SciTech

Cancer drug research

Credit: Art Staff Credit: Art Staff

During the month-long life of D. melanogaster, or the common fruit fly, the expression of its genes changes in dramatic ways to accommodate the differing needs of the organism.

In what is called a gene regulatory network, the expression of one gene may be in part responsible for the changes in the expression level of other genes. These networks, apart from coordinating the development of the organism, are also central to the functioning of certain diseases like cancers, in which the normal gene regulation of the cell is disrupted.

A better understanding of gene regulatory networks, and other complex systems, is being pioneered by the work of Eric Xing, an associate professor in Carnegie Mellon’s machine learning department, and his students.

In the July proceedings of the National Academy of Sciences, Xing and his student Amr Ahmed presented TESLA, an algorithm for inferring the hidden relationships between variables over time.

In the majority of existing techniques used to understand relationships over time, a static network is inferred for each time step, from samples of the directly observable relationships.

However, in many problems of practical interest, these relationships are not observable, and “this hidden relationship is something that does exist, and has an influence on the behavior,” Xing said.

In contrast to other approaches, TESLA infers a network of evolving hidden relationships without the requirement of being able to observe the relationships.

Xing and Ahmed applied their algorithm to the voting records of the U.S. Senate. Although the accounts of individual Senators may not be complete or truthful, “if people are related, then their behavior will be similar,” Xing said. The inferred network, indeed, showed two large cliques, the Democratic and Republican Senators, bridged by a few moderate and independent Senators.

Similarly, the ways that different sets of genes interact are not well understood, yet “the data is a proxy to the unobservable topology,” Xing explained.

Applying their algorithm to a set of samples of gene expression levels taken over the lifetime of a fruit fly, Xing and Ahmed were able to infer an evolving gene regulatory network. The network showed great changes in toplogy corresponding to the different developmental stages of the larva, pupa, and adult fly.

Such studies of life science are not new for Xing.

“After I graduated with a degree in physics,” he said, “I decided to use my knowledge to study life systems, because there are a lot of interesting open problems, and the solutions to those problems can improve human health.”

For a time, he studied cancer and several other diseases.

However, “it wasn’t enough to let me study complex biological systems, because there is no way to complete all the experiments that you want to do with human life spans and limited resources, which is what brought me to computer science.”

Now, his formal training in biology helps him to understand the motivations and goals of computational biology.

The work in TESLA is a stepping stone to the even more complex issues in computational biology.

“For large organisms, a lot of cells have the same DNA, but act in a different way — we think that in different cell types, the gene networks are functioning differently,” said Song Le, a Lane Fellow in computational biology at Carnegie Mellon who has been working with Xing.

By extending TESLA to include not only variations in time, but variations over space, Xing’s group hopes to shed light on cell differentiation — the way that the individual cells of a multicellular organism know how they should develop, becoming neurons, liver cells, or other types of cells depending on where they are in the body.

The overarching aim of the group’s project, Le said, is “to provide a set of tools for biologists to use networks to generate hypotheses and do experiments — and to use experimental results to refine the model.”