Endpoints Explained—Part I

[Hopeful]

2012 Mar 14, Daniel J. Sargent, PhD

Daniel J. Sargent, PhD, is the Ralph S. and Beverly E. Caulkins Professor of Cancer Research at the Mayo Clinic Cancer Center in Rochester, Minnesota, and Group Statistician for the Alliance for Clinical Trials in Oncology. In Part I of his commentary on clinical endpoints, Dr. Sargent discusses validation of surrogate endpoints and biomarkers in oncology clinical trials. In Part II, Dr. Sargent will cover specific study design and endpoints used in phase II and phase III studies.

Surrogate marker validation

Surrogate markers have the potential to be very helpful in oncology drug development by allowing us to more quickly identify active agents. The challenge is that validating a surrogate marker is difficult. The criteria for validation require that a treatment effect seen on the surrogate predicts the treatment effect on the true endpoint, and the only way to determine that is by examining a series of previously conducted trials. There is no way to validate a new surrogate endpoint in a new trial because you don’t have the true endpoint until you have observed the true endpoint, so to speak. The validation process requires the analysis of a large number of previously conducted trials through a meta-analysis. That requires cooperation among many research groups who are willing to provide the data. There are many examples of endpoints that are thought to be useful, but, in the end, don’t show a strong association with a true endpoint. Because of these false-positives, we really do have to be cautious in our use of surrogate endpoints.

For a full drug approval, the FDA requires an advantage on an endpoint that measures either how well a patient feels or how long he lives. Therefore, the surrogate endpoint has to be validated as an accurate predictor of one of those two things.

The criteria for validating surrogate endpoints are conceptually straightforward: the treatment effect that’s observed on the surrogate must predict the treatment effect observed on the true endpoint. For example, if you see a hazard ratio for progression-free survival (PFS) in favor of one regimen, that must reliably predict an overall survival (OS) advantage for that same regimen. It doesn’t have to be a perfect prediction, but it has to be a statistically robust prediction that, with a specified degree of certainty, we can trust that the effects we see on the surrogate will result in a benefit on the true endpoint.

Progression-free survival

PFS is an endpoint of substantial current interest. There are many issues to discuss about PFS. One of the key questions is: does lengthening a person’s PFS (ie, delaying the time until his or her tumor grows) represent a direct benefit to the patient?

Because most progressions are radiographic only and not symptomatic, many people feel that delaying a radiographic event that has no immediate implications for the patient in terms of how she feels is not a clinically relevant endpoint. Others feel that when a patient progresses and her tumor grows, they need to change the therapy, that it is certainly a sign that the disease is worsening, and delaying that time could be considered to be of clinical benefit to the patient. If this issue could be resolved, then whether PFS was a surrogate or not wouldn’t actually matter because an extension of PFS time could be considered a clinical benefit to a patient. But that is not an established viewpoint in the oncology community.

Much of the consideration regarding clinical benefit depends on the disease setting and the line of therapy. The standards for a first-line trial may be quite different than those for, say, a third- or fourth-line trial. Also, the standards may well differ based on the disease site since progressions are more symptomatic in some diseases than in others.

When we talk about the advantages and disadvantages of using PSF vs OS, the advantage of OS is that it is clear, unambiguous, there is no measurement error, and it is the ultimate endpoint for a patient to extend her survival. The challenge with OS is that, fortunately, in many diseases, we have made substantial improvements, and patients live for an extended period of time. Breast cancer and colon cancer are two prominent examples in which, even in the metastatic setting, patients can live for several years. So trials that require OS as the endpoint take a long time. But, even more difficult than that is that, over the course of the patient’s disease, the patient will likely get many treatments. She will get an initial treatment and she will get a second-line, a third-line, and, in many cases, fourth- and fifth-line therapies, in ovarian cancer, for example.

Therefore, the ability for any one new agent to improve OS is difficult because subsequent therapy is out of the control of the protocol. If you can only impact one element of a patient’s treatment, then the other treatments that he’ll subsequently receive add variability to that patient’s outcome. For this reason, many advocate PFS as a measure of drug activity. That is, if an agent can prolong PFS by a clinically meaningful amount (for example, a median of 3 or 4 months), it is clear that the agent has activity. Some would argue that that should be sufficient to allow for regulatory approval. However, the FDA has been skeptical about that and really has, in many cases, looked for what they would consider a truly meaningful patient benefit on an endpoint that is a more direct measure of how a patient feels or functions, or on the OS endpoint.

I personally think, in historical cases when there was an absence of second- and third-line therapies, that in many cases PFS advantages did predict OS advantages. In current oncology practice, with many new agents coming on the market, the ability to formally demonstrate surrogacy in many of our major diseases—breast cancer, colon cancer, ovarian cancer, and now prostate cancer—may be gone. That is, we know that patients will receive many lines of therapy after the treatment received on the trial. In such a case, to require an OS benefit, or to require such a large PFS benefit that it will translate to an OS benefit, is a very high bar, which will likely only be achievable through highly selected trials in which we have a treatment specifically designed to hit a target in a subset population. This is certainly the way we want to go. In the standard paradigm of evaluating cytotoxic agents in an unselected manner, the ability to extend OS from one of many lines of therapy will become increasingly difficult.

Personalized medicine and biomarkers

Personalized therapy requires specific new strategies for clinical trial design. This is an area of very active research—searching for the best ways to assess activity of both a therapy and a biomarker, because they almost always go together.

The two concepts are fundamentally linked, because in most cases we have a biomarker to select a group of patients, and then we have a therapy designed to benefit that selected group of patients. The simplest approach, which has had success, is to do it right up front. Rigorously define the biomarker, use that biomarker to select only the patients who we think will benefit, and then do the trial only in those patients. This strategy goes by several names, an enrichment design, or a targeted design, but the principle is a trial in which only patients who express a certain level of biomarker are enrolled.

These are efficient trials. They reduce the sample size substantially, and they are relatively easy to explain to patients because we think we have a therapy that will help if they have the biomarker. They’re difficult because many biomarkers of interest have low prevalence. That is, they are only present in a small number of patients, say, 5% or 10%. So, these trials require a large number of patients to be screened in order to find the patients to go on the trial. The other down side of these trials is that we only gain evidence about the efficacy of the therapy in the specific population studied. This leaves questions about other patients who might benefit from the therapy.

There are several examples, the most prominent of which is HER2 status in patients with adjuvant breast cancer, where there are remaining questions about whether trastuzumab has efficacy in patients other than those studied in the trial. There are also some questions about crizotinib in non–small cell lung cancer. Early evidence is that it may have efficacy in patients other than those with ALK mutations, which leaves questions that then require further study.

A different class of designs for biomarker-driven therapy enroll a larger set of patients, and then assess whether a biomarker can identify subgroups of the trial population with more or less benefit. In such a strategy, one recommendation is to use interim monitoring or other adaptive strategies to possibly narrow down the population as the trial is ongoing. The choice of an optimal design is very circumstance-specific. How strong are the preliminary data about the efficacy being restricted to a certain population? What is the prevalence of the biomarker of interest? If it’s a very–low prevalence biomarker, 5% or 10%, then it is likely that the targeted or enriched design is the only possible way to evaluate the biomarker, because to enroll all patients on a trial when you expect the maximal benefit to occur in only a small number of patients is just not efficient. All of these trials are heavily reliant on having a reliable, reproducible, pre-specified biomarker assay available.

There are other designs in which patients would be randomized without selection of a biomarker, and, then, a post-hoc analysis would be conducted to try to identify a subgroup of patients who would have the maximal benefit. With these designs, we have to be very careful because the lack of pre-specified biomarkers greatly enhances the possibility of a false-positive finding due to exploration of many different subgroups. Clearly, the prospective specification of the biomarker of interest greatly enhances the strength of the conclusions we can draw from any of these clinical trials. That’s not to say that there are not appropriate uses of a carefully planned retrospective analysis. A prime example is the differential activity of the EGFR inhibitors cetuximab and panitumumab in patients based on KRAS mutation status. Based on phase II studies that indicated that the activity of these EGFR inhibitors was restricted to patients with wild-type KRAS tumors, a planned analysis of previously conducted trials in unselected patients was conducted; and multiple trials showed the same conclusion—that the benefit was restricted to patients with KRAS wild-type tumors. The need for this strategy will not stop, in general, because as our biologic knowledge increases, we will increasingly go back to old trials and be able to identify biomarkers predictive of which patients actually were the ones who were likely to benefit from the treatment. The key is to replicate such a finding across multiple previously conducted trials, and, then, if we find a high level of consistency, we would feel comfortable in taking that biomarker forward.

I think the most difficult markers to validate are not single biomarkers but panels in which a SNP array or microarray of tumor RNA or DNA are assessed, and the goal is to come up with a multifactor signature to predict responsiveness. In such cases, in which the biology is not as well known, clearly prospectively specified analyses, repeated in multiple studies, are critical. The level of rigor for validation needs to be high in these multi-marker panel studies, because when you’re considering multiple biomarkers together, the risk of a false-positive finding certainly is increased. Optimally, we need to have a strategy that allows prospective validation, such as the TAILORx trial that is ongoing in breast cancer, to validate multi-gene signature for prognosis and prediction.

Future clinical trials

I believe that increasing knowledge of biology, and associated biomarkers, will inevitably lead us to conduct targeted trials which will look for large treatment effects in patients who are selected to have a high degree, and a high likelihood, of benefit. While I think that this is probably inevitable over time, it also raises some challenges. As disease populations become more and more segmented into very small groups, the trials become harder to perform because the availability of patients with a highly specific biomarker status is limited. This will require trials, in many cases, to become international collaborations, which increases their complexity. The strategy also raises the challenges with regard to the ability to develop the next biomarker. For example, if we take a patient population and we find 5% who will benefit from a certain treatment, then we have to do further development in only that 5% of patients, and maybe we find another marker that segregates that population into two populations, each of which is even smaller and more difficult to test.

This biomarker-based paradigm will require us to consider the level of evidence necessary to declare activity for one of these agents. From an overall perspective, oncology research is moving towards a rare disease paradigm. In many diseases, it may just become impossible to perform traditional trials with requirements to achieve a P-value less than 0.05, high statistical power, and an OS advantage. When the patient population becomes small, we’re going to have to consider either other endpoints or other statistical philosophies, so to speak. Should we use a Bayesian strategy, in which we borrow information from other clinical trials to help make decisions? Or do we loosen the P-value requirements—that a P-value of less 0.1 or 0.2, for example, be considered a sufficient level of evidence for activity in a very rare population? These are active areas of research that need to be fully considered as we enter this era of truly personalized therapy with patient populations that are becoming smaller and smaller.

Hopeful