Curing cancer, one computer model at a time

By Patricia DaukantasGCN StaffA lot can happen between a couple of proteins in one-millionth of a second.Understanding such interactions, uncovered by hundreds of hours of number-crunching on a supercomputer, might someday pave the way for drugs to fight cancer at its chemical core.'We're simulating on the atomic level various processes that proteins undergo, to see if there's something that can be used to therapeutic advantage,'' said biophysicist Stanley K. Burt, director of the National Cancer Institute's Advanced Biomedical Computing Center (ABCC) in Frederick, Md.The center recently leased a Cray Research SV1 vector supercomputer to create detailed simulations of the interactions between proteins and other biochemical molecules.Under a three-year, $6.5 million agreement with SGI, the laboratory will swap its 8-year-old Cray Y-MP vector supercomputer for an SV1 rated at 115 billion floating-point operations per second. SGI announced this month that it will spin off the Cray supercomputing line as a separate business unit.Burt said the lab took delivery of the four-node, 64-processor SV1 at the end of June. Later this year it will be upgraded to 96 processors, each rated at 1.2 GFLOPS.'Technology is changing so fast, we thought leasing would be the prudent choice,'' Burt said.The arrangement will keep ABCC open to Cray's planned SV2 architecture or other options that might arise, said Bob Lebherz, computer operations manager at the Frederick facility.The SV1 has 96G of memory and 1.2T of Fibre Channel storage, Lebherz said. A high-speed SGI GigaRing input/output subsystem connects the SV1 cluster's four nodes.NCI officials chose the SV1 largely for its compatibility with early supercomputers such as the Y-MP, said biochemist Jacob Maizel Jr., chief of NCI's Laboratory of Experimental and Computational Biology.Lebherz praised the SV1's checkpoint restart utility, which lets system administrators stop the supercomputer temporarily for maintenance and then bring it back up without having to restart a lengthy simulation from scratch. 'Checkpoint restart is just essential,'' he said.The calculations used in Burt's quantum-chemistry simulations require lots of storage for intermediate results and lots of access to the stored data. The SV1's local-memory storage gives better performance than scalar computers can, Maizel said.A typical quantum-chemistry simulation involves 500,000 lines of Fortran or C code to simulate tiny motions for only about one microsecond of real time, Burt said.Less powerful hardware had restricted the center's scientists to simulating only a billionth or a trillionth of a second of time. Simulations of a millisecond-long chemical reaction are still several computer generations in the future, Burt said.ABCC scientists base many of their simulations on a Gaussian algorithm originally developed by John A. Pople of Northwestern University. Pople shared last year's Nobel Prize in chemistry for his work in developing computational methods in quantum chemistry.Gaussian algorithms calculate everything from first principles and require about 16G of memory to run, Maizel said. Other molecular-mechanical simulations start with more assumptions and use less memory, but they have to run for a long time, he said.To study the large data sets produced by the calculations, NCI scientists use data analysis and visualization tools from Molecular Simulations Inc. of San Diego, Maizel said.The elastic interactions between two biological molecules depend on how their shapes fit together, Maizel said. 'These are very lumpy, bumpy molecules, and the lumps have to fit into the holes and sockets on the complementary molecules,'' he said. 'Calculating all that is a big job.''In his own research, Burt studies proteases'enzymes that break down proteins into smaller groups of amino acids known as peptides.Several years ago Burt and an NCI colleague, chemist Yuri Abashkin, performed computer simulations of proteases made by the virus that causes AIDS. Their simulation took about 1,500 CPU hours on the Cray Y-MP. The same model would probably take only 300 hours on the SV1, Burt said.When cancer cells metastasize, they use proteases to enter various tissues within the body. 'If you could understand how to block it, that would be good,'' Maizel said.The malaria parasite uses similar proteases to dissolve the hemoglobin that it lives on in the human bloodstream, Maizel said. Many other viruses, including those that cause the common cold, need proteases to replicate themselves.The Cray SV1 and ABCC's other high-performance computers are available to scientists from NCI, other divisions of the National Institutes of Health and general biomedical researchers, Maizel said. A 12-member panel representing the different groups reviews proposals and allocates time on the big machines. Most researchers submit their computational jobs from their own institutions rather than make the trek to Frederick.Unlike the nation's other publicly accessible supercomputer centers, ABCC focuses exclusively on computing that's related to biology, Burt said. Only recently have computers become big and fast enough to simulate biological phenomena, but such models already have had a large impact on the invention of new drugs, he said.