Computer-Conceived Chemical Compounds Make A Debut

The fledgling field of computer drug design, viewed skeptically by some scientists, gained dramatic validation when scientists from Agouron Pharmaceuticals Inc. reported at a March 18 cancer symposium in Amsterdam that they have created a new compound. Agouron claims the compound, dubbed AG-331, demonstrates "significant anti-tumor activity" in animal tests. "This is a completely novel chemical entity that was not found in nature," says Mike Varney, a computational chemist at San Diego-based Agouron. "It is not a biotechnology product. It is not a product cloned or screened. We truly built it, atom by atom." The study of how the shapes of molecules influence their propensity to react is an interdisciplinary field sometimes called molecular recognition. Working under this broad heading are structural and cell biologists; computer scientists; and theoretical, experimental, and computational chemists. Computational chemists owe much of their progress in developing software for modeling molecules to the seminal work of Nobel laureate Donald Cram, who in the early 1970s began making models of molecules with the only tools then available--hand-held models in which balls depicted the atoms and sticks represented the bonds between them. After synthesizing the molecules conceived in this fashion, Cram then reacted them with a target molecule. By comparing the X-ray crystallographs to the models, Cram came to realize that the shapes of any two given molecules helped to determine whether they would react. "They actually fit together, like a ball fits into a catcher's mitt," he says. For his work in proving this concept, Cram shared the 1987 Nobel Prize in chemistry. But even as the field of computational chemistry is growing, it is difficult to find synthetic organic chemists who are experienced in the modeling process, says Mike Varney of San Diego-based Agouron Pharmaceuticals Inc., who is looking for, but not readily finding, such a multitalented chemist. Among the places to get training in computational chemistry, according to Varney, are William Clark Still's group at Columbia University, as well as Paul Bartlett's chemistry group at the University of California, Berkeley, and Hoffmann-La Roche & Co. AG's molecular modeling group in Basel, Switzerland. An introduction to computational chemistry can be found in "The Design of Molecular Hosts, Guests, and Their Complexes" (D.J. Cram, Science, 240:760-67, 1988). Another good primer is "Computer-Aided Molecular Design" (J.A. McCammon, Science, 238:486-91, 1987). A treatment of the algorithms used in computational chemistry can be found in "Free Energy Calculations: A Breakthrough for Modeling Organic Chemistry in Solution" (W.J. Jorgensen, American Chemical Research, 22:184-89, 1989). The use of structural nuclear magnetic resonance as a complement to crystallography can be found in back-to-back articles dealing with the structure of cyclosporine A as revealed by NMR (Biochemistry, 30:6563-83, 1991). The drug-design methods followed by scientists at Agouron are outlined in an article entitled "Design of Enzyme Inhibitors Using Iterative Protein Crystallographic Analysis" (K. Appelt, et al., Journal of Medicinal Chemistry, 34:1925-34, 1991). The announcement by Agouron marks the successful convergence of significant advances in computer-based drug modeling with improved visualization of chemical structure via X-ray crystallography and nuclear magnetic resonance. Following its announcement, Agouron began preclinical toxicity testing of AG-331 and plans to file an Investigational New Drug application with the Food and Drug Administration later this year. In a traditional pharmaceutical approach, candidate drug compounds are found by screening available compounds in the laboratory. In computerized drug design, the screening is done using molecular modeling programs. These powerful programs help predict whether a proposed clinical structure is likely to work. For every false lead ruled out by this computerized process, drug designers can save the estimated $1 million cost of synthesizing a trial compound. "We are right on the edge of predicting the properties of molecules before we synthesize them," says William Clark Still, a computational chemist at Columbia University. Still is one of the leading researchers in a seven-year effort to use computers to predict how molecules will react. "That's what you have to do before you can really design a drug on the computer," he says. "That's the Holy Grail everyone is after." With AG-331, Varney and Agouron may already have found that grail. Varney says computerized drug designers begin by solving the structure of a target protein or enzyme using X- ray crystallography. They conceptualize the proposed drug molecule and then model it on a computer to see if it is likely to bind to the desired receptor. They then synthesize the most promising candidates and introduce them to the target receptors. By repeating the X-ray crystallography, they can observe how effectively the candidate binds and, if necessary, model a new compound and repeat the synthesis to achieve a greater affinity. In the case of AG-331, Varney says he wanted to inactivate the TS enzyme, associated with abnormal cell proliferation in solid tumors. To accomplish this, he designed a benz[cd]indole-containing inhibitor that binds at the folate-binding site of TS (Varney, et al., Journal of Medical Chemistry, 35:663-76, 1992). "State-of-the-art X-ray crystallography goes hand-in-hand with this process," says Michael Hudson, a biochemist at the University of Houston's Institute for Molecular Design. "We start with the characteristic we want to influence and work backward." But the difficulties inherent in crystallography are among the obstacles to further progress in molecular modeling, says Robert Sweet, a structural biologist at New York's Brookhaven National Laboratory. "A crystal must be absolutely pure; it needs conformity," Sweet says. "Biological molecules start with a fermentation broth. We are faced with the trouble of purifying what we want from a horrible brew." Sweet says a decade of interest in molecular structure has improved crystal-growing techniques and has led to the use of X-ray devices that reduce sampling time by using more light. He says one exciting innovation in the field was the emergence of dynamic crystallography, which Sweet likens to time-lapse photography. He cites David Eisenberg at the University of California, Los Angeles, and Greg Petsko at Brandeis University among those pushing developments in this area. At the same time, structural nuclear magnetic resonance (NMR) imaging has gained favor as an alternative to crystallography for some tasks. Ad Bax, a National Institutes of Health physicist, says NMR could reveal the atomic structure of a molecule in distilled water, avoiding the need to grow a crystal. "The assay is much easier to prepare with NMR," he says. Two-dimensional NMR has been used for 15 years, Bax says, but within the last five years the technology has taken off, as researchers like 1991 chemistry Nobel laureate Richard Ernst of the Swiss Federal Polytechnic Institute in Zurich found ways to add carbon-13 and other isotopes to molecular specimens. "Carbon-13 is NMR sensitive," Bax says. "It allowed us to spread the NMR spectrum into a three-dimensional picture." But Bax says current limitations in NMR technology cause images to grow fuzzy when the target molecule has an atomic weight greater than 30,000 to 40,000 daltons, which he describes as a small or medium-sized protein. "NMR is going through a very fashionable phase right now, but crystallography will always have its place," says Bax. "The core of a protein is better resolved by X-ray crystallography," and the technique can handle larger molecules than can NMR. Whether through crystallography or NMR imaging, once the structure of a target protein has been solved, the chemists go to work. At Columbia University's computational chemistry group, Still says, experimental chemists study the target protein and conceive of molecules that might bind to its activation site. These candidate molecules are tested in modeling programs that try to predict whether the candidate is likely to bind to the target. "Computer modeling became possible about five or six years ago as powerful workstations came into being," says Still. These machines use the reduced instruction set computing (RISC) microchip, which has vastly increased the speed of computing, letting chemists perform calculations that once required a supercomputer like those made by Minneapolis- based Cray Research Inc. "Workstations have about 10 percent of the computational power of a Cray [XMP] at 1 percent of the cost," Still says. Powerful as these new machines were, Still says the real break in computational chemistry involved a series of algorithms written by chemists like William Jorgensen of Yale University (see accompanying story), among others. These algorithms reduced the infinitely complex behavior of molecules to a problem that a fast machine could handle."It's a huge calculation to determine the binding energy and the affinity of every electron in your system," says Lynn TenEyck, a crystallographer and computer scientist at the San Diego Supercomputer Center at the University of California, San Diego. "You have to make a lot of assumptions and ignore a lot of the quantum mechanics." Molecules designed on computer and synthesized in the lab sometimes have an atomic structure that differs from that predicted by the model, and computational chemists start looking for flaws in their software. "That's where you learn the most," he says, "when your models fail and you have to figure out why." Ten-Eyck will be using part of a recent National Science Foundation grant to design a database to store data on molecular shapes, so researchers can draw upon each other's work. "The whole business of structural biology needs an enormous amount of database work, or we will drown in unrelated data," TenEyck says. TenEyck, who began in science as a crystallographer, is typical of the researchers who started the field of computational chemistry by becoming programmers themselves. But as the science matures, and the workstations grow more powerful and complex, computer scientists are increasingly being drawn into the field to wring more power out of existing hardware. "We have about 20 workstations in our group, and we usually use them alone," Still says. "Recently we have begun linking them in parallel." Using computers in parallel is a fairly new development requiring software that gives each machine a piece of the problem and combines these partial solutions into a complete answer. "The kinds of problems in computational chemistry are fairly easy to divide into pieces," Still says. "There are different phases of the problem whose solutions are independent." One example of the cross-disciplinary ferment enriching computational chemistry is the work of Tomas Lozano-Perez. As a computer scientist at the Massachusetts Institute of Technology, Lozano-Perez designed artificial intelligence programs to help robots learn how to assemble objects in outer space. He is now on sabbatical and working with Arris Pharmaceutical Corp. of South San Francisco, Calif., to adapt artificial intelligence technology to help screen molecular shapes in the search for pharmaceuticals. "The basic questions of biology and robotics are similar," says Lozano-Perez. "How do you recognize shapes? How do you define your data? The nice thing about this area for me is that everything is familiar and yet it is a lot different." Improved computer graphics have also increased the power of computational chemistry. "The traditional things you saw on a computer were sticks of light to represent the bonds between atoms," says Julie Newdoll, a University of California, San Francisco, microbiologist who specializes in the emerging field of computer graphics. "What's happening now is that people are using colors and three-dimensional shapes to make more realistic representations." UCSF's Computer Graphics Laboratory has produced a widely used program called MidasPlus, which takes the data that describe a molecule and translates this raw information into a color graphic that can be displayed on screen (T.E. Ferrin, et al., Journal of Molecular Graphics, 6:13, 1988). Newdoll says another UCSF program called Dock searches databases of synthesized compounds for a molecule with the shape and atomic characteristics that might fit the active site on a target protein. When Dock finds a possible match, she says, it loads descriptions of the candidate and target molecules into memory, displays them on screen, and allows them to be "docked" together. Some researchers, however, say that while molecular modeling on computers offers a valuable new research tool, it should not be overhyped. It will be five to 10 years before these methods deliver big payoffs, they say. "The methods are fairly new," says Bernard Brooks, a research chemist at NIH. "There has to be some caution." Biophysicist Ray Salemme, senior director for computational chemistry at Sterling Winthrop Pharmaceutical Research Division, based in Malvern, Pa., says he expects computational models will continue to evolve to take advantage of the more powerful computers becoming available. He says more research will be needed in the years ahead to understand systems far more complex than a single protein. "Alas, many of the biological functions that we would like to influence have to be affected over larger areas," Salemme says. "An autoimmune disease problem would hopefully be controlled by treating several parts of the system at once. A different kind of molecule is required, one that can affect several receptors or processes at once." Tom Abate is a freelance science writer based in Brooklyn, N.Y.

Computer-Conceived Chemical Compounds Make A Debut

Interested in reading more?

Become a Member of