On Monday mornings, Bob Michaels walks into the infusion center at Weill Cornell Medical College in New York City and takes a seat in a comfortable barcalounger.
An oncology nurse connects the port implanted in the retired university professor’s chest to a portable IV pump. The device will deliver a continuous supply of an experimental therapy over the next four days, as he carries it around in a small shoulder bag.
Doctors diagnosed Michaels, 70, with bladder cancer in the late summer of 2011 (we’ve changed his last name for medical privacy reasons). Despite several rounds of surgery and chemotherapy, the cancer continued to metastasize. His doctors were running out of treatment options.
Which is what brings Michaels to that barcalounger. Weill Cornell is one of several facilities participating in a clinical trial for the promising cancer drug, known as BPM 31510.
The study itself is largely indistinguishable from the hundreds of cancer trials under way at any given time in the United States — but the drug development process was unorthodox. It wasn’t a scientist who spotted the potential of BPM 31510. It was an artificial intelligence program, running on the servers of a Framingham, Mass., startup named Berg Pharma.
It’s among a growing number of companies and researchers applying smart algorithms and massive amounts of data to sift through gigantic stacks of medical research or the biology of the body itself for clues that could save the lives of cancer patients.
The techniques in question cross the fuzzy boundaries of AI, machine learning, computational medicine, quantitative pharmacology and plain old big data (and any practitioner will happily argue at length about which is what and why their approach is superior). But institutions as big as IBM, Merck, Memorial Sloan Kettering, UC Berkeley and the U.S. Food and Drug Administration are eager to explore the potential — and in most cases are investing millions to do so.
The research efforts roughly break down into two tracks: Those using these computational tools to improve personalized medicine, pinpointing the most effective existing drugs against an individual’s specific cancerous mutations — and those, like Berg, attempting to develop brand new treatments.
Therapies such as BPM 31510 will ultimately have to pass the same hurdles as any drug candidate. But the company hopes the approach vastly accelerates drug discovery and dramatically reduces the cost. The average price of developing a successful treatment easily surpasses $1 billion — but can exceed $4 billion when failed drug candidates are taken into account, as Forbes has noted.
“We think we’ll cut the drug development time at least in half and cut costs at least by 50 percent or more,” said Niven Narain, president and chief technology officer at Berg. “Our goals are to really make a tremendous impact on changing the American health care system.”
Michaels’ goal is more immediate and personal.
“My real hope, the deep hope, is that this is a home run,” he said. “That it wipes out the metastatic cancer and I live another thirty years like a healthy person.”
Grist For The Mill
There are plenty of reasons to keep expectations in check for any single drug candidate — or the computational approach in general. More than 95 percent of drug candidates fail in clinical trials and there’s a long history of premature claims in medicine, particularly when it comes to cancer.
“This is certainly promising, but it doesn’t mean we’ll be successful,” said David Patterson, a professor of computer science at UC Berkeley developing machine learning tools for cancer research. “Many people have gotten excited about new technologies many times in cancer.”
But he and other observers strongly suspect that digitizing biology will ultimately represent a fundamental step forward. A convergence of forces may have finally put medicine onto the trajectory of Moore’s Law, promising accelerating advances in understanding and treatments, including: the plummeting cost of DNA sequencing, the accelerating power of computational tools, improving understanding of the genomic basis of cancers and growing mounds of medical data.
In a strong signal of the perceived promise of these approaches, the FDA last month announced funding for the UCSF-Stanford Center of Excellence in Regulatory Science and Innovation, a joint effort designed to leverage diverse data sets and computational tools to accelerate drug development.
“We’re seeing this ‘big data’ trend in everything, but at least in this area it’s very exciting,” said Michael Keiser, an instructor at the UC San Francisco School of Medicine and founder of SeaChange Pharmaceuticals. “The traditional problem was just not having the grist for the mill.”
My Dear Watson
Glioblastoma multiforme is an aggressive brain cancer that, by some estimates, kills more than 13,000 people in the United States each year (including, in 2008, my father).
The existing standard of care is a combination of “debulking” surgery, radiation and chemotherapy. It can add months to a patient’s life, but typically the cancer makes a swift return.
In March, IBM announced a clinical study in partnership with the New York Genome Center aimed at improving the odds by using the tech giant’s Watson artificial intelligence system to hit upon more personalized treatment plans.
Cancers are often classified into location types — lung, breast, cervical — that suggest they’re homogenous. But tumors are unpredictable bundles of mutations, as many as a million among the three billion nucleotide base pairs in the human genome. And different mutations respond in different ways to different treatments.
For the approximately 25 patients in the study, the genome center will conduct whole DNA sequencing of normal cells and tumor cells to ascertain the precise mutations at work.
What typically happens at this point is technicians will tediously scan for known mutations and match them to drugs with demonstrated success in treating them. But the vast number of possible mutations and endless cancer studies add up to a big data problem that’s outstripping the capacities of human minds.
“You’re really trying to find the needle in the haystack,” said Steve Harvey, global technology and analytics leader at IBM.
Enter Watson, which can gobble up entire databases of medical literature and easily handle terabytes of genomic data, drawing lines between mutations and treatments that might have been missed before. The system can literally cut down the process from weeks to minutes, while covering a broader swath of the scientific literature.
Glioblastoma patients are in a race against time. Most will die within 15 months and many will pass away much sooner — so cutting down the amount of time it takes to identify treatments is critical in itself.
Many drug candidates are abandoned if they didn’t directly attack the principal target of a particular study, even if they were shown to be effective against some other mutations. But any drug candidate that made it through a Stage 1 clinical trial could potentially be used for patients in the current study.
IBM and the New York Genome Center are in the early stages of developing the study, which will begin later this year. Separately, oncologists at the Memorial Sloan-Kettering Cancer Center are collaborating with IBM to develop a Watson application that could help doctors identify treatments for patients with lung cancer.
A New Frontier
But others say that identifying the best known treatments isn’t the real problem — it’s finding novel ones.
“It’s not that the answer is in the literature,” UC Berkeley’s Patterson said. “We need to do new experiments to find things that haven’t been tried before.”
His team is collaborating with Dr. Brian Druker at Oregon Health and Science University on a study of acute myeloid leukemia, a deadly form of blood cancer.
A growing number of scientists believe the most effective treatments against certain cancers will not be a single silver bullet — but a cocktail of treatments, much as with HIV.
For Druker’s study, which is just getting under way, the researchers will sequence numerous patient blood samples. Then they’ll feed that information plus clinical data about various compounds into the machine learning algorithms at UC Berkeley, hoping to spot a handful of drugs likely to help.
In turn, they plan to rapidly test different combinations in varying proportions against the blood samples in Druker’s laboratory.
That data will flow back into the software. It promises to create a positive feedback loop that makes the algorithms increasingly smarter, more adept at finding the customized combination that will be most effective for any given person.
In a sense, the study straddles the two major approaches mentioned at the start: Taking known treatments but applying them in a novel way.
“We’re going from a biologically-intensive field to a more computer science-intensive field,” Patterson said. “There’s been a million-fold improvement in the cost of sequencing. It’s turning that information into bits — and that’s why we can step in.”
“Computer scientists can deal with terabytes or petabytes of information,” he said. “That’s in our wheelhouse.”
A terrible case
Michaels felt rundown throughout the summer of 2011, uncomfortable in his own body. But he was determined to enjoy the season and put off the inevitable round of doctors visits until fall. Then before dawn one morning, he woke up and rushed to the bathroom.
“It came out like burgundy wine,” he said.
After the initial biopsy, Michaels’ doctors decided to attack what they believed was a contained tumor with a six-week course of Bacillus Calmette-Guerin, a vaccine that stimulates the immune system. It was delivered through a urinary catheter, which was “no fun,” he said.
Worse, it didn’t work. They next performed a radical cystectomy, the removal of the bladder, at which point his urologist discovered the cancer was more widespread than originally thought.
Michaels underwent a round of chemotherapy — and another and another. The treatments alleviated symptoms, but the cancer continued to metastasize, most recently surfacing in the omentum, part of the lining of the abdominal cavity.
Dr. Scott Tagawa, Michaels’ oncologist at Weill Cornell, finally counseled him to consider the Berg Pharma clinical trial.
“He’s a very straight shooter, who was willing to say, ‘this is a really terrible case and (BPM 31510) looks pretty good,’” Michaels said. “I decided to go for it.”
Berg Pharma, cofounded in 2006 by Silicon Valley real estate magnate Carl Berg, is squarely in the camp of looking for new treatments.
For any given condition, the company begins by running a broad array of tests on a large number of healthy and diseased samples. They analyze tissue, blood, urine and more to identify the constituent DNA, proteins, lipids, metabolites and more.
Then the company creates artificial intelligence computer models that compare the two, ideally highlighting what has gone wrong in the diseased cells, what heightened level of proteins or missing lipids might be at work.
For cancer, Berg Pharma homed in on mitochondria, a structure within cells that influences a process known as apoptosis — or programmed cell death.
It’s the mechanism that usually causes old or damaged cells to self-destruct.
“But cancer cells unfortunately have subverted this normal metabolic process,” said Peter Yu, the oncologist overseeing the Berg trial at the Palo Alto Medical Center in Silicon Valley.
This allows deranged cells to invade healthy tissue and often spread throughout the body, which is precisely why the disease is so deadly.
BPM 31510 appears to switch mitochondria back on, restoring apoptosis — at least in the lab and in animals. That’s especially intriguing because it would seem to work across cancer types — no matter the genetic mutation — by addressing the underlying mechanism itself.
Altogether, nearly 40 patients are enrolled in the study, which spans Weill Cornell, Palo Alto Medical Foundation and the University of Texas MD Anderson Cancer Center.
“We’re one of the first companies, if not the first, to bring artificial intelligence into medicine for the purpose of developing drugs,” Berg’s Narain said.
Ultimately, though, it doesn’t matter how novel the approach is, or whether it happens in academia or private industry. What matters is whether it leads to the identification or creation of treatments that save or extend lives.
At this point, Yu said that nothing can be said scientifically about the effectiveness of the drug in humans. The current trial phase will continue through 2015, and the drug would still have several stages to go before the FDA could approve it.
Michaels is only a case study of one and he’s on chemotherapy as well. But he personally feels like the drug is working. He says his abdominal pain has subsided, his energy has improved and he’s working out again for the first time in more than a year.
“It’s really extraordinary,” he said. “I feel it’s like mentioning a no-hitter in the ninth inning, I don’t want to jinx it. But I’m hoping it’s going to do the job.”
This article originally appeared on Recode.net.