Skip to main content
Premium Trial:

Request an Annual Quote

Tumor Forecasting Using Mechanistic Rules of Biology May Improve Predictive Models, Study Suggests

Premium

NEW YORK – Doctors someday may be able forecast a cancer patient's chances of responding to treatments, developing resistance, and relapsing based on mechanistic modeling that complements existing spatial biology insights and biomarkers.

The tumor forecast, similar to a weather forecast, would simulate the tumor's behavior based on fundamental rules of biology rather than rely on statistical data generated from clinical datasets like conventional predictive models use. As envisioned by Elana Fertig, a computational biologist at Johns Hopkins University School of Medicine, such a model could yield insights not available through standard data-driven methods.

Fertig, who started her career not in cancer biology or biostatistics, but as a NASA research fellow in numerical weather prediction, became fascinated with the idea of applying the principles of weather forecasting to cancer biology. "If we know the laws of the system, we can write down the equations that describe them and [predict] where they're going to go in the future," Fertig said during a presentation at the American Association for Cancer Research in April. As an example, she said physicists can calculate the path of a ball thrown in the air and predict where it will fall.

In the past, the large number of variables governing changes in a biological system as complex as the human body have made these calculations challenging to simulate. However, advances in computer processing power in the 21st century have made much more complex modeling feasible.

In applying a mathematical modeling approach to tumor spatial biology, Fertig said the first step is to identify which variables in the system are the most relevant and develop equations that describe the interaction of those variables over time. "I would argue that one of the biggest limitations we have in our field is that we have no idea what the most relevant variables are," Fertig said.

To explore the use of such a model in a spatial tumor biology application, Fertig's group collaborated with a team led by Aleksander Popel, a professor of biomedical engineering at Johns Hopkins, to develop a specialized type of quantitative systems pharmacology (QSP) model. QSP models mechanistically simulate disease progression, pharmacokinetics, and pharmacodynamics for selected drugs, and are used to conduct virtual clinical trials that support drug discovery and clinical trial design.

Popel's group coupled a QSP platform with an agent-based model, creating a spatial QSP model that simulates a clinical trial testing the neoadjuvant activity of Bristol Myers Squibb's checkpoint inhibitor Opdivo (nivolumab) in patients with hepatocellular carcinoma. The agent-based component of the model is designed to reproduce spatial features from hepatocellular carcinoma tumor imaging and transcriptomic sequencing datasets. The result is a model that can track tumor progression at the organ scale while preserving information on spatial tumor heterogeneity at the tissue scale.

The researchers validated the model using spatial multiomics data from a clinical trial of neoadjuvant Opdivo and Exelixis' tyrosine kinase inhibitor Cabometyx (cabozantinib). In the clinical trial results, four out of 12 patients who underwent successful surgical resection had a major pathologic response and one had a complete pathologic response, translating to an overall response rate of 42 percent. When the researchers conducted a virtual clinical trial using the same patient data, 32.2 percent of the simulated patients were predicted to have a pathologic complete response, which the researchers described as consistent with the original clinical trial results.

In addition to predicting results, however, Fertig and her collaborators were able to simulate the response of the tumor over time and characterize the response of different cell types and their spatial relationships. In an example of one such tumor simulation that Fertig described in her AACR presentation, she noted that over time, the dark mass of tumor cells in a sample from a responding patient visibly faded away and were replaced with immune cells, while in a sample from a non-responding patient, the tumor mass remained relatively intact.

In comparison, with real-world surgical samples there is only one opportunity post-surgery to visualize those spatial relationships. "With these mathematical models, if we are accurately capturing the endpoint, we have the ability to run backwards in time, have a whole series of models, and start to compare the characteristics pre-treatment that [distinguish] responders and non-responders, and start thinking about biomarkers that we can calibrate to patients' samples that live across the lifecycle of the tumor," Fertig said.

In the virtual trial, the researchers noticed that proximity between cancer stem cells and T cells correlated with better response compared to tumors in which there was more distance between these cell types. They also identified low Arg1+ macrophage counts and higher ratios of M1-like to M2-like macrophages as potential biomarkers of response to therapy. Fertig said these results could inform future experiments and help researchers understand how liver cancer responds to immunotherapy.

In an interview, Fertig expanded on her vision of where mechanistic modeling fits in the landscape of artificial intelligence and machine learning for biological models. "There's a whole burgeoning field of interpretable AI, and you can use this mechanistic information instead of just a black box AI prediction," she said.

A challenge so far, Fertig noted, is that in her group's experience, the predictions can become a bit worse when mechanistic modeling is incorporated into those machine learning models. That could be because the resulting combined model is overfit, meaning there are so many parameters that the results only work for that specific dataset. Another possibility is that the interpretation of biological rules is not yet accurate enough.

If researchers can overcome those challenges, however, Fertig believes a forecast model could help bridge gaps in drug development, clinical trials, and patient care not addressed by existing statistical models.

For example, a mechanistic model could more accurately identify the drivers of cancer, rather than the pathway in general or the genes that happen to be overexpressed. The usual approach when developing targeted treatments for cancer involves identifying a gene with elevated expression, trying to knock out its activity with a drug, and seeing what happens to cancer cells. "That might not be the gene that's driving later metastasis," Fertig noted. "I would much rather see us using these models to predict what gene is most likely to be driving progression and then try to target that, as opposed to one that just happens to be there at the time you're measuring the tumor."

Tumor forecasts could also help researchers plan biomarker strategies for clinical trials, as her group did in the hepatocellular carcinoma virtual study. Beyond that, Fertig also sees a role for mechanistic forecasts in patient care.

"One of the things I hate about cancer care is basically you go to patients with a Kaplan-Meier mindset, [which provides] a view of the probability of dying many years out," Fertig said. "Living with that burden over your head, especially if it's going to be five years [or more], is just not great."

In contrast, a tumor forecast, as Fertig envisions it, has the potential to provide more accurate predictions in the near term and become less accurate as time goes on. Fertig's hope is that in the future, patients will receive updated individual forecasts throughout their cancer journeys that capture how well they're likely to respond to treatment or pick up signs of drug resistance and provide a more personalized picture of what their disease is doing. Armed with this information, patients could try a new treatment sooner than they would with other types of surveillance.

"Clinically, that's what I would love to see," Fertig said, estimating that it might take another five or 10 years of research before tumor forecasts are being used in patient care. If they are, "in addition to following the biology of the tumor, [tumor forecasts] would be psychologically a lot healthier for patients," she said.