Su-In Lee’s father had few treatment options when he was diagnosed with biliary tract cancer. No drugs are known to be effective against this rare disease. So his care team tried a common chemotherapy regimen for cancer of the neighboring pancreas.
The drugs didn’t work. Just three months after his diagnosis, he died.
Lee, an associate professor of computer science & engineering (CSE) and genome sciences, reflected on how distant the dream of personalized cancer treatment seemed to be.
“I kept thinking how wonderful it would be to match a patient to a particular drug based on a cancer cell’s molecular property,” Lee said.
She was uniquely positioned to help do just that. With her joint appointment, Lee sits at the intersection of computer science and medicine. Her research until then had primarily focused on basic biology. But after her father’s death in 2013, Lee searched for ways to match patients and drugs through machine learning and artificial intelligence (AI).
She just helped medicine move one step closer to realizing that dream. Last month Lee was first and corresponding author on a paper in Nature Communications that described a new machine learning algorithm called MERGE. The algorithm mines publicly available genomic data from patients with acute myeloid leukemia (AML) to find the best drugs for AML patients — and opens promising avenues to improve patient care for many other diseases, Lee said.
“This is going to be a decades-long effort, but Su-In’s work is a state-of-the-art attempt to make progress in our ability to predict drug responses based on genomic profiling,” said Dr. Tony Blau, a professor of medicine (hematology) and co-author of the Nature paper. “Computation is going to be absolutely essential to realizing the promise of precision medicine.”
Here’s MERGE in a nutshell: The algorithm takes the molecular profile of a patient, predicts the response to a large number of anti-cancer drugs, and helps clinicians select the best combination of drugs for that particular patient. (Take a deeper dive into the nitty-gritty at the Allen School news blog, or read the paper.)
The project began several years ago when Blau reached out to Lee and asked about a way to personalize cancer treatment. It took several years for Lee’s group to land on the MERGE approach — i.e., mining large amounts of publicly available genomic data to predict drug sensitivity for a particular patient.Together with Dr. Pam Becker in hematology, they created a pipeline of patient data for Lee’s group to work with.
At first Lee and a CSE doctoral student in her lab, Safiye Celik, the co-first author of the Nature paper, tried conventional machine learning methods. But the results were disappointing, Celik said; the algorithm couldn’t filter out enough noise to identify truly relevant genes as drivers of disease progression and robust biomarkers for drug response.
So they invented a novel machine learning method that can automatically learn from big data how to weight multiple factors in a gene to identify the most reliable biomarkers for drug sensitivity. The algorithm is named MERGE, short for “Mutation, Expression hubs, known Regulators, Genomic CNV, and mEthylation,” because it uses these five characteristics of genes obtained from big data.
The same approach can be applied to many other diseases, Celik said. Lee’s group has already had colleagues reach out to discuss potential drug targets for Alzheimer’s as well as personalized treatment for other types of cancer.
As exciting as that prospect sounds, there are limitations to MERGE, Lee said. The algorithm can tell patients and doctors which drugs are good matches, but it can’t explain why.
“All we can say is, ‘AI magic!’” Lee said.
She’s working on that parallel problem, too. Lee and another doctoral student, Scott Lundberg, developed a general AI algorithm that can explain why a certain prediction was made. They’re now applying this algorithm to develop a system that can predict patients’ progress and explain why. Lee is already collaborating with colleagues from Anesthesiology and Pain Medicine to predict adverse events during surgery, and she is working on a project to predict chronic kidney disease in collaboration with the Kidney Research Institute at the UW.
Lee is optimistic that computing’s contributions to healthcare will accelerate.
“It won’t happen tomorrow,” she said. “But in a year or two, when we have better computers and better AI techniques, we’ll find a way to better match patients to drugs.”
