PAGE 1 OF 2
by Karyn Hede
Drawing on computer power from volunteers worldwide, David Baker sees "a community-based solution to a long-standing scientific problem."
Most scientists cultivate collaborations to advance their work. But we're betting only David Baker has colleagues in Andorra, Belarus, and the Pitcairn Islands. Not to mention the rest of the world.
An HHMI investigator at the University of Washington, Baker relies on collaborators worldwide to help him uncover nature's rules for protein folding—the process by which a protein shapes itself to fulfill its function. Determine a protein's structure, researchers believe, and you can learn how this essential biological machine works. But getting to that point, from a mere amino acid sequence, requires computer power on a gargantuan scale. That's where Baker's far-flung friends come in.
Software that Baker and colleagues created using the BOINC (Berkeley Open Infrastructure for Network Computing) distributed computing platform taps participants' PCs during downtime—the computers perform protein-folding calculations while their owners, in effect, sleep. Harnessing that vast capacity, Baker has made considerable progress in the quest to compute protein structures from their sequences of amino acids. Progress has been so good, in fact, he now predicts that many, if not most, protein structures and interactions will one day be computable—a level of confidence that protein researchers have previously lacked.
Three recent accounts from the Baker lab (published in the September 16 and October 28, 2005, issues of Science and the August 1, 2005 issue of Proteins) report that given enough computing power his protein-modeling software, called Rosetta, can produce protein models that look like their natural counterparts at least about a third of the time. And Baker's results with determining the structure of a protein once it "docks" onto a partner are even better. Together, the papers demonstrate that it's possible to achieve high-resolution prediction of protein structure by first sampling a large number of potential variations at low resolution and then refining the best candidates with modeling that accounts for all of the atoms in the molecule.
Photo: Brian Smale