The Battle to Stop Bird Flu

Wired Magazine
Jan. 01, 2006

The pandemic has hit New Mexico. Inside the Los Alamos weapons lab, massive computer simulations are unleashing disease and tracking its course, 6 billion people at a time.

On a cold January day in 1976, Private David Lewis came down with the flu. Struck with the classic symptoms - headache, sore throat, fever - Lewis was told to go to his barracks at Fort Dix, New Jersey, and get some rest. Instead, he went on a march with other grunts, collapsed, and, after being rushed to the base hospital, died on February 4. He was the first - and, as it would turn out, the only - fatality of the great swine flu epidemic of 1976.

Lewis' death came just as health officials were starting to worry about an influenza outbreak in the US. The best science at the time held that flu epidemics erupted in once-a-decade cycles; since the last epidemic had occurred in 1968, the next one should be on the near horizon. As an article in The New York Times put it just days before Lewis fell ill: "Somewhere, in skies or fields or kitchens, the molecules of the next pandemic wait."

At Fort Dix, a few other soldiers developed flu symptoms. When lab tests revealed that perhaps 500 on the base had caught the virus, officials at the Centers for Disease Control and Prevention faced a quandary. Was this the epidemic they'd feared, in which case they should call for mass inoculation? Or should they play the odds, hoping the disease would go away as often happens?

They had little information to go on: the outbreak in New Jersey, isolated cases from Minnesota to Mississippi, and a flu virus that looked suspiciously like the strain that killed half a million Americans in 1918. Estimates for the chance of an epidemic ranged from 2 to 35 percent. Indeed, there was much the scientists didn't know about influenza, period. Flu viruses hadn't been isolated until the 1930s, and they are moody, fast-mutating pathogens. "The speed with which [mutation] can happen," the Times wrote, "is mystifying." When a strain was identified, there was no telling how virulent it was. At the time, the best computer models were in Russia, where health authorities were doing a fair job predicting the spread of flu from city to city. But those models took advantage of the Soviets' penchant for tracking the movements of their citizens; in the US, where travel was open, it was impossible to create such a forecast.

So on March 24, 1976, President Gerald Ford convened a "blue-ribbon panel" of experts from the CDC, the Food and Drug Administration, and the National Institutes of Health. After a few hours, Ford emerged with Jonas Salk, the doctor behind the polio vaccine, by his side and announced a plan "to inoculate every man, woman, and child in the United States." It was to be the largest immunization drive in US history.

The inoculations began on October 1. As of mid-December, 20 percent of the population had received a shot. But by then, it had become clear that an epidemic was not, in fact, at hand. Lewis remained the only fatality - unless you count the 32 other people who died from the vaccine. Soon the program, the last major inoculation effort in the US, was canceled.

The 1976 swine flu scare has become enshrined as "the epidemic that never was," one of the great fiascoes of our national health care system. But in truth, government officials performed well enough. In just a few months, they went from isolating a strange new flu virus to delivering a vaccine to every American who wanted one. The problem was, all they had were blunt instruments: crude mathematical models, rough estimates of infection rates, and a vaccine that often packed too strong a punch. They were fairly well equipped to react to a worst-case scenario - they just weren't equipped to determine if one was imminent. Forced to guess, they chose "to risk money rather than lives," as Theodore Cooper, an assistant secretary of Health Education and Welfare, said at the time. "Better to be safe than sorry."

All of which raises a question: With the specter of an actual flu epidemic looming, are we any better equipped today? H5N1, the strain of avian influenza currently festering in Asia, has yet to pull off the mutation that would customize it for human-to-human transmission. But we know it's an especially lethal virus; most health experts expect it will make that jump soon enough. So the task for experts is to devise a plan that pinpoints how the virus might spread through the US population - a plan that draws more from the Soviet approach to disease forecasting than from the CDC's approach in 1976.

Thirty years on, a new science of epidemiology is at hand. It's based on sophisticated computer models that can get ahead of a virus and, in a sometimes dazzling demonstration of computer science, provide exacting prescriptions for health care policy rather than best guesses. It's an approach pioneered not by physicians but by physicists. And it owes a lot to the nuclear bomb.

In 1992, the US announced a moratorium on nuclear testing. The move meant that the Pentagon could not use underground test explosions to "certify" its arsenal of weapons - to establish that its nuclear stockpile would work when called upon and be safe until that day. That forced the guardians of the stockpile - the nuclear scientists at Los Alamos National Laboratory in New Mexico - to devise new ways to do their jobs. And that meant massive supercomputer simulations.

Computer simulations have a long history at Los Alamos. They were first deployed at the lab during the Manhattan Project in the 1940s to model nuclear explosions - among the first computer models ever attempted. Early on, they were a coarse tool and no substitute for physical experiments; the physicist Richard Feynman, who worked at the lab in its earliest years, called them "a disease" that would lead scientists into computerized daydreams tangential to the task at hand. But over the next 50 years they became an important instrument at Los Alamos, indispensable to the study of nuclear fusion and rocket propulsion. The 1992 moratorium simply codified that role, making computer models the only game in town. Since then, the lab has built one of the world's largest supercomputing facilities, amassing a total of 85 teraflops of processing power.

These tools are now being used in research that goes far beyond weapons work. Among the lab's 6,000 scientists, you'll find astrophysicists modeling white dwarf stars, chemical engineers replicating the effects of Florida hurricanes, geologists modeling the Earth's core, and biologists constructing microbial genomes. "All science is simulation these days," says Stephen Lee, the deputy division leader of computational sciences at Los Alamos.

The most promising application of sim science to real-world policy targets epidemic disease. A mile from the main compound at Los Alamos, in a grade school turned research lab, half a dozen physicists and computer scientists (and one mathematical biologist) are grinding out disease like pepper from a mill. This is EpiSims, an ambitious computer-simulation project that has released anthrax in Houston, sown the bubonic plague in Chicago, and, most recently, spread the flu in Los Angeles.

In 2000, EpiSims let loose smallpox in Portland, Oregon. Programmers started by creating a computer model of the city that's accurate down to the individual high school, traffic light, and citizen. In EpiSims, as in life, people go about their daily business. So on Tuesday morning, John Doe leaves his apartment in the Pearl District at 6:45, stops at Starbucks at 7:08, gets to his office parking lot at 7:45, greets his colleagues in the elevator at 7:49, and is at his desk checking email by 8:02. There are three-quarters of a million John Does in EpiSims' Portland, and just as many Jane Does, each with their own routines and encounters. This is the secret of EpiSims: its insatiable appetite for minutiae. EpiSims is the closest we've come to a huge city living inside a computer - or more specifically, several hundred computers. James Smith, who runs the EpiSims project at Los Alamos, describes his tools as "giant data fusion engines." Tapping the scientists' sophisticated computing algorithms and the lab's supercomputer clusters, it takes about 300 parallel processors and less than 24 hours to run a one-year simulation.

Smallpox is an opportunistic virus, eager to take advantage of incidental encounters. It spreads through the respiratory system and incubates for as long as 10 days before the onset of fluish symptoms - coughing, fever, stomachache. Only days later do victims develop a pustular rash - the pox. It is vicious; in an untreated smallpox epidemic, 30 percent of those infected will die.

In the 2000 smallpox sim, the EpiSims team tracked the virus as it climbed toward its 30 percent fatality rate not all at once, but person by person: schoolteachers and shop clerks first, then office workers and hospital staff. As smallpox leapt from one unwitting victim to another, the EpiSims team watched disease ooze out of schools and shopping malls, erupt in downtown office buildings, and take root in neighborhoods.Within 90 days, Portland was teeming with smallpox. The epidemic was at hand.

But simulating the spread of disease is only half the job. EpiSims also had to evaluate how officials should respond. So, researchers rebooted the sim and Portland was once again alive and disease free. And this time the city had a plan of action. Four days after the first sign of virus, the authorities closed the schools, kicked off a mass vaccination program, and generally shut the city down. And with it, the disease: In 100 days, it had run its course. That sim was followed by another with a slightly different response strategy, and then another. EpiSims eventually ran through hundreds of smallpox models, sometimes vaccinating only exposed individuals, other times targeting the so-called superspreaders, individuals who transmit more than their share of disease, sometimes putting the entire city in quarantine. With every tweak, the disease would peter out or gain steam accordingly.

The EpiSims smallpox models led to a handful of contrarian conclusions about epidemic disease. The first: "The superspreader hypothesis isn't necessarily true," Smith says. This rule holds that in any population, the more social individuals - the hubs - are the principal conduits for spreading disease. Shatter the network by inoculating or removing these hubs, the theory goes, and you'll stand a better chance of knocking out the disease. But EpiSims has shown that we're all more popular than we might think. Even the most reclusive of us runs to Walgreens for toothpaste or drops by Boston Chicken for takeout. For a highly communicable disease like smallpox or influenza, these incidental interactions spread disease just as well as extended encounters. So chasing after the hubs can mean chasing after 80 percent of the population - a huge waste of time and energy. Better simply to inoculate the entire city.

A second revelation: With a lethal pathogen like smallpox, response time is all. As the delay stretches from 4 to 7 to 10 days before officials move into action, EpiSims found that the outbreak becomes increasingly lethal. It turns out that, in the ticking moments after an epidemic strikes, when health officials act is more important than what they actually do. Start with inoculation. Or quarantine. Or school closings. It doesn't matter. What does matter is reducing the time between first outbreak and first response. At the same time, EpiSims warns against overreacting to a less-lethal disease - as in 1976, when standard health measures would have sufficed. (How to tell the difference? Run a simulation.)

These sorts of precise, real-world conclusions are the payoff of the EpiSims approach. They are, to use Smith's term, "actionable" - worthy of consideration not just by scientists but by policymakers. Such relevance has made EpiSims a darling at Los Alamos and an integral component of a Department of Homeland Security project called Nisac (for National Infrastructure Simulation and Analysis Center), an effort to model a range of disasters and plot recovery strategies. Born in 2000 as a tiny $500,000 joint project between Los Alamos and its sister lab in Sandia, New Mexico, Nisac got a $20 million infusion after September 11, 2001, and a mandate to measure how the nation would fare after another deliberate attack, be it a dirty nuke or a bioweapon like smallpox.

More recently, as attention has turned to DHS's responsibility for acts of God as well as acts of terrorists, EpiSims has begun assessing the threat of avian influenza. With these new simulations, Smith's team is adding even more granularity. They're modeling the health care system down to the hospital bed, to see what happens if flu victims flood hospitals, fill the beds, and then spill back into their homes. They're taking into account slight behavior changes, so if people start wearing surgical masks, SARS-style, disease transmissions in the sim will fall off according to the masks' particulates-per-million filtration rate. The results go to the DHS and straight up the chain, helping inform the ultimate question that looms behind all of Nisac's work. "What do we tell the President?" says DV Rao, who directs the lab's Decision Applications division. At these highest levels, this sort of predictive science is an entirely new and unfamiliar decision-making tool. "I don't know what they make of it now," says Rao. "But in a year, hopefully they're going to say, 'All right. Tell us what we should be doing.'"

On a long flight to Maui in February 2003, Tim Germann, a physical chemist at Los Alamos, was reading Richard Preston's Demon in the Freezer. A vivid account of what's at stake if the last samples of smallpox escape US or Russian labs, Preston's warning struck Germann as real enough. But the scope of the danger was unquantified and apparently unquantifiable. Germann also had in his bag a copy of Science that included a piece coauthored by Emory University biostatistician Ira Longini. ("It was a long flight," Germann says.) Longini was investigating different vaccination strategies in the event of a smallpox outbreak. But his simulation sample totaled just 2,000 people - not large enough to extrapolate his conclusions to a larger population.

The reading made Germann wonder: Sure, a national outbreak of smallpox would be bad. But how bad, and how likely? And who would be at risk? Then Germann realized that he had a way of finding out. His day job involved computational materials science, specifically, how metal atoms - copper and iron - would react under stress or shock. In an epidemic, Germann thought, people might behave like the atoms in his simulations. "Atoms have short-range interactions," Germann explains. "Even though we're doing millions or billions of them, every one just moves in its local neighborhood. People work in the same way." So, just as a cooling metal slows down atoms, a quarantine slows down people. By bolstering these physics models with experimental data on things like how viruses circulate from children to adults, and he could conceivably model the entire US population, or even the entire global population.

Germann cranked up the simulation, adjusted the software, and added the parameters Longini had used to model his smallpox outbreak. That marked the birth of what would be called EpiCast - a combination of epidemic and forecast. "It's basically still the same code," Germann says. "We use it one day for running atoms and the next for people." As it turned out, the model showed that smallpox may not be the cataclysm many imagine. Because of the long lagtime between successive generations of an outbreak (two weeks or more), and the tell-tale symptoms, smallpox would be quickly identified. That, plus the stockpiling, post-9/11, of large quantities of vaccine, means that "it should be possible to contain" an outbreak after the first few waves, Germann says.

The flu, by contrast, has a very short generation time (days instead of weeks) and generic symptoms. What's more, it's nearly impossible to stockpile a vaccine because the virus is so quick to mutate. Add the fact that people can be infected and contagious without knowing it, and you've got one vexing virus. So Germann called Longini and described how his molecular models could be adapted to epidemics. "I thought it was a bit preposterous," recalls Longini, who has been modeling epidemics for 30 years, most recently with a National Institutes of Health research program investigating the risk of pandemic influenza. "When Tim told me he could model the whole country or the whole planet, 6 billion people, it sounded very impressive. But I wondered if it was really possible."

So Germann set to work creating flu scenarios to augment Longini's NIH work. With nearly 300 million agents representing every man, woman, and child in the US, EpiCast doesn't bother to track minute-by-minute behaviors as EpiSims does. Instead, Germann puts his computing power to work detailing how slightly different parameters - various antivirals or different isolation policies, for instance - have slightly different national repercussions. So far, the project has run about 200 simulations of an avian flu epidemic, models that have helped Longini's group reach provocative conclusions that fall along two lines: how a nationwide outbreak might take hold, and what policies would best combat it.

EpiCast reveals that, in contrast with flu epidemics of decades past, an outbreak today won't progress "like a wave across the country," spreading from town to town and state to state. Instead, no matter where it erupts - Seattle, Chicago, Miami - it will swiftly blanket the nation. "It starts in Chicago one day," Germann says, "and a couple of weeks later it's everywhere at once." Thank the airlines. Even though disease has piggybacked on air travel for decades, we generally had only isolated outbreaks of low-transmission viruses - like when SARS leapt from Hong Kong to Canada in 2003 but failed to spread beyond Toronto. In an epidemic of a highly communicable disease, the airlines' hub network would effectively seed every metropolitan area in the country within a month or two - and then reseed them, repeatedly.

EpiCast showed that local intervention measures can have some impact: Close the schools, enforce a quarantine, and the disease will slow down. That buys the federal government time to develop and mass-produce a vaccine. But Germann quickly adds a caveat: Acting locally may not be enough. In a worst-case outbreak, without a viable vaccine, "the disease will climb, and eventually go exponential. And once it's on the exponential curve, it's very difficult to contain." Cue Richard Preston.

In November, the Department of Health and Human Services released its pandemic influenza plan. The report offers a thorough and frank assessment of the havoc a full-fledged pandemic would wreak. The nation, the report says, "will be severely taxed, if not overwhelmed." Disease will break out repeatedly, for as long as a year. Hospitals will run out of beds and vaccines. Doctors and nurses will be overworked to the point of exhaustion. Mass fatalities will overwhelm mortuaries and morgues with bodies. Before it has exhausted itself, the report estimates, the disease could spread to as many as 90 million Americans, hospitalizing 10 million and killing almost 2 million.

The report also sketches out how the federal government should respond in such a scenario. In effect, officials face what Bruce Gellin, director of HHS's National Vaccine Program Office, describes as a reverse Hurricane Katrina: Rather than an all-out response focused on one particular region, a flu epidemic would force the government to ration its resources to serve the entire nation. How best to do that - tactically, quickly, and effectively - is now the focus of EpiCast's work.

After the HHS plan was released, Germann and Longini were called to Washington for a strategy session with officials from the NIH, DHS, and the White House. Plenty of Los Alamos scientists, starting with Oppenheimer and Feynman, have made the trek to the corridors of the Capitol. But those trips were concerned with fighting wars, not disease. During the HHS meeting, the officials talked about how to apply EpiCast to the problem at hand. Germann explained the power of the tool. If HHS wants to know where to stockpile antivirals, EpiCast can pinpoint optimal locations. If the government wants to slow down the spread of disease, EpiCast can suggest whether to screen airline passengers by body temperature - and determine just how high a fever is too high to fly. If the first outbreak is in, say, Los Angeles, "do you send doctors from around the country to the West Coast," Germann says, "or keep them where they are because it'll be everywhere in a few weeks?"

Germann assured the group he could help. Then he returned to Los Alamos. Every question means a new sim, and every sim helps answer questions that are otherwise unanswerable.