By Mark D. Uehling
December 15, 2003 | JOHNSON & JOHNSON recently decided to kill a promising target—a protein against which a drug could work—for diabetes. The company obtained its crucial data from neither a clinical trial nor animal experiments. Rather, the insight came from computational models. They showed the target was important in only acute metabolic disease, not the chronic, day-to-day grind of diabetes. "Other pharmaceutical companies have been working on this target for the last five years at least," says Richard Ho, head of medical informatics at J&J. "I presume they are still."
J&J relied on software and human expertise from Entelos, a California firm of engineers and biologists that is re-creating human physiology and clinical outcomes with flowcharts and differential equations. A key thought leader at J&J was initially skeptical, even hostile, to the notion that any computer could help him do his job. That's changed. "Some of the team leaders would like to use these simulation results and insights as part of their internal submissions to the company," Ho says.
|Supermodels: A Gallery
|What was simulated, simulator, and details...
Simulating clinical trials is a rapidly maturing science. Many big pharmaceutical companies routinely simulate large numbers of clinical trials with thousands of virtual patients. The results of those simulations are fed back into the pharmacological models that underpin them. And then the simulations are used to sharpen the designs of trials in real patients.
Simulations performed with software from Globomax, Entelos, and Pharsight (see "A Virtual Pharmacopeia," Nov. 2002 Bio·IT World, page 58), seldom show up in the peer-reviewed literature. So academics in the modeling and biostatistics field are unable to kick the tires. But the work is expanding. It's common to simulate much of what happens in early-phase clinical trials—optimizing dosing regimens—and computers are increasingly being used to predict side effects and adverse events in much larger, more complex trials.
Such compute-intensive projects may take days or weeks to run and help sponsors decide which drugs should be put into living patients at all. Indeed, the savings from a more aggressive use of simulation could be one of the few bright spots for an industry that has embraced a procession of sexy technologies (genomics, proteomics, combinatorial chemistry, high-throughput screening) that have only exacerbated swollen research budgets. Still, there is hesitation about taking simulation beyond isolated therapeutic teams. One casualty of cool enthusiasm for simulation: Physiome Sciences, which threw in the towel by merging with Predix Pharmaceuticals.
Simulating trials on a companywide basis, according to some observers, could revolutionize clinical trials that are often designed and executed haphazardly. "They would save people a lot of money," says Tony Greenfield of Greenfield Research in England. "You can run a trial in a matter of minutes and find out if your trial has problems." Greenfield's MetaGen software is specifically designed for clinicians who can put in a few basic assumptions—blood pressure will drop, or it will drop even more steeply in women—and see how they play out. Greenfield notes that many clinicians (and humans in general) have difficulty holding more than one variable in their minds at once.
"The thing is to try to predict what these unforeseen events might be," Greenfield says. "I have made assumptions, but they might be wrong. Let's see how wrong they can be and still get a satisfactory answer to the trial." His software exports data into the usual programs—Excel, SPSS, Minitab, SAS—used to analyze real-world clinical results.
|A Single Phase III Trial—And You're Done?
|We talk to Carl Peck, director of Georgetown University's Center for Drug Development Science about his proposal that by using biomarkers and other technologies, including simulation, just a single clinical trial might be necessary for FDA approval.
Others make predictions about the ultimate impact of modeling and simulation that are far more bodacious and audacious. Former FDA official Carl Peck, now director of Georgetown University's Center for Drug Development Science, believes that a variety of simulation techniques and biomarkers could make a relic out of the excruciatingly expensive 75 to 100 separate clinical trials conducted for each drug submitted to the agency (see "A Single Phase III Trial—And You're Done?"
Peck can tick off modeling titans at virtually every pharmaceutical company in one breath, acknowledging the field is increasingly accepted within the industry. But even as jetliners and cars are routinely built in silico, drug executives hesitate to fully exploit what simulation can do. "This industry is a real latecomer to this technology," Peck says. Caustically, he notes it is routine for the results of dozens of real-world clinical trials in each new drug application (NDA) to be discounted by expert reviewers at the FDA. That implies billions of dollars in clinical budgets are wasted every year.
FDA: Model More, Waste Less
Peck believes simulation could be part of the answer to reducing the optimal number of Phase III trials to ... one. That's right: one trial. Long an exponent of simulation, he believes it could expedite Phase I-II safety and dose-finding research. Although modeling may not be well appreciated among pharma management, Peck says, the FDA is comfortable with simulation data and has been using and promoting it for years. "The FDA is ahead of the industry," Peck says, adding that the cancer field is especially prone to concoct un-blinded, un-randomized, underpowered trials that generate few clinically relevant results. "They ought to fix their medieval processes," he says of oncologists and those who fund their research.
"Many new drugs have benefited from modeling and simulations, in many ways," says Peck, a former director of the FDA's Center for Drug Evaluation and Research. Amgen and Wyeth's "Enbrel, approved for once-per-week dosing in adult rheumatoid arthritis, benefited from simulations of the new once-weekly dosing regimen that predicted its safety and effectiveness relative to the already-approved twice-weekly dosing. FDA agreed and required only one small confirmatory safety study."
At the FDA, associate director of pharmacometrics Peter Lee says a new initiative at the agency, announced last month, will allow two companies a month to come to Maryland to discuss their models and simulations. "This is exploratory," Lee explains. "It's optional, voluntary." The idea is to have a drug's sponsor and the FDA in rough agreement about the validity of a model while there is time to tweak the design of a trial. Too often, Lee says, the FDA sees marginal models when it is too late to affect what happens in the real clinical world.
Lee says the FDA is working to improve the understanding of simulation in the scientific community. He stresses that real patient or animal data are the bedrock. "A lot of people think simulation—'Oh, you're making up data,'" Lee says. "'You simulate something that doesn't exist.' We have been trying to change that perception. Every time we present simulation results, we tell them, 'This is the study, this is the data we are basing our simulation on,' so that people don't have any doubts about the source of the information we're using."
The field is probably more advanced than many managers and scientists in industry realize. At one tiny company, Optimata, based in Israel, Zvia Agur and her colleagues recently modeled red-blood-cell generation in mice, monkeys, and humans. Their simulated results match physiological records almost exactly, reflecting hundreds of biological parameters that were rendered mathematically. Some of the results have been published (Skomorovski, K. et al. Br J Haematol 123, 683-691; 2003). Novartis is a customer of Optimata's software.
|How Entelos renders metabolism in mathematical terms
In general, there are too many approaches to modeling to describe in any meaningful detail. At their most basic, however, some models incorporate pharmacokinetics (factors that affect the amount of chemical agents over time) and pharmacodynamics (the effects of chemical agents over time). Other models are purely statistical, making almost no attempt to quantify biology. Still others (see "Digital Diabetes"
) drill into the most intricate niches of biochemical pathways, relying heavily upon systems biology and quantifying the rates by which every last protein, enzyme, or sugar comes into being or is chemically transformed.
Often, many of the tools are used in concert. A common starting point is NONMEM, software originally developed at the University of California at San Francisco and now licensed by Globomax. Thomas Ludden, vice president of pharmacometric research and development at the company, says the most complex pharmacokinetic and pharmacodynamic models may take as long as a month to run.
Digital Dosing Tool
Some consultants specialize in little more than running NONMEM jobs. Ludden has heard of customers running large modeling efforts on 200 computers. Like everyone else in the industry, he stresses the iterative nature of working with models: "Some people are using these models very formally or informally in determining first-in-man dosages or Phase I dosages," he says. "Those data can be fed into the model and hopefully define what you see in the next set of experiments."
At Entelos, meanwhile, there is a bit of frustration that the world at large has not grasped the practical nature of the company's tools. "It really is ready for prime time," says James Karis, Entelos president and CEO. But in the next breath—despite high-profile customers such as Pfizer and J&J—Karis acknowledges a difficult sales process. "It's a hard sell. There is a lot of dialogue about systems biology. There is not a widespread understanding of where some companies are in terms of their progress."
Again and again, customers have tested Entelos' programmers by asking them to model something that has already been measured in vivo, in real animals or patients. Again and again, Karis says, for the few select diseases the company has modeled—asthma, diabetes, obesity and rheumatoid arthritis—the simulated answers match the ones in the real world. The company's specialty: helping customers figure out dosages, trial durations, and patient populations.
But the real power of the software is in exploring 20 or 30 extra hypotheses that no drug company could afford—or that no real trial will ever be able to address. Says Karis: "We can make 'knockout' humans—knock out a pathway and say, what happens? We do double, triple doses. Things you could never do in a clinical trial ethically. Simulation becomes a whole investigative research capability. It's not just to replicate what could happen in a clinical trial. We take it way beyond that."
Digital Meets Flesh and Blood
Clearly, that sort of thing is already happening at companies like Eli Lilly. The digital patients in Sandy Allerheiligen's computer mirror flesh-and-blood counterparts. As director of pharmacokinetics, pharmacodynamics, and trial simulations at Lilly, Allerheiligen is in charge of simulating clinical studies to make trials in the real world more effective.
In developing Lilly's anticancer drug Gemzar, the company had plenty of clinical data. Even then, the simulations helped, illuminating effects the company had not fully understood in the clinic. It was the computer, Allerheiligen says, that correctly predicted that the drug would have very different effects on elderly females. "We might have been underdosing the younger men," she says, adding that the insights from the population-based approach were included in the FDA-approved label for the drug.
Lilly has been using pharmacokinetic and pharmacodynamic analyses for a decade, venturing lately into the computational prediction of adverse events. Lilly's data from population modeling and in silico simulation are routinely presented to the FDA. The data origins are clearly marked, even when a bit of information from real patients and from digital patients has been combined. The FDA seems to be sophisticated enough to interpret the data.
Lilly uses NONMEM and Pharsight software. The latter company's tools are especially useful in helping executives or nonstatisticians visualize complex results. "A person who is not used to looking at a series of differential equations and translating that likes that picture," Allerheiligen says, adding that the simulations are run on heavy iron—Sun and Linux workstations.
At its most basic, the computer allows Lilly to surmount the chronic, eternal problem of clinical research: the shortage of patients. Using simulated data, or data from real trials, or a combination of both, Lilly's statisticians and programmers can spin scenarios that explore what effects a drug might have in, say, a population of elderly Hispanic men—when it might be next to impossible to recruit a real trial filled with such patients.
Allerheiligen says the technology both saves time and amplifies scientific insights that are not available in clinical and animal studies: "We cannot do a clinical pharmacology study for each and every population. You couldn't recruit the patients. This enables us to do a screen and look at all the patients and look at the co-variates that are in the trial. To say, 'Oh, renal function looks like it makes a difference.' Then you can do an additional study that really nails that down. It's a useful exploratory tool because sometimes you find factors you didn't expect."
Another Pharsight customer, Switzerland's Hoffmann-La Roche, has also used the company's software in tandem with inhouse tools. Roche used in silico modeling to show that Boniva, an osteoporosis drug, could be tolerated on a monthly dosage. "The study provided reassuring evidence of the tolerability of higher doses of ibandronate [Boniva], allowing monthly administration," says Ronald Gieschke, a senior researcher in the modeling group within clinical pharmacology at Roche in Basel.
Gieschke plays down the notion that modeling and simulation are somehow cutting-edge. All of the largest pharmaceutical companies have their own modeling and simulation departments, he says. At Roche, most models typically take a day or two to run on Pentium desktop computers in batches of 50 trials, each with 150 patients. The company explores giving a drug once a day, twice a day, four times a day. Says Gieschke: "In each patient, we would have measurements on many variables like blood pressure and heart rate." Simulating 500 trials for a single drug is no problem, as long as the electric bill has been paid.
"From our perspective at Roche, we gave [simulation] a high value four years ago," Gieschke says. "It is not controversial." The main challenge, both within the company and in presenting such data outside the company, is rhetorical. Says Gieschke: "That is a challenge for the future—to make people more comfortable with what the modelers propose to them. We need to build the comfort level in the people who are still skeptics."
The question is not whether virtual clinical trials could replace actual clinical trials—that is premature, even absurd. But virtual and real clinical trials can be combined to minimize the number or duration of trials in living patients. Each batch of data improves the predictive power of the model (or trial) that produced it. "Instead of doing many trials, we use the knowledge from a very selective number of clinical trials to integrate all our knowledge in a model—and do one pivotal clinical trial," he says.
Pharsight's president and CEO, Shawn O'Connor, says that after some belt-tightening, customers are starting to buy again. The company sells consulting services as well as software, with employees evenly divided between supporting each revenue stream. Simulation's day is coming, he says. "The ebb and flow and appetite for our service in pharma is on the increase," O'Connor says. "The issue is changing people's mindsets. The industry is characterized by slow movement."
For key customers such as Aventis, he says, Pharsight can remove a lot of the million-dollar disappointments that arrive late in the development cycle. "The opportunity for fallout in Phase III should be minimal," he notes.
PHOTO OF GREEN FACE BY JOHN WEBER
|Optimata wanted to simulate an animal's response to an immune system protein. The graph shows a rhesus monkey's production of platelets and the computer's prediction.
Source: British Journal of Haematology