© 2005 AAAS

Schematic of Bayesian network inference using multidimensional flow cytometry data. Bayesian network analysis of the data from flow cytometry of 11 phosphoproteins and phospholipids in individual cells extracts an influence diagram reflecting the causal relationship between signaling network molecules.

Researchers have spent decades determining how proteins interact with each other in complex signaling networks by studying these relationships one at a time in isolation. This approach may have been necessary, but the resulting maps are accordingly suspect. "The average signaling map is a composite of data from everything from nematodes to yeast ... and although we draw an arrow between two molecules, whether it works in any given cell type and the strength, and timing and conditions are very contextual," says Stanford University researcher Garry Nolan.

Using a largely computational approach, Nolan and colleagues at the Massachusetts Institute of Technology quickly and successfully reverse-engineered...


Nolan says the study's success was due to the nature of the dataset. Flow cytometry enabled simultaneous measurement of multiple protein levels in individual cells, eliminating noise via population averaging and providing a much larger dataset than could be obtained from traditional techniques such as Western blotting. In addition, perturbations to the system generated data indicating that changes in the levels of one protein affect other proteins in the system. This data allowed the statistical system to add directionality into the arrows it placed between molecules.

Using truncated datasets the researchers showed not only that the Bayesian analysis thrives on sheer data quantity, but also that the perturbation data was key to predicting causality. Relationships derived from Bayesian analysis of datasets that lacked the perturbation conditions lacked accuracy in predicting the directionality of associations between molecules. Nolan likens the system to a delicate spider's web: "To understand a spider web you have to step back from it and pull on the individual strands and measure the effects at a distance to understand the relationships between individual components."

The networks in this study could not predict cyclic relationships, such as those obtained in the many signaling systems that rely on feedback loops. Larry Lok, of the Molecular Sciences Institute in Berkeley, Calif., explains that this is due to the theoretical framework on which it is based. "Causality is a kind of one-directional thing, and a thing can't normally cause itself."

But Lok was pleased with the mathematical application. "It's very satisfying when a new mathematical analysis yields results that look reasonable and reassuring, but also uncover a few new and unexpected things as well." Lok says he believes the system might be adapted to take time into account. "You can expand the data in time and still apply a similar Bayesian analysis, called dynamic Bayesian networks. Computationally it's more intense; whether it's practical or not remains to be seen."

Nir Friedman, a computer science professor at Hebrew University in Jerusalem, agrees that the temporal aspect "would be nice, but we can still learn a lot without it." He adds that the study has promise for determining how signaling systems in human cells differ in diseased and normal states.

Interested in reading more?

Magaizne Cover

Become a Member of

Receive full access to digital editions of The Scientist, as well as TS Digest, feature stories, more than 35 years of archives, and much more!
Already a member?