There are approximately 7,000 distinct rare diseases that affect 350 million people worldwide. Scientific advances such as the CRISPR/Cas9 genome-engineering system have simplified the pharmaceutical and biotech industry’s ability to develop gene therapies especially for single gene mutation disorders. The Food and Drug Administration (FDA) now has more than 700 active Investigation New Drugs (INDs) for gene and cell therapies and in 2017, approved two cell-based gene therapies (chimeric antigen receptor T-cells [CAR-T]) and the first gene-therapy product to be administered in vivo. Collins and former Commissioner Gottlieb, of the National Institute of Health and Food and Drug Administration respectively, have stated that ‘it seems reasonable to envision a day when gene therapy will be a mainstay of treatment for many diseases²’. The time appears right for businesses exploring rare diseases and their partners should consider response-adaptive over traditional randomized controlled trial designs.

Moving from randomized control to response-adaptive trials

Randomized controlled trials (RCTs) are the traditional approach most often used in Phase II-III clinical trials. In RCTs, the probability of being assigned the experimental drug and the control (placebo or comparator) is fixed throughout the trial, normally at 50%, so that each drug is given to a similar number of patients. This leads to a high chance of identifying whether one treatment is better than the other and understanding any pharmacological effect.

This approach is perfectly acceptable for common diseases. If we take cardiovascular disease as an example - of the 7 million people in the UK with cardiovascular disease, if 200 of them are in a clinical trial then 99.997% of the population will benefit from the results of a clinical trial where half of the participants receive the trial drug.³

Whereas in a rare disease such as cystinosis⁴, which only affects an estimated 660 people in the UK, if there are 200 patients in a clinical trial, only 70% of the population will benefit from the results of an RCT. Where rare disease treatments are being trialled, greater emphasis on the health of the trial population than on the general population is required and for this reason response-adaptive clinical trials are a more suitable design.

Response-adaptive trials differ from a randomized approach as they use information gained from previous patients within the study, to decide which treatment to allocate to the next patient or next group of patients. They vary the treatment allocation probability in order to favor the treatment, which is estimated, to be more effective in order to maximize the number of successful outcomes in patients.⁵

Bayesian adaptive designs

Bayesian statistical methods provide a framework in which information beyond that collected in a particular clinical trial can be used to make statistical inferences about the treatment outcomes. Prior information (from previous trials, scientific research or expert opinion) can be combined with information as it is accrued during a trial, as well as with the usual data available on completion of the trial, to make efficient and timely inferences about the safety and/or efficacy of a treatment or therapy.

Making use of relevant prior information can reduce sample sizes (or shorten trial lengths) required to meet objectives and so reduce overall development costs. But there are advantages beyond the costs. Use of data as they are collected (either through interim analyses or continual reassessment methods) allows the trial design to be adapted to improve design efficiency. For example, ineffective treatment arms could be dropped, further treatment arms could be introduced, the trial could be stopped early (due to established futility /efficacy), or randomization to treatment could be altered to favor the more effective treatment. Such adaptations are attractive to both researchers and patients, by making more efficient use of patient resource and potentially treating patients more effectively. And again, more efficient use of data can lead to lower overall costs. In general, adaptive trial designs are easier to implement within the Bayesian framework.

Due to the expense of clinical trials, more and more pharmaceutical companies are becoming interested in Bayesian methods-in 2006, the FDA called for advances in trial designs, specifically the use of adaptive design methods and Bayesian approaches in R&D⁶. More recently, both the Pharmaceutical Research and Manufacturers of America (PhRMA) and Biotechnology Industry Organization (BIO) established adaptive design working groups with the aim of proposing strategies, methodologies, and implementations for regulatory consideration.

In late 2018, the FDA issued draft guidance on the use of adaptive designs as part of its push to facilitate the adoption of more efficient trial designs and lowering costs in drug development. Extolling the value of adaptive designs, the former FDA Commissioner Scott Gottlieb, MD, said:

Adaptive clinical trials can give sponsors the flexibility to react to clinical evidence as it’s being collected, and modify the design and enrolment in trials by including more patients with characteristics that help predict that they’re more likely to derive a benefit. Or exclude patients with characteristics that suggest that they’re more likely to suffer a side effect. By enriching the enrolment in the trial for patients with characteristics that are likely to predict clinical success, it has the potential to make the development process more efficient. This approach also allows us to potentially learn much more about the characteristics that can inform safer prescribing.⁷

No legally binding guidance has been provided to date but given the on-going algorithmic development and improved computational speeds, these methods are becoming increasingly accessible and accepted, particularly in the area of oncology^7,8, and more robust regulatory input will be needed at some point. In addition, Bayesian methods have particular advantages in rare disease scenarios where traditional sample sizes can be difficult - if not impossible - to achieve. The standard frequentist methodology of hypothesis testing in a clinical trial may not always be the best approach and Bayesian methods allow alternative approaches to be considered.

Personalized medicine

Personalized medicine is a step away from the ‘one-size-fits-all’ medical approach and instead tailors the treatment to an individual to produce the best response and ensure more effective medical care. Since the 2003 human genome project, one can now map out human DNA and it is plausible to individualize medicine so that it targets a certain gene. As described by Vogenberg⁹, personalized medicine usually targets groups of patients which do not respond well to normal medicine due to certain characteristics such as age, genetics and environmental exposure.

Trial managers can also use these patient characteristics (also known as covariates) in order to allocate patients to a certain treatment. If one treatment is identified to work better on a patient with certain characteristics (e.g. female patients), then the probability of allocating that treatment to the next person can be adjusted depending on the next patient’s gender. This will lead to improved outcomes for patients within the trial.

Proposed adaptive trial design

Assume there are K ≥ 2 treatments and a total of N patients in a clinical trial. Each patient arrives into the trial sequentially, such that the outcome of a patient (patient n) is known before the next patient (patient n+1) enters the trial. Here it is assumed each patient only receives one treatment.

Yang¹⁰ assumes each treatment within the trial has an unknown probability of producing a positive outcome in each patient. This probability, which is dependent on the covariates of the patient, can be estimated using information from previous patients.

When a patient arrives into the trial, a covariate value, x, is observed for each patient. This covariate could be binary (e.g. gender) or it could be continuous (e.g. weight), or it could even represent multiple covariate values being observed. This patient’s covariate value along with information from previous patients, will then be used to assign this patient to a treatment.

The algorithm shown below is a slight variation of that proposed by Yang.

• Allocate the first 5 × K patients who enter the trial to each of the K treatments equally

• Estimate the probability of each treatment producing a successful outcome in a patient, based on information from the previous patients

• Given that the covariate is known, x of the next patient n+ 1, find the treatment with the highest probability of producing a successful outcome in that patient

• Select currently the best treatment with high probability and select each of the other treatments with low probability

• Trial managers now know the outcome of patient n + 1, update the probability estimate of that treatment producing a successful outcome in a patient

• Repeat steps 3-5 for the next patients n + 2, n + 3 and so on.

This algorithm selects the currently estimated ‘best’ treatment for patient n with high probability, therefore it will produce more successful outcomes in patients than a randomized control trial with the same treatments. Thus, patients are more likely to join the trial as they are more likely to have successful outcomes. Hence recruitment for the trial is expected to be easier and can happen quicker than an equally randomized trial.

However, the algorithm still gives a small probability of allocating a patient to one of the ‘worse’ treatments. Thus, it maintains a good chance of identifying significant differences between treatments if present. So, when the experimental drug is proven to be the better treatment, it can be moved to the next phase with high certainty it is better than the control drug. When the experimental drug is proven to be worse, the trials on the drug can be stopped, as it is known with high certainty it will not pass the next phase of trials. Hence pharmaceutical companies save money by only investing in and progressing the best treatments to the next phases of trials.

At the start of the trial, when little is known about any of the treatments, the difference in probability of being assigned the estimated best treatment and worse treatment is small. However, as information on the treatments is accumulated, the estimates of each treatment producing a successful outcome in the next patient becomes more certain. Hence as more patients enter the trial, the probability of being given the estimated best treatment increases and the probability of being given the estimated worse treatment decreases.

Conclusion

The design of clinical studies is imperative to the continued success of the pharmaceutical industry. As pharma businesses shift their focus away from ‘blockbusters’ to treatments for rare and orphan indications, clinical trial design must also go through a period of change as sponsors and their contract research partners ensure the most appropriate design is chosen to maximize returns.

Adaptive designs are a promising option for rare disease trial sponsors, offering more control and better-informed decisions around progressing a drug through trial phases while improving patient outcomes, recruitment and retention.

Holly Jackson is a PhD student at the University of Lancaster who is being sponsored by Quanticate.

References

1. Barrangou R, Doudna JA. Applications of CRISPR technologies in research and beyond. Nat Biotechnol 2016; 34: 933-41.
2. Collins FS., Gottlieb S. The Next Phase of Human Gene-Therapy Oversight. N Engl J Med 2018; 379:1393-1395
3. (2019, January 23). Retrieved from British Heart foundation: https://www.bhf.org.uk/what-we-do/our-research/heart-statistic
4. (2019, January 23). Retrieved from What is cystinosis?: https://www.cystinosis.org.uk/learn-more/what-is-cystinosiss/
5. Cheung, Y. (2006). Continuous Bayesian adaptive randomization based on event times with covariates. Statistics in Medicine, 55-70.
6. https://ojrd.biomedcentral.com/articles/10.1186/1750-1172-3-11
7. https://ojrd.biomedcentral.com/articles/10.1186/1750-1172-3-11 - section 4
8. https://www.sciencedirect.com/science/article/pii/S095980490700010X
9. [4] Vogenberg, F. R. (2010). Personalized medicine: part 1: evolution and development into theranostics. . Pharmacy and therapeutics, 560.
10. Yang, Y. (2002). Randomized allocation with nonparametric estimation for a multiarmed. The annals of Statistics, 100-121.