Algorithm Predicts Interactions Between Proteins Whose Structures Are Unsolved

BUFFALO, N.Y. -- A promising new algorithm that can predict interactions between proteins whose structures are unsolved has been developed by Jeffrey Skolnick, Ph.D., University at Buffalo Distinguished Professor and director of the Buffalo Center of Excellence in Bioinformatics.

The research is published in today's (Nov. 15, 2002) issue of Proteins.

Called MULTIPROSPECTOR, the new algorithm takes protein interaction prediction to a new level because it works on proteins on which little structural information exists, providing three-dimensional models of the protein-protein complex and identifying the amino acid residues that interact.

According to Skolnick, the new method takes the entire field of structural genomics an important step closer to the ultimate goal of using detailed information about genes and the proteins they encode to design more effective pharmaceuticals.

"The overall goal," he said, "is to develop personalized medicine, which is based on understanding how a drug affects you versus how it affects me."

He noted: "With this paper, we are moving toward an understanding of how the whole system works, what's known as systems biology, which is the key revolution in the post-genomic era," he explained.

According to Skolnick, complexes of interacting proteins provide exciting and novel targets for potential new drugs.

"Right now, very few drugs exist that inhibit protein-protein interactions; most work against single molecules," he said.

But, he noted, the Protein Data Bank, the international "public library" of solved protein structures from which scientists draw data, contains not just isolated molecules, but in many instances solved compounds of two or more proteins interacting.

"Lots of cellular signals are mediated by these protein-protein interactions," he said, "and we want to know exactly who's interacting with whom. Often, the function of one protein can be deduced by studying the proteins with which it interacts."

Skolnick conjectured that perhaps there are millions of these interactions, a seemingly intractable problem.

But, he said, the process is greatly accelerated if you have a computational method that helps pinpoint the sites on the interacting proteins that will help scientists discover their role in biochemical pathways.

"That's what our method aims to do," he explained. "So, using our supercomputer, we can start to see how the path fits together, how this enzyme interacts with that small molecule or functions in a cascade of cellular processes."

The paper describes how MULTIPROSPECTOR was able to correctly predict protein-protein interactions between many thousands of proteins in brewer's yeast, a model organism in structural genomics.

Skolnick and his colleagues took what is known as a threading approach to the problem, in which an amino acid sequence is "threaded" through a library of protein structures that already have been solved.

But they take the threading process a step further. After finding matches for an amino acid sequence, the process goes through a second threading phase for both proteins, but this time a value is assigned for the interfacial energy, the surface energy between the proteins, revealing the stability of the interaction and thus, the likelihood that these are the structures that are interacting.

"We have built a sensitive interfacial potential that appears to often work at assessing interaction stability," said Skolnick.

He explained that predicting interactions between proteins provides scientists with an additional and important tool in reaching the point where genotype (what's happening genetically) can be linked to phenotype (what's happening clinically, i.e. what is the physiological manifestation of specific protein structures in a particular cellular pathway).

The research was conducted while Skolnick was at the Danforth Plant Science Center in St. Louis. The paper is co-authored by Long Lu, Ph.D., and Hui Lu, Ph.D., both of Danforth.