Intracellular pathway interactions can be convoluted, which makes modeling them particularly challenging. Speaking at the sixth annual RECOMB Satellite on Systems Biology, held at Columbia University in New York City in November, Harvard Medical School's Peter Sorger put the problem simply: "Signalosomes are highly dynamic," he said.
Modeling protein-protein interactions to unravel intracellular signaling invokes a trade-off. While "the most sophisticated models give you the best estimate [of the interactions] ... the more [biochemical] realisms you include, the less identifiability of the system," Sorger said. Identifiability is a statistical property that substantiates inferential power; a model is identifiable if, theoretically, its underlying parameter can be deduced with infinite observational data. "As we get more sophisticated with the underlying hypothesis, we have a greater difficulty building a rigorous framework" with which to construct identifiable models, he added.
One of the Sorger lab's goals is to determine how certain cancer therapeutics — like the EGFR inhibitors gefitinib and lapatinib — affect intracellular signaling differentially.
Ultimately, "the problem is experimental," Sorger said. "It's not either systematic or intuitive." To that end, he and Julio Saez-Rodriguez, now at the European Bioinformatics Institute, developed a computational approach for modeling protein signaling networks and predicting cellular responses to stimuli.
Sorger, Saez-Rodriguez, and their colleagues demonstrate the power of training models on experimental data in a December 2009 Molecular Systems Biology paper. At the meeting, Sorger said that his team mined the literature to construct a consensus model of human hepatocellular carcinoma cell signaling that they then subjected to a series of experiments. As it turns out, they found very few of the interactions included in the literature-derived network. When calibrated against their experimental data, the consensus model showed a "fairly large region in which the goodness of fit remains relatively large," Sorger said, but it "collapsed." Somewhat ironically, the team realized that the data-trained model in which "nothing is interacting is actually a better fit to the data than the model we started with," Sorger said.
Going forward "it should be possible to increase model identifiability by adding additional data" the team noted in its paper. At the meeting, Sorger said that "more sophisticated models will be required to understand the differences between drugs" in order to improve cancer therapeutics.