AI model predicts effective immunotherapy combinations for liver cancer

Researchers at the Johns Hopkins Kimmel Cancer Center and the Johns Hopkins University School of Medicine have developed a method using computer tools to predict which patients with a primary liver cancer called hepatocellular carcinoma (HCC) will benefit most from combination therapy using immunotherapy and targeted therapies that block growth signals on which tumors depend. The study was supported in part by the National Institutes of Health and was published online July 14. Proceedings of the National Academy of Sciences.

Many cancers progress very quickly, and doctors may not necessarily have time to try surgery or other treatments, so our idea was to create a computational model that could simulate trying different doses and combinations of cancer treatments to help doctors figure out the best options for their patients. ”

Dr. Atul Deshpande, senior study author, assistant professor of oncology, Johns Hopkins University School of Medicine

The tool needs further validation before it can be incorporated into clinical care, he says.

The spatial QSP modeling platform was developed in the lab of Dr. Alexander Poppel, professor of biomedical engineering and oncology at Johns Hopkins University, and combines two tools. One is quantitative systems pharmacology (QSP). This is a mathematical model that uses equations to capture the systemic response to treatment and can simulate tumor progression and drug effects. The second is an agent-based model that tracks how individual cells behave in specific scenarios. Together, these map not only the number of cells present but also where they reside, predicting the dynamics of the tumor and surrounding microenvironment.

In this study, Deshpande and colleagues extended a platform for modeling fibroblasts (a cell type previously associated with resistance to immunotherapy in liver cancer), developed a machine learning calibration workflow that tailored the simulation to real-world clinical trial data, and generated virtual patients that could match predicted responses to actual outcomes.

In a kind of “war plan,” Deshpande says, “we generate virtual tumors to see what happens in the microenvironment. Will the cancer cells fight back? If we change the structure of the tumor, will that help the cancer cells and the immune cells?”

One advantage of computational models is scale, he says. A small, early-stage (Phase I) study of 15 patients can generate a virtual population of Phase III size, allowing researchers to estimate how a treatment will work in a much larger trial quickly and without putting anyone at risk. When the team simulated treatment with the targeted therapy cabozantinib and the immunotherapy nivolumab, alone or in combination, the predicted response rates closely matched those reported in real-world clinical trials, proving that the virtual patients behaved like real patients. The research team also validated the model’s predicted tumor structure against actual post-treatment tissue and compared the microenvironments of responders and non-responders.

Researchers found that fibroblasts remodeled the tumor microenvironment, causing immunosuppression. In particular, Deshpande says, among the hypothetical patients who did not respond to treatment, the researchers noticed that fibroblasts formed a kind of physical wall. “Even if the immune cells were near the tumor, the fibroblasts would block the immune cells from reaching the tumor,” he says.

Ideally, Deshpande says, modeling would get to the point where, when multiple treatments are available for a particular cancer, it can help researchers decide which treatments are most effective or which should be avoided. Structural features like these – both the fibroblast barrier flagged by the model and the structural patterns the team measured in patients’ tumors – are visible before treatment begins and could ultimately help predict who will benefit, the researchers say.

The study was co-supervised by Poppell and Elana Fertig, Ph.D., director of the Institute for Genomic Sciences at the University of Maryland School of Medicine. Co-authors were Shuming Zhang, Hanwen Wang, Yeonju Cho, Wendy Wong, Mark Yarchoan, Elizabeth Jaffee, Won Jin Ho, and Luciane Kagohara of Johns Hopkins. Heber Rocha of Indiana University.

This research was supported by grants from the National Institutes of Health (grant numbers U24CA284156, U01CA253403, and U01CA212007). Department of Defense-MEDCOM Authorization 144517-CA220654P. Maryland Cancer Moonshot Research Grant; and the Maryland Tobacco Reparation Fund.