TRAFIC: Statistical Plan for a Pragmatic Early Phase 1/2 Bayesian Adaptive Dose Escalation Trial in Rheumatoid Arthritis

Background : Adaptive model-based dose-finding designs have demonstrated advantages over traditional rule-based designs but have increased statistical complexity resulting in slow uptake especially outside of cancer trials. TRAFIC is a multi-centre, early phase trial in Rheumatoid Arthritis incorporating a model-based design. Methods : A Bayesian adaptive dose-finding phase I trial rolling into a single arm, single stage phase II trial. Model parameters for phase I were chosen via Monte Carlo simulation evaluating objective performance measures under clinically relevant scenarios and incorporated stopping rules for early termination. Potential designs were further calibrated utilising dose transition pathways. Discussion : TRAFIC is an MRC funded trial of a re-purposed treatment demonstrating that it is possible to design, fund and implement a model-based phase I trial in a non-cancer population within conventional research funding tracks and regulatory constraints. The phase I design allows borrowing of information from previous trials; all accumulated data to be utilised in decision-making; verification of operating characteristics through simulation; improved understanding for management and oversight teams through dose transition pathways. The rolling phase II design brings efficiencies in trial conduct including site and monitoring activities, and cost. TRAFIC is the first funded model-based dose-finding trial in inflammatory disease demonstrating that small phase I/II trials can have an underlying statistical basis for decision-making and interpretation.

Rheumatoid arthritis (RA) afflicts 0.5-1.0% of adults globally. It presents polyarthritis with coincident morbidities including vascular bone and cognitive deficits and together these substantially impact quality of life, disability and long-term survival. Around a third of patients have stopped working within two years of onset and around a half by ten years with significant costs to the economy 5 . Advances in RA management and biologic therapies have contributed to an improved prognosis but up to 50% fail to achieve remission, with 5-10% being refractory to all treatments. Consequently, there remains a significant unmet need in RA management requiring novel therapeutic approaches. Prior therapeutics have focussed primarily on immune based therapeutics. We sought to evaluate a stromal targeting strategy and since this is an entirely novel approach we required alternative methods to minimise patient exposure pending estimation of safety and initial indication of potential efficacy.
There is a paucity of early phase adaptive designs across all fields of medicine. A systematic review in rheumatology 6 identified just one adaptive early phase design from 62 trials considered. In this paper we report the statistical design, calibration and implementation of the Targeting the Rheumatoid Arthritis Synovial Fibroblast with Cyclin-Dependent Kinase Inhibition (TRAFIC) trial. TRAFIC is a non-commercial, multi-centre, phase I/II trial incorporating a Bayesian adaptive model-based dose finding phase I design to determine the safety, tolerability and efficacy of seliciclib as an addition to existing therapy in patients with RA. Rolling the trial from phase I to phase II allows site activity and momentum to be retained bringing trial conduct efficiencies including internal and external trial monitoring activities, and associated cost savings. It also allows continuity for clinical, trial and monitoring teams.
These advantages, and also potential advantages of 'sharing' patients from one phase to the next, make this approach attractive when conducting challenging early phase trials.
Seliciclib (R-roscovitine) is an orally available cyclin dependent kinase inhibitor with an acceptable toxicity profile 7 repurposed from the oncology setting. Determining the toxicity profile of seliciclib used in combination with a biologic plus or minus conventional synthetic disease-modifying anti-rheumatic drugs (sDMARDs) is an essential component of TRAFIC, providing important insight into its potential acceptability as an adjunctive therapy in RA.

Statistical Design
TRAFIC has a phase I dose finding trial rolling into a single arm, single stage phase II trial.
The primary objective of phase I is to determine the maximum tolerated dose (MTD) of seliciclib over a 4 week treatment period when given in addition to an existing TNF inhibitor with or without sDMARDs. Phase I is planned to roll into phase II for which the primary objective is to assess the potential efficacy of seliciclib following 12 weeks of treatment when administered at the MTD established in phase I; efficacy is based on a composite response measure.

Phase I Continual Reassessment Method Dose Finding Design
The MTD will be established using a modified one-stage Bayesian Continual Reassessment Method (CRM) model-based phase I design 8 . The primary outcome is identification of MTD based on the occurrence of dose limiting toxicities (DLTs). A DLT is defined as the cessation of the Investigational Medicinal Product (IMP) due to adverse events or reactions (AE/AR) occurring during the four week treatment period. These can be either symptomatic (e.g. nausea) or abnormal laboratory parameters or investigations. A DLT may be based upon the patient's request to stop treatment, regardless of severity, or an abnormal laboratory parameter necessitating cessation of treatment. In the event of several AEs/ARs contributing to the decision to discontinue IMP, only a single DLT will be recorded for the purposes of dose pathway decision making (see below). A flare of RA will not be considered an AE. The primary objective is to identify the MTD associated with a target 0.35 DLT probability. The definition of a DLT and its target probability was agreed at a TRAFIC Investigators consensus meeting. The MTD is defined as the dose level that is closest to the level at which 35% of patients experience a DLT over the treatment period (4 weeks) of seliciclib, comparable to the rate of sustained tolerated treatment on conventional synthetic sDMARDs such as Methotrexate 9 .
Phase I will include a maximum of 21 patients. Up to seven cohorts of three participants each will be treated. The decision to recruit three patients per cohort was logistical and with respect to the proposed 35% DLT rate. The number of cohorts was based on the selected dose range which encompasses up to five doses of IMP: participants will receive either 200mg, 400mg, 600mg, 800mg or 1000mg seliciclib daily for 4 consecutive days (one cycle) every week for 4 weeks. The dose range and schedule is based on healthy control data and oncology experience 7 . The prior estimate of MTD is 600mg (Dose Level 3) of seliciclib (Table 1)  The recommended dose (the dose with estimated DLT probability closest to the target of 35%) for each of the subsequent cohorts is determined using the CRM incorporating all of the accumulated DLT outcomes but for added safety the design includes a restriction to prevent skipping of untested doses when escalating. Recruitment continues until either the maximum sample size is reached, the trial is stopped early due to unacceptable levels of DLT at the lowest dose or the MTD is considered sufficiently well estimated. Two early stopping rules are included to allow for early termination: 1. If there is a high probability (> 0.7) that the posterior probability of DLT at the lowest dose is greater than the target DLT rate of 35%, indicating that the lowest dose is too toxic.
2. If four consecutive cohorts (three patients in each cohort) have already been allocated at the current MTD, which would also be the recommended dose level for the next cohort if the trial continued.
The value of 0.7 was selected so that the design will recommend stopping early for excessive toxicity when we observe 2 or 3 DLTs out of the first 3 patients at the lowest dose At trial completion, the number of patients experiencing DLT at each dose level, together with the proportion of patients with DLT at that dose level, will be reported. The Bayesian posterior probability of DLT at each dose level (with 90% probability interval) will be reported graphically and in tabular form. The MTD will be reported as well as the posterior probability that the DLT rate at dose level 1 (200mg) is greater than the target level of 35%. Secondary outcome measures, including PD biomarkers and PK parameters, will be presented descriptively and graphically.

CRM Model Details
The one-stage, one parameter Bayesian logistic dose-toxicity model used in TRAFIC is given by: where x is the scaled dose and ϑ(x, β) is the probability of a DLT at dose x (10). The model parameter b is assumed to be random with prior distribution N+0, σ . ; /. The Bayesian CRM model is completed by specification of the prior probability of a DLT associated with the five test doses (the skeleton) and σ . ; , the prior variance of b.
Early termination is allowed if there is a high probability that the posterior probability of DLT at the lowest dose is greater than the target DLT rate. The posterior distribution of the DLT occurrence probability at the lowest dose requires the posterior distribution for b to be evaluated. After the first n patients, this is given by: where X > = {x L , x ; , … , x > } are the scaled doses and Y > = {y L , y ; , … , y > } are the DLT outcomes (0 or 1).
The posterior probability that the DLT occurrence probability at Dose Level 1 is greater than some rate ϑL is given by b and x cL is the scaled dose for Dose Level 1. The integral ∫ p(b |X > , Y > ) . R BS dβ is analytically intractable and so is obtained by numerical integration (13). The Bayesian CRM model is calibrated (specification of the skeleton and prior variance of b) using the following algorithm of Cheung 10 which was implemented using a modified version of the R function mtrials part of the dfcrm R library 14 : a) For a given indifference interval half-width d, the associated skeleton is obtained using the dfcrm function getprior. The indifference interval is an interval into which the DLT probability of the selected dose will eventually fall given a sufficiently large sample size.
b) For each d, the least informative prior standard deviation for the slope parameter β, Ud (e), is determined. The SD is defined to be least informative if the prior probability that a dose level is the MTD is approximately equal for all dose levels. Table 1 shows the effect of varying f g for TRAFIC when d=0.06; a large fh does not necessary correspond to an uninformative prior for the MTD dose level.
c) The pair {e, σ LI β (e)} was chosen based upon evaluation of the following performance measures: i) the ability to correctly select the true MTD -the risk-adjusted average accuracy, AN (weighting the probability of selecting a dose level by the absolute discrepancy between the true probability of toxicity at that dose level and the target probability of toxicity), and the unadjusted probability of correctly selecting the true MTD; ii) optimal allocation defined as the mean proportion of patients treated within one dose level of the true MTD; iii) the mean proportion of patients treated at an overdose (a dose above the true MTD); and iv) the mean number of patients treated 10 .

Simulations
The performance measures in c) above were estimated by Monte Carlo simulation: 20,000 trials were simulated for different values of d (0.02 to 0.20 in steps of 0.01). Trials were simulated assuming the underlying probabilities of DLT (Table 2), these are the appropriate plateau calibration configurations with five test dose levels when the target DLT rate is 35%.
The performance of the design also was assessed via simulation under several clinically relevant scenarios.
Dose level 1 (200mg) was thought unlikely to be efficacious, hence calibration of the CRM focused on achieving: 1) good accuracy, optimal allocation and minimising the proportion of patients treated at an overdose when the true simulated MTD was at dose levels 2 to 5; and 2) minimising the number of patients treated when the true MTD was at dose level 1.
Performance measures implementing the early stopping rules 1 and 2 are plotted against d ( Figure 2) for each of the calibration curves ( Table 2).
The mean of the performance measures taken over the calibration curves 2, 3, 4 and 5, i.e. when the true simulated MTD is at dose levels 2, 3, 4 and 5 are given in Table 3 and  Table 5.

Phase II Fleming A'hern Design
Phase II is a single arm, single stage early phase trial based on a Fleming-A'Herns design 15 recruiting a total of 18 patients at the recommended phase II dose who provided written informed consent, with baseline and 12-week outcome data. Efficacy is assessed via clinical assessments, analysis of synovial biopsies, and contrast enhanced MRI of an affected hand and wrist. The primary outcome measure is a composite response rate at 12 weeks defined as achieving two of the following three criteria: i) EULAR moderate response or ACR20 response, ii) histological reduction in macrophage number in the sub lining layer of the synovium ≥20% and iii) reduction of Rheumatoid Arthritis MRI Scoring System score, on MRI, of ≥0.5 units or osteitis score of ≥0.2 units. Secondary outcome measures are as for phase I and, additionally, changes in pharmacodynamic biomarkers in synovial tissue and response rate after 1, 2, 3, 6 and 9 weeks of therapy.
Response rate will be calculated as the total number of patients responding as a proportion of all patients who start treatment. Any patients who are not assessable at 12 weeks will be classed as a non-responder. Individual components of the composite response outcome will be reported descriptively. Adverse events will be reported as the number of patients experiencing an event as a proportion of the total number of patients starting treatment, reported descriptively. Pharmacokinetic (PK) parameters and pharmacodynamic (PD) biomarkers will be presented graphically.
The Fleming A'Hern design assumes a composite response rate to reject Seliciclib (p0) <25% and a response rate to investigate Seliciclib further (p1) >50%. The justification to investigate Seliciclib further is based on observing a critical minimum number of responses, as specified in the Statistical Analysis Plan. As an early phase trial the error levels are inflated but restricted to an acceptable level of <15% a (type 1) and <20% b (type 2). With these stated parameters the target recruitment for phase II is calculated as 18 patients.
At the conclusion of phase II, the Fleming A'Hern design would indicate no further investigation is warranted if the observed number of clinical responses is less than the critical number, retained in the statistical trial master file. As this is the first trial investigating seliciclib in this indication (as a repurposed drug with a novel mechanism of action), the decision to continue will also be based on PD biomarkers and PK parameters, since the

DECLARATIONS Funding
This research is supported by the MRC DPFS award MR/L005123/1.

Ethics Approval and Consent to Participate
North East Ethics Approval Reference: 14/NE/1075.
All patients provide written informed consent to participate.

Consent for Publication
Not applicable.
No individual level details or images are included.

Availability of Data and Materials
Final trial dataset will be held by Newcastle University. Data sharing requests may be submitted following publication of results.

Competing Interests
The authors declare that they have no competing interests.

Authors Contributions
MC is trial statistician, developed the protocol and is author of the statistical analysis plan; CY is statistical advisor; CB WN IM AF SS AP are clinical leads and developed the protocol; JDI is Chief Investigator, conceived the trial, led the grant application and protocol development; DDS is statistical co-applicant, led the study design and has protocol oversight. All authors have read and approved the final manuscript.