ARTIFICIAL INTELLIGENCE AND CRACKING THE CODE TO CANCER
Abraham Heifets of Atomwise is using AI to speed up the development of medications targeted at specific cancer cells
IN ANTICIPATION OF THE 50TH ANNIversary of NASA astronauts landing on the moon, Newsweek is spotlighting pioneers in science and technology, highlighting their very own moonshots and how they hope to change the world.
Abraham Heifets is the CEO and cofounder of Atomwise, a biotech company using patented deep learning artificial intelligence technology to predict and discover which drugs will be better, safer and more potent for cancer patients.
Q _ What is your moonshot?
A_ To make novel, better and safer drugs, with the ultimate goal to get medicines into the hands of patients faster.
Q _ How do you do that?
A_ We’re trying to solve how you modify a cell that’s in the runaway disease process, and figure out what’s causing a cell to keep growing and dividing. Think of proteins in your body as machines on the assembly line. If the machine governing cell growth and division breaks and goes haywire, then the cell will keep growing and dividing. That’s a tumor and how cancer happens. If you see a machine going haywire, you’d want to throw in a monkey wrench so the machine is busy chomping on that instead. Today, it takes about 15 years and several billion dollars to find a new drug. Every day that you don’t have good treatment, that’s real people, patients, lives and health in the balance.
Q _ How does Atomwise go about searching for the right drugs? A_ Every other industry uses computers for design. But in pharma, you have to physically make and test every one of those prototypes. If you think about designing a new airplane, you’ll simulate a thousand wings before you ever build one. And only after the computer says wing #88 will fly, be fuel efficient and be quiet, and only after you simulate thousands of wings do you then go build the prototype that you take to the wind tunnel for the test flight. Atomwise is about bringing that efficiency and that design into biology and drug discovery.
Q _ Using deep learning artificial intelligence?
A_ That’s exactly right. My co-founder, Izzy, and I were grad students at the University of Toronto when deep learning and this current era of Artificial Intelligence was being invented. Our computational biology group was on the same hallway as Jeff Hinton’s deep learning group. He just won the Nobel Prize of computer science, the Turing Award, for inventing deep learning. We saw pretty early on that the kind of work that was happening for image recognition and speech recognition could be applied to molecular recognition.
Q _ How exactly does AI enable safer, more effective and potent drugs?
Frankly, patients shouldn’t have to be patient. It’s our job to get medicines to them as fast as possible.” — ABRAHAM HEIFETS
A_ Imagine you’re a biologist and you’ve been studying pediatric cancer. You’ve done tons of experiments and you’ve determined that if you could just block protein X, that would halt the disease. Rather than trying to kill off every rapidly dividing cell, you want to be able to arrest the disease without harming the healthy cells. Now you need the drug that is effective and safe. AI lets us begin by testing 2,000 times as many molecules as has been tested before. Once you find some sets of molecules that look pretty good, you try to make variations that will improve the molecule. The computer lets you evaluate billions— instead of tens or hundreds—of molecules in one go, which means you’re going to find better answers and you’re able to discover that winning lottery ticket. Our 10-to-the-10 project is the next step in that, which is running 10 billion molecules against pediatric cancer targets.
Q _ Does it feel amazing getting that “winning lottery ticket”? A_ Absolutely. I think everyone goes into this because they want to help people. And frankly, patients shouldn’t have to be patient. It’s our job to get medicines to them as fast as possible.
Q _ What is “success” to you, and are you close to achieving it?
A_ Success for everyone in this field is helping patients. A measure of success is that if you look at our large pharma partnerships, you can see they’re embracing this new approach; you can see there’s trust in Atomwise’s AI systems. We recently announced a deal with Eli Lilly for over half a billion dollars. You’ll see Bayer and Pfizer. The previous big deal we announced was with Charles River Labs. These industry-standard players have embraced AI approaches.
Q _ How do you picture the industry in 20 years if you succeed?
A_ The industry is shifting toward AI. We’re actually running the biggest application of AI-to-drug discovery in history, and we have over 200 projects in every therapeutic area. So, I think the potential application is huge. About 35% of those projects are in cancers. At the end of the day, our success is patient success.
Cancer Center in Houston. “The progress is astounding.” Eventually those images, too, are likely to help AI systems go beyond diagnosing cancer to spotting hints of the vulnerability of a patient’s unique cancer.
Data Dilemma
deep-learning algorithms look at more data and analyze it more thoroughly than machine learning programs do. They are a bit like Seymour, the ravenous plant in Little Shop of Horrors, whose appetite never stopped growing. Although researchers and clinicians now have access to databases that contain information from as many as 250,000 cancer patients, it’s not nearly enough.
Thousands of different mutations in a patient’s genome can shape the development of cancers and determine which treatments are effective. Each cancer cell is a moving target, continually developing new mutations that can help it evade immune cells and survive powerful cancer drugs. Since AI software needs thousands of examples of a particular pattern before it can begin to recognize it, and since a particular pattern of mutations may come up in only a few thousand patients altogether, the software may well need access to the data of millions of patients to make faster progress. “We can make predictions now about how tumors will evolve and what treatments will work, but right now a significant fraction of those predictions are wrong,” says UCLA’s Paul Boutros, a physician who heads up cancer data science for the UCLA Jonsson Comprehensive Cancer Center.
A number of collaborations—with names like the International Cancer Genome Consortium, the Oncology Research Information Exchange Network, and the Actionable Genome Consortium—have sprung up among research centers and hospitals to share patient data. Gathered with patients’ permission and with personally identifiable information stripped out, that data could eventually help researchers reach the needed critical mass of information. “We need to get to the point where all these different data networks are tied together into a network of networks,” says City of Hope’s Caligiuri. Clinicians need access to that data, too, to find patients like the ones they’re treating to see what might work. “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,” he says.
Easing the Bottleneck
medicine is of no use if patients don’t have access to it. To get new drugs out faster, researchers are using AI to speed the process of drug development. One of the biggest causes of delay in testing new drugs is recruiting enough patients for a trial. Researchers not only need a group to try the new drug, but another “control” group to get the standard treatment, for purposes of comparison. Even when a new precision drug is promising, it can take
To anyone who has just received a diagnosis of cancer: You need to get a from an oncologist
SECOND OPINION who is a specialist in your type of cancer before you start any treatment.
years to run the trials that demonstrate the drug actually works for an identifiable group of patients.
To speed things along, researchers are starting to use high-powered statistics and computer models to avoid having to recruit a control group at all. Instead, they use a mashup of data from past studies to predict how a real control group would fare. “The results you get from a synthetic control arm are as reliable as if you had actually enrolled control-group patients in the trial with the same physicians and protocols,” says Glen de Vries, president of Medidata Solutions, which has designed the statistical tools.
That won’t be enough to ease the trial bottleneck for clinicians and researchers hoping to come up with precision treatments for the deadliest, most aggressive cancers. For instance, glioblastoma, the brain cancer, has the lowest median survival time from diagnosis—15 months—of any major cancer. It’s challenging enough to design a drug that can make it through the blood-brain barrier to get at a glioblastoma tumor. The disease works so quickly that there’s barely time to give an experimental drug a chance to show whether or not it is effective.
To give more experimental precision drugs a better shot at glioblastomas, the newly created Ivy Brain Tumor Center at the Barrow Neurological Institute in Phoenix has developed “accelerated trials” for its brain-cancer patients. A newly diagnosed patient is first given a dose of an experimental precision drug. The dose is too small to harm the patient (in case it turns out to be toxic, always a risk with new drugs) but big enough to reach the tumor. After surgery, doctors test the tumor to see if the drug had any effect. If it did, the patient continues with an increased dose. If not, the patient and doctor find out in time to take another course of treatment. “Speed is the key to finding drugs that work,” says Ivy director Nader Sanai. The approach has already turned up a personalized treatment that in one patient’s case beat back a form of brain cancer called malignant meningioma.
While all these approaches together are likely to bring us closer to the day when most cancers succumb to precision treatments, no one thinks that day will be here soon. Still, the move to personalized treatments is benefitting almost all cancer patients by sparing them the ordeal of a treatment that has little chance of working. “If you can look at a genomic or other test and know ahead of time whether or not a patient’s tumor will respond to a treatment, then even if only one out of 100 patients responds you’ve saved 99 patients from unnecessary complications and expense,” says Stanley Robboy, vice-chair for diagnostic pathology at the Duke University Cancer Center. “These drugs can cost $100,000, and can bankrupt families.”
Even that modest benefit, however, is being denied to most advanced cancer patients today. Health insurance companies frequently balk at paying for the genetic tests, which can cost as much as $10,000. “Medicare and some companies are starting to provide some coverage,” says Roychowdhury. “But it’s an arduous process to get reimbursed for the testing, and it’s hard to get the cutting-edge tests covered at all.” That’s one reason most of the top cancer centers in the country don’t routinely provide the testing to all their