to the target of interest and what one might be able to do with chemistry to modify the molecule to get the desired properties," Milburn says.

A major development that has given structure-based methods a place of prominence in drug discovery has been increased speed. "Maybe 10 years ago, structure-based drug design was slower than some would wish in discovery in terms of being able to be on the same timescale of moving a project forward," Milburn says. Now, he notes, hundreds of crystals of complexes can be analyzed rapidly.

To have a "profound impact" on the iterative process of optimizing compounds through medicinal chemistry, structures of ligands bound to target must be solved within one to two days, says Kent Stewart, an associate research fellow in molecular modeling at Abbott Laboratories, Abbott Park, Ill. The longer it takes to solve the structures, the less impact they have on the drug discovery process.

STRUCTURE-BASED drug design can help lead to better compounds more quickly, asserts Regine Bohacek, the former head of drug design at Ariad Pharmaceuticals and now the president of Boston De Novo Design. (Ariad is one of her major clients.) Ariad has had success in using structure-based methods to discover inhibitors for such refractory targets as the SH2 domain of Src, a tyrosine kinase implicated in osteoporosis and other bone-related diseases. "Companies, including Ariad, have spent a lot of time doing combinatorial chemistry and screening internal databases to come up with ligands that bind to the SH2 domain, with little success" Bohacek says.

"Using our methods, we were able to increase the binding affinity and improve our compounds dramatically in a very short time," Bohacek says. "If you enlist thousands of chemists and let them make anything they want, certainly eventually you'll come up with tight-binding compounds, but it might take years and years. We can screen compounds on the computer. We're beginning to be able to predict which ones will bind well to the binding site." At Ariad, the binding mode of synthesized compounds to the target is quickly determined with X-ray crystallography and NMR. The agreement between the predicted and experimental structures validates their computational methods.

 ALTHOUGH SPEEDING UP drug discovery is certainly one of the goals of structure-based drug design, Structural Genomix' Milburn cautions against thinking of it in those terms. "What's important to remember is that structure-based drug design helps you get to quality molecules, molecules that have better pharmacological chemical properties. You have a much greater awareness of how those compounds bind to the pocket, so you have a much greater awareness of what parts of the molecule you can modify," he says.

	DOCKED Using structure-based methods, Ariad has designed high-affinity compounds for Src SH2. Shown is the disphosphonomethyl group of one of Ariad's compounds docked in the receptor binding site. This picture was made several months before the molecule was synthesized. Subsequent experimental data indicated that this predicted binding mode to Src SH2 was correct [J. Med. Chem., 44, 660 (2001)].

3-D structures produced by X-ray crystallography have become an integral part of the drug discovery process because of changes that have increased throughput. High-throughput crystallography involves several stages: protein expression, purification, and crystallization, followed by the X-ray structure determination and the necessary data analysis. Syrrx, Structural Genomix, and Cambridge, England-based Astex Technology are basing their structure-based drug design programs on high-throughput X-ray crystallography. They are looking at each step in the process, trying to figure out how best to streamline the process.

"The approach we've taken [at Structural Genomix] is to look at the whole process from gene to structure," Milburn says. "At each step, we've looked at what automation, what processes, can be put into place to streamline it as much as possible."

Syrrx has focused on miniaturizing, parallelizing, and automating the crystallization process. "One of the things we try to be careful about is automating where it helps us and not automating where it's not necessary," Stevens says. Connecting every portion of the process is unnecessary, he says. Instead, Syrrx takes a modular approach. So far, Syrrx has focused on technology development, and each module is at least second generation.

Syrrx' "protein production factory" incorporates robotics in several parts of the process, which was designed by engineers from the Genomics Institute of the Novartis Research Foundation (GNF), La Jolla, Calif. "We sat down with the engineers and looked at the whole pipeline," Stevens says. The engineers pointed out which stages would be prime for automation. For example, Syrrx produces 96 proteins in parallel. "Growing up one protein versus growing up 96 proteins isn't that different if you have the technology that allows you to parallelize the whole thing," Stevens notes.

Syrrx is able to use smaller volumes—requiring milliliters rather than liters of cell culture—for protein production because crystallographic information can be obtained from smaller crystals than was previously possible. Syrrx obtains structures from crystals approximately 40 m in size. In particular, improvements in synchrotron beam lines, which have become more focused and more intense, have enabled more rapid structure determination.

The availability of dedicated synchrotron beam lines has been very important to both Syrrx and Structural Genomix. Syrrx shares a dedicated beam line with GNF at the Advanced Light Source at Lawrence Berkeley National Laboratory. Structural Genomix has its own synchrotron beam line at Argonne National Laboratory.

"One of the problems with synchrotron beam lines is that the time is very scarce," Stevens says. "We decided to automate taking frozen crystals and mounting them on the beam line. You used to have to go in one at a time, set off alarms, and mount your samples. With the engineers at Lawrence Berkeley Laboratory, we developed a system that's able to take a large number of crystals all at once and put them on one at a time."

Syrrx is interested not just in structure determination but in what those structures can lead to. "We're interested in drug discovery, and it's a vital, absolutely critical component that we be able to get not just de novo structures but large numbers of cocrystal structures in a timely manner," Stevens says.

 HIGH-THROUGHPUT crystallography allows previously unthought of experiments to be carried out, Structural Genomix's Milburn claims. For example, libraries of compounds can be screened against targets. "The whole idea of lead optimization, which is currently a big driver for structure-based drug design, changes when you think about having a platform that can [determine structures] at that level," he says.

Originally, Structural Genomix focused on developing a comprehensive database of protein structures and models of structures. Although the company still adds to the database, the company's thinking is "more expanded," Milburn says. The company will use the structures it generates to become an expert in particular protein classes.

 IN ADDITION to crystallography, molecular modeling will play a stronger role at Structural Genomix in the future. The company acquired San Francisco-based Prospect Genomics, a computational genomics company, in April. "We're very interested in using all of the structures that we're determining in-house with what's available in the public to design as many homology models as accurately as possible," Milburn says.

Structural Genomix has focused on obtaining targets in bacterial genomes for development of anti-infectives. They have solved many crystal structures, which are in the company's database. In addition, the company is moving forward on solving structures of eukaryotic targets. However, the company does not yet have any potential therapeutics moving forward, Milburn says.

Interpretation of electron density data is a key area of development for Astex Technology, says Harren Jhoti, founder and chief scientific officer. The company has developed software called AutoSolve that analyzes and interprets electron density data automatically. Many of the structures of interest in structure-based drug design are those of ligands with targets whose structures are already known. The challenge is to determine the part of the structure where the ligand binds. "We've developed software that actually identifies electron density due to the ligand without the need for an experienced X-ray crystallographer having to look at the data, which is quite time consuming," Jhoti says.

The rapid data interpretation allows Astex to use crystallography in a different way than people have previously, Jhoti says. "Most people have used crystal structures of proteins for lead optimization. That's because it's such a slow technology when you're using conventional tools to be able to generate many protein-ligand structures," he says. "Because we can solve many hundreds of structures in a couple of days, we can actually use X-ray crystallography almost as a screening tool."

At Astex, computational methods are used for virtual screening to predict compounds that should bind to a target, which are then experimentally screened. These two components—computational and experimental—make up what Astex calls structural screening. The company prefers not to be viewed as a structural genomics company focused on the acquisition of crystal structures, but rather as a drug discovery company that uses crystallography as a tool.

Astex is using a combination of computational methods and high-throughput X-ray crystallography to detect the binding of small fragments, each one with a molecular weight of only 100 or 200 daltons, to protein targets. "If you fragment large drugs or druglike molecules, it's almost like a Lego approach to drug design. You can fit them all together and build large molecules," Jhoti says.

To improve protein production, Astex is developing small-molecule chemical agents that assist in the proper folding of the proteins. "Quite often when you're trying to express mammalian proteins in bacterial expression systems, you get high expression, but the protein is misfolded. We're able to recover a lot of that protein due to these small-molecule refolding agents," Jhoti says.

PREDICTING BINDERS Using fragment affinities, Locus Discovery identifies protein binding sites computationally and assembles potential inhibitors in the binding site. Using elastase as the target, the molecule ONO-5046, which is in Phase III clinical trials in Japan for emphysema, was among the predicted inhibitors. The structure shown here depicts the predicted binding within the active site.

Crystallography is not the only technique being used for structure-based drug design. NMR spectroscopy is the other main instrumental method in use. NMR is slower than crystallography for obtaining structures. In addition, NMR cannot generally be used to obtain three-dimensional structures of proteins larger than 30,000 Da.

Vertex Pharmaceuticals, Cambridge, Mass., is developing novel NMR methods for structure-based drug design. Vertex also has large groups in crystallography and computational chemistry, in addition to its work in NMR. "We've always been an industry leader in structure-based design," says Jonathan M. Moore, a senior research fellow and NMR spectroscopist at Vertex. "We're trying to apply that knowledge and those technologies that we've developed across entire gene families, as part of what we call a chemogenomic approach." Knowledge of one target is used to find potential inhibitors of other targets that are in the same gene family, such as kinases, but are involved in different biological pathways and therefore different diseases.

 ONE OF VERTEX'S NMR methods is a fragment-based approach known as SHAPES [Chem. Biol., 6, 755 (1999)]. (Meant as a joke about the tortured acronyms that are often used to describe NMR methods, the name SHAPES doesn't stand for anything.) The original SHAPES library was based on work done by Guy W. Bemis and Mark A. Murcko in Vertex's computational chemistry group. They were trying to design computational filters that would select druglike compounds. To help determine what made a molecule druglike, they combed the Comprehensive Medicinal Chemistry database and cataloged the compounds according to their frameworks.

At the simplest level, compounds were classified by shape descriptors that included their ring structure and linker with no regard for atom type or bond order. At this basic level, 32 frameworks encompassed half of all known drugs. When atom type and bond order were included, 41 frameworks represented nearly a quarter of known therapeutics [J. Med. Chem., 39, 2887 (1996)].

Up to that point, SHAPES was strictly a computational method. Then NMR got involved. "We realized that we could take this library of a very small number of compounds and use NMR to determine binding," Moore says. "One thing about NMR is that the technique itself can be insensitive at times, but it's very sensitive for detecting weak interactions, particularly between small molecules and a macromolecule."

Considering that the goal in developing therapeutics is to find inhibitors that work at nanomolar concentrations, a desire to find weak interactions might seem counterintuitive. However, weak binders can be modified to improve both their binding affinity and their druglike characteristics. Astex's Jhoti spent 10 years in a large pharmaceutical company where, he says, "we all got seduced by potency. It became apparent that once you found those [nanomolar] hits, those really potent leads, you couldn't go anywhere with them anyway. They tended to be too large. You were getting very high binding because you had many interactions going on. The downside was that you couldn't modify that lead particularly well to overcome other problems such as pharmacokinetics." Weak binders are difficult to detect with bioassays, but Jhoti says that they can be detected using Astex's high-throughput crystallography methods.

Starting with weak binders doesn't always work, but sometimes it can result in a new class of lead compound. One such example at Vertex was its program to find inhibitors of JNK-3, a type of kinase. "What we found in that program was that early high-throughput screens didn't give us satisfactory results with respect to number of tight-binding hits and the chemical classes represented by those tight-binding hits," Moore says. "After the SHAPES screening, we identified a number of compounds based on the SHAPES hits. We came up with some low micromolar inhibitors, which were then through straightforward medicinal chemistry optimized into nanomolar inhibitors with a few simple modifications."

Triad Therapeutics, a San Diego biotech company, uses NMR as the centerpiece of its drug discovery platform, which it calls integrated object-oriented pharmacoengineering, or IOPE. Triad focuses on biligand enzymes, which require cofactors for catalysis. Rather than study one enzyme at a time, Triad develops libraries for entire families of enzymes. The first step of IOPE is to select a gene family and find a druglike mimic for the cofactor. "The characteristic of any given gene family is that all members use the same cofactor for their catalysis," says Stephen Coutts, president and chief operating officer of Triad. "Since each member of the gene family uses the same cofactor, we call it a common ligand."

The next step in the IOPE process is to use NMR SOLVE—which stands for structurally oriented library valency engineering—to figure out where to attach a linker to the mimic so that it's directed into the enzyme's second site, where the substrate binds. This second site gives the enzyme its specificity.

"With that information in hand, we then build a library of molecules that are biligand. Each member of the library has something in common, the common ligand mimic, and something unique—a diversity element at the end of the linker that we so carefully placed. We end up with a library of biligands, most of which bind weakly to every member of the gene family but only a few of which will bind specifically and with high affinity to a given target within the gene family. Taking advantage of two-site binding, we can get into the nanomolar range of affinity very quickly," Coutts says.

 "APPROACHING drug design in an object-oriented way enables a massively parallel approach to drug discovery, where inhibitors are designed for many [hundreds] of protein targets at once," says Daniel Sem, Triad's vice president for biophysics. "Object-oriented strategies have revolutionized the software industry by increasing efficient development of software. Triad is translating these concepts into the realm of drug design to make the process more efficient when dealing with a large number of related protein targets. The concept is to reuse discovery and development efforts when possible by defining molecular objects that are starting points for drug leads for many targets."

To increase the likelihood that its libraries contain druglike molecules, Triad plans to do preliminary ADME (absorption, distribution, metabolism, and excretion) studies on the common ligand mimic to make sure that it is somewhat druglike and that it causes no toxicity problems, Sem says.

Triad's use of NMR is unique, claims Sem. "There's a big focus in the industry to use NMR as a screening tool or in the traditional way where complete structures are done. Although we do the former, we're using it more to get structural information in a very fast way on protein-ligand interactions," he says. "The way we approach the problem, we essentially remove the molecular weight limitation for most proteins. We've looked at proteins as large as 170 kDa and obtained structural information. We're focused just on the binding site, so we don't have to assign all of the atoms in the proteins."

Triad's NMR method depends on identifying reporter residues in the interface between the two binding domains and using those reporter residues to guide the chemistry of linking the diversity element with the common ligand mimic. The reporter residues are identified by selective isotope enrichment of particular amino acids with deuterium, carbon-13, and nitrogen-15, explains Maurizio Pellecchia, an NMR spectroscopist at Triad.

	SUCCESS STORY Agouron's AIDS drug nelfinavir (brand name Viracept) is one of the few drugs on the market that can be traced directly to structure-based methods. Here, the molecule is shown in the active site of HIV-1 protease.

"If we know the three-dimensional structure, we may have a good idea which amino acids we want to selectively label," Pellecchia says. "If we don't know the three-dimensional structure, we do random selective labeling and find the amino acids that are in the binding site." The SEA (solvent-exposed amides)-TROSY method, which was developed at Triad, also eliminates the problem of spectral overlap and simplifies NMR spectra [J. Am. Chem. Soc., 123, 4633 (2001)]. SEA-TROSY is an extension of the TROSY method invented by Kurt Wüthrich (C&EN, Aug. 10, 1998, page 55).

"We record a series of reference spectra with known inhibitors or the known cofactors. We compare the spectra of unknown ligands with the spectra that we measured with the known cofactors. We don't need to know the three-dimensional structure of the enzyme because we can orient an unknown ligand with respect to the known cofactor," Pellecchia says.

"It's as if we're using the reference ligand to map out the binding site. Then we learn how other ligands bind relative to that reference ligand," Sem explains further.

Another novel aspect of NMR SOLVE is the speed with which structural information is obtained, Sem believes. "We can get crude structural information faster than we can generate potential inhibitors for assay. We can really be part of the SAR [structure-activity relationship] process with medicinal chemists," he says. 

MASS SPECTROMETRY, although a popular technique in the pharmaceutical industry, is not often associated with structure-based design. Yet Ibis Therapeutics, Carlsbad, Calif., a division of Isis Pharmaceuticals, is using mass spectrometry to find compounds that interact with RNA.

Ibis originally targeted RNA using traditional high-throughput screening approaches, with little success. "We weren't getting very high hit rates, and we weren't getting good structure-activity information about the molecules that we were screening," says Eric Swayze, director of medicinal chemistry at Ibis. "We weren't able to get anywhere in a drug discovery program. We weren't finding good leads. We started thinking that perhaps our assay methodology wasn't providing us with the right information because of the unique nature of the target. Some very bright people—not me—started thinking about using mass spectrometry as an assay methodology. It turns out to be ideally matched for RNA targets," Swayze says.

With RNA, a subdomain of the RNA molecule can be snipped out and maintain its three-dimensional structure. A mixture of the RNA substructure and the potential small-molecule ligands is transferred into the mass spectrometer by electrospray ionization. Ligands that bind to the target are identified by mass shifts. The observed mass equals the mass of the target plus the mass of the ligand. Multiple ligands and even multiple targets can be screened in a single assay as long as they have different masses.

The mass spectrometry assay can also be used to map the binding location. A complex between the target and ligand is isolated in the mass spectrometer and is fragmented with a laser. The fragmentation of the complex is compared to the fragmentation of the unbound target.

"You can see where the small molecule protects the macromolecule from breaking apart," Swayze says. "Where it's protected from breaking apart, that's a pretty good indication that the molecule is bound in that vicinity. It's not like a high-resolution NMR structure or a crystal structure, but it gives you an idea of where the molecule is binding to the target. The advantage is that the mapping experiment takes only a few seconds. This allows us to map the binding locations of many ligands, and we can focus our efforts on ones that bind in the most promising locations."

Ibis uses the mass spectrometry assay as a primary screening tool. Its advantages are that it provides high throughput and high-value information, Swayze asserts. As many as 50 compounds can be screened simultaneously against three to five targets, with a total analysis time of less than one minute per mixture. Furthermore, the MS-based assay has a wide dynamic range and can detect both strong and very weak (millimolar) interactions with RNA, in contrast to typical high-throughput screening assays directed against RNA targets, Swayze says. This allows structure-activity relationship patterns to emerge, even for relatively weak ligands—information that can be used to design better compounds.

Mass spectrometry can also be used to find two compounds, both of which bind a target, that can then be linked to provide a higher affinity compound. "The mass spectrometer is particularly well-suited for that," Swayze says. "If two molecules are bound to the same target at the same time, they must be binding at different locations. If you can figure out roughly how far apart those locations are or where they are on the target, you can think of ways to make one molecule that uses both of those binding motifs." 

MEANWHILE, Locus Discovery is trying to eliminate much of the experimental aspects of structure-based drug design. The Princeton, N.J., company—soon to move to Blue Bell, Pa.—is a spinoff from Sarnoff Research Institute. The company's drug discovery platform is based on proprietary computational algorithms that predict compounds that will bind to targets.

"Starting with the crystal structure of a protein, the algorithm correctly identifies the relevant binding site on a protein and starts building small molecules to agonize or antagonize the functional activity of the protein," says William R. Moore, Jr., Locus' vice president of research and development and chief scientific officer.

Starting with the crystal structure of the protein, Locus uses a fragment-based approach to identify potential binders. "We have a palette of fragments of organic molecules, and we let the fragments combine with the proteins," Moore says. "We accurately calculate the free energy of binding of the fragments to the proteins. It's not a scoring function; it's not enthalpy. It's truly free energy." In addition, the algorithms take into account the water layer around the protein. The compounds are ranked in terms of relative affinity, and the free energies of all can be calculated if the affinity of one compound is known. 

IN THE MODEL, the fragments are usually observed to cluster on the protein surface at the binding site. Interestingly, the fragments lie in close proximity without superimposing. "These fragments cluster together in the binding site in close enough proximity that they can be connected chemically into unique small molecules. It's actually like doing combinatorial chemistry within the active site of a protein of interest," Moore says. "A chemist with a chemist's intuition could look at the fragments and the way they bind in the active site and connect them in a way that makes chemical sense, preserving their orientation within the active site. We also have software that does that, called Locus Chemistry Definition. We impose the chemical constraints from within the software. It takes into account bond length, bond angles, torsional angles."

A palette of 20 fragments can result in upward of 10⁵⁰ molecules, Moore says. Currently, Locus runs more than 100 fragments against a protein. "We've gone into the chemical literature and tried to cull all of the unique chemical fragments that are available. The chemistry that we can produce within the active site of a protein is immense and diverse," Moore says. "For many refractory targets, where the pharmaceutical industry has not been able to find suitable leads, we can find leads."

Locus is trying to ensure that the molecules its algorithms generate are viable leads by using fragments that appear in known therapeutics. "We think that if we bias the fragments toward the fragments that appear repeatedly in drugs, we will be building molecules that are more druglike," Moore says. "We'd like to get to the point that the molecules coming out are almost clinical candidates."

In an attempt to ascertain the attainability of such a goal, Locus will synthesize some of the molecules predicted by its software and determine their pharmacokinetic properties. The company has already started a database of the ADME properties of the fragments themselves.

Locus has verified that its methods generate molecules that are real drugs. For example, when the fragment library was run against HIV protease, the Merck drug Crixivan was one of the compounds generated. "The interesting result with HIV protease is that we came up with many unique structures that are predicted to have higher affinity to the target than Crixivan," Moore says. "It was actually more difficult to go in and find the structure of Crixivan from all the molecules that came out than to come up with completely unique high-affinity molecules." Similarly, methotrexate was one of the compounds generated when the fragments were run against dihydrofolate reductase. 

LOCUS IS CURRENTLY focused on ramping up its computational power. The algorithms are run on a supercomputer cluster of 196 parallel processors. By the end of the year, Locus will have 1,000 computers.

Sometimes structure-based methods can rescue a project from being terminated. That happened several times during the course of a project at Abbott Laboratories, says Stewart. The enzyme neuraminidase had been chosen as a target in the company's influenza program.

"In the neuraminidase project, there were some confusing elements to the structure-activity relationship that in the absence of 3-D structure really did not make sense at all," Stewart says. "Over a period of a few months of effort, we were able to make sense of that data and to put forth a working hypothesis for explaining that data with specific suggestions for workarounds of the problems we were facing."

Every project requires making closely related analogs and distantly related analogs, Stewart says. "You hope that the SAR makes sense. When we made little baby steps, sure, the SAR made sense. When we made larger structural changes, the SAR did not make sense. 3-D structure showed us that larger changes in the inhibitor molecule were causing new binding modes to occur, something that would have been unanticipated in the absence of 3-D structure."

Another area where structure-based methods are being applied is in the prediction of ADME and toxicity properties (ADME/Tox) (C&EN, June 5, 2000, page 63). "We've put a lot of effort already in trying to see if we can't get a shortcut in the ADME/Tox area" at Astex, Jhoti says. Astex recently announced a research collaboration with AstraZeneca in which they will solve crystal structures of small molecules interacting with cytochrome P450, which is involved in drug metabolism in the liver. "We'll be able to see exactly how these compounds are interacting with some of these P450s," he explains. "Then we'll be able to see how we can change the compounds so that we can mediate that interaction. You don't want to obliterate it because you'll have other problems."

In similar efforts, Syrrx scientist Duncan McRee has solved the structure of one P450 molecule, 2C5, which is being used by several pharmaceutical companies for in silico metabolism studies, Stevens says.

Despite all the efforts in structure-based drug design, only a few drugs currently on the market can be traced directly to these methods. Two examples are Viracept (nelfinavir) from Agouron and Agenerase (amprenavir) from Vertex.

The people working on structure-based drug design expect their methods to shave time off the drug discovery timeline. "So far, the biggest successes have been in the initial phase, which involves the discovery of potent compounds," Bohacek says. "Optimization of these leads to produce candidates with oral bioavailability and low toxicity remains a major opportunity and challenge for computational chemistry. Considerable effort is being directed toward solving these problems."