Trauma surgeons know how to fix gunshot wounds, lacerations and broken bones. It's what comes afterwards that really worries them. Even after the initial injury is treated, patients are at risk for secondary issues such as infection, sepsis and organ failure. While the biological pathways involved in these processes have been well studied and characterized, effective interventions to reliably stop the dangerous cascade have yet to be discovered.
"It was very frustrating for me to not have the drugs and tools necessary to fix what I thought was actually going wrong with those patients," said trauma surgeon and CI senior fellow Gary An, in his University of Chicago Alumni Weekend UnCommon Core talk. "Often we know what will happen, but we have no way to stop it."
The current fashionable approach to such intractable problems in medicine and other fields is Big Data, where answers hiding in massive datasets will be uncovered by advanced analytic methods. But quoting Admiral Ackbar, An warned the audience that this approach alone "is a trap," generating a multitude of correlations and hypotheses that don't always translate into real world applications.
"What it wants to appeal to is magic...if you can get enough data and a big powerful computer, an answer will magically appear," said An, an associate professor of surgery at University of Chicago Medicine. "That's fine if you want to diagnose or characterize. But if we want to engineer interventions to be able to manipulate systems, we need to have presumptions of mechanistic causality; we need to be able to test hypotheses."
To address this new bottleneck -- how to rapidly and inexpensively test hypotheses in complex systems -- An turned his own research focus to the same technology used to create battle scenes in Lord of the Rings: agent-based modeling. These models are made up of independent agents that operate and interact according to a set of rules, a method that can be used for both orcs in a fantasy battle and proteins in a gut epithelial cell.
An and his collaborators have built models for diseases such as inflammatory bowel disease, breast cancer and necrotizing enterocolitis, a gastrointestinal disease that afflicts premature infants. Each one uses the breadth of scientific literature on the disease to build a minimally sufficient disease model that recapitulates experimental observations, and subsequently can be used to test different hypotheses. If researchers think that increasing or decreasing the activity of a particular gene can reduce the probability of developing breast cancer, they can test it in the model, in silico, ruling out faulty hypotheses without (or before) investing in the considerable cost and time of a traditional experiment.
Though the technology involved in constructing and running these models is relatively new, the principles behind them are very old, An said. In the basic cycle of science, a scientist makes observations, forms a hypothesis and rigorously tests that prediction. Those experiments produce new observations that either support or argue against the hypothesis, and the cycle begins another lap. Using a computer model doesn't change this cycle, it just gives researchers a tool to better organize the process of selecting hypotheses with the most potential while incorporating an unprecedented flood of scientific publications and data.
"A good researcher does everything we talked about in their head, but not explicitly," An said. "The problem then is that when [their hypothesis] breaks, we don't know what went wrong, and if it works, it doesn't mean it's right, it just means it's plausible."
To help researchers that don't have the programming experience to build their own models, An is constructing a "computational model assistant," that automates many of the steps and speeds up the path to testing hypotheses in silico. The hope is to push scientists beyond big data, restoring the cycle of the classic scientific method but increasing its rotational speed to produce demonstrably effective interventions faster and more efficiently.
"If all you want to do is characterize your system, no problem," An said. "But for me as a physician, just telling me who's going to die, how someone's going to do, is not good enough. I want to be able to do something to make them better, and to do that, you have to close the loop, you have to have those mechanistic hypotheses and move forward."