Study Points Way to Better Structure Modeling and Docking
By Claire Attwooll
A team led by scientists at The Scripps Research Institute has evaluated the ability of computational modeling laboratories to accurately predict protein structures when given just the protein sequence and the structure of a small molecule bound to it. The novel assessment shows that modeling and docking methods are advancing, and highlights new directions for this challenging endeavor.
An article published in the June 2009 issue of Nature Reviews Drug Discovery reports on an assessment of computational modeling using a newly solved G protein-coupled receptor (GPCR) protein, part of a class of molecules critical to health and medicine.
"More than 50 percent of therapeutic drugs target GPCRs, including allergy and heart medication, drugs that target the central nervous system, and anti-depressants," explains team leader Raymond Stevens, a professor in the Department of Molecular Biology at Scripps Research. "Yet only a few GPCR structures have actually been solved, meaning our ability to accurately model these structures is of extreme importance."
Although many mysteries remain about the structure of GPCRs, the functions of these extremely prevalent proteins are well known. Without GPCRs we would have no vision, no smell, no behavior or mood regulation. GPCRs are signaling molecules that span the membranes of cells, "sensing" chemical messages outside the cells and converting them into action within the cell. GPCR signaling is brought about by chemical messengers binding to the GPCR and causing a change in its shape. Much like ringing a doorbell, this lets the cell know what is going on outside.
Despite their importance as a class of molecules, the first human GPCR structure wasn't solved until breakthrough work led by Stevens and Stanford University Professor Brian Kobilka in 2007, and only one other GPCR structure (bovine rhodopsin) had been solved before that. So when Stevens solved a third GPCR structure—the human adenosine A2A receptor, also known as the "caffeine receptor"—last year (Science 2008; 322(5905):1211-7), he realized he had a unique opportunity to test the accuracy of predictions about the molecule's structure generated by computer modeling techniques against the experimental structure.
"There are hundreds more GPCRs that are all potential drug targets," Stevens says, "so we need to know how accurately their structures can be predicted."
Putting Computer Prediction to the Test
Drug design without target structure is a bit like carpentry without measurements, but if accurate computational modeling can correctly predict the structures of drug targets, rational drug design would become faster and easier. Since solving actual crystal structures can take months to years and is technically very challenging, computational models could potentially provide fast answers—but how good are they?
To find out, Stevens set up the new study, dubbed GPCR Dock 2008 Assessment, with colleagues Charles Brooks at the University of Michigan, Scott Dixon at Daylight Chemical Information Systems Inc., California, and John Moult at the University of Maryland Biotechnology Institute. Keeping Stevens' new structure confidential for 30 days in line with the Protein Structure Initiative guidelines, they alerted the computational modeling laboratories of the assessment to search for participants.
Each participant was given the amino acid sequence for the human adenosine A2a receptor and the chemical structure of the AstraZeneca Parkinson's disease drug candidate ZM241385 that was bound to the receptor when crystals were formed, a bit like being told the ingredients in a dish and then trying to figure out what it is without any further information. Participants were allowed to submit 10 models and had to rank each model in their predicted order of accuracy. In total the GPCR Dock 2008 organizers received 206 models from 29 groups.
Models were then assessed for accuracy based firstly on how many correct contacts were predicted between the protein and the drug bound to it; secondly on the difference between the position of the drug (the ligand) in the models relative to the crystal structure; and thirdly on how well the modeled receptors overlapped with the experimental structure.
Not surprisingly, the scores showed a range in their predictions. One key factor was the orientation of the ligand. In the two GPCRs whose structure was known, the ligand sits somewhat parallel to the membrane. In the human adenosine A2A receptor however, the ligand sits perpendicular to the membrane.
This assessment gives computational modelers clues about successful general approaches. "Models based on homology gave more accurate predictions than de novo modeling," says Stevens. "In addition, key insight into the importance of the extracellular loops and 4 disulfide bonds were under-estimated in terms of receptor structure and active site shape."
Perhaps more importantly however, the assessment allowed each participant to review their own ability to predict the structures.
Ruben Abagyan, a professor at Scripps Research whose lab focuses on computational structural genomics and molecular design, and a participant in the new study, notes, "There is an enormous need for accurate prediction and it is very useful to know if there is a method that can consistently do this."
Abagyan collaborated with two different associates, Vsevolod Katritch and Polo Lam at Molsoft LLC, submitting ten and three models, respectively. Their models ranked the best and the third among all submitted models by the number of correctly predicted atom contacts, and were just behind a model from Stefano Costanzi at the National Institutes of Health, by a combined quality measure. Furthermore, each group ordered their models by confidence. Abagyan is particularly pleased, that, in each of the two cases, his team's highest confidence models were indeed the best, and these models exhibited a strong predictive power in virtual screening for new receptor antagonists.
In addition to Stevens, Brooks, Dixon, and Moult, authors of the paper "Community-wide assessment of GPCR structure modeling and ligand docking: GPCR Dock 2008," are Mayako Michino and Enrique Abola, both of Scripps Research, and all GPCR Dock 2008 participants. For more information, see http://www.nature.com/nrd/journal/v8/n6/abs/nrd2877.html.
This work was supported by the Protein Structure Initiative grant, the National Center for Research Resources, and the National Institutes of Health.
Send comments to: mikaono[at]scripps.edu