Improving early phase oncology clinical trial design: The case for finding the optimal biological dose

Historically early phase oncology drug development programmes have been based on the belief that “more is better”. Furthermore, rule‐based study designs such as the “3 + 3” design are still often used to identify the MTD. Phillips and Clark argue that newer Bayesian model‐assisted designs such as the BOIN design should become the go to designs for statisticians for MTD finding. This short communication goes one stage further and argues that Bayesian model‐assisted designs such as the BOIN12 which balances risk‐benefit should be included as one of the go to designs for early phase oncology trials, depending on the study objectives. Identifying the optimal biological dose for future research for many modern targeted drugs, immunotherapies, cell therapies and vaccine therapies can save significant time and resources.


| INTRODUCTION
Bayesian model-based and model-assisted designs have been used in recent years to improve the estimation of the Maximum Tolerated Dose (MTD) in First in Human early phase oncology studies. Phillips and Clark 1 present a case study that supports the use of the model-assisted Bayesian Optimal Interval Design (BOIN). 2 They argue it is easy to use in practice since the escalation rules are intuitive, and the design provides flexibility with respect to a variety of different design features including cohort sizes. The BOIN design also does not require in-trial data processing. Finally, the design performs well when limited information is known about the expected dose-toxicity curve and is optimal for MTD finding per Bayesian decision-theoretic criteria. Phillips and Clark argue that these newer Bayesian model-assisted type of designs which are readily available via existing shiny apps should become the go-to design for statisticians for identifying the MTD, rather than the classical standard "3 + 3" design which is still being used today.
Historically, oncology drugs have been developed based on the belief that "more is better". However, this presumption is not true for many modern targeted drugs, immunotherapies, cell therapies and vaccines. For example, efficacy may plateau or even decrease at high doses. Further, some of the therapies demonstrate minimal toxicity even in high doses, making the MTD unlikely to be reached. 3,4 As a result, designs aimed at finding the MTD are inappropriate for such therapies generally and, when used in that context, are likely to lead to finding inappropriate doses for future research. Subsequently Project Optimus, an FDA OCE regulatory initiative, is encouraging sponsors to identify the Optimal Biological Dose (OBD). 5 Consequently, study designs such as the BOIN12 6 are increasing in importance since they simultaneously assess the benefit-risk of new compounds.
This short communication compares the BOIN and BOIN12 study designs in selecting doses for future research. After briefly discussing both study designs, the associated operating characteristics for a range of different scenarios are explored using simulation. The BOIN12 design is less sensitive to the assumption that more is better and identifies an OBD which can potentially save time and resources. That is, rather than focusing on the MTD and then having to refine the therapeutic dose in subsequent trials, there are significant advantages in identifying the OBD in First in Human early phase oncology studies for many modern targeted drugs. Subsequently designs such as the BOIN12 which balances risk-benefit should be included as one of the go-to designs for early phase oncology trials, depending on the study objectives.

| BOIN STUDY DESIGN
In early phase oncology drug development, model-assisted Bayesian designs such as the BOIN study design can be deployed to identify the MTD. Other non-Bayesian designs such as the Modified Toxicity Probability Interval design, 7 although the decision boundaries for the design are not optimised according to the Bayesian rule, or nonmodel-assisted interval based designs such as the "i3 + 3" 8 are also available. Once identified, the MTD is then used in subsequent clinical trials to establish proof of efficacy, albeit it could be at a lower dose than the MTD. FDA OCE under Optimus are encouraging a range of doses to be explored.
The proportion of patients experiencing Dose Limiting Toxicity (DLT) events at specific doses is compared with a pair of fixed, pre-specified set of escalation and de-escalation boundaries based on the target toxicity rate. Other study design parameters needed to investigate the operating characteristics of a BOIN design using simulation include: 1. Number of doses to be studied 2. Planned cohort size 3. Maximum sample size for the study 4. Maximum number of patients to be treated at any one dose, typically set between 9-12 In the BOIN design, criteria are also established that prevent unnecessary escalation to doses that are potentially unsafe. Table 1 and Figure 1 summarise the decision rules and the study design process flow diagram for a BOIN study for a MTD target toxicity rate of 30% (i.e., ϕ ¼ 0.3, which is frequently used in practice), using cohort sizes of 3, 5 doses to be studied, and a maximum sample size of 30 patients with a maximum of 12 patients per dose level. The escalation and de-escalation boundaries are then ≤0.236 and >0.359. Table 1 and Figure 1 were generated from the web-based application: https://trialdesign.org/. 9

| BOIN12 STUDY DESIGN
The BOIN12 study design is an alternative BOIN type design but used when the primary objective is to identify the OBD by using a utility to measure the risk-benefit trade off, whose value is elicited from clinicians to reflect the clinical desirability of each possible toxicity and efficacy outcome. As an example of the framework for the utility to measure the toxicity-efficacy trade-off, as per Lin et al, 6 suppose toxicity and efficacy are binary. Given any patient in the trial, Table 2 presents the four possible different outcomes. The most desirable outcome (no toxicity, efficacy) is denoted u 1 and is usually assigned a score of 100. The least desirable outcome (no efficacy, toxicity) is denoted u 4 and usually assigned a score of 0. Specifying u 2 > u 3 means that toxicity is judged more important than efficacy and vice versa. A higher value of u indicates a higher desirability of that dose in the risk-benefit trade-off.
Given a specific dose, d, let p 1 , …, p 4 denote the respective probabilities of observing outcomes u 1 , …, u 4 which typically vary across dose. Averaging across the four possible outcomes, the desirability (or mean utility) of dose d is During the trial, first the BOIN12 study design escalates/de-escalates based on observed DLT rates. If the current dose, say dose j, is greater than or equal to the escalation and lower than the de-escalation limits, then the dose for the next cohort of patients is selected as either the current dose ( j) or the previous dose level (jÀ1) based on the desirability score determined by counting the number of patients: 1. Treated at doses jÀ1, j (and j + 1) 2. Who experienced toxicity at each dose level 3. Who experienced efficacy at each dose level At the end of the trial, the dose with the highest desirability is the OBD. Figure 2 summarises the BOIN12 process flow diagram from the web-based application: https://trialdesign.org/ 9 for a design with a MTD target toxicity rate of 35% and efficacy 25%. Dose j + 1 = Dose above the current dose; Dose jÀ1 = Dose below current dose. Because BOIN12 considers the toxicity-efficacy trade off, the target MTD value is typically set slightly higher than the target toxicity rate in conventional toxicity-based designs such as the BOIN design. For example, 30% is frequently used for the BOIN design, compared with 35% for BOIN12. 10 The higher toxicity rate for the BOIN12 design permits a dose with slightly higher toxicity but significant improvement in efficacy to be selected.
A key innovation of BOIN12 is that its calculation of dose desirability can be pre tabulated and included in the trial protocol. As an example, Table 3 gives the desirability scores for a target toxicity of 35% and a target efficacy of 25%, a utility specification of u 1 = 100, u 2 = 40, u 3 = 60 and u 4 = 0, cohort sizes of 3 and a maximum of nine patients treated at any one dose.
More details of the BOIN12 and other model-assisted phase I/II designs can be found in a recently published book by Yuan, Y., Lin, R., & Lee, J. J. entitled "Model-Assisted Bayesian Designs for Dose Finding and Optimization: Methods and Applications". 10

| BOIN AND BOIN12 OPERATING CHARACTERISTICS
To investigate the performance of the BOIN and BOIN12 study designs in terms of likely doses that would be selected for future research, simulations using the web-based application: https://trialdesign.org/ 9 were performed. More specifically, how often using a design to identify the MTD would likely lead to additional cost and time to explore doses from an efficacy perspective, instead of implementing an appropriate design to identify the OBD up front. For illustrative purposes, for the BOIN12 this paper has assumed the 2 Â 2 utility table comprising u 1 = 100, u 2 = 40, u 3 = 60 and u 4 = 0 and a target toxicity of 35%. That is, a study where the primary goal is to find the dose level that possesses the highest efficacy rate amongst the doses with toxicity rates not greater than 35%. The target efficacy level was set at 25%. A correlation of 0.1 was also assumed between toxicity and efficacy. As per Lin, 11 16 different scenarios were studied as outlined in Table 4. The target OBD are bolded in the table based on clinical and statistical considerations. Five 5 dose levels, cohort sizes of 3 and a maximum number of treated patients per dose of 12 patients were assumed. For the BOIN design, only the toxicity data from the 16 scenarios were used. The target toxicity rate was set at 30%, a value frequently used in practice. Although different, the target toxicity rates for the BOIN and BOIN12 reflect frequently used conventions, and Lin et al. 6 This permits the selected doses when using the BOIN and BOIN12 designs in practice to be compared.
T A B L E 2 Framework for the desirability of doses based on the utility score assigned to four different outcomes Toxicity Efficacy

Yes No
No The results of the simulations are presented in Table 5, specifically for each scenario the most frequently selected dose from the simulations together with the frequency of selection are presented. Not surprisingly, the BOIN12 design tends to outperform the BOIN design with respect to determining the OBD. When considering the most frequently selected dose for each scenario and design, in the simulations the BOIN design only aligned with the OBD 19% (Scenarios 3, 4, 14) of the time, compared with 100% alignment for the BOIN12 design. In the cases where the MTD is the OBD such as scenarios 3 and 4, the BOIN design can be used. However, the simulations show that BOIN12 still maintains a competitive performance in such scenarios. Scenarios 1, 8 and 9 represent cases where one might expect the BOIN study design to select an inappropriate dose for further research, when efficacy is not factored in the FIH study. That is, where the efficacy dose response curves flatten or decrease at the higher doses as is the case for many modern targeted drugs, immunotherapies, cell therapies and vaccines. Table 5 supports that implementing study designs that balances benefit-risk will lead to earlier identification T A B L E 3 Desirability score for a target toxicity of 35% and a target efficacy lower limit of 25%, a utility specification of u 1 = 100, u 2 = 40, u 3 = 60 and u 4 = 0, cohort sizes of 3 and a maximum of nine patients treated at any one dose of the OBD. If the objectives are not to identify the OBD in the FIH study, but the MTD, further studies resulting in additional time and cost are needed to explore the efficacy-dose response relationship. If the additional research is not undertaken, then either efficacy may not be demonstrated in subsequent trials and/or patients will be exposed to an unnecessarily high dose.  Figure 3 illustrates the benefits of the BOIN12 design by selecting the OBD more frequently for scenarios 8 and 9 where efficacy decreases at higher doses and that more patients are treated at the OBD. Furthermore, the BOIN12 design generally allocates fewer patients to higher doses than the BOIN design, suggesting the design is potentially safer.
Scenarios 15 and 16 are cases where the development of the compound should be stopped since the lowest dose is above the targeted toxicity level. However, on average both the BOIN and BOIN12 design suggest dose 1 or 2 should be researched in further clinical trials.

| DISCUSSION
Zhou et al 12 has shown model-based and model-assisted designs out-perform simple rule-based designs such as the standard or classical "3 + 3" design when estimating the MTD. Although there has been relatively slow up take, model-assisted designs such as the BOIN design are now starting to be used in practice to identify the MTD. The design allows tailored escalation towards a target toxicity dose level, whilst maintaining control of the specified maximal tolerable rate of toxic events.
Historically, oncology drugs have been developed based on the belief that "more is better". However, this assumption is not true for many modern targeted drugs, immunotherapies, cell therapies and vaccine therapies. Designs aimed at finding the MTD are inappropriate for such therapies generally and, when used in that context, are likely to lead to identifying inappropriate doses for future research.
F I G U R E 3 BOIN and BOIN12 for scenarios 8 and 9, assuming a 2 Â 2 utility table of u 1 = 100, u 2 = 40, u 3 = 60, u 4 = 0, target toxicity of 35% and target efficacy of 25%. A correlation of 0.1 was assumed between toxicity and efficacy, cohort sizes of 3 and a maximum number of treated patients per dose of 12 patients.
Statisticians have an opportunity to shape and improve early phase oncology drug development programmes by introducing newer, more efficient study designs and promoting the need to include more than a single dose in phase 1b studies and beyond. This manuscript supports that model-assisted designs such as the BOIN12 for identifying the OBD, balancing benefit-risk, should be included as one of the go-to designs for early phase oncology trials, depending on the study objectives. Identifying the OBD for future research for many modern targeted drugs, immunotherapies, cell therapies and vaccine therapies can save significant time and cost.
It is important to note that since efficacy endpoints can sometimes take much longer to observe than safety, designs such as the BOIN12 are not always appropriate. Researchers need to balance the additional development time and costs needed to explore efficacy in subsequent studies after identifying the MTD in the FIH study, versus identifying the OBD from the onset. Furthermore, when there is late onset of toxicity and efficacy, alternative designs that estimate the OBD in such situations should be considered. For example, the TITE-BOIN12 design 13 may be more appropriate. Other areas where early phase oncology development could be improved is the incorporation of PK/PD and/or biomarker data. Dose selection for future research should be based on a broad range of evidence, not solely based on the MTD.
In summary, statisticians need a range of different study designs for early phase oncology and deployed depending on the study objectives and properties of the therapy under investigation. The BOIN12 design which balances riskbenefit should certainly be included as one of the go-to designs for early phase oncology trials.