Cancer in the Time of Algorithms
The news hit Shirley Pepke like a ton of bricks. Sitting in the oncologist’s office that fall day in 2013 as he described limited treatments and survival outlooks, she conjured up the image of her husband with their son and daughter, then ages 9 and 3, at home, waiting. She wasn’t sure when she got home what she would tell them to offer them some solace at the prospect of facing life without her.
For Pepke, a non-drinker, nonsmoker, otherwise healthy Caltech genomics researcher in her early 40s, the late-stage ovarian cancer diagnosis was completely unexpected. No one in her family had cancer. On the way home she mulled it all. And her first thoughts, she recalled, were: I’m going to use whatever tools I have to fight this thing. Some people get cancer and do fundraisers; I’m good at doing computational research on complex systems. I’ll pursue every avenue to extend my life. I owe it to my children.
That survival instinct started Pepke on a yearlong journey to find a successful treatment — a journey that led her to fellow physicist Greg Ver Steeg at USC Viterbi, whose research was recently employed to help people find love online. Little did Ver Steeg know that he’d be using it to try to save her life.
Pepke began her career as a physicist and data scientist, developing artificial intelligence software for NASA’s launch vehicles and algorithms to analyze high-throughput biological data at Caltech. But with the cancer spreading fast, her research focus shifted abruptly.
Everyone who has ovarian cancer gets the same frontline treatment — chemotherapy that is meant to attack the cancer head-on and in the process kill everything in its path, even healthy cells. Most women will respond to this. However, if and when the cancer comes back, many patients fail to respond or respond only partially, so a different drug therapy is needed. But there is no second-line drug that works for most people.
“Then there are clinical trials,” Pepke explained. “So how do you decide? Right now there’s almost no information. Women have to choose without knowing how aggressive their cancer is, without know-ing how the toxicity of any of these drugs will affect them — and they’re really toxic. It’s like throwing darts at a dartboard. You pick a drug, you hope that it works a little bit, and you hope the toxicity isn’t bad for you. And you iterate on that till the end of the line, which usually happens fairly quickly.”
Pepke did the standard frontline treatment, but her doctors weren’t talking to a standard patient. She looks at information very differently. If she could get the right data, she believed, then maybe she could see something the doctors weren’t seeing.
She was in luck. The Cancer Genome Atlas, the most ambitious ge-nome project in the history of life sciences, had recently published the first genome-wide expression and mutation data specifically for ovarian cancer. Ovarian cancer isn’t common — only 1.3 percent of women will be diagnosed in their lifetimes, and it com-
prises 2.4 percent of cancer deaths, according to the National Cancer Institute — but the data essentially gave Pepke close to the periodic table for ovarian cancer.
Still, she needed someone to crunch the data in a very meaningful way, and do it fast. A friend, Kristina Lerman, a project leader at the USC Information Sciences Institute, led her to Ver Steeg, an associate professor in USC Viterbi’s Department of Computer Science who is known in machine learning circles for having developed correlation explanation (CorEx), a tool that teases out hidden patterns in large, high-dimensional data sets.
“The idea behind CorEx,” Ver Steeg ex-plained, “is to ask what are the hidden factors that explain correlations in a large data set. For example, why different parts of your brain light up at the same time, or how does your answer to question A predict the way you’ll answer question B. It could be anything.”
A physicist who was now applying principles of quantum mechanics to extract useful information from complex systems, Ver Steeg took inspiration from Einstein’s attempt to unravel quantum entanglement — a counterintuitive phenomenon in which a quantum particle in one location seems to instantly “know” what a particle somewhere else is doing. Einstein’s failed attempt to attribute this effect to hidden factors nevertheless led to new mathematical tools with applications beyond quantum mechanics.
Ver Steeg had previously used CorEx to fine-tune eHarmony’s matchmaking algorithms and to predict online extremism. The closest he ever got to medical applications was looking at hidden factors in brain activity. But he was confident that he could train CorEx to solve Pepke’s dilemma.
“Essentially, it’s a matching problem,” Ver Steeg said. “In one scenario you’re looking for the right person in a big set of unknowns; in the other you’re looking for the right treatment.”
Pepke was already part of a research protocol with the Translational Genomics Research Institute (TGen) in Phoenix. This wasn’t some mail-order transaction. Pepke knew the researchers at TGen were rigorous and focused on applying their research to make an impact. Not only did they give her the data she needed, but they also did their own full analysis of her tumor, which she then sent to Ver Steeg to help extract those hidden factors.
“The first step there is to get a gene expression profile of your tumor,” Ver Steeg explained. “So there were hundreds of thousands of numbers that she got in the mail after she had sent in a sample of her own tumor for analysis. These numbers indicate the exact signature of your specific tumor.”
Pepke kept sending Ver Steeg gene expression data and notes on how the USC ISI researcher could improve CorEx. This wasn’t research that was happening in a lab. It wasn’t funded by anyone, and it wasn’t done to add another line to their CVs. It was happening in coffee shops, through email exchanges, text messages and phone calls. Pepke and Ver Steeg were just two Ph.D.s who wanted to make order out of chaos. And they got each other’s jokes, they say, with Ver Steeg often wishing her “a monotonic recovery.”
Pepke had gone through surgery and standard frontline chemotherapy shortly after her diagnosis, and from the spring of 2014 the cancer had been in remission. But in January 2015, as the Pepkes took down holiday decorations and welcomed a new year, the tumor recurred aggressively. Her doctors offered up a menu of second-line therapies, but none appeared any better or worse than the other. They turned back to CorEx.
Then, a breakthrough — CorEx discovered Pepke’s immune system was responding to the treatment but not killing the tumor. Instead of doing more chemotherapy, as recommended, Pepke convinced her oncologist to put her on a brand new immunotherapy drug that had shown the potential to “reset” the immune system, though only in mice.
“I told him, if I don’t do this, I will die soon,” Pepke said. “It’s a standard almost no researcher has. I have to believe it enough to bet my life on it.”
After immunotherapy, she went through one more round of surgery and chemotherapy. Yet in June 2015, it appeared that the cancer was progressing despite the therapy. Pepke decided to suspend all treatment and focus on her family.
“I just wanted to enjoy the summer with my kids,” Pepke recalled, holding back tears.
Two months later, when she returned to her oncologist, no signs of her tumor could be found, and her MRI was clear. As of now, Pepke’s cancer has been in remission for nearly two years.
“In the end, no one can say what happened or which treatment it responded to,” she said. “The disease progression was very consistent with an immunotherapy-type response.”
The research duo published their findings in a paper titled “Comprehensive Discovery of Subsample Gene Expression Components by Information Explanation: Therapeutic Implications in Cancer.” Pepke hopes that the findings will give women hope.
“I have met so many brave women fighting ovarian cancer, but the absolute fiercest are the mothers,” she said. “Many go through harrowing treatments and then go back home, put on a brave face and pick up their kids from school. I hope that this research can help some of them.
“A future in which cancer patients have the ability to create a personalized treatment plan based on their gene expression data is very close,” Pepke said. “Think about it — we can now assess an individual tumor and see what’s driving it to behave this way.”