An In Silico Time Machine
Anton: A computer dedicated to molecular dynamics simulations.
In biology, many exciting events happen on the millisecond timescale—proteins fold, channels open and close, and enzymes act on their substrates. Atomic-level simulations of this duration are beyond the reach of current technology, but a new specialized computer called Anton—described in the July 2008 issue of Communications of the ACM—may change all this. Slated to be operational by the end of the year, the machine is projected to speed up molecular dynamics simulations 100-fold.
The basic goal is to be able to visualize, at the atomic level of detail, an entire biological trajectory, such as an anti-cancer drug (like Gleevec®) inactivating its target enzyme, says David E. Shaw, PhD, chief scientist of D.E. Shaw Research, the independent research laboratory that is creating Anton, and a senior research fellow at the Center for Computational Biology and Bioinformatics at Columbia University. Because it provides what might be thought of as a computational microscope, Anton is named after 17th century scientist Anton van Leeuwenhoek, known as the father of microscopy.
“Our machine only does molecular dynamics. It does it blindingly fast, but it’s pretty brittle and isn’t designed to do anything else,” Shaw explains. In molecular dynamics simulations, time is broken into discrete steps, each a few femtoseconds (10-15 of a second) of simulated time. At each step, the computer calculates the force exerted on each atom in the system (typically 25,000 to 100,000 atoms) and updates its position and velocity. The various time steps cannot be executed simultaneously since each is dependent on the previous, but Anton uses 512 highly specialized chips working in parallel to speed up the massive calculations within each step.
“They’ve done a beautiful job, and there are a lot of intellectually interesting aspects to the approaches they’ve taken,” says Vijay Pande, PhD, associate professor of chemistry at Stanford University and director of the protein folding distributed-computing project Folding@home. Still, Pande advocates a different approach. Rather than simulating one long trajectory, which could take a million days on one general purpose computer, he simulates a large number of shorter trajectories and then merges them together with a clever algorithm. This may take just 10 days on 100,000 computers. “The approach not only gives access to long timescales, but having many trajectories allows you to do statistical testing, which you cannot do on a single trajectory,” Pande says. “Most of the questions that people in the field are interested in are inherently statistical questions,” he says.
But according to Shaw, “The two approaches are very complementary and I think they may turn out to be useful for solving very different types of problems.” Combining many smaller trajectories is more efficient, he says. “But there are some cases in which you’d like to have confidence that what you’re seeing is one continuous, unbiased, physically realistic trajectory.”
Though other groups have previously attempted to develop specialized computers for molecular dynamics simulations, most efforts have failed to stay ahead of Moore’s Law, which says that the speed of general purpose computers doubles every 18 months.
“The Shaw group’s effort has been one of the most exciting examples of trying to do that to date,” says Pande. “Since the machine isn’t out yet, it’s too early to say whether they have succeeded or not. But they’ve got a reasonable shot.”