Professor of Biostatistics
Department of Biostatistics at the Harvard School of Public Health
Name: Victor DeGruttola
Position: Professor of Biostatistics, Department of Biostatistics at the Harvard School of Public Health, 1986 the present
Education: DSc, 1986 Harvard School of Public Health
Victor DeGruttola's research is concerned mostly with statistical methods required for public health responses to the AIDS epidemic. This focus has taken him into the realm of biomarkers, especially how advanced biomarkers which he calls higher-dimensionality markers can be used to provide more than one treatment guideline.
Ultimately, many biomarkers should eventually make their ways through the regulatory gauntlet to become surrogate endpoints one example is the use of viral load to determine the efficacy of an HIV treatment. Drug makers should be able to use other information yielded by a genetics- or genomics-based test for information about not only whether treatment is successful, but what strategy should be taken next.
Pharmacogenomics Reporter caught up with DeGruttola a week after his talk on "Considerations in Evaluation of Surrogate Endpoints in Clinical Trials" at the April 28th FDA Science Forum in Washington, D.C.
Is there an accepted way to validate a surrogate endpoint?
The issue of validating surrogate endpoints and showing that a treatment effect on a surrogate endpoint reliably predicts a clinical benefit is a very controversial subject, and most statisticians believe and agree that it's very challenging to be able to demonstrate convincingly that effects on a surrogate will predict effects on a clinical endpoint.
There has been a fair amount of work that's been done on methods for doing that kind of demonstration. And in fact, there's a Proceedings paper from an NIH workshop on surrogate endpoints that was published a year or two ago, that kind of laid out those issues.
Most of the work that's been done has looked at low-dimensional surrogates, like blood pressure, terminal response, viral load, and so on. What I'm actually interested in is looking at high-dimensional surrogates, and that's kind of an area that I'm just getting started working in instead of a single measurement, you might have a genetic sequence. Particularly in the area of virology, where I do most of my work as a statistician.
People look at the amount of virus that returns after an anti-viral treatment fails, and then look at the resistance mutations that are accumulated by the virus as a result of treatment pressure. All of the drug development in HIV uses surrogates viral load. But viral load appears not to be a perfect surrogate, because one thing it doesn't tell you about is, after you have virologic rebounds, what treatments remain for a patient. And clearly, two patients could have the same amount of virus, but have very different treatment options depending on the resistance mutations that the virus has accumulated.
That's one of the areas that I'm working on how best to characterize the options that remain, and how to use not only information about the viral load, but also on the viral genome as a kind of surrogate for a longer-term clinical effect.
As both a prognostic and diagnostic marker?
That's right. It would be both. And increasingly the FDA is requiring information about the impact of antiviral drug on the development of resistance, which essentially means using a higher-dimensional surrogate, because when you say that one of the impacts of the treatment will be affecting the drug options that remain to a patient, you're basically saying that those options are kind of a surrogate for the longer-term clinical effect. Because the clinical effect isn't just suppressing the virus, it's also changing the virus.
How close is the FDA to accepting higher-dimensional surrogates as clinical endpoints?
I don't think there's a single case of that, I think it's something that needs to be investigated. How best to do that, and how to make use of those analyses to understand the role of genomics in investigating the impact of a drug.
But what the FDA does require is information about the resistance mutations that are induced by an antiviral drug, and also information about what kind of strains the drug will have activity against.
But in those cases, it hasn't been posed as a problem of surrogacy, but, 'What is the relevant patient population? And second of all, what are the adverse consequences of the drug (the development of resistance being an adverse consequence, because it affects other drugs)?' But you could, if you wanted to think about it that way, characterize the development of resistance mutations in the viral genome as kind of a surrogate for what's going to happen to the patient later on.
But there aren't any studies that I know of that have actually done that, so it's not an area that has so far been developed, but one I think we need to start thinking about.
What do you see as the barriers to developing surrogate endpoints like these?
I think the biggest barriers are collecting the amount of information that is necessary.
A point that Tom Fleming [chairman of the department of biostatistics at the University of Washington] has made a number of times is that you need a lot more information to validate a surrogate to show that drug is beneficial for one clinical endpoint compared to another drug. If you want to show that not only is that true Drug A is better than Drug B for some clinical endpoint but that to show that some marker is also a surrogate, you need more information because the kinds of estimations that are needed to [prove] that something is a surrogate have inherently more variability.
Also you want to look across a number of studies, because what you want to see is a consistent pattern that the drug effects on the marker will predict the drug effects on the clinical endpoint not only in one study, but across a range of studies, look for that kind of consistency. So, the first challenge is just getting that kind of data for any marker.
But if you're talking about a high-dimensional marker, not only are there challenges regarding getting the information, but I think we also need to think about, 'What are the best approaches to the statistical analysis?' And there are a number of issues. One is, 'How is that high-dimensional information, like a genetic sequence, going to be used?' One way might to be to find some sort of metric. Could you take a genetic sequence, for example, and turn that into a statement about the drug options that remain available to a patient. So the goal is to reduce the dimensionality down into a single metric. Or is what you're trying to do look for certain exceptions. An example might be if you said, 'In general viral load is a good surrogate for a longer-term clinical benefit, except, if these mutations have developed, the surrogate is no longer reliable. That's a different use of high-dimensional data, to sort of sift through all the information that's been collected on the mutations that were induced as a result of the drug pressure, and figure out, 'Are there subgroups or special cases in which your statement about surrogacy for something like viral load doesn't apply, And so I think that as well as the challenges in getting the information, it is a challenge to develop appropriate statistical methodology. But I think it's worth doing, because a lot of what's collected is higher-dimensional. And should there be immune-based therapies that are also successful, there will be the same issues regarding characterizing immune responses, which are also complex.