22 days to 5 minutes: a faster way to trust cancer simulations

By Eric Shaw

It takes about 10 minutes to simulate a tumor growing inside a computer. Millions of cells divide, migrate and recruit blood vessels, all governed by biological rules. Testing whether that simulation is trustworthy is a fundamentally different problem.

In one representative tumor model, a standard stress test would require more than 3,000 separate runs and nearly 22 days of continuous computing. A team led by University of Michigan mathematician Trachette Jackson has compressed that process to under five minutes.

The method is called SMoRe GloS (Surrogate Modeling for Recapitulating Global Sensitivity). It was published in September in PLOS Computational Biology. The problem it addresses extends well beyond cancer: the simulations best suited to capture the complexity of biological systems are often too expensive to rigorously test. That gap between building a model and trusting its predictions is a real barrier for researchers trying to inform treatment strategies.

Cancer modelers frequently rely on what are called agent-based models, or ABMs. Rather than treating a tumor as a uniform mass, an ABM assigns every individual cell its own behavioral rules: how rapidly it divides, how far it migrates, how it responds to oxygen or signals from its neighbors. The approach resembles a simulated city where every resident independently decides where to go and whom to interact with, and the city’s character emerges from millions of those choices at once.

When those millions of simulated cells interact, patterns emerge that no one programmed directly. Tumors develop distinctive shapes. Blood vessels extend toward oxygen-deprived regions. Successive runs produce subtly different outcomes. That resolution lets researchers explore “what if” scenarios that simpler models cannot match. It also makes agent-based models costly to run.

“With complex models, you can’t just ask, ‘Does it run?'” said Jackson, a professor of mathematics and associate vice president for research at U-M. “You have to ask which underlying assumptions are driving the results you’re seeing.”

What “stress testing” a model means

The established way to answer that question is called global sensitivity analysis. Jackson compares it to perfecting a recipe.

“I’ve got all these ingredients I control: the spices, the temperature, the timing,” Jackson said. “Global sensitivity analysis tells me which one matters most for the flavor. But testing them all the conventional way means preparing the entire dish from scratch every single time I adjust an ingredient.”

In practice, the procedure involves changing each uncertain input, such as cell division rates, migration speeds and drug response thresholds, and measuring how the model’s predictions shift. Which inputs exert the most influence on the outcome? Which barely matter? The answers determine how much confidence researchers can place in the simulation.

For agent-based models, this stress test requires thousands of individual runs. Each one can take minutes to hours on a computing cluster. In the study’s harder test case, a 3D model of tumor growth with blood vessels, one method would need 3,120 runs and about 22 days. A simpler screening approach still required 450 runs and roughly 75 hours.

“Global sensitivity analysis is one of the best ways to understand a model,” said Daniel Bergman, who developed the SMoRe GloS method as a postdoctoral researcher in Jackson’s group. “But for agent-based models it can be so costly that people skip it entirely.”

Bergman and Jackson worked on the project with Harsh Vardhan Jain at the University of Minnesota Duluth and Kerri-Ann Norton at Bard College. Bergman is now at the University of Maryland School of Medicine.

“We wanted to make it practical,” Bergman said.

“A lot of models are used without the kind of rigorous testing we’d ideally want, because it’s simply not feasible,” said Norton, a computational biologist at Bard College. “Tools that lower that barrier can raise the bar for the whole field.”

Building a surrogate

Rather than running the full ABM thousands of times, SMoRe GloS constructs a far simpler surrogate. It’s a compact set of equations that tracks the tumor’s overall growth over time, tuned to approximate how the complex model’s output changes as its inputs vary. The surrogate doesn’t track individual cells. It captures the system’s broader behavior, which is enough to reveal which inputs matter most.

The surrogate does not replace the full model. It fills in during the heavy lifting of sensitivity testing, where the goal is to explore many input combinations quickly.

Building the surrogate requires some initial computing investment. For the tumor model with blood vessels, the team ran 486 simulations on a university computing cluster. That upfront cost produced a simplified model capable of handling both testing methods. The more demanding procedure, which would have consumed 22 days on the full ABM, ran in under five minutes on an ordinary desktop.

The team also discovered something unexpected. The surrogate could reliably gauge which inputs mattered most even for biological processes it did not represent. Cell migration and contact inhibition are spatial phenomena; they happen because cells physically encounter one another. Those dynamics are absent from the surrogate’s equations. Yet its sensitivity rankings closely matched those of the full agent-based model.

“Our goal wasn’t just speed,” Jackson said. “Our goal was trust. We have to make sure we have a clear link between our model’s parameters and real biological processes.”

That emphasis on trust is what sets SMoRe GloS apart from “black box” alternatives, including some machine-learning surrogates. A black-box model can match a target curve with precision but offer little insight into why. In SMoRe GloS, each parameter in the surrogate maps to a specific biological quantity, such as a cell division rate or a growth limit, rather than blending variables in ways that hide their individual roles.

“It’s tempting to pick the surrogate that matches the curve most closely,” said Jain. “But if five different configurations all produce the same output, you can’t determine which one reflects the actual biology. We chose a surrogate whose parameters can be individually identified, so the results carry genuine meaning.”

What it could change

The SMoRe GloS code is freely available on GitHub. Bergman is now extending the method for pancreatic cancer at Maryland, and other groups have begun applying it to infectious disease, ecology and immunotherapy models.

Jackson’s group is currently using SMoRe GloS to investigate treatment resistance in bladder cancer and the timing of immune therapies, work that could help researchers explore when to start a drug or how to sequence multiple treatments.

Jain frames the broader promise in terms most people already understand: weather forecasting. If computational models eventually help doctors plan cancer treatments for individual patients, this kind of stress testing would draw the boundaries of what those predictions can and cannot reliably say, much like the cone around a hurricane’s projected path.

“Think of it like a hurricane forecast,” Jain said. “The model gives you the predicted path, and our work is what draws the cone of uncertainty around it, so a doctor could see not just what might happen but how confident we should be.”

That clinical future remains distant. These models generate hypotheses and help researchers understand how tumors respond to treatment. They are not yet informing bedside decisions. But the gap between a model that runs and one that has earned enough credibility to guide research is precisely the gap SMoRe GloS was designed to close.

“When this kind of testing takes weeks, it becomes something you only do once in a while,” Jackson said. “But trust can’t be optional. Our goal is to make the analysis that builds trust as routine as model development itself.”

The team ran its simulations on the Great Lakes high-performance computing cluster, a campus-wide resource available to researchers across the University of Michigan.