Mayo researchers and NASA Frontier Development Lab data scientists are embarking on a research sprint this month to optimize an artificial intelligence (AI) algorithm for colorectal cancer and possibly other cancers.
The algorithm was developed by researchers within the Mayo Clinic Center for Individualized Medicine and shows potential in detecting spatio-temporal patterns of colorectal cancer progression solely from tumor snapshots.
“Our research shows the algorithm is able to predict the evolutionary trajectory by which colorectal cancer is going to occur,” says Nicholas Chia, Ph.D., the Bernard and Edith Waterman co-director for the Mayo Clinic Center for Individualized Medicine’s Microbiome Program. “We have information from this algorithm in terms of what event came first, what events are most important and exactly what the path to cancer was or what it will be.”
For eight weeks, Mayo will work with a team of NASA engineers, computer scientists and software developers in order to get the algorithm optimized for multi-omics data integration and causal modeling.
Dr. Chia says the project will also use complex multi-omics data, including the microbiome, to sequence the path of cancer — starting with what causes it, to what drives it, and potentially, how to prevent it.
"Our research shows the algorithm is able to predict the evolutionary trajectory by which colorectal cancer is going to occur.” - Nicholas Chia, Ph.D.
John Kalantari Ph.D., a machine learning scientist within the Center’s Microbiome Program, says he was inspired to develop the algorithm by contemporary applications of an AI technique known as reinforcement learning — popularized by its use in autonomous driving and defeating human experts in computer games, such as chess, Go, and StarCraft.
"We had a eureka moment when we realized that if we viewed our patient cancers as the result of an optimal game of cell evolution, then we could use inverse reinforcement learning techniques to learn the optimal 'moves' and environmental conditions that enable cancer progression, metastasis, recurrence, immune system evasion, and/or changes in treatment efficacy,” Dr. Kalantari explains. “By reverse-engineering how a tumor survived and thrived to become cancer in each individual patient, we are able to enhance our understanding of cancer systems biology and also improve our ability to predict treatment outcomes, discover early biomarkers of progression and identify new therapeutic/preventative targets in a more holistic manner."
Recent results inspire confidence
Dr. Chia says his team chose to test the algorithm on colorectal cancer because of its canonical pathway of development, which was first discovered in 1993 by Bert Vogelstein, M.D., a pioneer in cancer genomics who found the gene responsible for inherited colon cancer and discovered how genes influence the susceptibility, leading to a blood test that can help identify people who are genetically predisposed to colon cancer.
"We had a eureka moment when we realized that if we viewed our patient cancers as the result of an optimal game of cell evolution, then we could use inverse reinforcement learning techniques to learn the optimal 'moves' and environmental conditions that enable cancer progression, metastasis, recurrence, immune system evasion, and/or changes in treatment efficacy.” - John Kalantari Ph.D.
In a recent study, Dr. Chia’s team tested the algorithm on 27 patients with colorectal cancer, using whole genome sequencing and DNA methylation to see if it could accurately identify spatial and temporal patterns of cancer progression. Many machine learning methodologies that strive to predict system dynamics require access to time-series data. Longitudinal patient tumor samples are few and far between, making existing techniques impractical to use.
“In order to innovate in both medicine and AI, we need to rethink how we can leverage our existing knowledge and data more efficiently,” Dr. Kalantari explains.
The result was a new Bayesian nonparametric algorithm called the ‘Pop-Up Restaurant for Inverse Reinforcement Learning’ (PUR-IRL). With their patient cohort and novel algorithm, the team was able to re-identify mutations associated with colorectal cancer progression and also predict the correct causal ordering in which they occurred.
“We were able to show a path to the same exact canonical pathway that Bert Vogelstein outlined by using just data from tumor samples,” Dr. Chia explains. “So this is promising because lots of cancers don’t have precursor lesions, and there is not a way of assessing the order like in colorectal cancer. The fact that we got the same answer as what’s already established tells us we’re on the right path and this algorithm works.”
The ultimate goal is to demonstrate the feasibility in order to study a larger cohort of patients with colorectal cancer and possibly apply the algorithm to other cancers.
“When the results come out, we hope in the near future we will be able to provide a report that can be interpreted by a clinician to understand what genes are driving this particular tumor in this particular patient,” Dr. Chia explains. “Through this inference, we can better understand the causal role of the microbiome in colorectal cancer and potentially other cancers.”
The Mayo Clinic Center for Individualized Medicine was selected for the prestigious NASA AI accelerator program after its AI cancer algorithm paper was awarded “Outstanding Paper Honorable Mention” out of 9,000 submissions at the Association for the Advancement of Artificial Intelligence conference in February. The project was also recently selected to be funded $100,000 by the Amazon Web Services Machine Learning Research Awards program.
Read more stories about advances in individualized medicine.
Register to get weekly updates from the Mayo Clinic Center for Individualized Medicine blog.