The computational scientist who founded the world’s largest hedge fund says he is two years away from building a new breed of computer that could solve the Holy Grail of structural biology – simulating the process of protein folding and protein-drug interactions.
David Shaw was an assistant professor in computer science at Columbia University in the early ’80s, until he was lured to Wall Street by Morgan Stanley. In 1988, he launched his own investment bank, D.E. Shaw & Co, which he calls “a computational hedge fund.” Today, Shaw & Co. manages more than $21 billion in assets. The company uses closely guarded computer algorithms to detect and capitalize on price inequalities in international stock markets. (Among the firm’s alumni is Jeff Bezos, founder of Amazon.com.) Shaw also served as a scientific advisor to former President Clinton.
But after turning 50 a few years ago, Shaw relinquished day-to-day control of the company in search of a new challenge. “As the firm grew, I felt myself getting stupider” and bogged down in administration, Shaw told a packed lecture theater at MIT this week. Fascinated by the potential of applying advances in computational science and microprocessor design to biology, Shaw returned to the laboratory, joining the Center for Computational Biology and Bioinformatics as an adjunct professor of biomedical informatics. He has assembled an interdisciplinary group of some 40 computational scientists, architects and engineers, working alongside mathematicians, biologists and chemists, to produce a quantum leap in molecular dynamics simulations. This, he hopes, will help him to emerge as “one of the soldiers in the war against cancer and other diseases.”
Shaw’s research agenda that combines major advances in computational algorithms and architecture to provide the requisite thousand-fold increase in processing speed to simulate protein folding and protein-protein interactions in a biologically meaningful timeframe. This is a field in which “computer scientists and architects can make an impact,” Shaw said. “We don’t know protein structures, especially membrane proteins, or how they fit together, or how the whole machine works.”
Unlike experimental methods, which are “tough, wet and sloppy – very hard work,” Shaw prefers the computational approach to producing the “gold standard” of molecular dynamics simulations.
In simulating protein folding, Shaw explained that time must be divided into a large number of discrete slices (on the order of femtoseconds). Then integrate Newton’s Laws of Motion; calculate the pairwise forces of atoms (with separate terms for stretching, bending, torsion, electrostatic and Van der Waals forces): move the atoms; and repeat or iterate thousands of times, for thousands of atoms in a typical protein, to simulate the natural curved motions of the molecules.
Shaw presented a movie of one such example – the interaction of a defensin protein with a bacterial membrane. It took “3.4 CPU days to simulate 2 nanoseconds” – simply “too slow,” said Shaw, to learn how the protein disrupts the bacterial cell, but providing some preliminary clues nonetheless. Other applications include protein folding, protein-protein interactions, drug-target binding (such as Gleevec), and mechanisms of intracellular machines.
The goal is to produce simulations of complex proteins (64,000 atoms), over biologically relevant timescales of a millisecond. But to do so requires an enormous number of discrete steps, and a 1,000- to 10,000-fold speedup in processing, depending on the processor configuration. Today, it takes 100 milliseconds of compute time to model 1 femtosecond. “What takes so long?” Shaw asked rhetorically. Modeling interactions between 64,000 atoms, repeating the process 10*12 times. The goal is to reduce that requirement to 10 microseconds.
To achieve this, Shaw’s team is working on novel algorithms and designing new hardware.
New Architectures
Shaw says a new computer, based on parallel architecture, special ASICs (application specific integrated circuits), and designed specifically to tackle molecular dynamics problems, should be complete in 2008. Supercomputers such as Blue Gene earned Shaw’s praise. “Blue Gene is much more ambitious, a great project,” said Shaw, “one of the best things I’ve ever seen.” But it’s still not fast enough for molecular dynamics.
A key component of the new architecture is “choreographed communication,” which Shaw explained sees that “data flows just where needed,” dispensing with external memory storage and caches. New circuit designs are devoted for modeling specific forces and interactions, utilizing the areas that would be used for caches, wires and so on. The computer should ultimately have 512 nodes (one ASIC/node).
New Algorithms
Last fall, Shaw published a new method for the parallel evaluation of pairwise interactions that slashes the volume of data shared between processors. (The report appeared in the Journal of Computational Chemistry.) The NT (neutral territory) algorithm takes advantage of a novel formula for calculating atomic interactions that greatly reduces the computational load.
Reluctant to divulge too many details, saying a paper had just been submitted for publication, Shaw presented a brief video simulation of the export of sodium ions by the Na/H antiporter membrane protein. The charge of specific transmembrane residues has major consequences on the direction of sodium ion movement, Shaw said.
Shaw acknowledged there were many problems to be addressed, and said he’s always worrying about something. But he hoped that his work could provide new insights into drug development, helping to solve the problems such as Gleevec resistance or angiogenesis.
“It’s very risky, lots can go wrong, but very interesting,” said Shaw.