How can sites and sponsors come up with trial designs that are efficient, effective and flexible enough to adapt to challenges that might pop up along the way? One clinical trial technology expert suggests that effecting such designs in the modern age could call for a less-modern approach.
Outsourcing-Pharma (OSP) recently spoke with Pantelis Vlachos (PV), director and strategic consultant with Cytel, about Bayesian trial design, how the method differs from others and how the approach could benefit current and future trials.
OSP: Could you please tell us a little about Cytel?
PV: Cytel provides advanced analytics and biometrics to support confident decision-making in the life sciences. We use a variety of expert methods to unlock the power of data, ranging from statistical methods for small or skewed data, to real-world analysis and innovative clinical trial design.
Cytel is rare in that we are a scientific community of statisticians and epidemiologists who live and breathe data, so to speak. Importantly, we have experts across many different methodologies all in one place, as well as a commitment to making these complex methods more widely accessible through our software.
OSP: Could you please explain what we mean by Bayesian trial design?
PV: When most people think of clinical trials, they are thinking about the use of “frequentist” statistics without even realizing it. Frequentist methods were developed in the middle of the 20th century and took off in the 1980s and 1990s in no small part because computers made doing frequentist calculations much easier.
Bayesian statistics are a much older form of statistics developed by Thomas Bayes in the 1700s. They enable much more rapid learning, more flexible trial designs, and oftentimes can accommodate smaller sample sizes.
But, until recently, the technology needed to design and implement these trials simply was not available to the average statistician. The computations required for Bayesian designs can still take powerful computers several hours to complete.
Frequentist and Bayesian methods understand the scientific question posed to statisticians a little bit differently. Frequentists begin with a hypothesis like, this therapy will have no effect on alleviating that disease. Then data is amassed, and Frequentists ask, what is the likelihood that we would see this data if the hypothesis were true? Based on this likelihood, the null hypothesis is either accepted or rejected.
Bayesian methods reverse this question. Bayesian methods begin with the question of how every piece of new data should help us update what we already know. Bayesians want to know what the data implies the hypothesis should be, rather than what the data reveals about existing hypotheses.
Bayesian methods are good for trials that require learning quickly, like rare disease trials, because each datum can help us update our beliefs. They are also great for therapeutic areas where there is already a lot of historical data on one or more of the trial arms being tested. They have always been important for dose-escalation trial designs, as these trials work with small sample sizes. Over the past ten years, Bayesian designs have really taken off.
OSP: What benefits do Bayesian methods offer over more traditional trial designs?
PV: Bayesian trials offer flexibility, learning by doing, and often smaller sample sizes. Whereas traditional trials might require a sponsor to enroll hundreds of patients before we can look at the data, Bayesian methods allow us to look as we go along. This means that when things are not going well, we can intervene early, and if things are going extraordinarily well, we can capture that information and accelerate the trial.
Recent Bayesian trials have demonstrated the ability to reach statistically sound conclusions using hundreds fewer patients, or with timelines accelerated by up to ten months, or with some combination of these objectives. Given the rapid pace of learning, Bayesian designs were also instrumental in early COVID-19 exploration, when very little was known about the disease, and trials needed to learn rapidly and stop quickly if a therapy was not working.
Another important feature of Bayesian methods is the ability to build what are called meta-analytic priors. These priors can aggregate historical data across several studies, and then use this data to substantiate aspects of a new study.
When several therapies are being tested for a single disease, but say for slightly different populations or with different endpoints, then Bayesian statistics allow us to combine these studies using what are called hierarchical methods.
Finally, it is important to note that Bayesian methods are not just important for Bayesian studies, but also for certain features of Frequentist studies, like interim looks in complex innovative designs, the calculation of benefit-risk in early phase trials, and so forth.
OSP: Similarly, what challenges do trials opting for Bayesian methods need to anticipate, and how might they meet these challenges?
PV: Bayesian methods require more complex calculations. One of our goals in providing software for Bayesian methods is to ensure that more people can use them. This has always been a part of Cytel’s ethos.
Complex statistical designs will only benefit patients and clinicians if they can be widely used. Bayesian methods in dose-escalation and dose-finding have been a part of Cytel’s East trial design software package for over a decade. They can also be found in U-Design, the software produced by Laiya, Cytel’s newly acquired subsidiary.
Still, much Bayesian calculation depends on strong, powerful computers that the average statistician would not have been able to access. Now with East Alloy and Solara, both of which utilize cloud-computing, these complex methods are going to be even easier for statisticians to use.
Further, regulatory bodies are opening up to the widespread use of Bayesian methods. But there is also a different set of regulations for Bayesian methods, so working with knowledgeable consultants for regulatory support is essential.
OSP: Please tell us about East Alloy, U-Design and Solara—how the platforms work, how they benefit clinical trial teams, etc.
PV: East Alloy aims to make computationally intensive Bayesian calculations more easily available. Similarly, U-Design enables easier adoption of a range of familiar and advanced dose-finding designs—many of which include Bayesian methods—and is offered via a software as a service (SaaS) model.
Right now, even on-premise statisticians don’t necessarily have software available to test multiple possible designs quickly, and some might feel a bit unsteady using Bayesian methods in the absence of verified software. East Alloy is computationally intensive, verified, and graphically intuitive for those who know the methods but want more reliable technology.
Solara, on the other hand, is a cloud-powered decision-support platform for clinical development teams to collaboratively improve R&D efficiency and give new therapies a better chance of getting to market. Frequently, development teams lack sufficient time and resources to fully explore, compare and select the most promising design options to meet both clinical and business goals. So, Solara was created to remove that obstacle—by enabling statisticians to explore exponentially more trial designs, like complex Bayesian designs, in any given period of time.