By Alissa Poh

Oct. 16, 2008 | By taking an engineering approach and “breaking” a computational model designed to illuminate cell signaling pathways, scientists at MIT have made some surprising observations about cellular communication that could prove useful in improving disease treatments such as chemotherapy. Their findings appear in the October 17 issue of Cell.

“We asked if we could apply traditional failure analysis to a biological model, using a variety of computational tricks,” says Michael Yaffe, the paper’s senior author and a faculty member at MIT’s Koch Institute for Integrative Cancer Research. Engineers are very familiar with failure analysis, using it to explore, for example, the ways in which a bridge might get overloaded or otherwise crumble, so as to improve its overall design.

Yaffe and colleagues had previously designed their computational model to allow simultaneous investigations of relationships between five key signaling pathways: MAP kinase, p38, JNK, NFkB, and PI-3K/AKT.

“Unlike a typical computational model, which would be one where you pretended you understood everything about some biological process, then wrote a bunch of equations to describe it, we took what we knew were key components and just collected a lot of data,” Yaffe explains of the actual model-building. “There were no prior assumptions; we simply let the data speak for itself. And our data-driven model turned out to be remarkably able to predict cell survival or death.”

But even though it worked, their model was still essentially one of correlation, says Yaffe, who is also affiliated with the Broad Institute of MIT and Harvard, and with Beth Israel Deaconess Medical Center. “It might say, ‘I see a signal at a certain time, and it seems to correlate strongly with survival,’ but it told us nothing about causation.” So Yaffe and the paper’s first author, Kevin Janes, decided to make use of their training in classical engineering and perform a “model breakpoint analysis.”

To “break” their model, they simply made the data inputs increasingly implausible. “Say, for instance, that the range of NFkB signals went from one to 10, depending on different conditions,” Yaffe elaborates. “We kept that full range, but either saturated the signal so it started off at one and went up to 10 almost immediately, so all the intermediate values were basically eliminated; or we desensitized the signal so it stayed very close to one until the end, when it suddenly shot up to 10.”

The researchers wanted to know whether their model could still predict life or death, even with bad data. What they found was “surprising,” Yaffe says. “I expected that if the data got worse, the model would just worsen along with it. But it didn’t happen that way at all. The model worked fine even as we made the data worse, then suddenly at one point – either through desensitization or saturation, depending on the stimulus used – it would fail catastrophically. Different types of bad data made the model fail in different ways.”

The team gained their biggest insights, in taking this approach, by not merely shrugging off the model’s failure. Instead, they asked whether it was telling them something important about the biology involved, analyzing what had changed from the point just before failure to the actual breaking point. “We found that in each case, our increasingly implausible data led to the model leaving something out – probably an important signal, so the model was no longer able to follow this particular molecule – or inadvertently including something else,” Yaffe says.

Perhaps their most surprising finding was the model’s prediction that, looking in particular at MAPKAP-kinase 2 (MK2) – a gene “classically associated with driving cell death” – less cell death occurred with both saturation and desensitization “breaking” approaches, than if the molecule was simply left to its own devices.

“I could understand how, if the molecule caused cell death, desensitizing it would lead to less cell death,” Yaffe says, “but it didn’t make any sense to me why increasing its activity would also produce the same result. So I didn’t believe [the model]; that might be what it said, but it wasn’t what common sense was telling me.”

Yet, when the researchers came up with an experimental way of doing what the model had predicted – using mutant forms of MK2 that were overactive, or very difficult to activate – they had to concede that the model was right, as they observed a decrease in cell death either way. So perhaps in tumor cells, depending on the molecule(s) involved, both inactivating and hyperactivating mutations can interfere with the process of cell death.

“I think what this means is that in situations where you have a lot of signaling networks acting together, you can’t make simple conclusions like, ‘If I over-express a protein that causes cell death, I’ll get more cell death,’” Yaffe says. “It may be true in certain cases, but I don’t think it’s a general statement.”

In other words, cellular signaling networks may be adjusted so that what a cell measures isn’t necessarily the absolute activity of any one protein, the researchers say. Cells primarily care that a molecule has a wide range of activities, uniformly distributed across a lot of cellular space. Or, to use Yaffe’s analogy: “It [the cell] doesn’t actually care whether the speedometer says 60 MPH, or 5 MPH; what it actually interprets is whether or not the car is capable of going from 5 to 60 MPH.”

Yaffe and colleagues regard their modified engineering approach as an interesting way of coaxing a lot more information out of a single computational model by simply playing mathematical games, so experiments aren’t necessary. From a biological standpoint, while it may seem paradoxical that the same cellular phenotype could occur when a signaling pathway is either turned on or off, “it could help explain why a range of different mutation types in certain genes could ultimately end up causing the same disease, because of the way the cell interprets the information.”