Scientists are delegating the task of finding the weaknesses in cancer tumors to software. “DEEP LEARNING”
We should be able to go to a computer, type in INFORMATION ABOUT A PATIENT’s cancer, and up will pop 50 cases around the world that are similar at the molecular level.
20,000 genes of a typical human genome include three billion DNA nucleotides, or bits of information, any of which can be mutated, repeated or moved in any number of ways to cause cancer. Each of the human body’s billions of cells has its own copy of the genome, subject to its own mutations.
But DNA is only part of the picture: Whereas DNA is a blueprint, the real work in our cells is carried out by proteins—complex molecules that control almost everything in our biology. Proteins govern both the growth of a cancer tumor and the work of the immune system in fighting it. There are as many as 6 million basic proteins and variations on them, and researchers are now measuring thousands of them directly in cancer-tissue samples and feeding that information to the deep-learning programs.
“Drugs don’t target genes, they target proteins,” says David Spetzler, chief scientific officer of Caris Life Sciences in Irving, Texas. “That’s where we’re seeing the most progress in understanding cancer, and it’s what’s going to be the most useful information we gather in the next five years.” Says Jeffrey Balser, a physician who heads the Vanderbilt University Medical Center: “That’s a lot of incredibly deep knowledge coming to the table.”
Deep-learning algorithms don’t work the way scientists do—they never “understand” the biology behind the cancer they’re analyzing. Instead, they digest reams of information from tissue samples of patients that had certain kinds of cancer, and correlate that information with the ultimate fate of those patients—who responded to which treatments and who didn’t. It’s a kind of hit-or-miss association exercise, but one that’s conducted thousands of times, using vast amounts of data. Computers can tease out patterns in the data that a human could never see—linking, say, the presence of the FGFR gene to a particular cancer of the bile duct.
Spetzler’s company, for instance, is working to crunch protein-fortified data with deep-learning software. To wring useful insights out of the data from 170,000 cancer patients that Caris has access to, the company enlists hundreds of different deep-learning algorithms. The programs essentially compete with one another to find patterns in the data that indicate which drugs will work best with which patients. “Different algorithms will miss different patients, but together they can do a better job,” says Spetzler.
AI is helping provide yet another critical set of clues
to how to match patients to new drugs by learning to read slides of tissue samples taken in biopsies. Those slides have always been read under a microscope by pathologists, who come up with a cancer diagnosis based on the cells’ appearance. So-called “machine learning” programs are starting to step in. An Israeli company called Nucleai has trained its software with 20 million digitized biopsy slides to recognize cancer, and it already performs with 97 percent accuracy.
Diagnosing cancer is just the start, says Nucleai CEO Avi Veidman. The goal now is to use AI to extract more information from slides than pathologists can—information that can help match patients to new drugs. “Most of the information in that tissue isn’t being used when doctors or software are trying to predict the treatments that will work,” says Veidman, who spent two decades with Israel’s intelligence forces developing AI software to recognize missile bases and terrorist activity in satellite images before turning his attention to cancer three years ago. “AI can analyze the different types of features in the image much more efficiently and find hidden patterns.” He notes, for example, that subtle signs of the battle between the patient’s cancer cells and immune-system cells can be spotted by the software, and those signs can provide essential clues to whether or not the cancer might be vulnerable to one of several new immunotherapy drugs—that is, drugs that work not by fighting the cancer directly, but by boosting a patient’s immune system so it can attack the tumor.
South-Korean firm Lunit has developed AI software that can analyze pathology slides to predict, for example, which patients will respond to a relatively new type of cancer drug called checkpoint inhibitors, which can prevent cancer cells from blocking a patient’s immune cells. Lunit claims that the software is 50 percent more accurate than tests that use genetic data alone. “That’s going way beyond what human eyes can do,” says CEO and physician Beomseok Brandon Suh. “The software is finding patterns that are too complex for people to recognize, but that have biological meaning.”
Similar advances are taking place with AI-based systems that are reading X-rays, MRIs and other image data. “There are already algorithms that are as good at reading a mammogram as a highly trained radiologist, or at recognizing skin cancer as a dermatologist,” says Chi Young Ok, a pathologist at the MD Anderson