Take the protein challenge

A Santa Cruz team ranked among the world's best at a recent contest to predict how proteins fold

Next time you cook fettuccine, grab a handful and hurl it at a wall. The resulting mess will give you some idea of what a folded protein looks like. It's a drippy tangle of ribbons and loops, seemingly without order.

That metaphor goes soft when one considers real proteins. Their shapes, it turns out, are anything but random. Of the 100,000 flavors of protein in a human body, each has a distinct shape, and it always crumples into its own unique pattern. For decades, researchers have hungered to learn how proteins turn these tricks, because a protein's convoluted form dictates its precise function in the cell.

Bioinformatics offers a potentially powerful approach: predicting, with computers, how proteins fold. Such knowledge might help chemists design proteins for specific tasks, such as jamming the works of a nasty virus. It's a mightily complex problem, but a recent contest revealed solid progress toward solving it.

The international contest, called CASP2, drew 76 teams of researchers. Each team started with the same raw data--protein chains, spelled out as long strings of letters--and tried to predict how the chains would fold up. Independently, biochemists solved each protein's shape exactly with traditional and time-consuming lab techniques.

UCSC computer scientists spent long nights last summer comparing the new sequences to their huge library of known protein shapes using hidden Markov models (see main article). The team trained its intelligent programs to recognize familiar patterns within the protein chains, as well as subtle variations lurking within those patterns.

"Life tends to constrain itself to certain configurations," says graduate student Christian Barrett. "Our models reveal the statistical relatedness among proteins. They aren't scattered over infinite possibilities."

The work yielded several probable matches, which the team submitted to the contest. The results,
announced in December, placed UCSC on the top tier--even among competitors who have studied protein prediction far longer than David Haussler's group has existed.

Endocrinologist Olivier Lichtarge of UC San Francisco analyzed the contest. By the simplest statistical measure, he says, "UCSC predicted five folds extremely accurately, more than any other team." However, no team predicted an accurate model of an entire protein at the atomic level, the precision needed for useful drug design. That may change at CASP3, set for 1998.--Robert Irion

Computer scientist Kevin Karplus (left) and graduate students Kimmen Sjolander (center) and Christian Barrett made a splash at a recent protein structure prediction contest. For the protein shown, their prediction matched the real structure within the blue sections. Researchers from the European Bioinformatics Institute collaborated on the work.

Return to Summer '97 home page