Personalized medicine proponents have long argued that the use of pharmacogenomics can reduce the number of adverse outcomes associated with prescription drugs, but it's been hard to gauge the actual cost and time that would be required to conduct the studies necessary to drive broad adoption of such methods.
A team of pathologists at Beth Israel Deaconess Medical Center and Harvard Medical School recently took a stab at such an estimate by building a probabilistic model to predict the time and investment required to cut the rate of drug-related adverse outcomes in half using pharmacogenomics. The Monte Carlo model, described in a recent paper in Clinical Chemistry, takes as input a number of parameters, including data on the associations between genetic variants and adverse outcomes, the time required to discover and confirm these genetic associations, the cost of this work, the number of people taking the drugs, and other factors, in order to calculate the overall costs and time required to achieve a specified reduction in adverse events.
As described in the paper, the researchers, led by Ramy Arnaout, co-founder of Beth Israel's Genomic Medicine Initiative, ran the model for eight genetic associations involving six prescription drugs — clopidogrel, warfarin, escitalopram, carbamazepine, the nicotine-replacement patch, and abacavir. Using this approach, they determined that it would require a research investment of $1.5 billion to $6 billion and take approximately 20 years to develop guidelines that would cut drug-related adverse outcomes in half using current genomic methods.
This estimate includes a "pump-priming" phase of five to seven years that would require an investment of "hundreds of millions of dollars, with little apparent return on investment," the authors note.
Among a number of findings from the study, Arnaout and his colleagues concluded that the single most important determinant of the total investment required in PGx is the extent to which genomic variants are responsible for adverse outcomes. For drugs in which most of the risk of adverse events can be attributed to genetic factors, less investment will be required, and vice versa.
The reason for this, they note, is that "the higher the percentage attributable risk, the larger the role genomics plays, and therefore the more high-impact associations there are likely to be, given the frequencies of the associations already found. Conversely, the smaller the role of genomics, the more rare associations will have to be found and confirmed to cut adverse outcomes in half, at a higher overall cost."
Pharmacogenomics Reporter recently caught up with Arnaout, assistant professor of pathology at Beth Israel, to discuss the study and his team's findings. The following is a transcript of the interview, which has been edited for length.
The paper looks at how the development of genetic testing guidelines would help eventually reduce genetically associated adverse events. Can you go into a little detail about what you mean by genetic guidelines? Would these be FDA labeling changes or society- or community-recommended guidelines?
If guideline development for other drugs goes the same way that it’s started to go with warfarin and clopidogrel, then it's going to be a little bit of both. There are various levels of evidence that practitioners will decide that they want to use in order to do genetics-based, guidelines-based prescribing. This will probably happen on a hospital-by-hospital basis.
At a number of hospitals, you will come in and they will have decided that it is better for you and for them — both healthwise and financially — to test you to see, for instance, what CYP gene polymorphism you have before deciding to put you on a high dose or a low dose of warfarin. Or similarly for clopidogrel.
Those guidelines can be society based, but they are ultimately put into practice by the hospital, with input from whatever the insurance organizations are that the hospital works with. So there is a continuum of evidence that clinicians and hospitals will use in making their ultimate decision on what the evidence is, without always waiting for a society or the FDA.
So clinicians tend to provide these tests when there's enough evidence across all these different parameters that weight the decision in favor of testing instead of not testing.
Right. You can see that today. Vanderbilt University has been on the forefront of this kind of testing and they test for warfarin and clopidogrel. As far as I know there is no national sort of guideline for doing that, although the FDA mentions it, but they say that the balance of the evidence favors testing people for VKORC1, CYP2C9, CYP2C19, so they do it.
Some hospitals will say, 'Yeah, we totally agree with what Vanderbilt is doing and we ought to be doing it too,' and others will say, 'Well, we're going to wait until we see a little bit more evidence or hear more from the FDA or this or that.'
So there has to be evidence, but you also have to be convinced that the evidence is sufficient. There's no one magic bullet that is going to guide everyone to test or not test at this point.
Who do you think would get the most use out of a paper like this? Would it be insurers? Regulators?
Hospitals and patients, definitely. I would hope it would include the pharma industry and funders of pharmacogenomic-based research, too. Because when you're trying to do a high-level study like this — where it's not giving a thumbs up or a thumbs down on the wisdom of investigating any particular genetic cause for any particular adverse outcome, but just looking at the field as a whole — you'd hope that people who concern themselves with the big picture would be the ones who would pick up on this and say, 'Well, let's think about what the implications are here.'
The conversations I have been having, especially with colleagues in clinical pathology, which is my specialty, but also some people in the VC community, [have been that] there ought to be a lot of interest in the pharmaceutical industry for taking drugs that they know are safe, but have only been effective in a small number of people, and resurrecting those drugs based on pharmacogenomics studies. So the logic would go, 'We have this drug that wasn't released widely because we wanted to use it for, say, cholesterol, and it definitely reduces cholesterol, but only in about 5 percent of the people who it was given to.' So you can treat that as a 95 adverse outcome of no effect, and if we can figure out what that 5 percent is, and if it's pretty cheap to do, then that might be worth it in order to bring this drug to market.
The model is focused on the research investment necessary to cut the number of adverse events in half, but the paper didn't address the cost savings associated with that reduction in adverse events. Is it safe to assume that the cost would be half of what the cost is now, or would it be more complicated than that?
For a good approximation it would probably be around half, but we realized we didn't have to model out the savings in order to demonstrate impact. In 1995, researchers estimated the cost of adverse events — just on inpatients — and found that to be $80 billion a year. The main reason why we focused where we focused was to see what we could say on the cost side of the equation. But we realized, if adverse events cost $80 billion a year, and the numbers that we found for developing guidelines are on the order of hundreds of millions of dollars a year, then you've got a cost that’s such a small fraction of the potential savings that if you get even a small percent improvement on just the adverse events, developing guidelines pays for itself pretty quickly.
Were you surprised at all by the numbers you eventually came up with? The range of $1.5 billion to $6 billion?
You know, we were. It kind of goes back to where this project came from. My lab does genomic sequencing and clinical informatics and I was sitting next to the chief academic officer of the institution at a celebration we threw for the tenth anniversary of the draft human genome, and we were hearing about all the amazing things genomics was going to be doing for everybody soon. [The chief academic officer] leans over to me and he said, ‘Your lab does genomics, right?' And I said yes. And he said, 'So is this actually going to happen?' And I started to say something to him and then I stopped because I didn’t know what the answer to that big-picture question was. I thought, it kind of depends what the 'it' is, and what your timeline is, and how much it's going to cost. And I thought, if didn’t know, then probably not too many other people did, either, and that this was something to really think about.
As I thought about it, I realized that I could convince myself that it would cost anywhere from hundreds of millions to half a trillion, and I really had no instinct with which to ground that. So after a while I said, 'Well, I don't know the answer but I know how we can find out.' And of course I was incredibly fortunate to have this great team on my side — not just [coauthor Vikas] Sukhatme but also [co-authors Thomas] Buck and [Paulvalery] Roulette. So that's really how this paper was born, and why we were surprised. It came in on the low end.
Is this something you see being modified as more data becomes available to get a more accurate picture of the broader market? Or could someone take it and apply their own parameters to use it, say, on a case-by-case basis for a new drug and a new test?
The former. The model is robust enough that as you find other drug/adverse outcome genetic associations, you can just plug that data into the model and see how things change, come up with ever narrower confidence intervals.
At the same time, it is vital to investigate more deeply the assumptions in the model that we found were the most important determinants of the total cost. So, for example, we're never going to be able to cut adverse outcomes in half using genetics if it turns out that overall genetics can't explain half the adverse outcomes. We know there are all kinds of things that explain adverse outcomes, especially things like societal factors. For instance, if a drug wasn't labeled well so the patient took the wrong pill. Or if the patient's insurance only covers some part of the prescription so they didn't get it refilled. Neither of those are genetic problems, and both can lead to adverse outcomes. And things like that happen, a lot. They may not be cool in the same way that genomics is but they're very important. A lot of people are working on solving them in various ways.
Nevertheless, it appears that genetics does contribute — at least in the cases we have — to somewhere over half of the adverse outcomes. That said, you wonder if there's a bias there in the sense that, of course, the ones that work so far — the ones that are the examples we've used in our paper — are examples for which that must be the case. For warfarin, you have a set of seven associations that combined explain close to 80 percent of the variability. That's huge. But what if we just found the low-hanging fruit? We don't think that bleeding on warfarin is going to be all that atypical of other results, but that really is a fundamental assumption. So understanding how your genotype impacts that phenotype of adverse outcomes is hugely important. So in my mind, our study really highlights that as something that deserves more attention.
In the paper, you recommend a '50,000 Pharmacogenomes Project.' Are you in discussions with anyone about making that a reality? How would you envision something like that taking shape?
There are people out there who have very real-world, very practical experience with large-scale projects like what we proposed. I wouldn't pretend for a second that I'm the expert on how to put such a project together, only that our paper highlights pharmacogenomics as a place where such a project would be useful. You have all of these amazing projects out there — the Personal Genomes Project, the Million Veterans program, UK 10K — and I think that piggybacking on those projects is the way to go, and those are conversations that I do hope to have. The basic idea wouldn't be, 'Let's start from zero and start finding new people,' but saying, 'Let's start with the people who have enrolled in these [projects already] and simply make a point of [tracking] their prescriptions and adverse outcomes in some computer- and researcher-friendly — and of course patient-friendly — way.'
The model in the paper focuses on germline variants and prescription drugs. Do you have plans to create a similar model for studies that are using somatic variants to guide cancer therapy?
We've been thinking about it and it's more daunting. As we were going through this, a pair of axes emerged for us in how to think about pharmacogenomics. One axis is the number of people who are affected. So many more people are on prescription drugs than, for instance, suffer from rare inborn errors of metabolism. The number of people with cancer falls somewhere in between.
But there was another axis that turned out to be very important in guiding our thinking going forward, which is evolution. For all practical purposes, your germline genome doesn’t change. But cancers and infectious agents have genomes that change and adapt all the time. And what that means is that you're in an arms race. You could come up with a drug that will work against the particular genetics of your cancer, and it might work amazingly well for three months, only to [stop working] and the cancer recurs. If you're lucky enough at this point to have the resources to look at the genetics of it, you'll find out that the reason is because there's a mutation and be on a combination of drugs to make it harder for those mutations to escape. The drug that you provided put selective pressure on the cancer and the cancer has evolved with an escape mutation. It wasn't because you picked poorly, it’s just a feature of evolution.
So you can't model that evolution.
You can, but there's this entire other type of data that you need, which to my knowledge we're only now getting experience with in cancer. We probably have more experience with it in infectious disease. With infectious disease, we've seen organisms become resistant to a greater or lesser extent. We saw it with vancomycin; we see it with methicillin-resistant Staphylococcus aureus; we see it with tuberculosis drugs, malaria, HIV. So we've seen it and it's within the memory of publications to say, 'Yes, things were 90 percent sensitive in this time and place, and now they are 12 percent sensitive,' but that's an entirely different thing that one would have to put into the model. With cancer, I think they're only just now starting to internalize that.
So this idea that evolution is going to be a major opponent for us as we try to use genomics for better health is something that emerged loud and clear from our study.
What are your next steps? It seems like you could go in a lot of different directions.
The study only just came out, so we’re really at the phase of getting reaction and response from our peers and from the wider community to get everyone's opinion and see what they think. That will really inform how things go moving forward. There's this idea of the 50,000 Pharmacogenomes Project, and that number is an estimate. The statisticians will have to weigh in on what that number should be.
So we'll see how things go. I hope to be getting more feedback soon.
I guess it's good to get a number out there and see if it meets people's expectations or not.
That actually has been the most amazing single thing. You put this number out of $5 billion or $6 billion. And I talked to one person who said, '$5 billion or &6 billion? What a waste of money.' And literally that same day someone else said, '$5 billion or $6 billion? What a bargain. What an unbelievable future.' So it depends who you're talking to.
But once one has a number, one can discuss whether that number is worth what you expect to get out of it. Without one, it's pure emotion. And hopefully in at least some small way we've moved things forward with this little study from pure emotion to at least being able to say, 'Alright, $6 billion. Is that worth it to you? Is 20 years too long or okay?' As opposed to simply saying, 'Genomics good’ or ‘genomics bad.' Those things aren't constructive. What's more constructive is saying, 'What specifically are you looking to get out of this? Here's how much it's going to cost and how long it's going to take, and what to do to get there.'
That has been very refreshing to us, to have people now walking around with a number in their head where previously they didn't have one.