NEW YORK – A new initiative within the Medical University of South Carolina Hollings Cancer Center will leverage artificial intelligence to identify and recommend individualized treatment strategies to cancer patients.
Hollings researchers are building a supervised machine-learning algorithm that will incorporate genomic, transcriptomic, and proteomic data to predict the sensitivity of a tumor to drugs on the market and make personalized therapy recommendations. Thai Ho, an oncologist specializing in genitourinary cancers and director of precision medicine at Hollings, is spearheading the program. He said his team is drawing on data from patients at MUSC, including patients' treatment histories and outcomes, to build the treatment-predictive algorithm.
Ho and colleagues will initially focus on patients with cancers with defective DNA repair mechanisms, particularly those with prostate and kidney cancer. They are building on existing algorithms provided by partners including Caris Life Sciences, Tempus, and Guardant Health and incorporating patient data from MUSC patients.
For now, the MUSC team is assembling training sets and identifying the type of data they need, according to Ho. To date, the group has collected data from about 300 kidney cancer patients. "We're still trying to identify what kind of information is needed to parse out those training sets," he said.
One training set will be unbiased and will include every patient with a specific tumor type and the various drugs they are treated with. That group will train the algorithm to identify tumors and predict their sensitivity to drugs targeting DNA repair mechanisms. Another group will comprise "extreme responders," or those who have responded for very short or long periods of time to the treatments they received, which Ho said will help the AI algorithm to refine therapy suggestions by characterizing the key differences between the two groups.
In addition to molecular alterations, the model will incorporate information from patients' medical records and pathology reports, such as the size and grade of the tumor, degree of necrosis, cancer staging, and histological changes. Once the algorithm has been trained, the researchers will validate it retrospectively using data from the Cancer Genome Atlas or external hospitals and then validate the algorithm prospectively in a future clinical trial.
Ho said he chose to study kidney cancer because there has been relatively little research into the drivers of those tumors and biomarkers of response. There are only a handful of driver mutations identified in kidney cancer, including some genes involved in chromosome remodeling, DNA methylation, and DNA damage repair. However, current standard treatment for kidney cancer does not involve genomic profiling or biomarker-directed precision therapies. Instead, common therapy options for kidney cancer include tyrosine kinase inhibitors, angiogenesis inhibitors, and mTOR inhibitors.
Ho's research has focused on resistance to immunotherapy in kidney cancer and mutations in the SETD2 gene, which are associated with more aggressive disease. Currently, there are no approved drugs known to target SETD2.
Ho envisions the algorithm he and his collaborators are developing as a tool that helps physicians interpret complex genetic testing results. "Any time you get molecular profiling, it's about 50 pages of data," Ho said. "The first page is key because that's generally the only page that physicians have time to look at." That page lists drugs ranked according to how likely they are to benefit patients based on the molecular changes detected in their tumors and the available evidence for therapies targeting those mutations. However, the AI algorithm Ho's team is developing might provide a different ranking of drugs that could provide a better chance of success, giving doctors an alternative to the standard rankings they might find on a genomic test report.
Ho emphasized that the algorithm would not be picking the treatments for patients. That's still up to the treating oncologist and a molecular tumor board of bioinformaticians, molecular biologists, ethicists, and other experts that MUSC has set up to guide precision oncology decisions. The recommendations put forth by the algorithm are meant to supplement all the information and molecular profiling reports doctors and MUSC's molecular tumor board already have when deciding how to treat patients.
Ho acknowledged that an AI algorithm can make mistakes, which is why his group will limit the recommendations to marketed drugs that have been tested in patients and have a known toxicity profile. "We don't want to bypass standard-of-care therapy," Ho said. "If the patient should be receiving some [standard therapy], we always rank that higher than something the genomics may direct us to."
Researchers within MUSC's precision oncology initiative are also developing an AI algorithm to match patients to clinical trials assessing investigational treatments. Ho said the aim of that program will be to direct patients to trials based not only on their molecular alterations but also on the extensive inclusion and exclusion criteria of those studies, as well as factors like how close patients live to trial sites. "A patient is more likely to enroll in a clinical study if they live within a short distance of a specific cancer center," Ho said. "That's a way of helping us enroll more patients on clinical studies, particularly patients that are not well represented in clinical studies."
Along with increasing representation in clinical trials, minimizing financial toxicity for patients is an important part of the initiative, Ho said, noting that out-of-pocket costs associated with genomic screening and targeted therapies can add up for patients, even those with insurance. Because MUSC has secured a commitment from its partners to provide genetic testing to patients irrespective of their ability to pay, he estimated that less than 5 percent of patients who receive molecular profiling at MUSC pay more than $100 out of pocket.
Ho and his team at MUSC hope eventually to extend the advantages of its precision oncology initiative and AI-assisted molecular tumor board recommendations to doctors and patients throughout South Carolina and match patients to genomically driven clinical trials at Hollings. "In the future, we hope to encourage physicians [outside of Hollings] to submit their patient cases to our tumor boards if their patients are already within the MUSC system to better serve patients outside the Charleston area," he said.