The Patient Equation

by Glen de Vries

  Dr. David Fajgenbaum, an assistant professor of medicine at the University of Pennsylvania and the co‐founder and executive director of the Castleman Disease Collaborative Network, is not just the most prominent researcher and advocate working on the condition. He's also a patient, who has survived five life‐threatening hospitalizations with the deadliest form of the illness, idiopathic multicentric Castleman disease (iMCD), which kills more than a third of sufferers within five years of initial diagnosis.

  One recent study found that incorporating additional patient data into a predictive model—powered by my colleagues at Medidata—was able to boost the percentage of cases in which the only traditional drug treatment for iMCD was effective. The study found six distinct subtypes of iMCD “that were completely novel to the medical community,” said Medidata's chief data officer David Lee, talking to Pharmaceutical Technology magazine.1 For one group, the traditional treatment had a 65% effectiveness rate, as opposed to 19% for the other groups—an obviously huge difference. But it means that instead of talking about a drug that works in one out of five patients—out of a population of tens of thousands—you have a drug that works for three out of five patients, but with a market size an order of magnitude smaller. It's a different ballgame if you're the life sciences company trying to develop a go‐to‐market strategy with a more targeted therapeutic.

  As Lee said, the finding shows that data management is critical. “[I]f we didn't take the time to integrate the ‐omic data with the clinical trial data properly then we would never have been able to find these signals no matter what types of algorithms we had used.”2

  Dr. David Fajgenbaum's Quest for a Cure

  That traditional treatment, a drug called siltuximab, the only FDA‐approved treatment for Castleman disease, didn't work for Dr. Fajgenbaum in his own battle. Instead, he is leading a trial for another drug, sirolimus, after experimenting on himself and finding that it has been able to keep his iMCD at bay for more than five years.3

  Fajgenbaum documented his journey in his memoir, Chasing My Cure: A Doctor's Race to Turn Hope into Action, published by Random House in 2019. He was in his third year of medical school in 2010, and suddenly he was tired, sweating at night, and losing weight. He eventually went to the emergency room at the University of Pennsylvania, after taking his OB‐GYN exam at school.4 Doctors told him his liver, kidneys, and bone marrow weren't working, and, after being hospitalized, a retinal hemorrhage caused him to temporarily go blind. He was there for seven weeks before recovering, and still didn't have a diagnosis. Over the next few years, he suffered four relapses, discovering along the way that it was Castleman disease, which is diagnosed in just 5,000 people in the United States each year (about the same number as ALS, also known as Lou Gehrig's disease). Instead of becoming an oncologist, Dr. Fajgenbaum went to business school after medical school and began diving headfirst into the study of Castleman disease.

  As he waded through the research—and his own personal data—he decided to take it upon himself to find a treatment that would keep him out of the hospital, or at least buy him a little time. As he told me in an interview, he examined his blood samples in the months leading up to his most recent relapse of the condition and found markers of T cell activation and blood vessel growth, measured proteins during his flare‐up, and ran the data through three different pathway analysis software systems to find that everything pointed toward the mTOR intracellular signaling pathway being involved in the disease.5

  He performed an experiment in the lab that confirmed what the databases were proposing: mTOR was activated. He went to a database of FDA‐approved drugs and found the mTOR inhibitor sirolimus, which was a drug approved 25 years ago for kidney transplants, and he tried it on himself. To his surprise, it has worked—in fact, he hasn't had a relapse since starting on the drug more than five years ago.

  The question Dr. Fajgenbaum is trying to answer now is whether his finding is transferable to other patients—and he is currently running a trial to figure out just that, with 24 patients whose disease has been refractory to all other therapies. Anecdotally, he has seen some good response from previous nonresponders (although some patients still haven't responded), but the right step is a legitimate trial, as he hopes to figure out if there is a particular subgroup of iMCD sufferers who will benefit from sirolimus just like he has.

  One of the biggest challenges in his research, Dr. Fajgenbaum told me, is a logistical one—the data he needs is simply all in different places. Patient medical records, proteomic data, clinical trial results, and more—answers may be out there, but you get so much more when you can see different layers of the layer cake. It's the fundamental patient equation story in a nutshell: for us to see the patterns that we may finally have the ability to see, we need all of the data to be accessible. It needs to be harmonized, sortable, accurate, and complete. Not just genes alone, not just proteomics alone, not just trials alone. How well do people do when they get one drug versus another? It's a question Dr. Fajgenbaum is trying to answer, but can't always answer it well.

  We've done well with some of these questions in the cancer space, Dr. Fajgenbaum explains, but with rare diseases like Castleman disease, it's harder. There are simply fewer patients, fewer sources. Ninety‐five percent of the 7,000 known rare diseases don't have a single FDA‐approved drug, but there are 1,500 drugs out there—like sirolimus—that might have positive effects without our even knowing it. And the only way we're going to find them is through smart analysis of trial data, real‐world data, as many layers of data as we can collect. “We need to think about creative ways to leverage data,” he says, “and try new approaches out on patients with no options, look for promising hits, and see if we get real results.”

  Dr. Fajgenbaum talks about his frustration with serum samples being stored away—samples that would have been destroyed if he hadn't rescued them to analyze for the study he did that ended up finding a subtype of patients who had a superior response to siltuximab. The reality is that for a fast‐moving disease like iMCD, any insight into whether someone will be a siltuximab responder can mean the difference between life and death. Patients like him don't have three weeks to wait to see if the drug works. If their profile says they might respond, they can give the drug three to five days before starting chemotherapy, just to see—but if they're likely not going to respond, now doctors can know to start chemotherapy right away, and avoid the increased risk that the patient tanks quickly.

  Armed with the siltuximab study results, Dr. Fajgenbaum now plans to return to the samples and see if he can find other possible drugs that might work. Perhaps one group will show a benefit from siltuximab, one from sirolimus, and other drugs they haven't yet tried might show positive effects with other nonresponders. But without Dr. Fajgenbaum and his motivated work, Castleman disease would likely not be on the radar screen for many, since the population is just so small.

  Rare Diseases, Common Problems

  There are three big problems most rare disease researchers share, Dr. Fajgenbaum tells me: funding is hard to find, sample sizes are small, and the data just isn't as useful as it should be. Researchers may be interested and capable, but samples need to be well‐curated—especially when there are so few, and data is hard to come by. If there's one action item Dr. Fajgenbaum wants life sciences industry readers to take from his work, it's to unlock their samples. If we can get samples out of freezers and into the hands of researchers who care about the diseases, we have the tools now to do so much more than we used to be able to. From large‐scale proteomics analysis to RNA sequencing technologies, there is so much computing power we can exploit if the samples are there.

  Further, Dr. Fajgenbaum believes clinical trials have to add some of these layers in order to maximize what we can do with the results. If proteomics and transcriptomics were part of the standard trial, and if more lab tests associated with proteomics and pathways were part of the standard trial, then we could find out so much more about these patients
and how we might be able to subtype them. Researchers need the most fine‐grained data they can have—in the spirit of eventually figuring out how little data we really need to make the best prediction. Hidden in the data may be simple answers—but we need rich data sets at the beginning in order to find them.

  The work on Castleman disease suggests that there may be other diseases that sort themselves out in a similar way, with patterns revealing responders and nonresponders, and different drugs becoming the best treatments for various subgroups. Rheumatoid arthritis, lymphoma, and HIV are just three that come to mind as sharing some features that indicate the analysis may be transferable. We can only hope to find more and more data that supports this suggestion.

  I've talked throughout this section about a range of conditions being looked at successfully through the patient equation lens, and how various doctors, researchers, and life sciences companies are approaching ovulation, asthma, diabetes, the flu, sepsis, cancer, bacterial infections, and rare conditions like Castleman disease. In the next section of the book, I want to shift focus just a bit, moving us from the work already being done to the important work we in the life sciences industry can and should be doing. I want to talk about how we can all start to build our own patient equations and apply the potential of these new data technologies to our businesses. We'll examine the importance of good data and data science, and about how we approach research and clinical trials. After that, we'll look at building disease management platforms that can actually turn data into actionable insights, to get treatments out into the world and to tell if they are working, to properly motivate and incent patients, and ultimately to put ourselves in the best position to achieve economic benefits from all of this critically important work.

  Notes

  1. Allie Nawrat, “Castleman Disease: Can Machine Learning Help Drug Development?,” Pharmaceutical Technology, February 26, 2019, https://www.pharmaceutical-technology.com/features/castleman-disease-machine-learning-cdcn-medidata/.

  2. Ibid.

  3. John Kopp, “Penn Doctor Makes Research Strides into His Own Rare Disease,” PhillyVoice, January 24, 2019, https://www.phillyvoice.com/penn-medicine-doctor-research-strides-own-rare-disease-castleman-disease-immune-system-disorder/.

  4. David Fajgenbaum, Chasing My Cure: A Doctor's Race to Turn Hope into Action. (Random House Publishing Group, 2019).

  5. David Fajgenbaum, interview for The Patient Equation, interview by Glen de Vries and Jeremy Blachman, June 26, 2019.

  SECTION 3

  Building Your Own Patient Equations

  9

  The Steam Table

  It may not be since high school physics that you've seen a steam table (see Figure 9.1), a chart of the pressures and temperatures at which water turns from liquid into gas.

  If you plot these numbers on a graph, you get a view of where you are in the world, based on these pressure and temperature variables, a point that tells you, with definitiveness, whether water will be water, ice, or vapor (see Figure 9.2). You know, because you have all the information you need for the equation.

  Perhaps you see where this is going. In a mathematically‐idealized world, we would build that “steam table” for every condition, every disease—so that you would know who to treat, who not to treat, and what we should be treating them with based on empirical results. Using the example of PSA, prostate cancer, and radical prostatectomy—the surgical removal of the prostate—the columns of our steam table could chart PSA results, the rows could chart the patient's age, and in each box we would look at the ratio of the number of people who didn't suffer from prostate cancer versus those who did, after their prostates were or were not removed.

  Figure 9.1 A steam table

  Steam tables show the amount of a substance—typically water, hence the name—that is vapor versus liquid at a given set of combinations of temperature and pressure.

  Figure 9.2 Phase diagram

  Phase diagrams graphically illustrate the transition of matter from one phase to another (phase in this instance meaning liquid, solid, or gas). Illustrated here for H2O.

  This should start to sound both impractical and wildly unethical—but the concept is important. If we looked at that data, we'd find that at certain combinations of age and PSA (more advanced age, relatively low PSA numbers), treating the patient doesn't drive better outcomes than leaving them alone. That would define a set of patients we shouldn't treat. There would be another group of patients who, if we choose not to treat them (likely those with excessive PSA scores at relatively young ages), the data would show they will suffer, and perhaps die, from prostate cancer. That would start to form a shape defining the patients we should treat.

  By adding a treatment alternative, say an experimental drug instead of surgery, we can add more data to the table, and start to look at the shapes forming in the diagram. Now we'll start to see shapes forming around which patients to treat with this new drug versus which patients to treat surgically (see Figure 9.3). However, the lines might not be that clear—not as crisp as those between water and steam. In fact, they probably won't be nearly as crisp in this example, because treating prostate cancer is vastly more complicated than boiling water. This is where we need to start to take these two simple ideas—the table of results (the steam table), and the graphical representation of that raw data (the phase diagram)—and open our minds and our analyses to more and more dimensions of data.

  We could add the grade of the cancer based on a pathologist's examination of a tumor biopsy. Now we have a three‐dimensional space, and yet another way to calculate the ratios for patients untreated or treated with different therapies. We could add a dimension for the RT‐PCR for PSA assay that I worked on at Columbia in the 1990s. We could add a dimension for activity, or for patient territory. We could add even more treatment options—different kinds of surgery and drugs, experimental and not experimental.

  It may sound overwhelming, but conceptually there is no change. When we move from two dimensions to three, the lines of transition become sheets. When we move to four and more dimensions, the shapes become harder to visualize—but they are still there. In the patient equation‐driven future, we need to catalog our data so that we can fill in as much of the steam table as possible, and impute the rest. And we need to look for the transitions—the phase shifts—that define how to optimally treat patients.

  Figure 9.3 A phase diagram for treatment choices

  Can the same principles visualized in a phase diagram be used to delineate the transitions between when or how a patient should be treated to receive what is computationally the best possible outcome based on existing data? This two‐dimensional example is based on two biomarkers, instead of temperature and pressure from the purely chemical phase diagram in Figure 9.2. An actual phase diagram for a real disease and possible therapies would be much more complicated in structure, extending into many dimensions and likely with multiple existing and experimental therapies. But any given combination of biomarkers points to a region—or the transition between regions—corresponding to the patient equation output for the best treatment choice.

  It's more than just pressure and temperature that we need to know—and so that graph, in reality, is far more multidimensional than the one on the page. But for all of this talk about sensors and data collection, this graph is the goal. This graph is where we are headed. This graph is what we need to build in order to make more granular and accurate decisions than ever before about who to treat and how. Our job in the life sciences industry is to make these graphs better and better, to figure out the layers that matter for every condition, the characteristics that need to be plotted, the codification of doctor's intuition, data‐supported and scalable.

  It's not about one good therapy—it's about putting a system in place to build all the good therapies, and then propagating them out into the real world, collecting more and more data to verify and sharpen our predictions, and ultimately generating consistent patient outcomes like we've
never been able to generate before. It's an easy science experiment to see when water freezes or turns into gas—but for people it's far harder, with an uncountable number of variables, some of which matter some of the time, and some of which we don't even know exist yet.

  Progressing Toward Alzheimer's Disease…or Maybe Not

  Standing in front of a lecture hall at Columbia University not long ago, I found myself trying to establish for the audience of undergraduates and graduate students why the idea of patient equations would have such a huge impact for the future of health care. A visceral example, originally given to me by Paul Herrling of Novartis, popped into my head. It has proven to be incredibly effective—particularly with a room full of young people. It's a “good news, bad news” kind of scenario, applicable to every student in the room.

  It's usually better to start with the bad news: the existence of beta‐amyloid plaques—clumps of protein that are thought to clog our cognitive circuitry—has been widely debated as to its causal or coincident relationship with Alzheimer's disease.1 However, even if their existence is merely something present in greater quantities as one progresses towards dementia, its utility as a biomarker for disease progression means that beta‐amyloid plaques should be worrisome to all of us. That's the bad news for the students. There is evidence to show that even in the students' relatively young brains, the march toward dementia has begun.2