Researchers Use AI to Improve Immunotherapy Outcomes

A research team from the University of Maryland is utilizing artificial intelligence (AI) to determine the most effective ways to treat cancer patients with immunotherapy.

Immunotherapy takes advantage of the body’s own immune system to fight off cancer, and in the past decade, it has been very successful in treating patients with localized and advanced disease. These resounding positive outcomes have even allowed certain immunotherapies to become the standard first-line treatment for many cancers, oftentimes combined with chemotherapy. But while these cutting-edge treatments have offered hope for many patients, still only about 20 in 100 will respond to them.

As immunotherapy research continues to develop and more drugs are made available, scientists are still trying to find ways to allow current treatments to be effective in a larger number of patients. The best way to do this is to pin point more precisely how patients will respond to the treatments. Mark (Max) Leiserson, assistant professor of Computer Science at the University of Maryland and his team from Microsoft Research and Memorial Sloan Kettering Cancer Center are using a branch of artificial intelligence known as “machine learning” to accomplish this goal.

The team is working to create more accurate predictions of treatment outcome by using a new computer modeling system that examines data from both cancer patients and their cancer at the same time. A recent study published in PLOS One shows Leiserson and his team using such data from a bladder cancer clinical trial to show that the technique is effective at locating the features that predict effective immune responses in patients. Such a system will also allow doctors to treat fewer patients who likely won’t respond.

The computer model was created when Leiserson and his team took data from the bladder cancer trial, which contained a large amount of information about tumor cells, immune cells, and relevant patient demographics and history. Designed to determine the features associated with certain drug responses, the trial worked with enough data for Leiserson’s team to turn to machine learning.

After introducing 36 different features into the model, they let the computer come up with patterns that could anticipate a patient’s blood increase in the kinds of immune cells that fight tumors. The algorithm that resulted consisted of 20 identified features that could account for 79% of variations in patient immune responses. Leiserson says that such a complete group of features is enough for scientists to accurately predict a patients immune response.

The data from the clinical trial was useful in showing how effective the model could be. It proved that it could pick up 100% of patients who would benefit from the immunotherapy treatment, and as low as 38% of patients who would ultimately not end up benefiting from it.

“What’s also exciting about the study,” said Leiserson, “is that we were not just looking at patient outcome, but at a specific marker of immune response, which gave us a much better picture of what’s going on.”

The work done by Leiserson’s team helps augment the recent focus on precision oncology, a branch of medicine that aims to modify treatments so that they work individually for patients and their specific tumors. The long-term goal in these areas is to utilize the prediction tool offered by the model so that clinicians can really start to accentuate the benefits of immunotherapy treatments.