The road less traveled. That describes how two research teams are approaching the fight against cancer, using GPU-accelerated computing and deep learning to explore new genetic pathways for clues to more effective treatments.
One team is building on its surprising discovery that many of the most potent “driver” genes — the culprits for cancer progression — aren’t located on the chromosomes, where scientists normally search for them, but on circular pieces of extrachromosomal DNA, or ecDNA.
The researchers, led by Dr. Paul Mischel of the Ludwig Cancer Research Institute, in San Diego, plan to use high performance computing and GPU-accelerated deep learning to understand these genes and their role in cancer.
The other team, from the University of Toronto (U of T), will use GPU-accelerated deep learning and DNA sequencing data to rapidly build a cancer’s “family tree” — revealing the mutations turning healthy cells into malignant tumors, and enabling us to predict how they’ll evolve.
Support for Cancer Research
The grants are part of the NVIDIA Foundation’s Compute the Cure effort, which supports researchers using innovative computing techniques to yield breakthroughs in cancer diagnostics and treatment.
A group of NVIDIA employees, with the support of researchers at the National Cancer Institute, chose the recipients from among nearly 70 proposals submitted from around the world.
Finding Cancer Hot Spots
For Paul Mischel, the fight’s personal — he was 14 when he lost his father to cancer.
“I became a pathologist to look the enemy in the eye,” said Mischel, the head of Molecular Pathology at the Ludwig Cancer Research and a professor of pathology at University of California, San Diego.
In February, Mischel and his UCSD colleague,Vineet Bafna, found that driver genes in more than half of human cancers are located outside the chromosome on ecDNA. These genes mutate more quickly than other types, so they’re more likely to become resistant to drugs and harder to treat.
“These cancers changed at a rate that made no sense. It was all happening way too fast,” Mischel said.
Now the team, which includes post-doctoral researcher Elizabeth Brunk, plans to use deep learning to more accurately detect ecDNA and determine how it changes cell behavior. By analyzing vast quantities of cancer data with deep learning, researchers hope to identify molecular hot spots that are pivotal in cancer development, and eventually target these with treatments.
Cancer’s Family Tree
Species evolve through a series of mutations. So does cancer. And cancer’s mutations can give cells the ability to evade the immune system or grow in ways normal cells don’t.
Quaid Morris, professor of computer science and molecular genetics at U of T’s Donnelly Centre for Cellular and Biomolecular Research, and David Duvenaud, professor at U of T’s Department of Computer Science, want to explain how this happens.
With GPU-accelerated deep learning, they plan to detect patterns in DNA sequencing data to trace when mutations occur and how the disease develops. They’ll also examine how the cancer responds to treatment and how it’s likely to evolve, which could lead to new treatments. Both Morris and Duvenaud are also members at the Vector Institute for Artificial Intelligence.
The researchers will build what amounts to a family tree within each tumor. This detailed work is a laborious weeks-long process, but with the Compute the Cure grant, Morris and Duvenaud plan to develop GPU-accelerated deep learning software to turn those weeks into seconds. That would be fast enough for doctors to use to predict prognosis and choose treatments. Reconstructing thousands of these trees will also shed light on patterns in mutations across patients, the researchers said.
“The better we understand how cancer develops and progresses, and which cancers are dangerous (or likely to become so), the easier it will be to design new treatments, and to match cancers to existing treatments,” Morris said.
The lead image for this story shows a cancer cell in which chromosomes are labeled blue and the potent “driver” genes that lead to cancer are red. Dr. Paul Mischel of the Ludwig Cancer Research Institute found that these genes are overwhelmingly not on chromosomes, but on pieces of extrachromosomal DNA, or ecDNA. Image courtesy of Paul Mischel.