Abstract
PURPOSE
Using novel Bayesian adaptive designs has great potential to improve the efficiency of early-phase clinical trials. A major barrier for clinical researchers to adopt novel designs is the lack of easy-to-use software. Our purpose is to develop a user-friendly software platform to implement novel clinical trial designs that address various challenges in early-phase dose-finding trials.
METHODS
We used R Shiny to develop a web-based software platform to facilitate the use of recent novel adaptive designs.
RESULTS
We developed a web-based software suite, called Bayesian optimal interval (BOIN) suite, which includes R Shiny applications to handle various clinical settings, including single-agent phase I trials with and without prior information, trials with late-onset toxicity, trials to find the optimal biological dose based on risk-benefit trade-off, and drug combination trials to find a single maximum tolerated dose (MTD) or the MTD contour. The applications are built using the same software architecture to ensure the best and a uniform user experience, and they are developed using a proven software development standard operating procedure to ensure accuracy, robustness, and reproducibility. The suite is freely available with internet access and a web browser without the need of installing any other software.
CONCLUSION
The BOIN suite allows clinical researchers to design various types of early-phase clinical trials under a unified framework. This work is extremely important because it not only advances the clinical research and drug development by facilitating the use of novel trial designs with optimal performance but also enhances collaborations between biostatisticians and clinicians by disseminating novel statistical methodology to broader scientific communities through user-friendly software. The BOIN suite establishes a KISS principle: keep it simple, but smart.
BACKGROUND
The main purpose of phase I clinical studies is to evaluate the safety of a new drug and identify a maximum tolerated dose (MTD), defined as the highest dose with an acceptable level of toxicity. The designs used to identify the MTD can be classified into three types according to their statistical foundations and implementation approaches: algorithm-based (also known as rule-based), model-based, and model-assisted methods.1,2 Algorithm-based designs, for example, the 3 + 3 design,3 guide dose assignment through a set of prespecified rules without any parametric assumptions on the dose-toxicity curve. They are simple to implement, but rigid and have poor accuracy to identify the correct MTD.4 Model-based designs assume a dose-toxicity model and guide the dose escalation and de-escalation by continuously updating the estimate of the dose-toxicity model based on accrued data. Typical examples of model-based designs include the continual reassessment method (CRM5) and its variations, such as the escalation with overdose control,6 the Bayesian logistic regression model,7 and the Bayesian model averaging CRM.8 Model-based designs yield better performance than the 3 + 3 design in identifying the MTD and allocating more patients to the MTD,9,10 but they are less transparent and more complicated to implement.
Model-assisted designs are a relatively new class of designs that enjoy both the simplicity of algorithm-based designs and the good performance of model-based designs. Similar to model-based designs, model-assisted designs use a statistical model (e.g., the binomial model for the number of dose-limiting toxicities) to assist the derivation of decision rules; similar to algorithm-based designs, model-assisted designs can tabulate the dose escalation and de-escalation rules before the onset of the trial, making it as simple as the 3 + 3 design to implement. Classical examples of model-assisted designs in phase I clinical trials include the modified toxicity probability interval (mTPI) design,11 Bayesian optimal interval (BOIN) design,12 and keyboard design.13 Zhou et al9,10 reviewed and compared various phase I trial designs and found that the BOIN design yields desirable performance that is comparable to CRM and better than mTPI, whereas it is still simple and transparent. The BOIN design has been used in various types of cancer trials (see Yuan et al1 for examples) and extended to accommodate various clinical settings, such as late-onset toxicity,14 drug combination trials,15 efficacy and toxicity-based phase I and II trials,16,17 and dose-finding with late-onset efficacy and toxicity.18,25 Therefore, in this paper, we focus on the software development for BOIN designs using the R package Shiny. The software platform for the BOIN designs is referred to as BOIN suite hereafter. The software we developed for BOIN designs and others (eg, CRM and keyboard designs) can be found in Ref. 19.
One distinct feature of the BOIN suite is that it seamlessly integrates various BOIN methods to allow users to design different types of early-phase trials under a unified framework. The suite includes apps to design single-agent phase I trials with no historical data (BOIN12) or by incorporating historical data (iBOIN19), trials with delayed toxicity (time-to-event [TITE]-BOIN14), trials to find the optimal biological dose (OBD) based on a utility function that considers the risk-benefit trade-off (U-BOIN17), and drug combination trials (BOIN COMB) to find a single MTD15 or the MTD contour.20
To achieve the best user experience, all the shiny apps in the BOIN suite share a similar graphical user interface and are built upon the same software architecture with the modules that naturally follow the trial design process: (1) generate decision tables; (2) simulate operating characteristics such as the probability of selecting the MTD or OBD depending on the purpose of a trial, percentage (or number) of patients allocated to each dose, and early stopping probabilities; (3) prepare trial protocol; (4) compute the recommended dose for the next cohort of patients; and (5) estimate MTD (or OBD) when a trial is completed. Lee and Chu21 suggested that the availability of accompanying software for the implementation of novel Bayesian methods is crucial for the use of these methods in clinical trials. Our objective is to promote and facilitate the use of novel trial designs to improve drug discovery and clinical research by developing reliable, robust, and easy-to-use software.
The rest of the text is organized as follows. The Methods section describes the dose-finding designs included in the BOIN suite. The Results section provides the specific features and functionalities in each shiny app, and a real trial example to illustrate the use of the BOIN app. The Summary section concludes the study with a brief summary.
METHODS
In this section, we briefly describe the key idea for each design in the BOIN suite and delegate the technical details to the Data Supplement.
BOIN for Single-Agent Trials
The BOIN design is straightforward. It makes the decision of dose escalation and de-escalation by simply comparing the observed dose-limiting toxicity (DLT) probability with a pair of escalation and de-escalation boundaries that are optimized to minimize the probability of exposing patients to overly toxic or subtherapeutic doses. More precisely, let denote the observed DLT probability at dose level j = 1, …, J, where yj is the number of observed DLTs and nj is the number of patients treated at dose level j. Suppose the current dose level is j, then the dose assignment for the next cohort of patients is determined as below:
If , escalate to dose level j+1.
If de-escalate to dose level .
Otherwise, stay at the current dose.
The BOIN design enjoys several desirable statistical properties (eg, optimality, consistency, long-memory coherence, and Bayesian-frequentist dual interpretation). Liu and Yuan12 provide the derivation of and a description of the statistical properties for the BOIN design.
For the purpose of overdose control, the BOIN design employs a dose elimination or early stopping rule as follows: eliminate doses j and higher levels from further examination if and at least three patients are treated on dose level j, where pj is the true DLT probability of dose level j, j = 1, …, J, ϕ is the target DLT probability, and δ is a probability cutoff (typical values are 0.9 and 0.95). This posterior probability is evaluated based on the beta-binomial model , with . When the lowest dose is eliminated, stop the trial for safety.
The simple structure of the BOIN design renders it some unique advantages. For example, the BOIN design guarantees de-escalating the dose when the observed DLT probability pj is higher than the de-escalation boundary λd. This makes it particularly easy for clinicians and regulatory agents to assess the safety of a trial using the BOIN design. Although remarkably simple, Zhou et al9,10 found that the BOIN design yields outstanding performance that is comparable to the CRM. The BOIN design has been used in many early-phase clinical trials, including trials for pediatric tumors, adult tumors, solid tumors, and liquid tumors (see Yuan et al1 for some examples).
BOIN for Single-Agent Trials Using Informative Prior (iBOIN)
The iBOIN design is an extension of BOIN that allows incorporating historical data to improve the efficiency of clinical trials by using an informative prior.19 The iBOIN design retains the same concise decision rule as BOIN: the decision of dose escalation and de-escalation is made by simply comparing with a pair of escalation and de-escalation boundaries . The only difference is that is dose-specific, as indicated by subscript j, because of the incorporation of the dose-specific prior information (eg, the prior estimates of DLT probability for J = 5 doses under investigations are [0.05, 0.1, 0.3, 0.45, 0.6]). Zhou et al19 provided an easy and intuitive procedure to determine : based on the historical data, they obtain the prior estimate of DLT probability at each dose (i.e., known as skeleton) and set a desirable prior effective sample size (ie, the number of patients intended to be represented by the prior information). Then, can be easily determined using a close-form formula (see the Data Supplement, Section S.1.2 for details).
Escalation and de-escalation boundaries of iBOIN are adaptive to the prior information. When the prior information indicates that the DLT probability of a dose is likely below the target, the corresponding escalation boundary will be lower to encourage dose escalation; when the prior information indicates that the DLT probability of a dose is likely above the target, the corresponding de-escalation boundary will be lower to encourage dose de-escalation. Zhou et al19 show that the iBOIN design can substantially improve the accuracy of dose-finding and is robust to prior misspecification. When no prior information is used, iBOIN becomes BOIN.
BOIN for Single-Agent Trials With Delayed Toxicity (TITE-BOIN)
With the advent of novel molecularly targeted agents and immunotherapy, late-onset toxicity has become a common problem in phase I trials. The BOIN design described above cannot directly handle late-onset toxicity. This is because when toxicity is late-onset, at the time to make the decision of dose assignment for a new patient, some patients’ DLT data are pending, thus disabling the direct calculation of to apply the BOIN decision rule. The same logistic difficulty arises when the accrual is faster than the toxicity evaluation window. To overcome this issue, Yuan et al14 proposed the TITE-BOIN design that allows real-time dose assignment decisions for new patients in the presence of pending toxicity data. The idea of TITE-BOIN is to impute the pending DLT data using the observed data from all patients (including the follow-up time of pending patients). After the imputation, we obtain the complete data and can calculate so that the BOIN escalation and de-escalation rules can be applied. The hallmark of the TITE-BOIN design is its simplicity. The decision rule of the TITE-BOIN design can be pretabulated and included in the trial protocol. To conduct a trial, no complicated model fitting or estimation is needed. We only need to count the number of DLTs observed and number of patients whose outcomes are pending and calculate the standardized total follow-up time, which is defined as the ratio of the total follow-up time for patients having pending outcomes over the DLT assessment window. Then, look up the predetermined decision table to determine the dose for treating the next cohort of patients. Yuan et al14 showed that TITE-BOIN yields similar accuracy as the more complicated model-based TITE-CRM,22 but it is substantially simpler, safer, and more robust. More statistical details are provided in the Data Supplement, Section S.1.3.
Utility-Based Phase I-II Design to Identify Optimal Biological Dose (U-BOIN)
Traditional dose-finding designs developed for chemotherapy aim to find the MTD. The underlying assumption is that both toxicity and efficacy monotonically increase with the dose, and thus, the MTD presents the most efficacious dose that is safe. This assumption, however, is often questionable for immunotherapy and targeted agents. To optimize the treatment benefit of immunotherapy and targeted therapy, it is important to consider toxicity and efficacy simultaneously and quantify their risk-benefit trade-off during dose-finding. The objective of such a dose-finding design is to identify the OBD, defined as the dose that has the highest desirability in terms of the risk-benefit trade-off. Zhou et al17 developed the U-BOIN design, a utility-based seamless phase I and II design, to find the OBD. The U-BOIN dose-finding procedure consists of two stages. In stage I, the BOIN design12 is used to quickly explore the dose space and collect preliminary toxicity and efficacy data. In stage II, in light of accumulating efficacy and toxicity from both stages I and II, the desirability of each dose is continuously updated, based on the posterior estimate of toxicity-efficacy trade-off, to guide dose assignment for new patients. At the end of the trial, the dose with the highest estimate of desirability is selected as the OBD.
U-BOIN measures the desirability of the doses using utility. Under this approach, each possible outcome is assigned a utility score to represent its desirability. For example, for binary toxicity and efficacy endpoints, given a patient, there are four possible outcomes: (efficacy, no toxicity), (no efficacy, toxicity), (efficacy, toxicity), and (no efficacy, no toxicity). We assign the most desirable outcome (efficacy, no toxicity) a score of 100 and the least desirable outcome (no efficacy, toxicity) that of 0. We then use these two as the reference to score the other intermediate outcomes, for example, (efficacy, toxicity) = 60 and (no efficacy, no toxicity) = 40. The scores should be elicited from clinicians to reflect the toxicity-efficacy trade-off in medical decision making, which depends on the disease, drug, and target patients.
Compared with complicated model-based designs, U-BOIN is simple to implement because its decision rule can be pretabulated in the protocol. Moreover, numerical studies show that U-BOIN often outperforms model-based designs in identifying the OBD, and it is substantially more robust as it does not make any parametric assumption on dose-toxicity and dose-efficacy curves.17
BOIN Designs for Combination Trials
Compared with single-agent trials, drug combination trials have some unique challenges: (1) the drug combinations are only partially ordered in terms of their DLT probabilities and (2) there could exist an MTD contour (ie, more than one MTD exists). To address the first challenge, Lin and Yin15 proposed the BOIN Comb design to find an MTD for drug combination trials. Zhang and Yuan20 developed the BOIN-waterfall design to simultaneously address both the challenges.
BOIN comb design to find a single MTD.
Suppose there are two drugs (A and B) under investigation. Drug A has J dose levels, whereas B has K levels. Let pjk denote the true DLT probability of the combination dose (j, k), where j, j = 1, …, J and k, k = 1, …, K, and denote the corresponding observed DLT probability as , where njk is the number of patients treated on the dose combination (j, k) and yjk is the number of patients who experience toxicity.
The BOIN Comb design determines dose escalation and de-escalation based on the same decision rule as the single-agent BOIN design described previously. Suppose the current dose combination is (j, k), the design escalates the dose if , de-escalates the dose if , and otherwise remains at the current dose. The difference is that in combination trials, there is more than one neighboring dose to move to when performing dose escalation or de-escalation. The BOIN Comb design uses the following rule to determine which drug to escalate or de-escalate: when escalating, escalate to (j + 1, k) or (j, k + 1), whichever has a higher posterior probability that the DLT probability locates within the stay interval ; when de-escalating, de-escalate to (j − 1, k) or (j, k − 1), whichever has a higher posterior probability that the DLT probability locates within . A prominent feature of the BOIN Comb design is that the dose-assignment rule can be equivalently translated into a rule based on a so-called desirability score. An example of the desirability score table for the BOIN Comb design is shown in the Data Supplement (online only). During the trial conduct, we only need to know the number of patients treated and the number of DLTs at the doses and look up the decision table (like Table 1) to determine if escalation or de-escalation is needed. To determine which dose to move to, we look up the desirability score as demonstrated in the Data Supplement. Despite its simplicity, the BOIN Comb design is a top performer in accuracy, safety, and reliability among existing drug combination designs.15 Data Supplement Section S.1.5.1 provides more details for the BOIN Comb design, and Section S.3.4.1 gives a trial example using this design.
TABLE 1.
Dose Escalation, De-Escalation, and Elimination Boundaries in the BOIN Design With a Target DLT Probability of 0.3
BOIN-waterfall design to find the MTD contour.
In some trials, it is of interest to find multiple MTDs (or the MTD contour), which can be further evaluated in subsequent phase II trials. Finding the MTD contour is more challenging than finding a single MTD, because the former requires a thorough exploration of the whole dose matrix using a limited sample size; otherwise, it is at the risk of missing some MTDs.
To overcome this challenge, Zhang and Yuan20 proposed the BOIN-waterfall design. The basic idea of BOIN-waterfall is to divide the two-dimensional dose-finding problem into a series of simpler one-dimensional dose-finding subtrials. In each of the subtrials, dose DLT probabilities are fully ordered and the standard BOIN is used to identify the candidate MTD. After all the subtrials are finished, BOIN-waterfall selects the MTD contour based on the data from all the subtrials using matrix isotonic regression.27 The innovation of BOIN-waterfall is that the subtrials are conducted in a certain order such that the precedent subtrial informs the conduct of the subsequent trial. Data Supplement Section S.1.5.2 provides more design details, and Section S.3.4.2 gives a trial example using BOIN-waterfall.
RESULTS
Shiny Software
The BOIN suite was developed using R programming,23 primarily with the Shiny package.24 The strengths of the BOIN suite lie in the following unique features:
(1) Contains multiple novel designs, each addressing a different challenge in phase I or phase I and II clinical trials.
(2) Runs under different operating systems.
(3) Does not require any installation.
(4) Is not limited by the hardware of users’ computers.
(5) Does not require statistical programming skills to use.
(6) Is regularly maintained and updated on the server so that users do not have to worry about software updates.
(7) Provides comprehensive user help files and is free to use.
Different from most published shiny applications that only contain a single app for a specific design, the BOIN suite contains all the aforementioned Bayesian adaptive designs (Fig 1) and includes a decision tree for users to determine which design to use based on their trial settings (Fig 2). All the app interfaces are user-friendly and self-explanatory. The evaluations of design performance can be done in a point-and-click manner. Table 2 summarizes the structures for the above shiny apps. All apps have the capability to implement the following three main tasks.
FIG 1.
BOIN suite available at www.trialdesign.org. BOIN, Bayesian optimal interval; CRM, continual reassessment method; MTD, maximum tolerated dose; OBD, optimal biological dose; TITE-BOIN, time-to-event-BOIN.
FIG 2.
The decision tree to guide design choice from the BOIN suite options. BOIN, Bayesian optimal interval; MTD, maximum tolerated dose; OBD, optimal biological dose.
TABLE 2.
Structures for the BOIN Suite Shiny Apps
Trial Design.
At the planning stage, users can use the apps to design a trial and conduct simulation for various configurations of trial scenarios and obtain decision rules based on their specified trial settings. A concise protocol template is provided in both html and Word formats.
Trial Conduct.
To determine dose assignment during a trial, users can simply compare the observed data with the decision boundaries obtained using the apps. Alternatively, for more complicated designs, such as U-BOIN, BOIN Comb, and BOIN-waterfall, users can upload trial data to the app to obtain the recommended dose for the next cohort of patients.
Trial Conclusion: MTD or OBD Estimation.
After a trial completes accrual and has all patients’ outcomes evaluated, the apps can be used to determine the MTD, MTD contour, or OBD, depending on the trial design implemented.
Reproducibility and Quality Control
Reproducibility and quality control are vital for software development. We employed a proven standard operating procedure (SOP) and technologies to ensure that the apps are reproducible, reliable, and stable.
(1) The source code of each app has been reviewed by both software engineers and clinical trial experts for accuracy, completeness, repeatability, rigor, and robustness.
(2) The apps have been stress-tested extensively in a wide variety of cases that arose from real-world research.
(3) All the apps are frequently upgraded based on the feedback received from experts worldwide. For example, many options of the apps (e.g., accelerated titration and an early stopping rule that stops the trial when the number of patients accumulated at a dose reaches a certain number), which are not included in the original methodology, are added based on users’ feedback to account for clinical and practical considerations.
(4) Every file used to develop the apps is carefully maintained in Microsoft’s Team Foundation Server source control system to track and document modifications.
Implementation With a Real Clinical Trial Example
Table 3 summarizes the design settings and required input for each application. We provide a concrete trial example using BOIN in this section, and examples for using iBOIN, TITE-BOIN, U-BOIN, and BOIN COMB apps are available in the Data Supplement, Sections S.3.1–S.3.4.
TABLE 3.
The Design Setting and Design Input for the Shiny Apps in BOIN Suite
We use a phase I-II trial conducted at the National Cancer Institute (ClinicalTrials.gov identifier: NCT02942264) to illustrate the use of the BOIN design. The primary objective of the trial was to identify the MTD of TG02, a pyrimidine-based multikinase inhibitor, combined with temozolomide in adult patients with recurrent anaplastic astrocytoma or glioblastoma. Due to the lack of effective treatments for this patient population, the target DLT probability was set at a relatively high value of 0.35. Four doses (i.e., 150, 200, 250, and 300 mg) of TG02 were investigated. The maximum sample size for the dose-finding was 24 patients, treated in cohorts of 3. Based on the toxicity profile of TG02 in patients with other cancer, the investigator chose 200 mg of TG02 as the starting dose.
For the purpose of overdose control, if the observed data suggest that there is more than a 95% chance that a dose is above the MTD, that is, , the dose (j) and higher doses will be eliminated from further examination. To save sample size, this trial added an early stopping rule: if the number of patients treated at a dose reaches 12, stop the trial and select the MTD.
The following text describes how the BOIN app was used to plan and implement the TG02 trial, as well as to estimate the MTD.
Trial planning.
Under the Trial setting tab of the BOIN app, enter the trial settings as shown on the left-hand side of Figure 3A. After clicking on Get Flow Chart and Decision Table, the app returns the BOIN design diagram under Design Flow Chart (right-hand side of Fig 3A) and the decision table (Fig 3B), which can be directly included in the trial protocol. The Animation tab (the fourth one in Fig 3A) is used to help clinicians and trialists understand how the BOIN design works. The Data Supplement shows an example of an animated trial. This function greatly facilitates the collaboration between statisticians and practitioners.
FIG 3.
Use the shiny app to obtain design flowchart (A) and decision table (B) for the trial example using the BOIN design. BOIN, Bayesian optimal interval; DLT, dose-limiting toxicity; MTD, maximum tolerated dose.
To assess the accuracy and safety of the design before using it in the actual trial, we use the Simulation tab to evaluate the probability of correctly selecting the true MTD and the percentage of patients allocated to each dose. A design with good operating characteristics should have desirable probability to identify the true MTD and treat a large proportion of patients on the MTD across various scenarios. We consider four scenarios with MTD located at dose levels 1, 2, 3, and 4, respectively, as shown in Figure 4A.
FIG 4.
Use the shiny app to conduct simulation (A), download trial protocol template (B), and select MTD (C) for the trial example using the BOIN design. BOIN, Bayesian optimal interval; DLT, dose-limiting toxicity; MTD, maximum tolerated dose.
The BOIN app allows direct input (ie, type in) of scenarios or upload using a prespecified template. If Upload scenario file is selected, users will see the template. We choose Type in to enter the scenarios manually. We can conveniently add (remove) a scenario through the Add a Scenario (Remove a Scenario) button and save our scenarios using the Save Scenarios button. After scenarios are entered, we set the number of simulated trials (eg, 1,000) for each scenario. To ensure reproducible results of our simulation, we set the random number generator seed as 6 under Set Seed. After clicking Run Simulation, we see the simulation results under the Operating Characteristics tab, including the percentage of selecting a dose as MTD (selection %) and the percentage of patients treated at each dose (% Pts treated). Additionally, the results include the average number of patients (number of patients) and the percentage of stopping the trial early without selecting MTD (% early stopping). If desired, we can also perform a comparison between BOIN and the commonly used 3 + 3 design by checking the box of Simulate 3 + 3 design for comparison.
The BOIN app provides a handy Trial Protocol tab for us to download a template to prepare the protocol (Fig 4B). The template includes a concise description of the dose-finding algorithm of BOIN, along with the design diagram, decision table, and simulation results. The template is available in both English and Chinese with html and Word formats. The English template for this trial example is provided in the Data Supplement, Section S.2.
Trial conduct.
To conduct the trial, the decision table (Fig 3B) obtained during our trial planning is all we need. We treat the first cohort of patients at the second dose, as specified by the physician during trial planning, and use the decision table to determine the dose for subsequent cohorts of patients. Suppose there are no toxicities observed for the first three patients, we will escalate to dose level 3 for the next cohort of patients. At any time, we simply count the number of patients treated on that dose and the number of patients who experience DLT to determine whether escalation or de-escalation is triggered or if the current dose should be eliminated. For example, suppose six patients were treated at current dose, we would de-escalate the dose if three of them experienced toxicity or eliminate the dose and its higher doses if five or more patients experienced toxicity. The trial will continue until the maximum sample size is reached or the number of patients treated at any dose reaches 12.
MTD estimation.
The BOIN design uses an isotonic estimate of toxicity probability to select MTD. Because of the confidentiality of trial data, we use a set of hypothetical trial data to illustrate how to select MTD using the BOIN app. Suppose that at the end of the trial, we have treated 6, 12, and 6 patients at dose levels 2, 3, and 4, respectively. The corresponding numbers of patients who experience DLT are 1, 4, and 3. No patients have been treated at dose level 1. After entering the data under the Select MTD tab (Fig 4C), clicking on Estimate the MTD produces the selection result on the right side of the app, where we see the estimated MTD (dose level 3) and the DLT estimates for all doses used to treat patients, along with their corresponding 95% credible intervals. The results are also visualized below the printed output. A concise report is also available for download through the Download report button.
DISCUSSION
We have developed the BOIN suite platform, based on R Shiny, to facilitate the use of novel adaptive designs for early-phase clinical trials. The suite includes an array of tools that can be used to tackle various challenges in phase I and phase I and II clinical trials. The suite was developed following proven SOP to ensure its accuracy, robustness, and reproducibility. It provides an interactive platform to help clinicians and regulatory bodies to understand the operating characteristics of novel designs without requiring programming skills. Our future work includes continuously refining and expanding the software suite to address emerging challenges in clinical trials. The upcoming software includes the recently proposed BOIN12 and TITE-BOIN12 (unpublished data), which are one-stage utility-based phase I and II designs suitable for targeted therapy and immunotherapies without or with delayed toxicity and efficacy.
ACKNOWLEDGMENT
The authors thank Jessica Swann for her editorial assistance.
SUPPORT
R. L. was supported in part by grants P30CA016672 and P50CA221703 from the National Cancer Institute, National Institutes of Health. J. J. L. was supported in part by the grants CA016672 and CA221703 from the National Cancer Institute and RP150519 and RP160668 from the Cancer Prevention and Research Institute of Texas. Y. Y. was supported in part by the grants CA016672, 1P50CA221707, and P50CA127001 from the National Cancer Institute.
AUTHOR CONTRIBUTIONS
Conception and design: All authors
Collection and assembly of data: Yanhong Zhou, Ruitao Lin, Ying Yuan
Data analysis and interpretation: Yanhong Zhou, Ruitao Lin, J. Jack Lee, Ying Yuan
Manuscript writing: All authors
Final approval of manuscript: All authors
Accountable for all aspects of the work: All authors
AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST
The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO’s conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/cci/author-center.
Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).
Ying Yuan
Honoraria: Ono Pharmaceutical
Consulting or Advisory Role: Boehringer Ingelheim Pharmaceuticals, Midas Medical Technologies, Amgen, Abbvie, SERVIER, Starpax, Vertex
Ruitao Lin
Consulting or Advisory Role: Abbvie
J. Jack Lee
Consulting or Advisory Role: Abbvie
No other potential conflicts of interest were reported.
REFERENCES
- 1.Yuan Y, Lee JJ, Hilsenbeck SG.Model-assisted designs for early-phase clinical trials: Simplicity meets superiority JCO Precision Oncol 31–122019 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Yuan Y, Zhou H, Zhou Y.Phase I cancer clinical trial design: Single and combination agents Biopharmaceutical Applied Statistics Symposium Springer; Singapore: 2018. pp 205–233 [Google Scholar]
- 3.Dixon WJ, Mood AM.A method for obtaining and analyzing sensitivity data J Am Stat Assoc 43109–1261948 [Google Scholar]
- 4.Yuan Y, Nguyen HQ, Thall PF. Bayesian Designs for Phase I-II Clinical Trials. Chapman and Hall/CRC; Boca Raton, FL: 2017. [Google Scholar]
- 5.O’Quigley J, Pepe M, Fisher L.Continual reassessment method: a practical design for phase 1 clinical trials in cancer Biometrics 46:33–481990 [PubMed] [Google Scholar]
- 6.Babb J, Rogatko A, Zacks S.Cancer phase I clinical trials: Efficient dose escalation with overdose control Stat Med 171103–11201998 [DOI] [PubMed] [Google Scholar]
- 7.Neuenschwander B, Branson M, Gsponer T.Critical aspects of the Bayesian approach to phase I cancer trials Stat Med 272420–24392008 [DOI] [PubMed] [Google Scholar]
- 8.Yin G, Yuan Y.Bayesian model averaging continual reassessment method in phase I clinical trials J Am Stat Assoc 104954–9682009 [Google Scholar]
- 9.Zhou H, Murray TA, Pan H, et al. Comparative review of novel model-assisted designs for phase I clinical trials Stat Med 372208–22222018 [DOI] [PubMed] [Google Scholar]
- 10.Zhou H, Yuan Y, Nie L.Accuracy, safety, and reliability of novel phase I trial designs Clin Cancer Res 244357–43642018 [DOI] [PubMed] [Google Scholar]
- 11.Ji Y, Liu P, Li Y, et al. A modified toxicity probability interval method for dose-finding trials Clinical Trials 7653–6632010 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Liu S, Yuan Y.Bayesian optimal interval designs for phase I clinical trials J R Stat Soc Ser C Appl Stat 64507–5232015 [Google Scholar]
- 13.Yan F, Mandrekar SJ, Yuan Y.Keyboard: A novel Bayesian toxicity probability interval design for phase I clinical trials Clin Cancer Res 233994–40032017 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Yuan Y, Lin R, Li D, et al. Time-to-event Bayesian optimal interval design to accelerate phase I trials Clin Cancer Res 244921–49302018 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Lin R, Yin G.Bayesian optimal interval design for dose finding in drug- combination trials Stat Methods Med Res 262155–21672017 [DOI] [PubMed] [Google Scholar]
- 16.Takeda K, Taguri M, Morita S.BOIN-ET: Bayesian optimal interval design for dose finding based on both efficacy and toxicity outcomes Pharm Stat 17383–3952018 [DOI] [PubMed] [Google Scholar]
- 17.Zhou Y, Lee JJ, Yuan Y.A utility-based Bayesian optimal interval (u- boin) phase I/II design to identify the optimal biological dose for targeted and immune therapies Stat Med 38S5299–S53162019 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Takeda K, Morita S, Taguri M.TITE-BOIN-ET: Time-to-event Bayesian optimal interval design to accelerate dose-finding based on both efficacy and toxicity outcomes Pharm Sta 19335–3492019 [DOI] [PubMed] [Google Scholar]
- 19.Yuan Y, Lee JJ. Integrated platform for designing clinical trials. www.trialdesign.org [Google Scholar]
- 20.Zhou Y, Lee JJ, Wang S, et al. Incorporating historical information to improve phase I clinical trial designs. arXiv preprint arXiv:2004.12972, 2020. [DOI] [PubMed]
- 21.Zhang L, Yuan Y.A practical Bayesian design to identify the maximum tolerated dose contour for drug combination trials Stat Med 354924–49362016 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Lee JJ, Chu CT.Bayesian clinical trials in action Stat Med 312955–29722012 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Cheung YK, Chappell R.Sequential designs for phase I clinical trials with late-onset toxicities Biometrics 561177–11822000 [DOI] [PubMed] [Google Scholar]
- 24.R Core Team . R: A language and environment for statistical computing. 2014. http://www.r-project.org [Google Scholar]
- 25.Chang W, Cheng J, Allaire J, et al. Shiny: Web application framework for R. R Package Version. 2017;1 [Google Scholar]
- 26.Lin R, Zhou Y, Yan F, et al. BOIN12: Bayesian optimal interval phase I/II trial design for utility-based dose finding with immunotherapy and targeted therapies JCO Precision Oncol 41393–14022020 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Bril G, Dykstra R, Pillers C, et al. Algorithm as 206: Isotonic regression in two independent variables J R Stat Soc Ser C Appl Stat 33352–3571984 [Google Scholar]