Karen Sachs and Douglas Lauffenburger, computational biologists at MIT (Cambridge, MA), wanted to use Bayesian networks to model biological systems. These graphical models have been used to predict transcriptional network structures based on mRNA levels.
The problem for those interested in signaling, says Sachs, was that the method is probabilistic, and thus requires many independent samples. This is what Omar Perez and Garry Nolan (Stanford University, Stanford, CA) could provide, with the help of a very fancy flow cytometer. They were able to label 11 reagents (antibodies recognizing phosphorylated proteins and lipids) with fluorophores that could be detected simultaneously in single cells. Having battled through that “tremendous challenge,” says Sachs, “in the blink of an eye you can get thousands of data points.”
The team measured the response to a number of perturbations, and used the results to build their model. The Bayesian algorithm threw out redundant linkages and assessed candidate models for simplicity and accuracy. The final model of T cell signaling had 17 linkages, all of which were reported previously, although two of those with little evidence. Three known linkages were missing from the new model, probably because of a lack of feedback in the Bayesian model.
Future challenges will include detecting how the model changes under different conditions, especially in response to diseases and drugs. In tackling less well-characterized pathways, the Bayesian system will be able to deduce connections successfully even if the linkages are indirect.