Optimizing Trial Efficiency with Bayesian Dose Escalation Designs
By: Vanessa Beddo, Vice President, Biostatistical Consulting, Allucent
Optimizing Trial Efficiency with Bayesian Dose Escalation Designs
By: Vanessa Beddo, Vice President, Biostatistical Consulting, Allucent
As acceptance of flexible trial designs continues to advance, many sponsors are increasingly exploring alternate methods of dose escalation to support the identification of a maximum tolerated dose (MTD) and recommended Phase II dose (RP2D) in open-label Early Phase or First in Human studies.
Bayesian dose escalation designs have also coincided with advances in oncology therapy development, such as immunotherapies and antibody-drug conjugates. These more flexible methods are viewed as potentially more appropriate than those initially designed for chemotherapy drug development.
Allucent has been working with clients to investigate and select flexible-dose escalation designs that are appropriate for their situation and provide for their operational execution. The broader use of these methods is an exciting development with the potential for a more precise determination of an MTD, increasing chances for future trial success while preserving patient safety.
Three common types of open-label adaptive ascending dose escalation designs include:
- Rule-based, which includes the 3+3 design
- Bayesian methods
— Model-based, such as the continual reassessment model (mCRM) or Bayesian logistic regression model (BLRM)
— Hybrid models, including keyboard/mTPI-2 and Bayesian optimal interval (BOIN) designs
Some of these Bayesian methods can easily be executed in a manner similar to that of a rule-based design.
A Brief Description of Common Bayesian Dose Escalation Designs
In general, the use of Bayesian methods involves planning and simulation to ensure that the design is appropriate for operational execution, based on assumptions around dose-level escalation plans and presumed toxicity rates. This information involves not only statistical, but also preclinical, clinical, pharmacokinetic, and/or pharmacodynamic areas of expertise. Simulation outcomes for candidate dose-escalation designs ensure that protocol reviewers understand the level or risk to patient safety (kept low) and provide insight as to what a team might expect with respect to the operational characteristics of the study.
Briefly, the BLRM and mCRM methods model the probability that a patient might experience a DLT based on an assumed dose-toxicity function, which is repeatedly estimated based on emerging data. After estimation, these posterior probabilities are used to assign subsequent cohorts of patients to an optimal dose. Dose optimality is defined and based on a target toxicity rate or acceptable toxicity interval; additional safeguards can also be incorporated to avoid cohort assignments to dose levels that may be associated with an unacceptably high toxicity risk (i.e., due to estimation variability). For these two methods, ongoing statistical support during the study in support of cohort dose-level recommendations is needed in order to determine the next optimal dose level for exploration. Further, cohort reviews often also involve an analysis of how the study will operationally proceed, given hypothetical outcomes. In short, for these model-based designs, statistical input is not only involved with respect to study setup but is also required for ongoing support with respect to cohort escalation decisions.
Hybrid models, such as the modified toxicity probability interval (mTPI-2) or Bayesian optimal interval (BOIN) methods, involve similar setup and exploration with respect to a review of the trial’s operational characteristics. It should be noted that the latter was identified by the FDA as Fit-for-Purpose in December, 2021 (see https://www.fda.gov/drugs/development-approval-process-drugs/drug-development-tools-fit-purpose-initiative for additional details). However, instead of assuming a single dose-toxicity curve for purposes of posterior DLT probability modeling, these methods consider modeling relative to individual dose levels during the trial. Escalation decisions are dependent on observed DLT rates at explored dose levels, alongside predetermined decision rules. This typically allows for minimal statistical support during operational conduct. The MTD is identified based on an isotonic regression using information accumulated during dose exploration.
All of these designs (model-based and hybrid) also allow for various stopping rules, defined as appropriate for each method, including the maximum number of patients treated at a dose or overall in the study. As mentioned previously, such assumptions would be included in the planning stage within the context of study setup simulations, allowing for an understanding of the operational characteristics of the intended design. Metrics to be explored can include the efficiency with which the MTD can be identified, the likelihood of patient overdosing, or the overall DLT rate, given each set of assumptions and dose escalation methodology.
Comparisons with the Standard Rule-Based Design
The standard 3+3 design assumes a predefined dose-escalation pattern with rules for identifying the MTD, static assumptions for targeted toxicity rate, and fixed cohort sizes for prospective enrollment. As a rule-based model, it does not consider information collected during exploration of previous doses, nor does it incorporate uncertainty. Further, there is no opportunity for dose rechallenge. Thus, in references of simulation studies exploring comparisons of Bayesian designs against the 3+3, the 3+3 design has demonstrated high variability in MTD identification, with a generally lower likelihood of finding the “true” MTD.
In conclusion, the efficiency with which the MTD is identified is improved using Bayesian methods. Determining which method is the best for each situation is based on individual assumptions around potential drug performance and any additional regulatory or safety requirements. Lastly, in comparison to the static 3+3 design, these Bayesian methodologies offer an opportunity to include additional cohorts, alongside a predefined method for including their associated DLT information, for further understanding of early drug performance. Such flexibility for exploration is often amenable to the development of recent, novel cancer therapies (in comparison to those underpinning traditional chemotherapy).
Additional Considerations and Summary
The selection of popular designs described in this article are used to select a safe, tolerable dose for one investigational agent dependent on full cohort completion of a fixed DLT period. Although this design scenario is supported at Allucent, it is worth mentioning that other design characteristics may be relevant to the selection of alternate methodologies that can meet one or more of the following requirements:
- Dynamic decision-making with respect to cohort DLT period completion (e.g., “rolling” or time-to-event designs)
- Dose exploration of combination therapies
- Dual exploration of dose levels against toxicity rates and efficacy outcomes
The Allucent consulting team can assist with an evaluation of your design needs and help you select a match that best aligns with your drug development objectives.
ABOUT Vanessa Beddo
Vanessa Beddo regularly provides innovative and complex study design solutions in support of efficiently navigated regulatory paths, building on her 15+ years of experience in clinical trial design and analysis. As VP of Biostatistical Consulting at Allucent, she is also responsible for regulatory body interactions on an ongoing basis throughout program lifecycles on behalf of her clients. Prior to this role, Dr. Beddo served in various management positions involving oversight of statistical/programming staff and leadership with respect to departmental initiatives, processes, and training. Her expertise also includes clinical development of pharmaceuticals and biologics, having served as the lead statistician on Phase I-IV clinical trials and drug submissions leading to successful drug approvals.