BOSTON — Virtual screening and other computational methods are slowly becoming core components of the drug development process, but the field still needs to overcome some considerable challenges before it can hold its own against high-throughput screening and other experimental approaches.
At Cambridge Healthtech Institute's Structure-Based Drug Design conference here this week, speakers from several pharmaceutical firms shared some promising results that they are seeing from the use of these tools — as well as a laundry list of challenges that they'd like to see addressed.
"In spite of its limitations, virtual screening for lead generation is very useful," said Diane Joseph-McCarthy, principal research scientist in the chemical and screening sciences department of Wyeth Research.
Joseph-McCarthy said that Wyeth currently has five "advanced hits," one discovery lead, and one pre-development candidate that can trace their origins to virtual screening. Nevertheless, she noted, the approach is "still an emerging technology," and there is plenty of room for improvement — primarily in methods for sampling the conformational space of ligands, modeling protein and ligand flexibility, and scoring functions.
"Predicting binding affinities is very complicated. There's an awful lot going on when a ligand in water binds to a protein in water."
A number of speakers cited very similar challenges, with a common refrain being the lack of methods that are able to account for the flexibility of proteins and ligands. While methods are improving in this area, most docking tools still assume that either the protein or the ligand remains rigid during the binding process. Several speakers said that software should also be able to account for so-called "induced fit" effects, in which the ligand changes the shape of the active site upon binding.
Scoring functions present another area ripe for improvement. "Predicting binding affinities is very complicated. There's an awful lot going on when a ligand in water binds to a protein in water," said William Jorgensen, professor of chemistry at Yale University. The idea of a "simple" scoring function that can rank binding affinities in a few seconds is "lunacy," he said. "Don't expect perfection if you're going to spend a matter of seconds on the problem."
Another challenge is the lack of high-quality crystal structures for many targets of interest to drugmakers. Ying-Duo Gao, a senior research fellow in molecular systems at Merck, described a project the company began in 2001 based on the protein target 11-beta hydroxysteroid dehydrogenase type 1.
At the time, Gao said, the crystal structure for this protein was not available, so the company opted for homology modeling. A Blast search identified only two homologs, however, with only 20 percent sequence identity, well below the 50-percent threshold that serves as the rule of thumb for building a good-quality homology model.
Since the company had no other options at the time, "we decided to try our luck and see what the homology model would give us," she said.
As it turned out, Merck was able to use the model to generate a very promising lead series, Gao said. When the crystal structure was resolved in 2004, the Merck researchers compared it to their predicted model to find that their predicted active site was fairly accurate, even though the model as a whole was far from perfect.
The conclusion, according to Gao, is that even low-homology models can be of use in drug development when no other options are available, and "a partly correct pocket can do a lot of good work for docking and design."
The idea of a "simple" scoring function that can rank binding affinities in a few seconds is "lunacy. Don't expect perfection if you're going to spend a matter of seconds on the problem."
Some researchers are using computational methods to identify novel binding sites on well-studied targets. Jeffrey Wiseman, vice president of technology and informatics at Locus Pharmaceuticals, said that his firm used its fragment-based computational platform to find a new binding site in p38 MAP kinase, a very well-known target that has a high risk of toxicity because one in every five kinase inhibitors targets the highly conserved ATP binding site, leading to off-target effects.
Wiseman said that Locus has identified an allosteric binding site on p38 for which only one in every 200 kinase inhibitors shows selectivity. The company has also used its computational platform to design a number of potential inhibitors that it is currently optimizing.
Not every pharmaceutical firm is sold on the promise of computational drug design, however. During his talk, Kenton Longenecker of Abbott Laboratories discussed the use of protein crystallography but omitted any mention of computational methods — a relative rarity in a program full of virtual screening and molecular dynamics. When asked in the Q&A session following his talk why he didn't discuss the company's computational tools, he said that "there's an emphasis on experimental feedback at Abbott more so than at some other places."
When pressed as to whether Abbott doesn't trust computational methods in the drug development process, Longenecker declined to comment.
— Bernadette Toner ([email protected])