click to enlarge

This image of an x-ray crystal structure of rapamycin bound to the FK506-binding protein includes RAD001 (a Novartis analogue in clinical trials) superimposed over the rapamycin. (Source: Matthew Wortman of the Drug Discovery Center at the University of Cincinnati) 

From quantum mechanical techniques to docking algorithms, drug discovery grows increasingly dependent on molecular modeling. Today, some work in this area continues to rely on custom software, but a range of commercially available products also exist. Some of these products are providing expert knowledge even for amateur users. Overall, the trend is to bring increasingly accurate modeling to a wider range of users. In addition, some of today’s commercial applications take advantage of cloud computing to bring researchers even more modeling resources.

“With the numbers of new molecular entities that successfully progress to market remaining static over the past five years, pharmaceutical companies want to learn from that challenge and put the knowledge back into the discovery process,” says Adrian Stevens, product marketing manager, life sciences at Accelrys in San Diego, Calif.

Making the most of molecular modeling for drug discovery, however, takes know-how. “The challenge is to know where modeling software can contribute to this process,” says Modest von Korff, PhD, senior research informatician at Actelion Pharmaceuticals (Allschwil, Switzerland). He believes that an ideal compound for preclinical development should possess many qualities—such as high activity and specificity, good solubility, and so on—and many modeling tools exist for predicting these features. As Korff notes, though, a modeler must “be able to assess the quality of a solution.”

A range of applications
Despite such challenges, the pharmaceutical industry relies on molecular modeling. As von Korff states, “Structure-based molecular modeling delivers explanatory models for medicinal chemists and guides synthesis programs.” He also discusses other applications, such as using virtual screening to create focused libraries for biological testing. He calls predicting toxicity, side effects, and metabolites “the most ambitious modeling tasks today.”

Beyond gathering more information from molecular modeling tools, research teams need information-sharing capabilities. “They want to collaborate with peers in different locations, different countries,” says Frank Brown, PhD, chief science officer at Accelrys.

Scientists outside of industry also use today’s molecular modeling tools. “From an academic point of view, it’s a very cost-effective way to start a small-molecule discovery program,” says Ruben Papoian, PhD, director of the Drug Discovery Center at the University of Cincinnati. “We use it quite extensively when we have structural information on a target. We start with modeling and then virtual screening followed by HTS.” When that screening identifies hits, Papoian and his colleagues use modeling techniques to determine structure-activity relationships (SAR).

Overcoming obstacles


click to enlarge

Schrödinger’s Maestro software was used to create this image of the binding site of the recently solved D3 GPCR structure. (Source: Schrödinger) 

Even with today’s advanced technology, molecular modelers face several obstacles. According to Pengyu Ren, PhD, assistant professor of biomedical engineering at the University of Texas at Austin, the key areas that always need improvement are: computer power, algorithmic efficiency, and the physical models that describe interatomic interactions.

In many ways, these features compete. For example, as the models get more accurate, they demand even more computational power and more-efficient algorithms. For docking simulations, for example, Ren points out the need for approximations in several characteristics, including “how to handle water, the vibration of the molecule, the flex of the protein.”

Ren is working with colleagues at the Texas Institute of Drug & Diagnostic Development to improve the physical models for the accurate evaluation of protein-ligand interactions. For example, he’s testing an electrostatic model where the results reveal that “polarization plays an important—but somewhat unexpected—role, because it makes the attraction weaker in docking.”

The cloud and beyond
While some researchers keep simulations close to home, others will take advantage of cloud computing. As Woody Sherman, PhD, vice president, applications science at Schrödinger (New York), says, “Jaguar—our quantum mechanics package—can run on the Amazon EC2 Cloud and let a customer access thousands of processors simultaneously. That way, someone can accomplish extremely computationally demanding discovery work in a week that might have taken a year on a traditional in-house computer cluster.” Beyond Jaguar, all of Schrödinger’s software can now run on the cloud.

In addition to accessing more computer resources on the cloud, there are also significant improvements in the underlying molecular modeling methods. For quantum mechanical approaches, for instance, Sherman points out that “new methods have recently been developed to speed up the calculations and make them more accurate by including higher-order effects like dispersion and electron correlation without significant additional computational overhead.”

Algorithmic advances also improve other modeling techniques, including docking and structure-based virtual screening. “With faster and more accurate docking methods we are able to improve the enrichment of active compounds found in databases, which means researchers can find more diverse hits that might not have been found otherwise,” Sherman says. Even when researchers work with targets where the protein structure or binding site is unknown, structure-based approaches such as Schrödinger’s Prime for homology modeling, SiteMap to detect potential binding sites, and Glide for structure-based virtual screening, can still be used.

Searching for simplification
Early in 2010, Tripos (St. Louis, Mo.) released an updated version of its SYBYL-X molecular modeling suite. According to Brian Masek, PhD, product manager and lead scientist at Tripos, “This release includes many improvements in simple tasks—things that a modeler might do 10 to 15 times a day—by making them doable in one or two mouse clicks.” These tasks include things like displaying hydrogen atoms, showing or hiding other parts of a structure, the style of rendering proteins, and so on.

A follow-up release in October 2010, added other features to SYBYL-X, including what Masek calls “reworking the QSAR [quantitative SAR] workflow.” As Masek explains, “It now guides a user through the steps of a QSAR study, from getting the data to making predictions and delivering the findings to the team.” He adds that this update also “makes 3D QSAR much more accessible to the scientist who doesn’t always do QSAR. Masek adds, “A number of customers are now reporting successes in their discovery projects based on the new QSAR science in SYBYL-X.”

As some pharmaceutical companies merge and reduce staff, it is difficult to have modeling experts in every location. “So they need automation in modeling,” says Brown. “We have been able to take some of these classical modeling approaches —like QSAR—and use Pipeline Pilot [Accelerys] to ‘build in the expert’ to create fully automated model learning and validation platforms.” The automation of these reproducible stepped tasks lets the experts focus on the more complex ones, such as modeling protein-protein interactions.

The latest version of Accelrys’ Discovery Studio even includes automated tools for dealing with antibodies. “You can rapidly construct either full length—immunoglobulin-G based—models or a framework structure using separate heavy and light chain templates,” Stevens says.

For such complex tasks, researchers previously used multiple software programs. To remedy this, Accelrys is working with other vendors, Stevens says, “so that they can embed their tools into Discovery Studio or Pipeline Pilot. That way, a scientist can do all of the work in one environment.”

Overall, software tools for molecular modeling offer many advantages to drug discovery, from improving the quality of a library of compounds for biological screening to sharing data with colleagues. As hardware and algorithmic capabilities improve, the software power will grow even stronger. As that happens, the software must also provide increasing expertise—helping drug-discovery scientists make the best decisions.

About the Author
Mike May is a publishing consultant for science and technology based in Austin, Texas.