Despite the flush atmosphere that Listwin cultivates, most Canary grants are in fact relatively modest. At an average of between $100,000 and $200,000, they are side projects for most researchers, covering their costs but hardly funding a full research lab. The festivities and the team-building, though, get the scientists committed to the greater mission of early detection. Soon enough, team members start assigning their junior scientists to early-detection work, and they engage outside resources and colleagues on the problem. In this way, Canary’s strategy is to “create leverage,” Listwin says, giving the foundation access to far more brain power and institutional muscle than its size might otherwise command.
The Canary approach comes at a time when the NCI is in the midst of what director Niederhuber calls “a big pivot” away from a single-minded war approach and toward a portfolio of strategies, including prevention and early detection. But 40 years into the war on cancer, he says, changing the course of the NCI is akin to turning around the proverbial aircraft carrier. That leaves the “well-informed higher-risk activities” to more agile groups like Canary. “New screening approaches are increasingly important. I think eventually you’ll go in for a pit stop on a regular basis,” Niederhuber says. “And with a little bit of blood, we’ll know what’s normal and what’s abnormal.”
The typical human body contains something less than 2 gallons of blood. The bloodstream is basically a transport system, a combination of plasma—the fluid itself—and a number of passengers, mostly red and white blood cells, which distribute oxygen and fight infection. Blood also contains thousands of proteins that serve a range of biological purposes, from distributing energy and nutrition to repairing injury and inflammation. The science of proteomics is trying to correlate each of these proteins with its specific metabolic function.
When the first Canary team came together in 2004, proteomics promised to be a powerful tool for early detection. All the teams needed to do was pump biomarkers through the testing process, identify a handful that link to early-stage cancers, corroborate the results with a CT scan or MRI, and then roll out the early-detection test. “It looked like a pretty simple problem,” says Patrick Brown, a molecular biologist at Stanford and member of Canary’s ovarian cancer team. “Get a molecule, make a test, and you’re done. It was just a matter of going out and finding them.”
Brown doesn’t think that anymore. “It’s gone from something that seems really simple and really boring scientifically,” he says, “to something that’s not at all simple and, therefore, really compelling scientifically.” He functions at Canary as something of a bug tester, probing for logical flaws, false assumptions, and wishful thinking. The complications that have turned up around blood protein biomarkers, he says, are riddles that must be solved before the way forward becomes clear. And two riddles stand out.
The first goes something like this: In the past decade, proteomics has been great at discovery—the eureka moment when a protein is identified and strongly associated with a cancer. The field has identified thousands of proteins in cancerous human tissue, and hundreds of research papers have claimed strong correlations between a particular new marker and a certain type of cancer. But there’s been a dearth of validation—the more laborious process of confirming the results and establishing that a protein actually does work as a biomarker for a particular cancer and isn’t the result of some unrelated condition like inflammation or anxiety.
The problem starts with the very structure of the proteomic investigations. Most of these are case/control studies, in which proteins extracted from known cancer patients (the cases) are compared with proteins extracted from healthy volunteers (the controls). In a perfect study, you want the cases and the controls to match up in every way—age, sex, diet, home town—except for the fact that half of the sample has cancer. That way, any differences that turn up are statistically likely to be due to the cancer. But in reality, good samples of cancer tissue are in short supply, so most research is done in a take-what-you-can-get mode. The controls are assembled afterward and matched as well as possible. The result is that the cases and controls often have little in common—they can come from people of different ages, different towns, or countless other variables. “So it’s not surprising that you find all sorts of differences between the cases and controls,” says Lee Hartwell, on whose watch the Hutchinson Center has become a leader in proteomics research. “But those differences could have nothing whatsoever to do with the fact that they have cancer.”
Take the case of prolactin. In 2005, a research group at Yale announced it had identified several biomarkers that together could work as a test for ovarian cancer. (More markers mean better odds of a true positive, since different people have different proteins in their blood at different times.) The Yale markers included CA125; osteopontin, a protein believed to be overexpressed in several cancers; and prolactin, a pituitary hormone found in the breasts, ovaries, and other organs. The test for early detection of ovarian cancer was released commercially by LabCorp last June under the name OvaSure.
The results troubled the Canary ovarian team, which had already taken stock of a few of these and other markers and ruled them as insufficient for a valid test. The inclusion of prolactin, in particular, stood out. “It looked wrong to me,” says Nicole Urban, head of gynecological cancer research at the Hutchinson Center. “It seemed highly unlikely that it was related to the cancer.”
So Urban ran her own study, comparing prolactin levels in women with ovarian cancer to those who were cancer-free. She also introduced further variables: when and under what circumstances the blood was drawn. It turned out that during a routine blood test, prolactin was present in normal levels among cases and controls alike. But the levels spiked dramatically when blood was drawn right before the patient went into surgery—whether it was surgery for ovarian cancer or another procedure. In other words, Urban concluded, it seems that prolactin isn’t a biomarker for cancer. It’s a biomarker for a stressed-out patient about to go under the knife. (Last fall, after the Food and Drug Administration warned that there were “serious regulatory problems” with the OvaSure test, LabCorp withdrew it from the market).
Photo: Alex Farnum, Retouching: Burkhard Schittny
The ambiguity over prolactin exemplifies the leap required to get from an apparent signal to a true signal. “A good biomarker will tell us something we don’t know,” says Martin McIntosh, who crunched the prolactin numbers with Urban. “But even worse is when you think you have a good biomarker, and it’s telling us something we don’t actually want to know.” And that’s the first riddle of biomarkers.
But assume that science eventually makes that leap and that a list of biomarkers with proven links to specific cancers is in hand. The next step is to find these markers in the blood. This is the second riddle: It’s one thing to find a biomarker in the research lab, using tissue known to be cancerous. But putting a test into clinical practice means finding a marker when it’s floating around in those gallons of human blood. Doing that accurately and consistently is a far more daunting proposition.
Patrick Brown first noted this problem in a presentation at the 2007 Canary Symposium. He started by laying out the yardsticks. The basic premise of early detection assumes there’s a window of opportunity when a would-be lethal cancer is germinating but potentially curable. For ovarian cancer, Brown put this window at about four years. Assuming annual or biannual screening, an effective test then must be able to detect a cancer when it’s too small to be lethal but large enough for a significant number of proteins to spill into the bloodstream. This boils down to a question of signal versus noise: Are current testing technologies, known as assays, accurate enough to catch those few extra molecules, or will they be mistaken for randomness?
Brown offered some preliminary calculations. He started by estimating the size of a pre-advanced-stage ovarian tumor during this window of opportunity. On average, these tumors are just 2 millimeters in diameter, or 4 milligrams in mass. “That’s less than one-ten-millionth of the mass of the average adult!” Brown noted. But with current assay technology, a tumor would have to be closer to 30 millimeters in diameter, he figured, to throw off enough biomarker molecules to exceed levels for normal women and to be reliably spotted amid all the other stuff in the blood. And at that size, he acknowledged, most ovarian cancers have already metastasized, so early detection wouldn’t likely save a life. According to these calculations, the prospects for blood-based early detection looked bleak.
For more than a year, Brown’s presentation hung over the project. It seemed to expose a paradox at the very core of early detection: What use is a biomarker if it doesn’t show up on a test until it’s too late?
The Canary approach may be collaborative, but it’s also competitive. Sam Gambhir, Brown’s Stanford colleague, had been working on a mathematical model to address the problem. Though Gambhir’s specialty is radiology and imaging, his PhD is in mathematics, and he thought some additional number-crunching might point the way. His model re-created the human bloodstream and sent some CA125, the known marker for ovarian cancer, into the mix. Soon enough Gambhir had his answer: According to his calculations, a blood test for a biomarker like CA125 can reveal a growth as small as one-half of 1 millimeter, “maybe even one-tenth of 1 millimeter,” says Gambhir, who published his calculations in PLoS Medicine this past August. “So it’s not out of the question to have a blood test that can detect a tumor that’s very small, small enough to work for early detection.” In other words: A biomarker test is possible. The cancer can be perceived.
Computerized tomography was developed in the 1960s in London at EMI, the electronics and recording giant. Legend has it that the Beatles made the technology, better known as CT scanning, possible; sales from their hit records allowed EMI to fund an engineer’s dabbling in medical imaging. The machines are like an x-ray machine in orbit. Where a traditional x-ray creates a two-dimensional image of the human body, a CT instrument rotates an x-ray on an axis around the body, producing a three-dimensional image or “slice” that’s much sharper and more detailed than a conventional x-ray.
Used at first for brain images, CT scans were a slow and tedious technology lagging behind x-rays, ultrasound, and MRIs for decades. In the 1990s, though, faster computation allowed for faster image processing, and several companies engaged in what came to be known as the slice wars. Image quality soon climbed along a geometric progression common to many technologies, from 16 slices per rotation to 32 to 64 to 128. The boom failed to reduce costs—the machines still run about $2 million apiece—but it made CT machines ubiquitous at American hospitals. Today, about 62 million scans are performed in the US annually, about twice as many as a decade ago. Even as warnings about overuse grow louder (the machines send 50 or more times the radiation into the body than a conventional x-ray), there’s an increasing call for putting CT scans to greater use, particularly as a potential screening tool for hard-to-see and hard-to-diagnose diseases like lung cancer.
While lung cancer kills more people worldwide than any other form of the disease, it remains comparatively under-researched. In part, this is because of the stigma it carries as a self-inflicted smokers’ disease. But it is also neglected simply because its location, deep within the body, makes it exceptionally hard to detect and treat. To this problem, CT scans offer a remedy. Compared with the foggy blur of an x-ray, a CT scan of the lungs is sharp and detailed. The lobes of each lung show up as a river system, the bronchioles that conduct air from the trachea fanning out into the alveoli, one tributary branching into a hundred more. Any unusual blip, be it from infection or cancer, shows up on this map as a well-defined land mass with a precise longitude and latitude.
In the mid-1990s, the International Early Lung Cancer Action Program began a 12-year study to examine the potential of CT scans as a screening tool for the disease. The study brought 30,000 smokers into hospitals and scanned their lungs, following up with another scan a year or so later. The scans turned up 484 cases of potential cancer, and subsequent biopsies confirmed that 85 percent of those patients did indeed have stage I lung cancer. It was a stunning result, far higher than many screening tests would have predicted. Even more remarkable was the survival rate: Of the 375 patients who opted for surgery, 92 percent were still alive 10 years later. The triumphant findings, published in 2006 in The New England Journal of Medicine, seemed to make a clear case for the widespread use of CT scans as a screening test for the early detection of lung cancer.
But there’s one question the study didn’t ask. “What if they’re finding things that look like cancer—even things that may be cancer under the microscope—but that aren’t the cancers that actually kill people?” asks Peter Bach, a pulmonologist at Memorial Sloan-Kettering Cancer Center in New York and a member of Canary’s lung cancer team. Though Bach and his Canary colleagues are eager to find a viable imaging test for lung cancer, they are wary about jumping onto the CT bandwagon. Their concern is that, by itself, a CT scan makes it too easy to rush to judgment. With no knowledge of a tumor’s molecular characteristics—the sort of information a biomarker test might provide—a CT scan offers an alluring but potentially deceptive image.