 |
|
| Back Issues |
|
|
|
|
ACS Members can sign up to receive C&EN e-mail newsletter.  |
|
|
 Join ACS |
|
||||||||||||||||||||
|
FROM THE ACS MEETING
|
||||||||||||||||||||
|
ADME/Tox computational screening early in the process could make drug discovery more effective and much less costly It's often said that the three most important factors to consider in a real estate transaction are "location, location, location." For modelers of drug pharmacokinetics, though, it's "data, data, data."
But although new in silico methods may be replacing some of the lab work, accurate models of something as complex as biological chemistry hinge on a high quality and quantity of experimental results. The value of computational screening becomes obvious when considering the cost and risks of drug development. Of the estimated $600 million cost of bringing a new drug to market, more than $400 million of that is wasted pursuing leads that turn out to be dogs. In fact, one-half of potential drugs fail because of ADME/Tox issues that crop up late in the game, noted Osman F. Güner, senior director of lead identification and optimization at Accelrys in San Diego and organizer of the symposium. Add to that the fact that high-throughput and combinatorial drug discovery methods are increasing the numbers of potential drug leads that com- panies want to explore, and the risk grows even larger. "IT'S TERRIBLE when you get [to the clinical stage], and after all the expense you incurred and all the hopes that you raised, something dies at that point because of something you hadn't anticipated," Jay T. Goodwin, a senior research scientist at Pharmacia in Kalamazoo, Mich., said at the symposium. "So the whole industry is focused on dealing with this issue."
Just recently, ADME/Tox modeling was a fledgling effort--as reflected in a symposium on the same topic at an ACS meeting two years ago (C&EN, June 5, 2000, page 63). Back then, the problems that computational chemists were tackling were much simpler--predicting solubility or absorption, for example. "Today," Güner said, "we are working on much harder problems--things like drug metabolism and toxicity; these are very, very difficult to do." Computational methods are already well developed for other aspects of drug discovery. Techniques such as modeling QSAR (quantitative structure-activity relationships), which help identify molecules that might bind tightly to drug targets, have provided a springboard for models of ADME/Tox properties. And a seminal tool for weeding out bad compounds early in the drug development process has been the so-called rule of five, developed by computational chemist Christopher Lipinski and his colleagues at Pfizer in 1997. Lipinski's group developed a set of criteria for molecules that are likely to be well behaved in the body. The name "rule of five" stuck because each of the four criteria is a multiple of five: A good drug molecule should have a molecular weight of under 500, fewer than 10 H-bond acceptors, fewer than five H-bond donors, and a C log P value (a measure of water solubility) of less than 5. This and other approaches represent a new "property-based design," as opposed to purely structure-based drug design, noted Han van de Waterbeemd, director and head of drug metabolism technology, at Pfizer Global Research & Development in the U.K. Chemists are now considering pharmacokinetics, "not just which molecule is the tightest binder," van de Waterbeemd said. "Since we have a much better understanding now of what basic molecular properties are important for pharmacokinetics and toxicology, we can talk about a property-based approach," he added. Guidelines like the rule of five are particularly useful during the early stages of drug discovery. "You want the predictive model to be very fast so that you can apply these predictions to very large numbers of compound libraries," Güner said.
BUT AS A POTENTIAL drug gets closer to clinical trials, speed is less important than accuracy, Güner said. And although models can be developed that are more applicable to a broader swath of chemistry, the more generic they are, the less accurate they may be in general. "This is like trying to do everything for everyone and ending up not satisfying anyone," he said. At the other end of the spectrum are attempts to rigorously model ADME/Tox for one particular class of compounds. "While this approach improves the accuracy of the prediction for the given class of molecules, we are then challenged to develop different and multiple models for different classes of compounds," Güner said. Chemists are now stepping up to that challenge with more complex and sophisticated models. A standard laboratory model of cell permeability is the human cancer cell line CACO-2, which behaves much like intestinal epithelial cells. Curt M. Breneman, who is a chemistry professor at Rensselaer Polytechnic Institute, is using a relatively new statistical pattern-recognition strategy known as a support vector machine to model the permeability of CACO-2. And combinations of long-used computational techniques include a blend of three-dimensional molecular interaction fields calculations with the pattern recognition program SIMCA (soft independent modeling of chemical analogy). This combination by Tripos chemist Philippa R. N. Jayatilleke and colleagues allows them to get a better handle on a molecule's structural properties. Other aspects of ADME/Tox need more attention than models, said Peter C. Jurs, chemistry professor at Pennsylvania State University. "Model-developing methods are probably ahead of where they need to be," he said. Where there's room for improvement, he said, is in generating better, more realistic descriptors--or collections of molecular properties or physical structures--that can be used to describe complex molecules in just a few "dimensions." There are innumerable varieties of descriptors from as many different research labs, using combinations of features such as hydrogen donators or acceptors, positive or negative charges, or an aromatic ring center.
For example, computational chemist Michelle L. Lamb at Deltagen Research Laboratories in Redwood City, Calif., and her colleagues devised a version of so-called shape-feature descriptors, which include not only chemical features but 3-D conformations. Lamb's group also applied descriptors known as pharmacophores to help weed out compounds that have an affinity for the membrane transport molecule P-glycoprotein [J. Med. Chem., 45, 1737 (2002)]. These molecules wouldn't make good drugs, because P-glycoprotein transports them outside of cells. Because drugs are metabolized in the liver, hepatotoxicity is of particular concern to ADME/Tox modelers. Liver problems are frequently the culprit when drugs are yanked off the market. For example, Rezulin (troglitazone), manufactured by Parke-Davis/Warner-Lambert--a drug used to treat type 2 diabetes--was discontinued in 2000 because of liver toxicity concerns. Various companies are working to find ways to predict whether a compound will damage a human liver. Computational chemist Ailan Cheng at Accelrys' Princeton, N.J., laboratory and Paul E. Blower, of LeadScope in Columbus, Ohio, reported methods for identifying hepatotoxic compounds. And Yulia V. Borodina with the Russian Academy of Medical Science in Moscow described her institution's software package for modeling drug metabolism. Researchers are also attempting to predict the inhibition of various human cytochrome P450 (CYP) enzymes--a class of enzymes responsible for the metabolism of more than 50% of all known drugs. For example, Carleton R. Sage of Lion Bioscience presented a model for the enzyme CYP 3A4, and Roberta G. Susnow of Accelrys presented one for the enzyme CYP 2D6. "Predicting a metabolic profile is one of the more challenging tasks in this field," Güner noted. "It's great to see that several groups around the world are now tackling this difficult problem." BUT THE MODELS are only as good as the data upon which they're based. Unlike ab initio modeling, which calculates molecular properties based solely on quantum mechanical principles, ADME/Tox calculations are semiempirical. Hence, everyone involved in developing models is clamoring for more data from experimentalists to flesh out ADME/Tox models. However, even when there's data to be had, getting access to it is an issue for some researchers. Because many companies want to protect their databases, large chunks of knowledge remain fragmented. "Companies need to release data so academics can develop better models," Johann Gasteiger, chemistry professor at the University of Erlangen-Nuremberg, told his largely industrial audience at the meeting. "From the point of view of an academic, getting access to data with the right characteristics is difficult, and it is a necessary step to doing something useful and meaningful," Jurs agreed. To that end, Güner said, Accelrys has been attempting to form a consortium with other companies in order to pool proprietary data for building models. But the concept presents difficulties. "Companies are not willing to share their data, and if they are, they want to control the data," he said. "Unless we have a way of somehow maintaining a proprietary relationship with each company's data, we may have a problem there."
PREPACKAGED ADME/Tox Predictions Move From Concept To Software Just a short time ago, drug developers could only imagine ways to toss out drug failures early in the game. Now, first-generation predictive models for adsorption, distribution, metabolism, excretion, and toxicity (ADME/Tox) screening are a standard part of the drug development process in many companies and have also found their way into the marketplace. "Things have come a long way in a few years," noted William Jorgensen, chemistry professor at Yale University. "Back then, the idea was more conceptual--we were more visionary," agreed Osman F. Güner, senior director of lead identification and optimization at Accelrys. "Whereas right now, these tools are being used in industry." In the effort to help whittle down the list of potential drug failures early on, numerous companies have designed software packages that calculate various ADME/Tox physical properties, particularly absorption and solubility. These include ClogP from BioByte; Absolv from Sirius Analytical Instruments; iDEA from Lion Bioscience; C2.ADME, part of the Cerius2 software package, and TopKat from Accelrys; and Solubility DB from Advanced Chemistry Development. And from Jorgensen comes a relatively new package, called QikProp, which will screen 150,000 to 200,000 compounds in an hour. Marketed by Schrödinger, QikProp calculates a wide range of properties based on the whole molecule, as opposed to fragments, thus enabling it to make predictions about new or unrecognized structures, Jorgensen said. Speakers at the ADME/Tox informatics symposium at the Orlando ACS meeting, organized by Güner, charted the steady ADME/Tox modeling progress of industry and academia. Chemistry professor Johann Gasteiger at the University of Erlangen-Nuremberg noted that his software package PETRA's ability to predict a molecule's pKa using empirical methods is on par with predictions using ab initio, or quantum mechanical, methods. Chemists are moving into new terrain with programs to predict liver toxicity, one of the most important properties to determine in the early stages of drug development. Ailan Cheng, a computational chemist at Accelrys, told attendees about the success of her program in identifying potentially hepatotoxic compounds. And Paul E. Blower, a computational chemist at LeadScope, has found ways to analyze drug databases to find chemical structures that may correlate with liver toxicity. Russian scientists are going a step further and taking a stab at predicting the result of a drug's first metabolic pass through the body. Yulia Borodina, a chemist with the Russian Academy of Medical Science, said PASS (which stands for prediction of activity spectra for substances) predicts the hundreds of properties for a metabolized compound on the basis of its structural formula.
|
||||||||||||||||||||
|
Chemical & Engineering News |
||||||||||||||||||||
|
||||||||
|
||||||||
|
||||||||
|
||||||||
|
||||||||
|
||||||||
|
||||||||
|
||||||||
