Skip to main content
Premium Trial:

Request an Annual Quote

USC-Led Team Uses Markov Chain Models to Study Cancer's Metastasis Mechanisms


Scientists from the University of Southern California and other institutions have used a Markov chain model-based approach to study the mechanisms cancers use to spread cells from primary tumor sites.

They published the results of their analysis in a recent issue of Cancer Research. Previously, they published a detailed description of the model in an article in PLOS One.

According to the Cancer Research paper, which describes the model's application to lung cancer data, cancers appear to metastasize in three ways: "self-seeding of the primary tumor; re-seeding of the primary tumor from a metastatic site; and re-seeding of metastatic tumors." This finding contradicts what the researchers claim is the traditional medical view that metastasis occurs in a single direction.

The model also showed, the researchers said, that the primary tumor site and the first site to which the metastasized cells spread play a key role in "future pathways," as well as the "timescales" of the disease's progression.

That’s because some destination sites in the body act as "sponges" and tumor cells that migrate to these sites are unlikely to spread further, the paper explains. Conversely, other areas are described as "spreaders" meaning that cells that migrate to these parts first are more likely to proliferate further.

Paul Newton, a professor in USC's department of aerospace and mechanical engineering and one of the study authors, told BioInform that knowing which sites are spreaders could help guide oncology drug development efforts.

For example, the researchers found that one of lung cancer's main spreader sites is the adrenal gland. With that information in hand, Newton said, drug developers could work on specialized treatments that target adrenal glands in lung cancer patients.

It could also improve the way clinical trials are designed, he said, explaining that their model could be used in computer simulations that test the outcomes of the potential treatments in silico prior to using live individuals.

For the Cancer Research study, the researchers applied their method to human autopsy reports of 163 lung cancer patients in the New England area from 1914 to 1943. They targeted this time period because it predates the use of radiation and chemotherapy, and provides researchers with a clear view of how cancer progresses if left untreated.

Newton et al's model is similar to internet search algorithms like Google PageRank in the sense that both use transition probabilities — the probabilities associated with changes in a system's state — and steady state distributions — a rank ordered list of possible changes in a system's state — in their computations.

However, while PageRank computes a steady state distribution based on known transition probabilities from one site to another — the likelihood that a person will go from one website to another — the cancer model works in reverse.

Newton explained it this way: Google tracks people's movements from one website to another and, based on data collated from millions of web surfers, it is able to rank order websites according to their level of importance, he said. But in cancer, "we have no way of knowing what the transition probabilities are for cancer to spread from the adrenal glands to the liver if you have primary lung cancer" because " we don't know how all the different sites in your body are connected," he said.

Because the researchers were analyzing a dataset that catalogued the "natural history of metastatic disease" in multiple patients, they knew what the steady state distribution of the disease was, or, in other words, the sites to which the diseased cells travel, he explained.

"We start with the steady state distribution and do the inverse of what Google does," he said. "We infer what the transition probabilities are going from site to site that will give us that [observed] steady state distribution."

Besides lung cancer, Newton and his colleagues have used their approach to build models for breast, colorectal, prostate, and ovarian cancer also based on autopsy data and they've found that each cancer type has different sponge and spreader sites.

The next step, he said, "would be to try to understand the biological reasons why a site might be a sponge or a spreader. That’s really unexplored territory right now."

He told BioInform that his group is currently "fleshing out the biological reasons that adrenal might be a spreader for lung cancer but not other cancers" and that they're "actively looking at data-assimilation methods which will allow us to tailor these kinds of models to more specialized sub-populations and individuals."

Joining the USC researchers on the study were scientists from Scripps Clinic; Scripps Research Institute; Billings Clinic; University of California, San Diego; and Memorial Sloan-Kettering Cancer Center.

Filed under